I asked AI to port the device libs reference implementation. It mostly worked, though it got the compares wrong and also missed a fold that happened in compiler. With that fixed I get identical DAG output, and almost the same globalisel output (differing by an inverted compare and select). Also adjusted some stylistic choices.
10243 lines
525 KiB
LLVM
10243 lines
525 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,SI,GCN-SDAG,SI-SDAG %s
|
|
; RUN: llc -global-isel=1 -global-isel-abort=2 -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefixes=GCN,SI,GCN-GISEL,SI-GISEL %s
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=GCN,VI,GCN-SDAG,VI-SDAG %s
|
|
; RUN: llc -global-isel=1 -global-isel-abort=2 -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=GCN,VI,GCN-GISEL,VI-GISEL %s
|
|
; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX900,GCN-SDAG,GFX900-SDAG %s
|
|
; RUN: llc -global-isel=1 -global-isel-abort=2 -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX900,GCN-GISEL,GFX900-GISEL %s
|
|
|
|
define double @v_exp_f64(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define <2 x double> @v_exp_v2f64(<2 x double> %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_v2f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[4:5], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s26, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v16, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v7, s26, v16, v5
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[8:9], v[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[7:8], v[8:9], -v[6:7]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[4:5]|, s[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v5, v8, v5, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v4, v7, v4, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[4:5], s[8:9], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, 0x3b39803f
|
|
; SI-SDAG-NEXT: v_mul_f64 v[14:15], v[2:3], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[10:11], v[7:8]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v7, s26, v16, v15
|
|
; SI-SDAG-NEXT: v_add_f64 v[16:17], v[14:15], v[6:7]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[14:15]|, s[6:7]
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[16:17], -v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s14, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v10, s12
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v7, v7, v15, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v6, v6, v14, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s15, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v11, s13
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], s[8:9], v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], s[14:15], v[10:11]
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], s[10:11], v[14:15]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[12:13]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], s[14:15], v[10:11]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s18, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s19, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], v[10:11], s[12:13]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[18:19]
|
|
; SI-SDAG-NEXT: s_mov_b32 s20, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s21, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], v[10:11], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[20:21]
|
|
; SI-SDAG-NEXT: s_mov_b32 s22, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s23, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], v[10:11], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[22:23]
|
|
; SI-SDAG-NEXT: s_mov_b32 s24, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s25, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], v[10:11], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[24:25]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], v[10:11], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v16, v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[14:15], v[10:11], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], s[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[14:15], v[4:5], s[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[14:15], v[4:5], s[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[12:13], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x40900000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[14:15], v[4:5], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v6, v[6:7]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[8:9], v[8:9], v16
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[8:9], v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[14:15], v[4:5], 1.0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v9, v7, v9, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v9, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v8, vcc
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[6:7], v[2:3]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[8:9], v[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v5, v7, v5, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v5, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v4, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_v2f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[4:5], v[0:1], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v9, 0x80000000, v5
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v9, 0x43300000, v9
|
|
; SI-GISEL-NEXT: v_add_f64 v[10:11], v[4:5], v[8:9]
|
|
; SI-GISEL-NEXT: v_mul_f64 v[6:7], v[2:3], v[6:7]
|
|
; SI-GISEL-NEXT: v_add_f64 v[9:10], v[10:11], -v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v12, 0x432fffff
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[4:5]|, v[11:12]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v4, v9, v4, vcc
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v9, 0x80000000, v7
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v9, 0x43300000, v9
|
|
; SI-GISEL-NEXT: v_add_f64 v[13:14], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v5, v10, v5, vcc
|
|
; SI-GISEL-NEXT: v_add_f64 v[8:9], v[13:14], -v[8:9]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[6:7]|, v[11:12]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v12, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v6, v8, v6, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v7, v9, v7, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], -v[4:5], v[8:9], v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], -v[6:7], v[8:9], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], -v[4:5], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], -v[6:7], v[12:13], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v12, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[16:17], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v16, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x40900000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v17, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[14:15]
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[16:17]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[8:9], v[12:13], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[8:9], v[0:1], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v0, v[4:5]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v4, v[6:7]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[10:11], v0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[8:9], v4
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[2:3], v[14:15]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[2:3], v[16:17]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v10, v5, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v4, s[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[4:5]
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_v2f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[4:5], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: v_mul_f64 v[6:7], v[2:3], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v13, s7
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v12, s6
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s8, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x40900000
|
|
; VI-SDAG-NEXT: s_mov_b32 s9, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[6:7], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[4:5], v[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[4:5], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[4:5], v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v4, v[4:5]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v6, v[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[4:5], v[12:13]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], s[4:5], v[12:13]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[8:9], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[8:9], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[10:11], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[4:5], v[8:9], v4
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[6:7], v[10:11], v6
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v5, v8, v5, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v0, v8, v7, s[6:7]
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v0, s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[8:9], s[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v5, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v6, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_v2f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[6:7], v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[4:5], v[2:3], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v12, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], -v[6:7], v[8:9], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], -v[4:5], v[8:9], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], -v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], -v[4:5], v[12:13], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v12, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[6:7]
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v7, v[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[16:17], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[14:15], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[8:9], v[12:13], 1.0
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[10:11], v6
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[6:7], v[8:9], v7
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x40900000
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[8:9]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[2:3], v[8:9]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[6:7], v[0:1], v[10:11]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], v[2:3], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v6, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v3, v0, v5, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v5, v0, v7, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v4, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v5, s[8:9]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_v2f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[4:5], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[6:7], v[2:3], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v13, s7
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v12, s6
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s8, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x40900000
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s9, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[6:7], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[4:5], v[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[4:5], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[4:5], v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v4, v[4:5]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v6, v[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[4:5], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], s[4:5], v[12:13]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[8:9], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[8:9], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[10:11], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[4:5], v[8:9], v4
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[6:7], v[10:11], v6
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v5, v8, v5, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v0, v8, v7, s[6:7]
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v0, s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[8:9], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v5, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v6, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_v2f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[6:7], v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[4:5], v[2:3], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v12, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v13, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], -v[6:7], v[8:9], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], -v[4:5], v[8:9], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], -v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], -v[4:5], v[12:13], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v12, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v13, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v7, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[16:17], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[8:9], v[12:13], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[14:15], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[8:9], v[12:13], 1.0
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[10:11], v6
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[6:7], v[8:9], v7
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[2:3], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[6:7], v[0:1], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], v[2:3], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v6, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v3, v0, v5, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v5, v0, v7, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v4, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v5, s[8:9]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call <2 x double> @llvm.exp.v2f64(<2 x double> %in)
|
|
ret <2 x double> %result
|
|
}
|
|
|
|
define <3 x double> @v_exp_v3f64(<3 x double> %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_v3f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[8:9], v[0:1], s[6:7]
|
|
; SI-SDAG-NEXT: s_brev_b32 s46, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v10, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v7, s46, v10, v9
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[11:12], v[8:9], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[11:12], v[11:12], -v[6:7]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[8:9]|, s[8:9]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v12, v12, v9, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v11, v11, v8, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[11:12], s[10:11], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e928af3
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[11:12], s[12:13], v[7:8]
|
|
; SI-SDAG-NEXT: s_mov_b32 s14, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v9, s5
|
|
; SI-SDAG-NEXT: s_mov_b32 s15, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v8, s4
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], s[14:15], v[8:9]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s18, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s19, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[18:19]
|
|
; SI-SDAG-NEXT: s_mov_b32 s20, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s21, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[20:21]
|
|
; SI-SDAG-NEXT: s_mov_b32 s22, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s23, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[22:23]
|
|
; SI-SDAG-NEXT: s_mov_b32 s24, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s25, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[24:25]
|
|
; SI-SDAG-NEXT: s_mov_b32 s26, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s27, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[26:27]
|
|
; SI-SDAG-NEXT: s_mov_b32 s28, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s29, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[28:29]
|
|
; SI-SDAG-NEXT: s_mov_b32 s42, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s43, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], s[42:43]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v7, v[11:12]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[13:14], v[15:16], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s40, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[11:12], v[13:14], v[15:16], 1.0
|
|
; SI-SDAG-NEXT: v_mul_f64 v[13:14], v[2:3], s[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s44, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s41, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s45, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[11:12], v[11:12], v7
|
|
; SI-SDAG-NEXT: v_bfi_b32 v7, s46, v10, v14
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[40:41], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[44:45], v[0:1]
|
|
; SI-SDAG-NEXT: v_add_f64 v[15:16], v[13:14], v[6:7]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v19, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v12, v19, v12, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v12, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v11, vcc
|
|
; SI-SDAG-NEXT: v_add_f64 v[11:12], v[15:16], -v[6:7]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[13:14]|, s[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[44:45], v[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v12, v12, v14, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v11, v11, v13, vcc
|
|
; SI-SDAG-NEXT: v_mul_f64 v[13:14], v[4:5], s[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[11:12], s[10:11], v[2:3]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v7, s46, v10, v14
|
|
; SI-SDAG-NEXT: v_add_f64 v[17:18], v[13:14], v[6:7]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[13:14]|, s[8:9]
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[17:18], -v[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[11:12], s[12:13], v[15:16]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v7, v7, v14, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v6, v6, v13, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[6:7], s[10:11], v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], s[14:15], v[8:9]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[6:7], s[12:13], v[17:18]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], s[14:15], v[8:9]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[26:27]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[28:29]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[26:27]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], s[42:43]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[28:29]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[15:16], v[13:14], 1.0
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v12, v[11:12]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], s[42:43]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[15:16], v[13:14], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], 1.0
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v6, v[6:7]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[10:11], v[10:11], v12
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[40:41], v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[17:18], v[8:9], 1.0
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v7, v19, v11, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v7, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v10, vcc
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[6:7], v[8:9], v6
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[40:41], v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[44:45], v[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v7, v19, v7, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v7, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v4, 0, v6, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_v3f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[12:13], v[0:1], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v7, 0x80000000, v13
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v7, 0x43300000, v7
|
|
; SI-GISEL-NEXT: v_add_f64 v[14:15], v[12:13], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[14:15], v[14:15], -v[6:7]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[12:13]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v18, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v12, v14, v12, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v13, v15, v13, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], -v[12:13], v[18:19], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v20, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_mul_f64 v[16:17], v[2:3], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], -v[12:13], v[20:21], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v24, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v26, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v25, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v7, 0x80000000, v17
|
|
; SI-GISEL-NEXT: v_fma_f64 v[28:29], v[14:15], v[24:25], v[26:27]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v30, 0x623fde64
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v7, 0x43300000, v7
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v31, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_add_f64 v[22:23], v[16:17], v[6:7]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[28:29], v[14:15], v[28:29], v[30:31]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v32, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v33, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_add_f64 v[22:23], v[22:23], -v[6:7]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[16:17]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[28:29], v[14:15], v[28:29], v[32:33]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v34, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v35, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v16, v22, v16, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v17, v23, v17, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[28:29], v[34:35]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v28, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v29, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_mul_f64 v[10:11], v[4:5], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[28:29]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v36, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v37, 0x3f811111
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v7, 0x80000000, v11
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[36:37]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v38, 0x555502a1
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v7, 0x43300000, v7
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v39, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_add_f64 v[48:49], v[10:11], v[6:7]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[38:39]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v50, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v51, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[48:49], -v[6:7]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[50:51]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v48, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v49, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[48:49]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[10:11]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[14:15], v[22:23], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], -v[16:17], v[18:19], v[2:3]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v6, v6, v10, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v7, v7, v11, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], -v[6:7], v[18:19], v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], -v[16:17], v[20:21], v[22:23]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], -v[6:7], v[20:21], v[18:19]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[22:23], v[24:25], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[18:19], v[24:25], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[22:23], v[20:21], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[18:19], v[24:25], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[22:23], v[20:21], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[18:19], v[24:25], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[22:23], v[20:21], v[34:35]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v26, v[12:13]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[22:23], v[20:21], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[18:19], v[24:25], v[34:35]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[20:21], v[36:37]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[18:19], v[24:25], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[12:13], v[38:39]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[14:15], v[8:9], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[12:13], v[50:51]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[18:19], v[24:25], v[36:37]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[12:13], v[48:49]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x40900000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[8:9], v[8:9], v26
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[18:19], v[20:21], v[38:39]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[12:13], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v16, v[16:17]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v17, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v8, 0, v8, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[18:19], v[20:21], v[50:51]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[12:13], 1.0
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v22, v17, v9, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[18:19], v[20:21], v[48:49]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v8, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[8:9], v[12:13], v16
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v22, vcc
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[2:3], v[10:11]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v16, v[6:7]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v8, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v13, v17, v9, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[18:19], v[20:21], 1.0
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[2:3], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[18:19], v[8:9], 1.0
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v12, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[6:7], v[6:7], v16
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v13, vcc
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[4:5], v[10:11]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v6, 0, v6, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v7, v17, v7, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[4:5], v[14:15]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v6, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v5, 0, v7, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_v3f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[6:7], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mul_f64 v[8:9], v[2:3], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s8, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s9, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: s_mov_b32 s10, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: s_mov_b32 s11, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s12, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v13, s11
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[8:9], v[8:9]
|
|
; VI-SDAG-NEXT: s_mov_b32 s13, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v12, s10
|
|
; VI-SDAG-NEXT: s_mov_b32 s10, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s11, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s14, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s15, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[6:7], v[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s16, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s17, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: s_mov_b32 s18, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s19, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[8:9], v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s20, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s21, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[8:9], v[14:15]
|
|
; VI-SDAG-NEXT: s_mov_b32 s22, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s23, 0x3fa55555
|
|
; VI-SDAG-NEXT: s_mov_b32 s24, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s25, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], s[12:13], v[12:13]
|
|
; VI-SDAG-NEXT: s_mov_b32 s26, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s27, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], s[12:13], v[12:13]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v20, v[6:7]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v21, v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[10:11]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[10:11]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[14:15]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[14:15]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[10:11], v[16:17], 1.0
|
|
; VI-SDAG-NEXT: v_mul_f64 v[16:17], v[4:5], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[14:15], v[18:19], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[8:9], v[10:11], v20
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[16:17], v[16:17]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[10:11], v[14:15], v21
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[6:7], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[8:9], v[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], s[12:13], v[12:13]
|
|
; VI-SDAG-NEXT: s_mov_b32 s12, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s13, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[12:13], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[12:13], v[2:3]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[12:13], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s10, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s11, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[10:11], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[10:11], v[2:3]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[10:11], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[14:15]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[18:19], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v12, v[16:17]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v13, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v9, v13, v9, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v0, v13, v11, s[6:7]
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v0, s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v8, vcc
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[6:7], v[6:7], v12
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[8:9], s[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v9, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v2, v13, v7, s[10:11]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v2, s[12:13]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v10, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[12:13], s[10:11]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v4, 0, v6, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_v3f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[8:9], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[10:11], v[2:3], v[6:7]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[6:7], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v12, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[8:9], v[8:9]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[10:11], v[10:11]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], -v[8:9], v[12:13], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], -v[10:11], v[12:13], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], -v[6:7], v[12:13], v[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], -v[8:9], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], -v[10:11], v[18:19], v[16:17]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], -v[6:7], v[18:19], v[12:13]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[14:15], v[22:23], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[16:17], v[24:25], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], v[14:15], v[20:21], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[16:17], v[22:23], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v20, v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[12:13], v[18:19], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v18, v[10:11]
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v19, v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v12, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v13, 0x40900000
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[6:7], v[14:15], v20
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[14:15], v[16:17], v18
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[8:9], v[8:9], v19
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[12:13]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[2:3], v[12:13]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[6:7], v[4:5], v[12:13]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], v[0:1], v[10:11]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[10:11], v[2:3], v[10:11]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[12:13], v[4:5], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v16, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v6, 0, v6, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v14, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v8, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v5, v16, v7, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v7, v16, v15, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v8, v16, v9, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v6, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[10:11]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v3, s[12:13]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v5, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v7, s[10:11]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v5, 0, v8, s[12:13]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_v3f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[6:7], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[8:9], v[2:3], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s8, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s9, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s10, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s11, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s12, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v13, s11
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[8:9], v[8:9]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s13, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v12, s10
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s10, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s11, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s14, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s15, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[6:7], v[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s16, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s17, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s18, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s19, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[8:9], v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s20, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s21, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[8:9], v[14:15]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s22, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s23, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s24, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s25, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], s[12:13], v[12:13]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s26, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s27, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], s[12:13], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v20, v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v21, v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[10:11]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[10:11]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], v[16:17], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[14:15], v[18:19], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[10:11], v[16:17], 1.0
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[16:17], v[4:5], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[14:15], v[18:19], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[8:9], v[10:11], v20
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[16:17], v[16:17]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[10:11], v[14:15], v21
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[6:7], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[8:9], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], s[12:13], v[12:13]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s12, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s13, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[12:13], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[12:13], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[12:13], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s10, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s11, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[10:11], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[10:11], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[10:11], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[18:19], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[18:19], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v12, v[16:17]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v13, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v9, v13, v9, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v0, v13, v11, s[6:7]
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v0, s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v8, vcc
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[6:7], v[6:7], v12
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[8:9], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v9, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v2, v13, v7, s[10:11]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v2, s[12:13]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v10, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[12:13], s[10:11]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v4, 0, v6, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_v3f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[8:9], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[10:11], v[2:3], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[6:7], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v12, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v13, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[8:9], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[10:11], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], -v[8:9], v[12:13], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], -v[10:11], v[12:13], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], -v[6:7], v[12:13], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], -v[8:9], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], -v[10:11], v[18:19], v[16:17]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], -v[6:7], v[18:19], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[22:23], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[16:17], v[24:25], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[14:15], v[22:23], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[16:17], v[24:25], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[12:13], v[18:19], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], v[14:15], v[20:21], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[16:17], v[22:23], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v20, v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[12:13], v[18:19], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v18, v[10:11]
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v19, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v12, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v13, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[6:7], v[14:15], v20
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[14:15], v[16:17], v18
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[8:9], v[8:9], v19
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[2:3], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[6:7], v[4:5], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], v[0:1], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[10:11], v[2:3], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[12:13], v[4:5], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v16, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v6, 0, v6, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v14, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v8, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v5, v16, v7, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v7, v16, v15, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v8, v16, v9, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v6, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v3, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v5, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v7, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v5, 0, v8, s[12:13]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call <3 x double> @llvm.exp.v2f64(<3 x double> %in)
|
|
ret <3 x double> %result
|
|
}
|
|
|
|
define <4 x double> @v_exp_v4f64(<4 x double> %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_v4f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[10:11], v[0:1], s[6:7]
|
|
; SI-SDAG-NEXT: s_brev_b32 s46, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v18, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v9, s46, v18, v11
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[12:13], v[10:11], v[8:9]
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0x432fffff
|
|
; SI-SDAG-NEXT: v_mul_f64 v[14:15], v[2:3], s[6:7]
|
|
; SI-SDAG-NEXT: v_add_f64 v[12:13], v[12:13], -v[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[10:11]|, s[8:9]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v9, s46, v18, v15
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v11, v13, v11, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v10, v12, v10, vcc
|
|
; SI-SDAG-NEXT: v_add_f64 v[12:13], v[14:15], v[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[14:15]|, s[8:9]
|
|
; SI-SDAG-NEXT: v_add_f64 v[12:13], v[12:13], -v[8:9]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v13, v13, v15, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v12, v12, v14, vcc
|
|
; SI-SDAG-NEXT: v_mul_f64 v[14:15], v[4:5], s[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_bfi_b32 v9, s46, v18, v15
|
|
; SI-SDAG-NEXT: v_add_f64 v[16:17], v[14:15], v[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[14:15]|, s[8:9]
|
|
; SI-SDAG-NEXT: v_add_f64 v[16:17], v[16:17], -v[8:9]
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, 0x3b39803f
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v15, v17, v15, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v14, v16, v14, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[10:11], s[10:11], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e928af3
|
|
; SI-SDAG-NEXT: v_fma_f64 v[19:20], v[10:11], s[12:13], v[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v17, s5
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v16, s4
|
|
; SI-SDAG-NEXT: s_mov_b32 s14, 0x623fde64
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], s[16:17], v[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s15, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[14:15]
|
|
; SI-SDAG-NEXT: s_mov_b32 s18, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s19, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[18:19]
|
|
; SI-SDAG-NEXT: s_mov_b32 s20, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s21, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[20:21]
|
|
; SI-SDAG-NEXT: s_mov_b32 s22, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s23, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[22:23]
|
|
; SI-SDAG-NEXT: s_mov_b32 s24, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s25, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[24:25]
|
|
; SI-SDAG-NEXT: s_mov_b32 s40, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s41, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[40:41]
|
|
; SI-SDAG-NEXT: s_mov_b32 s42, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s43, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[42:43]
|
|
; SI-SDAG-NEXT: s_mov_b32 s44, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s45, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], s[44:45]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v9, v[10:11]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[21:22], v[19:20], v[21:22], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s28, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s26, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s29, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s27, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_fma_f64 v[19:20], v[19:20], v[21:22], 1.0
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[28:29], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[26:27], v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[12:13], s[10:11], v[2:3]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[9:10], v[19:20], v9
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[12:13], s[12:13], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v19, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v20, v19, v10, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], s[16:17], v[16:17]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v12, v[12:13]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[14:15]
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[40:41]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[42:43]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], s[44:45]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], v[10:11], 1.0
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v9, vcc
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[9:10], v[10:11], v12
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[28:29], v[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v20, s[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[26:27], v[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v12, v19, v10, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v12, s[4:5]
|
|
; SI-SDAG-NEXT: v_mul_f64 v[12:13], v[6:7], s[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], s[10:11], v[4:5]
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v9, vcc
|
|
; SI-SDAG-NEXT: v_bfi_b32 v9, s46, v18, v13
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[14:15], s[12:13], v[10:11]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v20, v[14:15]
|
|
; SI-SDAG-NEXT: v_add_f64 v[14:15], v[12:13], v[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[12:13]|, s[8:9]
|
|
; SI-SDAG-NEXT: v_add_f64 v[8:9], v[14:15], -v[8:9]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], s[16:17], v[16:17]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v9, v9, v13, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v8, v8, v12, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], s[10:11], v[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[14:15]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[8:9], s[12:13], v[12:13]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], s[16:17], v[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[14:15]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[40:41]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[42:43]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], s[44:45]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[40:41]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[42:43]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[10:11], v[14:15], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], s[44:45]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[8:9]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[16:17], 1.0
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[10:11], v[10:11], v20
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[28:29], v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[26:27], v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[12:13], v[12:13], v[14:15], 1.0
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v9, v19, v11, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v9, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v4, 0, v10, vcc
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[8:9], v[12:13], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[28:29], v[6:7]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[26:27], v[6:7]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v9, v19, v9, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v7, 0, v9, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v6, 0, v8, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_v4f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v12, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[14:15], v[0:1], v[12:13]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v9, 0x80000000, v15
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v9, 0x43300000, v9
|
|
; SI-GISEL-NEXT: v_add_f64 v[16:17], v[14:15], v[8:9]
|
|
; SI-GISEL-NEXT: v_mul_f64 v[18:19], v[2:3], v[12:13]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, -1
|
|
; SI-GISEL-NEXT: v_add_f64 v[16:17], v[16:17], -v[8:9]
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v9, 0x80000000, v19
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x432fffff
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v9, 0x43300000, v9
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[14:15]|, v[10:11]
|
|
; SI-GISEL-NEXT: v_add_f64 v[20:21], v[18:19], v[8:9]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v16, v16, v14, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v17, v17, v15, vcc
|
|
; SI-GISEL-NEXT: v_add_f64 v[14:15], v[20:21], -v[8:9]
|
|
; SI-GISEL-NEXT: v_mul_f64 v[20:21], v[4:5], v[12:13]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[18:19]|, v[10:11]
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v9, 0x80000000, v21
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v9, 0x43300000, v9
|
|
; SI-GISEL-NEXT: v_add_f64 v[22:23], v[20:21], v[8:9]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v14, v14, v18, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v15, v15, v19, vcc
|
|
; SI-GISEL-NEXT: v_add_f64 v[18:19], v[22:23], -v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v22, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v23, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[20:21]|, v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], -v[16:17], v[22:23], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v26, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v18, v18, v20, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v19, v19, v21, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], -v[16:17], v[26:27], v[24:25]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v24, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v28, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v25, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v29, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[30:31], v[20:21], v[24:25], v[28:29]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v32, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v33, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_mul_f64 v[12:13], v[6:7], v[12:13]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[30:31], v[20:21], v[30:31], v[32:33]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v34, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v35, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v9, 0x80000000, v13
|
|
; SI-GISEL-NEXT: v_fma_f64 v[30:31], v[20:21], v[30:31], v[34:35]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v36, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v9, 0x43300000, v9
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v37, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_add_f64 v[38:39], v[12:13], v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[30:31], v[20:21], v[30:31], v[36:37]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v48, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v49, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_add_f64 v[8:9], v[38:39], -v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[30:31], v[20:21], v[30:31], v[48:49]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v38, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v39, 0x3f811111
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[12:13]|, v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[20:21], v[30:31], v[38:39]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v30, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v31, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v8, v8, v12, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v9, v9, v13, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[20:21], v[10:11], v[30:31]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v12, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[20:21], v[10:11], v[12:13]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v50, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v51, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[20:21], v[10:11], v[50:51]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v52, v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[20:21], v[10:11], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], v[20:21], v[10:11], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v20, v[16:17]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], -v[14:15], v[22:23], v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], -v[14:15], v[26:27], v[16:17]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[10:11], v[10:11], v20
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[24:25], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[34:35]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[36:37]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[48:49]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[38:39]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[12:13]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], v[50:51]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[16:17], v[20:21], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[16:17], v[20:21], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], -v[18:19], v[22:23], v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], -v[8:9], v[22:23], v[6:7]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], -v[18:19], v[26:27], v[20:21]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], -v[8:9], v[26:27], v[22:23]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[26:27], v[20:21], v[24:25], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[26:27], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[34:35]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[34:35]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[36:37]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v28, 0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[48:49]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[36:37]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[38:39]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v29, 0x40900000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[30:31]
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[12:13]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[48:49]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], v[50:51]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v53, 0, v10, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[38:39]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v10, v[18:19]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[24:25], v[22:23], v[24:25], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[20:21], v[14:15], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v18, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v26, 0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[24:25], v[12:13]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v19, v18, v11, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[10:11], v[14:15], v10
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[4:5], v[28:29]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v27, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[22:23], v[12:13], v[50:51]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v10, 0, v10, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v11, v18, v11, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[26:27]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[16:17], v[16:17], v52
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[2:3], v[28:29]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v14, v[8:9]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v53, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v19, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[2:3], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[22:23], v[12:13], 1.0
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v16, 0, v16, s[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v17, v18, v17, s[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v16, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v17, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[4:5], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[22:23], v[8:9], 1.0
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v10, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v5, 0, v11, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[8:9], v[8:9], v14
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[6:7], v[28:29]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v8, 0, v8, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v9, v18, v9, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[6:7], v[26:27]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v6, 0, v8, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v7, 0, v9, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_v4f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s16, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s17, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[8:9], v[0:1], s[16:17]
|
|
; VI-SDAG-NEXT: s_mov_b32 s18, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s19, 0xbfe62e42
|
|
; VI-SDAG-NEXT: s_mov_b32 s20, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s21, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s22, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[10:11], v[8:9]
|
|
; VI-SDAG-NEXT: s_mov_b32 s23, 0x3e5ade15
|
|
; VI-SDAG-NEXT: s_mov_b32 s24, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s25, 0x3ec71dee
|
|
; VI-SDAG-NEXT: s_mov_b32 s26, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s27, 0x3efa0199
|
|
; VI-SDAG-NEXT: s_mov_b32 s28, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s29, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[10:11], s[18:19], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s40, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s41, 0x3f56c16c
|
|
; VI-SDAG-NEXT: s_mov_b32 s42, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s43, 0x3f811111
|
|
; VI-SDAG-NEXT: s_mov_b32 s44, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s45, 0x3fa55555
|
|
; VI-SDAG-NEXT: s_mov_b32 s46, 0x55555511
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], s[20:21], v[8:9]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v9, s5
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, s4
|
|
; VI-SDAG-NEXT: s_mov_b32 s47, 0x3fc55555
|
|
; VI-SDAG-NEXT: s_mov_b32 s56, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s57, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s58, 0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], s[22:23], v[8:9]
|
|
; VI-SDAG-NEXT: s_mov_b32 s60, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s59, 0x40900000
|
|
; VI-SDAG-NEXT: s_mov_b32 s61, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[58:59], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[60:61], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[60:61], v[2:3]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[60:61], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[24:25]
|
|
; VI-SDAG-NEXT: s_and_b64 s[6:7], s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[40:41]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[42:43]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[44:45]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[46:47]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[56:57]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[12:13], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[10:11], v[12:13], v10
|
|
; VI-SDAG-NEXT: v_mul_f64 v[12:13], v[2:3], s[16:17]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v10, s[6:7]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[12:13], v[12:13]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[58:59], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], s[18:19], v[2:3]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v1, v[12:13]
|
|
; VI-SDAG-NEXT: s_and_b64 s[10:11], s[8:9], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], s[20:21], v[14:15]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], s[22:23], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[40:41]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[42:43]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[44:45]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[46:47]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[56:57]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[14:15], v[16:17], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[12:13], v[14:15], v1
|
|
; VI-SDAG-NEXT: v_mul_f64 v[14:15], v[4:5], s[16:17]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v12, s[10:11]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[14:15], v[14:15]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[58:59], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], s[18:19], v[4:5]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v1, v[14:15]
|
|
; VI-SDAG-NEXT: s_and_b64 s[14:15], s[12:13], s[10:11]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], s[20:21], v[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[22:23], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[40:41]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[42:43]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[44:45]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[46:47]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[56:57]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[16:17], v[18:19], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[14:15], v[16:17], v1
|
|
; VI-SDAG-NEXT: v_mul_f64 v[16:17], v[6:7], s[16:17]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[16:17], s[60:61], v[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v4, 0, v14, s[14:15]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[16:17], v[16:17]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[14:15], s[58:59], v[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[18:19], v[6:7]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v1, v[16:17]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x7ff00000
|
|
; VI-SDAG-NEXT: s_and_b64 s[18:19], s[16:17], s[14:15]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, v7, v13, s[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v5, v7, v15, s[10:11]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v5, s[12:13]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[20:21], v[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], s[22:23], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[40:41]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[42:43]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[44:45]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[46:47]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[56:57]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[8:9], v[8:9], v1
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, v7, v11, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v7, v7, v9, s[14:15]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v6, 0, v8, s[18:19]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v7, 0, v7, s[16:17]
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_v4f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[10:11], v[0:1], v[8:9]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[12:13], v[2:3], v[8:9]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[16:17], v[4:5], v[8:9]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[8:9], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[14:15], v[10:11]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[12:13], v[12:13]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[10:11], v[16:17]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[8:9], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v16, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v17, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], -v[14:15], v[16:17], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], -v[12:13], v[16:17], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], -v[10:11], v[16:17], v[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], -v[8:9], v[16:17], v[6:7]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], -v[14:15], v[26:27], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], -v[12:13], v[26:27], v[20:21]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], -v[10:11], v[26:27], v[22:23]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], -v[8:9], v[26:27], v[24:25]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v12, v[12:13]
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v13, v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[26:27], v[18:19], v[28:29], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[28:29], v[16:17], v[30:31], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[30:31], v[22:23], v[32:33], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[16:17], v[28:29], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v31, v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], v[24:25], v[20:21], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v26, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v27, 0x40900000
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v28, 0
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[10:11], v[18:19], v31
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v18, v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v29, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[8:9], v[16:17], v12
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[12:13], v[22:23], v13
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[4:5], v[26:27]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[14:15], v[14:15], v18
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[6:7], v[26:27]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[6:7], v[0:1], v[26:27]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[10:11], v[2:3], v[26:27]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], v[0:1], v[28:29]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[12:13], v[2:3], v[28:29]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[14:15], v[4:5], v[28:29]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[16:17], v[6:7], v[28:29]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v30, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v12, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v10, 0, v10, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v8, s[10:11]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v5, 0, v14, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v7, v30, v11, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v8, v30, v9, s[10:11]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v9, v30, v13, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v11, v30, v15, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v10, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[12:13]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v3, s[14:15]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v6, 0, v5, s[16:17]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v7, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v8, s[12:13]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v5, 0, v9, s[14:15]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v7, 0, v11, s[16:17]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_v4f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s16, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s17, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[8:9], v[0:1], s[16:17]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s18, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s19, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s20, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s21, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s22, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[10:11], v[8:9]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s23, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s24, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s25, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s26, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s27, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s28, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s29, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[10:11], s[18:19], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s40, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s41, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s42, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s43, 0x3f811111
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s44, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s45, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s46, 0x55555511
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], s[20:21], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v9, s5
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, s4
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s47, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s56, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s57, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s58, 0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], s[22:23], v[8:9]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s60, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s59, 0x40900000
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s61, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[58:59], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[60:61], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[60:61], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[60:61], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[24:25]
|
|
; GFX900-SDAG-NEXT: s_and_b64 s[6:7], s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[40:41]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[42:43]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[44:45]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[46:47]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], s[56:57]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[12:13], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[10:11], v[12:13], v10
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[12:13], v[2:3], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v10, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[12:13], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[58:59], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], s[18:19], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v1, v[12:13]
|
|
; GFX900-SDAG-NEXT: s_and_b64 s[10:11], s[8:9], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[12:13], s[20:21], v[14:15]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], s[22:23], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[40:41]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[42:43]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[44:45]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[46:47]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], s[56:57]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], v[16:17], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[14:15], v[16:17], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[12:13], v[14:15], v1
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[14:15], v[4:5], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v12, s[10:11]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[14:15], v[14:15]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[58:59], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], s[18:19], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v1, v[14:15]
|
|
; GFX900-SDAG-NEXT: s_and_b64 s[14:15], s[12:13], s[10:11]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[14:15], s[20:21], v[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[22:23], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[40:41]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[42:43]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[44:45]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[46:47]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], s[56:57]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], v[18:19], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[16:17], v[18:19], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[14:15], v[16:17], v1
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[16:17], v[6:7], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[16:17], s[60:61], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v4, 0, v14, s[14:15]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[16:17], v[16:17]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[14:15], s[58:59], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[18:19], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v1, v[16:17]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: s_and_b64 s[18:19], s[16:17], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, v7, v13, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v5, v7, v15, s[10:11]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v5, s[12:13]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[20:21], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], s[22:23], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[40:41]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[42:43]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[44:45]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[46:47]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], s[56:57]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[18:19], v[8:9], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[8:9], v[8:9], v1
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, v7, v11, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v7, v7, v9, s[14:15]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v6, 0, v8, s[18:19]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v7, 0, v7, s[16:17]
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_v4f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[10:11], v[0:1], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[12:13], v[2:3], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[16:17], v[4:5], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[8:9], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[14:15], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[12:13], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[10:11], v[16:17]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[8:9], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v16, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v17, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], -v[14:15], v[16:17], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], -v[12:13], v[16:17], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], -v[10:11], v[16:17], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], -v[8:9], v[16:17], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], -v[14:15], v[26:27], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], -v[12:13], v[26:27], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], -v[10:11], v[26:27], v[22:23]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], -v[8:9], v[26:27], v[24:25]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v12, v[12:13]
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v13, v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[18:19], v[28:29], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[16:17], v[30:31], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[32:33], v[22:23], v[32:33], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[26:27], v[18:19], v[28:29], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[28:29], v[16:17], v[30:31], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[30:31], v[22:23], v[32:33], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[24:25], v[20:21], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[18:19], v[26:27], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[16:17], v[28:29], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[22:23], v[30:31], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v31, v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], v[24:25], v[20:21], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v26, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v27, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v28, 0
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[10:11], v[18:19], v31
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v18, v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v29, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[8:9], v[16:17], v12
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[12:13], v[22:23], v13
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[4:5], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[14:15], v[14:15], v18
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], v[6:7], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[6:7], v[0:1], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[10:11], v[2:3], v[26:27]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], v[0:1], v[28:29]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[12:13], v[2:3], v[28:29]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[14:15], v[4:5], v[28:29]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[16:17], v[6:7], v[28:29]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v30, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v12, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v10, 0, v10, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v8, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v5, 0, v14, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v7, v30, v11, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v8, v30, v9, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v9, v30, v13, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v11, v30, v15, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v10, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v3, s[14:15]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v6, 0, v5, s[16:17]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v7, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v8, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v5, 0, v9, s[14:15]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v7, 0, v11, s[16:17]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call <4 x double> @llvm.exp.v2f64(<4 x double> %in)
|
|
ret <4 x double> %result
|
|
}
|
|
|
|
define amdgpu_ps <2 x i32> @s_exp_f64(double inreg %in) #0 {
|
|
; SI-SDAG-LABEL: s_exp_f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; SI-SDAG-NEXT: s_brev_b32 s2, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v2, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v3, s2, v2, v1
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v2, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[0:1], v[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[2:3], v[4:5], -v[2:3]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v3, s1
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v2, s0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[0:1], s[2:3], v[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[0:1], s[2:3], v[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: s_mov_b32 s2, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s3, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v6, v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[2:3], v[4:5], 1.0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v2, 0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[0:1], v6
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[0:1], s[0:1], v[4:5]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[0:1], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s0, v0
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s1, v1
|
|
; SI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; SI-GISEL-LABEL: s_exp_f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v3, 0x80000000, v1
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v3, 0x43300000, v3
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[2:3], v[4:5], -v[2:3]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], -v[0:1], v[4:5], s[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], -v[0:1], v[4:5], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[2:3], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[0:1], v6
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s1, v1
|
|
; SI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-SDAG-LABEL: s_exp_f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0xfefa39ef
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v3, s1
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v2, s0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v5, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[2:3], v[0:1], s[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[2:3], v[0:1], s[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v0, v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[2:3], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s2, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s3, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], 1.0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[0:1], v[4:5]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[2:3], v0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[2:3], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[2:3]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s1, v1
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s0, v0
|
|
; VI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-GISEL-LABEL: s_exp_f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[2:3], -v[0:1], v[2:3], s[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[2:3], -v[0:1], v[4:5], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v0, v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], 1.0
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[2:3], v0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[2:3]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[2:3]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; VI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-SDAG-LABEL: s_exp_f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v3, s1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v2, s0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v5, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[2:3], v[0:1], s[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[2:3], v[0:1], s[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v0, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[2:3], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s2, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s3, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], s[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], 1.0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[0:1], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[0:1], v[2:3], v0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[2:3], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[2:3]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s1, v1
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s0, v0
|
|
; GFX900-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-GISEL-LABEL: s_exp_f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[2:3], -v[0:1], v[2:3], s[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[2:3], -v[0:1], v[4:5], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v0, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[2:3], v[4:5], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], 1.0
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[0:1], v[2:3], v0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[2:3]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[2:3]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; GFX900-GISEL-NEXT: ; return to shader part epilog
|
|
%result = call double @llvm.exp.f64(double %in)
|
|
%cast = bitcast double %result to <2 x i32>
|
|
ret <2 x i32> %cast
|
|
}
|
|
|
|
define amdgpu_ps <4 x i32> @s_exp_v2f64(<2 x double> inreg %in) #0 {
|
|
; SI-SDAG-LABEL: s_exp_v2f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v0, s4
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v1, s5
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], s[2:3], v[0:1]
|
|
; SI-SDAG-NEXT: s_brev_b32 s26, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v12, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s26, v12, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, s3
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, s2
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[2:3], s[6:7], v[5:6]
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0x3b39803f
|
|
; SI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[8:9], v[5:6]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s26, v12, v1
|
|
; SI-SDAG-NEXT: v_add_f64 v[12:13], v[0:1], v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[12:13], -v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, s1
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, s0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[0:1], s[6:7], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v8, s10
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v9, s11
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[0:1], s[8:9], v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], s[12:13], v[8:9]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[12:13], v[8:9]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[10:11]
|
|
; SI-SDAG-NEXT: s_mov_b32 s14, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s15, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], s[10:11]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[14:15]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], s[14:15]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s18, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s19, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v12, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[8:9], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[18:19]
|
|
; SI-SDAG-NEXT: s_mov_b32 s20, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s21, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[20:21]
|
|
; SI-SDAG-NEXT: s_mov_b32 s22, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s23, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[22:23]
|
|
; SI-SDAG-NEXT: s_mov_b32 s24, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s25, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[24:25]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], s[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], s[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[10:11], v[6:7], v[10:11], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], 1.0
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v13, v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[10:11], 1.0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v10, 0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v9, 0x40900000
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v11, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[4:5], v[2:3], 1.0
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[6:7], v[6:7], v12
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[2:3], v[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[2:3], v[10:11]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v12, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[0:1], v13
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[0:1], v[8:9]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[0:1], s[0:1], v[10:11]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v7, v12, v7, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[2:3], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, v12, v1, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, 0, v6, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v7, s[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s0, v0
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s2, v3
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s1, v1
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s3, v2
|
|
; SI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; SI-GISEL-LABEL: s_exp_v2f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mul_f64 v[0:1], s[2:3], v[0:1]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v1
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[10:11], s[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], -v[0:1], v[10:11], s[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[4:5], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[10:11], v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[10:11], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v16, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v17, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[16:17]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v10, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[8:9], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[6:7], v[12:13], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x40900000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], v[16:17]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[6:7], v[6:7], v10
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v13, v[0:1]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v11, 0, v6, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v12, v10, v7, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[4:5], v[2:3], 1.0
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[6:7]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[0:1], v13
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[0:1], s[2:3], v[8:9]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v11, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v12, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[2:3], v[6:7]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[0:1]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e64 v1, v10, v1, s[0:1]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s0, v2
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s1, v3
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s2, v0
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; SI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-SDAG-LABEL: s_exp_v2f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v0, s4
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v1, s5
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], s[2:3], v[0:1]
|
|
; VI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v5, s3
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, s1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, s2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, s0
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v9, s7
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, s6
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[0:1], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[0:1], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v3, v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], s[4:5], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], s[4:5], v[8:9]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[10:11], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[8:9], 1.0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[6:7], v3
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[2:3], v[4:5]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[2:3], v[6:7]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[0:1], v[4:5]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[0:1], v[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, v8, v1, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[8:9], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, v8, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[6:7]
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s3, v1
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s1, v3
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s0, v1
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s2, v0
|
|
; VI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-GISEL-LABEL: s_exp_v2f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[0:1], s[2:3], v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[4:5], s[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[4:5], s[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[8:9], v[6:7]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[8:9], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v10, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], v[6:7], v[12:13], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[6:7], v[10:11], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[8:9], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[6:7], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v8
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v4, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, v6, v5, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v4, s[8:9]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v2, s[8:9]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s2, v3
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; VI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-SDAG-LABEL: s_exp_v2f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v0, s4
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, s5
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], s[2:3], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v5, s3
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, s1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, s2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, s0
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v9, s7
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, s6
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[0:1], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[0:1], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v3, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], s[4:5], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], s[4:5], v[8:9]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[4:5], v[10:11], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[6:7], v[8:9], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[10:11], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[8:9], 1.0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[0:1], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[6:7], v3
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[2:3], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[2:3], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[0:1], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[0:1], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, v8, v1, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[8:9], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, v8, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s3, v1
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s1, v3
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s0, v1
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s2, v0
|
|
; GFX900-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-GISEL-LABEL: s_exp_v2f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[0:1], s[2:3], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[4:5], s[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[4:5], s[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[8:9], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[8:9], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v10, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], v[6:7], v[12:13], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[4:5], v[8:9], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[6:7], v[10:11], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[8:9], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[0:1], v[6:7], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v8
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v4, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, v6, v5, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v4, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v2, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s2, v3
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; GFX900-GISEL-NEXT: ; return to shader part epilog
|
|
%result = call <2 x double> @llvm.exp.v2f64(<2 x double> %in)
|
|
%cast = bitcast <2 x double> %result to <4 x i32>
|
|
ret <4 x i32> %cast
|
|
}
|
|
|
|
define amdgpu_ps <6 x i32> @s_exp_v3f64(<3 x double> inreg %in) #0 {
|
|
; SI-SDAG-LABEL: s_exp_v3f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v0, s6
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v1, s7
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_brev_b32 s33, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v15, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v15, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[10:11]
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, s5
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, s4
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[2:3], s[8:9], v[5:6]
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x3e928af3
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[2:3], s[12:13], v[5:6]
|
|
; SI-SDAG-NEXT: s_mov_b32 s14, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v8, s7
|
|
; SI-SDAG-NEXT: s_mov_b32 s15, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, s6
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], s[14:15], v[7:8]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s18, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s19, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[18:19]
|
|
; SI-SDAG-NEXT: s_mov_b32 s20, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s21, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[20:21]
|
|
; SI-SDAG-NEXT: s_mov_b32 s22, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s23, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[22:23]
|
|
; SI-SDAG-NEXT: s_mov_b32 s24, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s25, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[24:25]
|
|
; SI-SDAG-NEXT: s_mov_b32 s26, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s27, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[26:27]
|
|
; SI-SDAG-NEXT: s_mov_b32 s28, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s29, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[28:29]
|
|
; SI-SDAG-NEXT: s_mov_b32 s30, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s31, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], s[30:31]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v11, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[5:6], v[9:10], 1.0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v19, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[5:6], v[9:10], 1.0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v9, 0
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v11
|
|
; SI-SDAG-NEXT: v_mul_f64 v[11:12], s[2:3], v[0:1]
|
|
; SI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v15, v12
|
|
; SI-SDAG-NEXT: v_add_f64 v[13:14], v[11:12], v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[11:12]|, s[10:11]
|
|
; SI-SDAG-NEXT: v_add_f64 v[5:6], v[13:14], -v[4:5]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v10, 0x40900000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v12, v6, v12, s[6:7]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v11, v5, v11, s[6:7]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, s3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, s2
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[11:12], s[8:9], v[5:6]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[9:10]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[11:12], s[12:13], v[5:6]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v15, v1
|
|
; SI-SDAG-NEXT: v_add_f64 v[15:16], v[0:1], v[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v20, v19, v3, vcc
|
|
; SI-SDAG-NEXT: v_add_f64 v[3:4], v[15:16], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[0:1]|, s[10:11]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[13:14], s[14:15], v[7:8]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v16, s1
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, v4, v1, s[6:7]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v0, v3, v0, s[6:7]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v15, s0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[17:18], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[0:1], s[8:9], v[15:16]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[0:1], s[12:13], v[15:16]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], s[14:15], v[7:8]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[18:19]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[26:27]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[20:21]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[28:29]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[22:23]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], s[30:31]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[24:25]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], 1.0
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v11, v[11:12]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[26:27]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[13:14], v[5:6], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[28:29]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[5:6], v[5:6], v11
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[15:16], v[7:8], s[30:31]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[6:7], s[2:3], v[9:10]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v3, 0
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v11, v19, v6, s[6:7]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[15:16], v[7:8], 1.0
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[4:5], v[3:4]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[15:16], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[2:3], v[3:4]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[0:1], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[0:1], v[9:10]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[0:1], s[0:1], v[3:4]
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[2:3], s[6:7]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, v19, v1, s[8:9]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, 0, v5, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[0:1], s[8:9]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v17, 0, v20, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v6, 0, v11, s[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s0, v0
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s2, v3
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s4, v2
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s1, v1
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s3, v6
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s5, v17
|
|
; SI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; SI-GISEL-LABEL: s_exp_v3f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_mul_f64 v[6:7], s[2:3], v[0:1]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v7
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[14:15], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], -v[2:3], v[10:11], s[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v16, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_add_f64 v[14:15], v[14:15], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[6:7]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v17, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], -v[2:3], v[16:17], v[12:13]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v6, v14, v6, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v7, v15, v7, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v18, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_mul_f64 v[0:1], s[4:5], v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[12:13], v[14:15], v[18:19]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v22, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v23, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v1
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[12:13], v[20:21], v[22:23]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v24, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v25, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_add_f64 v[26:27], v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[12:13], v[20:21], v[24:25]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v28, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v29, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[26:27], -v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[20:21], v[12:13], v[20:21], v[28:29]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v26, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[12:13], v[20:21], v[26:27]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v20, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3f811111
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[12:13], v[8:9], v[20:21]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[12:13], v[4:5], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v30, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v31, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[12:13], v[4:5], v[30:31]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v32, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v33, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[12:13], v[4:5], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[12:13], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[12:13], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], -v[6:7], v[10:11], s[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], -v[0:1], v[10:11], s[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], -v[6:7], v[16:17], v[12:13]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[10:11], -v[0:1], v[16:17], v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[14:15], v[18:19]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], v[18:19]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[22:23]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], v[22:23]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[24:25]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], v[24:25]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[14:15], v[10:11], v[14:15], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[20:21]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v18, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[30:31]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v18
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[12:13], v[16:17], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], v[10:11], v[14:15], v[20:21]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x40900000
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v17, v[6:7]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[10:11], v[12:13], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[14:15]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v16, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v5, v16, v5, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[8:9]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v17
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v12, 0, v4, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v13, 0, v5, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[10:11], v[6:7], v[30:31]
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[2:3], v[14:15]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[10:11], v[4:5], v[32:33]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v6, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v7, v16, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[10:11], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v4, v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[10:11], v[2:3], 1.0
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[2:3], v[8:9]
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s0, v12
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v6, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[0:1], v4
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v7, vcc
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[14:15]
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s1, v13
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v16, v1, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[8:9]
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s2, v2
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s3, v3
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s4, v0
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s5, v1
|
|
; SI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-SDAG-LABEL: s_exp_v3f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v0, s6
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v1, s7
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_mul_f64 v[4:5], s[2:3], v[0:1]
|
|
; VI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0xfefa39ef
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, s5
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v9, s3
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, s4
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, s2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v11, s1
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v10, s0
|
|
; VI-SDAG-NEXT: s_mov_b32 s8, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: s_mov_b32 s9, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[6:7], v[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[6:7], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], s[6:7], v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v13, s9
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v12, s8
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[6:7], v[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[6:7], v[8:9]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], s[6:7], v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v4, v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], s[6:7], v[12:13]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], s[6:7], v[12:13]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], s[6:7], v[12:13]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[16:17], 1.0
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v14, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[2:3], v[10:11], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[0:1]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[6:7], v14
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[4:5], v[8:9], v4
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x40900000
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v9, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[14:15], s[4:5], v[8:9]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[6:7], s[2:3], v[6:7]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[12:13], s[2:3], v[8:9]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v10
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[0:1], v[6:7]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[0:1], v[8:9]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v10, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[14:15], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[14:15]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[12:13], s[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v5, v10, v5, s[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, v10, v3, s[8:9]
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s5, v1
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v4, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[10:11], s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v5, s[12:13]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[10:11]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s1, v3
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s3, v5
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s0, v2
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s2, v1
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s4, v0
|
|
; VI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-GISEL-LABEL: s_exp_v3f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[4:5], s[2:3], v[0:1]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[0:1], s[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v12, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], -v[2:3], v[6:7], s[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], -v[4:5], v[6:7], s[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], -v[0:1], v[6:7], s[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], -v[2:3], v[12:13], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], -v[4:5], v[12:13], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], -v[0:1], v[12:13], v[6:7]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v12, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[14:15], v[12:13]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[14:15], v[12:13]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[14:15], v[12:13]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], v[8:9], v[16:17], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[18:19], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], v[8:9], v[14:15], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[16:17], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v16, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[2:3], v[6:7], v[12:13], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v14, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v15, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[6:7], s[2:3], v[14:15]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[8:9], v16
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[10:11], v6
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[4:5], v[14:15]
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[14:15]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v17, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, v17, v1, s[6:7]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[2:3], v[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, v17, v3, s[8:9]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[4:5], v[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v5, v17, v5, vcc
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[6:7]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s2, v0
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[6:7]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[8:9]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s4, v2
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v3, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v5, vcc
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s0, v4
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s5, v2
|
|
; VI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-SDAG-LABEL: s_exp_v3f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v0, s6
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, s7
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[4:5], s[2:3], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, s5
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v9, s3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, s4
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, s2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v11, s1
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v10, s0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s8, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s9, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[6:7], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[6:7], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], s[6:7], v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v13, s9
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v12, s8
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[6:7], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[6:7], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[0:1], s[6:7], v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v4, v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], s[6:7], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], s[6:7], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], s[6:7], v[12:13]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[6:7], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[10:11], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[16:17], 1.0
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v14, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[2:3], v[10:11], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[0:1], v[6:7], v14
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[4:5], v[8:9], v4
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v9, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[14:15], s[4:5], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[6:7], s[2:3], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[12:13], s[2:3], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v10
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[0:1], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[0:1], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v10, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[14:15], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[14:15]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[12:13], s[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v5, v10, v5, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, v10, v3, s[8:9]
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s5, v1
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v4, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[10:11], s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v5, s[12:13]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[10:11]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s1, v3
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s3, v5
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s0, v2
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s2, v1
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s4, v0
|
|
; GFX900-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-GISEL-LABEL: s_exp_v3f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[4:5], s[2:3], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[0:1], s[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v12, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v13, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], -v[2:3], v[6:7], s[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], -v[4:5], v[6:7], s[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], -v[0:1], v[6:7], s[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], -v[2:3], v[12:13], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], -v[4:5], v[12:13], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], -v[0:1], v[12:13], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v12, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v13, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[14:15], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[14:15], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[14:15], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[18:19], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], v[8:9], v[16:17], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[10:11], v[18:19], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], v[8:9], v[14:15], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[16:17], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v16, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[2:3], v[6:7], v[12:13], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v14, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v15, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[6:7], s[2:3], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[8:9], v16
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[0:1], v[10:11], v6
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[4:5], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v17, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, v17, v1, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[2:3], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, v17, v3, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[4:5], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v5, v17, v5, vcc
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v0, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s2, v0
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s4, v2
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v3, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v4, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v5, vcc
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s0, v4
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s5, v2
|
|
; GFX900-GISEL-NEXT: ; return to shader part epilog
|
|
%result = call <3 x double> @llvm.exp.v3f64(<3 x double> %in)
|
|
%cast = bitcast <3 x double> %result to <6 x i32>
|
|
ret <6 x i32> %cast
|
|
}
|
|
|
|
define amdgpu_ps <8 x i32> @s_exp_v4f64(<4 x double> inreg %in) #0 {
|
|
; SI-SDAG-LABEL: s_exp_v4f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v0, s8
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v1, s9
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: s_brev_b32 s33, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v19, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v19, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s12, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s13, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[12:13]
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, s6
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, s7
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[2:3], s[8:9], v[5:6]
|
|
; SI-SDAG-NEXT: s_mov_b32 s10, 0x3b39803f
|
|
; SI-SDAG-NEXT: v_mul_f64 v[8:9], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s11, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[10:11], v[5:6]
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v19, v9
|
|
; SI-SDAG-NEXT: v_add_f64 v[10:11], v[8:9], v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[8:9]|, s[12:13]
|
|
; SI-SDAG-NEXT: v_add_f64 v[10:11], v[10:11], -v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v9, v11, v9, vcc
|
|
; SI-SDAG-NEXT: v_mul_f64 v[11:12], s[2:3], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3e928af3
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v19, v12
|
|
; SI-SDAG-NEXT: v_add_f64 v[13:14], v[11:12], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s14, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v8, v10, v8, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v15, s16
|
|
; SI-SDAG-NEXT: v_add_f64 v[13:14], v[13:14], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[11:12]|, s[12:13]
|
|
; SI-SDAG-NEXT: s_mov_b32 s15, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v16, s17
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v12, v14, v12, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v11, v13, v11, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], s[14:15], v[15:16]
|
|
; SI-SDAG-NEXT: s_mov_b32 s16, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s17, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[16:17]
|
|
; SI-SDAG-NEXT: s_mov_b32 s18, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s19, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[18:19]
|
|
; SI-SDAG-NEXT: s_mov_b32 s20, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s21, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[20:21]
|
|
; SI-SDAG-NEXT: s_mov_b32 s22, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s23, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[22:23]
|
|
; SI-SDAG-NEXT: s_mov_b32 s24, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s25, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[24:25]
|
|
; SI-SDAG-NEXT: s_mov_b32 s26, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s27, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v18, s5
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[26:27]
|
|
; SI-SDAG-NEXT: s_mov_b32 s28, 0x55555511
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v17, s4
|
|
; SI-SDAG-NEXT: s_mov_b32 s29, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[8:9], s[8:9], v[17:18]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[28:29]
|
|
; SI-SDAG-NEXT: s_mov_b32 s30, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s31, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[8:9], s[10:11], v[17:18]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], s[30:31]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], s[14:15], v[15:16]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[13:14], v[6:7], v[13:14], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[5:6], v[6:7], v[13:14], 1.0
|
|
; SI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[18:19]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[6:7], v[5:6], v10
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s33, v19, v1
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[20:21]
|
|
; SI-SDAG-NEXT: v_add_f64 v[19:20], v[0:1], v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[22:23]
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[19:20], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, s[12:13]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[24:25]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, s3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, s2
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[26:27]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[11:12], s[8:9], v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[28:29]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[11:12], s[10:11], v[4:5]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], s[30:31]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[19:20], v[4:5], s[14:15], v[15:16]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[19:20], v[4:5], v[19:20], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[17:18], v[2:3], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[4:5], v[19:20], s[18:19]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v13, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[17:18], v[4:5], v[17:18], s[20:21]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[8:9]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v14, 0x40900000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], v[17:18], s[22:23]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[6:7], v[13:14]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v19, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v17, v19, v7, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[4:5], v[8:9], s[24:25]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v10
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v10, s1
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v9, s0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[4:5], v[7:8], s[26:27]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[0:1], s[8:9], v[9:10]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[4:5], v[7:8], s[28:29]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[9:10], v[0:1], s[10:11], v[9:10]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[4:5], v[7:8], s[30:31]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[9:10], s[14:15], v[15:16]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[4:5], v[7:8], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[15:16], v[9:10], v[15:16], s[16:17]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[7:8], 1.0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[15:16], s[18:19]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v15, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[7:8], s[20:21]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v11, v[11:12]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[7:8], s[22:23]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v16, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[7:8], s[24:25]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[4:5], v[13:14]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[7:8], s[26:27]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[4:5], v[15:16]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[7:8], s[28:29]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v3, v19, v3, s[8:9]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v12, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[3:4], v[4:5], v11
|
|
; SI-SDAG-NEXT: v_fma_f64 v[7:8], v[9:10], v[7:8], s[30:31]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[10:11], s[2:3], v[13:14]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[6:7], v[15:16]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v11, v19, v4, s[10:11]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[9:10], v[7:8], 1.0
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v7, v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[9:10], v[4:5], 1.0
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[2:3], v[15:16]
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[6:7], vcc
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[0:1], v7
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[0:1], v[13:14]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[0:1], s[0:1], v[15:16]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v5, 0, v6, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], s[8:9]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[2:3], s[10:11]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, v19, v1, s[12:13]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, 0, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[0:1], s[12:13]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v17, 0, v17, s[6:7]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v4, 0, v11, s[2:3]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s0, v0
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s2, v3
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s4, v2
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s6, v5
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s1, v1
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s3, v4
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s5, v12
|
|
; SI-SDAG-NEXT: v_readfirstlane_b32 s7, v17
|
|
; SI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; SI-GISEL-LABEL: s_exp_v4f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mul_f64 v[10:11], s[2:3], v[0:1]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v5, v2, vcc
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v11
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[10:11], v[4:5]
|
|
; SI-GISEL-NEXT: v_mul_f64 v[14:15], s[4:5], v[0:1]
|
|
; SI-GISEL-NEXT: v_add_f64 v[5:6], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[10:11]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v12, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v10, v5, v10, vcc
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v15
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[16:17], v[14:15], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v13, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v11, v6, v11, vcc
|
|
; SI-GISEL-NEXT: v_add_f64 v[5:6], v[16:17], -v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], -v[2:3], v[12:13], s[0:1]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[14:15]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v18, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v15, v6, v15, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[18:19], v[16:17]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v16, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v20, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v17, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v21, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[6:7], v[16:17], v[20:21]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v24, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v25, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_mul_f64 v[0:1], s[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[6:7], v[22:23], v[24:25]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v26, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v27, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v14, v5, v14, vcc
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v1
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[6:7], v[22:23], v[26:27]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v28, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v29, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_add_f64 v[30:31], v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[6:7], v[22:23], v[28:29]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v32, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v33, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[30:31], -v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[22:23], v[6:7], v[22:23], v[32:33]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v30, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v31, 0x3f811111
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[8:9], v[6:7], v[22:23], v[30:31]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v22, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v23, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[6:7], v[8:9], v[22:23]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[6:7], v[4:5], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v34, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v35, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[6:7], v[4:5], v[34:35]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[6:7], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], v[6:7], v[4:5], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], -v[10:11], v[12:13], s[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], -v[10:11], v[18:19], v[6:7]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v10, v[10:11]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[16:17], v[20:21]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[24:25]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[22:23]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], v[34:35]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], v[6:7], v[36:37], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[6:7], v[36:37], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], -v[14:15], v[12:13], s[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], -v[0:1], v[12:13], s[6:7]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[36:37], -v[14:15], v[18:19], v[36:37]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[12:13], -v[0:1], v[18:19], v[12:13]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[16:17], v[20:21]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[20:21]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[24:25]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[24:25]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[26:27]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[28:29]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v20, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[28:29]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[32:33]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[22:23]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v20
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[8:9]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[30:31]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[18:19], v[36:37], v[18:19], v[34:35]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[6:7], v[6:7], v10
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[36:37], v[18:19], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v18, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v19, 0x40900000
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[18:19]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[16:17], v[12:13], v[16:17], v[22:23]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v20, 0, v4, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v10, v11, v5, vcc
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[2:3], v[18:19]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v21, 0, v6, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v22, v11, v7, vcc
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v14, v[14:15]
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[36:37], v[2:3], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[12:13], v[16:17], v[8:9]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v8, 0, v20, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v9, 0, v10, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v14
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v10, 0, v21, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v14, 0, v22, vcc
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[18:19]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[12:13], v[6:7], v[34:35]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v15, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v16, v11, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[12:13], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[12:13], v[2:3], 1.0
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[4:5]
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s0, v8
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v15, vcc
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[0:1], v6
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, 0, v16, vcc
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[6:7], v[18:19]
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s1, v9
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v11, v1, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s2, v10
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s3, v14
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s4, v2
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s5, v3
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s6, v0
|
|
; SI-GISEL-NEXT: v_readfirstlane_b32 s7, v1
|
|
; SI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-SDAG-LABEL: s_exp_v4f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_mov_b32 s8, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s9, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v0, s8
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v1, s9
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s8, 0xfefa39ef
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, s6
|
|
; VI-SDAG-NEXT: s_mov_b32 s9, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, s7
|
|
; VI-SDAG-NEXT: v_mul_f64 v[4:5], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s10, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s11, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: s_mov_b32 s14, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: s_mov_b32 s15, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s12, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v10, s14
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s13, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v11, s15
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[8:9], v[6:7]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v9, s5
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, s4
|
|
; VI-SDAG-NEXT: s_mov_b32 s14, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s15, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[8:9], v[8:9]
|
|
; VI-SDAG-NEXT: s_mov_b32 s16, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s17, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[10:11], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s18, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s19, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: s_mov_b32 s20, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s21, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[10:11], v[8:9]
|
|
; VI-SDAG-NEXT: s_mov_b32 s22, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s23, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], s[12:13], v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s24, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s25, 0x3fa55555
|
|
; VI-SDAG-NEXT: s_mov_b32 s26, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s27, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[12:13], v[10:11]
|
|
; VI-SDAG-NEXT: s_mov_b32 s28, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s29, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[14:15]
|
|
; VI-SDAG-NEXT: v_mul_f64 v[16:17], s[2:3], v[0:1]
|
|
; VI-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v19, s3
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v18, s2
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[14:15]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v20, v[2:3]
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v21, v[4:5]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[16:17]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[16:17], v[16:17]
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[8:9], v[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[10:11], v[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[12:13], 1.0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v13, s1
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v12, s0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[0:1], s[8:9], v[12:13]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], s[12:13], v[10:11]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[4:5], v[6:7], v20
|
|
; VI-SDAG-NEXT: v_fma_f64 v[12:13], v[0:1], s[10:11], v[12:13]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[14:15]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], s[12:13], v[10:11]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[14:15]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[16:17]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[20:21]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[22:23]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[24:25]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[26:27]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[28:29]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], 1.0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v18, 0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v19, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[0:1], v[18:19]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[6:7], v[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], 1.0
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[4:5], v[18:19]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[10:11], s[2:3], v[18:19]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[2:3], v[12:13], v[10:11], 1.0
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v11, v[0:1]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[8:9], v21
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v8, 0
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[16:17]
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v9, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[6:7], v[8:9]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[14:15], s[0:1], v[8:9]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v11
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[4:5], v[8:9]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[2:3], v[8:9]
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[6:7], v[14:15], v10
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v12, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v5, v12, v5, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[6:7], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, v12, v1, s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, v12, v3, s[12:13]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[14:15]
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s1, v3
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, 0, v4, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], s[8:9]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[2:3], s[10:11]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v7, v12, v7, s[10:11]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v4, 0, v6, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[14:15], s[12:13]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v5, s[6:7]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v7, 0, v7, s[2:3]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s0, v2
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s2, v4
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s4, v0
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s6, v3
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s3, v7
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s5, v1
|
|
; VI-SDAG-NEXT: v_readfirstlane_b32 s7, v5
|
|
; VI-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; VI-GISEL-LABEL: s_exp_v4f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[4:5], s[2:3], v[0:1]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[6:7], s[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_mul_f64 v[0:1], s[6:7], v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v16, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v17, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], -v[2:3], v[8:9], s[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], -v[4:5], v[8:9], s[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], -v[6:7], v[8:9], s[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], -v[0:1], v[8:9], s[6:7]
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v26, v[4:5]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], -v[2:3], v[16:17], v[10:11]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], -v[4:5], v[16:17], v[12:13]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[14:15], -v[6:7], v[16:17], v[14:15]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[8:9], -v[0:1], v[16:17], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v16, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v17, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[18:19], v[16:17]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[18:19], v[16:17]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[18:19], v[16:17]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[18:19], v[16:17]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[20:21], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[20:21], v[12:13], v[22:23], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[24:25], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v24, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v25, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[18:19], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[12:13], v[12:13], v[20:21], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[2:3], v[14:15], v[22:23], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[8:9], v[16:17], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v14, v[6:7]
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v18, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v19, 0x40900000
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[10:11], v24
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[18:19]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v14
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[10:11], s[4:5], v[18:19]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[6:7], v[12:13], v26
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[2:3], v[18:19]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v8
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[6:7], v[18:19]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v20, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v21, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v25, v1, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[10:11]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, v25, v3, s[10:11]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[20:21]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[4:5], v[20:21]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v6, 0, v6, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v7, v25, v7, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v4, s[12:13]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v5, v25, v5, s[12:13]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[2:3], v[20:21]
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[12:13], s[6:7], v[20:21]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[10:11]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s4, v2
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v3, s[10:11]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v6, 0, v6, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v4, s[12:13]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v7, s[8:9]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v5, s[12:13]
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s2, v6
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s6, v4
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s5, v2
|
|
; VI-GISEL-NEXT: v_readfirstlane_b32 s7, v3
|
|
; VI-GISEL-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-SDAG-LABEL: s_exp_v4f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s8, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s9, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v0, s8
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v1, s9
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s8, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, s6
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s9, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, s7
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[4:5], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s10, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s11, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s14, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s15, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s12, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v10, s14
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s13, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v11, s15
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[8:9], v[6:7]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v9, s5
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, s4
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s14, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s15, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[8:9], v[8:9]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s16, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s17, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[2:3], s[10:11], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s18, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s19, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s20, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s21, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[4:5], s[10:11], v[8:9]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s22, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s23, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], s[12:13], v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s24, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s25, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s26, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s27, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], s[12:13], v[10:11]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s28, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s29, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[16:17], s[2:3], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[0:1], s[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v19, s3
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v18, s2
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v20, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v21, v[4:5]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[16:17], v[16:17]
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[8:9], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[18:19], v[16:17], s[10:11], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[6:7], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[8:9], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[6:7], v[12:13], 1.0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v13, s1
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v12, s0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[0:1], s[8:9], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[8:9], v[8:9], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], s[12:13], v[10:11]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[4:5], v[6:7], v20
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[12:13], v[0:1], s[10:11], v[12:13]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], s[12:13], v[10:11]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[14:15]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[16:17]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[20:21]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[22:23]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[24:25]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[26:27]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], s[28:29]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[14:15], v[18:19], v[14:15], 1.0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v18, 0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v19, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[0:1], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[6:7], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[10:11], v[12:13], v[10:11], 1.0
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[4:5], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[10:11], s[2:3], v[18:19]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[2:3], v[12:13], v[10:11], 1.0
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v11, v[0:1]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[0:1], v[8:9], v21
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v8, 0
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v10, v[16:17]
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v9, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[6:7], s[6:7], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[14:15], s[0:1], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v11
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[4:5], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[2:3], s[2:3], v[8:9]
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[6:7], v[14:15], v10
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v12, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v5, v12, v5, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[6:7], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, v12, v1, s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, v12, v3, s[12:13]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v3, s[14:15]
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s1, v3
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, 0, v4, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], s[8:9]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[2:3], s[10:11]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v7, v12, v7, s[10:11]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v4, 0, v6, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[14:15], s[12:13]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v5, 0, v5, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v7, 0, v7, s[2:3]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s0, v2
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s2, v4
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s4, v0
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s6, v3
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s3, v7
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s5, v1
|
|
; GFX900-SDAG-NEXT: v_readfirstlane_b32 s7, v5
|
|
; GFX900-SDAG-NEXT: ; return to shader part epilog
|
|
;
|
|
; GFX900-GISEL-LABEL: s_exp_v4f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], s[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[4:5], s[2:3], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[6:7], s[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[0:1], s[6:7], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v16, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v17, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[4:5], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[6:7], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[0:1], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], -v[2:3], v[8:9], s[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], -v[4:5], v[8:9], s[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], -v[6:7], v[8:9], s[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], -v[0:1], v[8:9], s[6:7]
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v26, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], -v[2:3], v[16:17], v[10:11]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], -v[4:5], v[16:17], v[12:13]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[14:15], -v[6:7], v[16:17], v[14:15]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[8:9], -v[0:1], v[16:17], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v16, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v17, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[18:19], v[16:17]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[18:19], v[16:17]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[18:19], v[16:17]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[18:19], v[16:17]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[10:11], v[20:21], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[12:13], v[22:23], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[24:25], v[14:15], v[24:25], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[18:19], v[10:11], v[20:21], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[20:21], v[12:13], v[22:23], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[22:23], v[14:15], v[24:25], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[16:17], v[8:9], v[16:17], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v24, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v25, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[10:11], v[10:11], v[18:19], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[12:13], v[12:13], v[20:21], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[2:3], v[14:15], v[22:23], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[8:9], v[16:17], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v14, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v18, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v19, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[0:1], v[10:11], v24
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, s[0:1], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v14
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[10:11], s[4:5], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[6:7], v[12:13], v26
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[8:9], s[2:3], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v8
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 s[12:13], s[6:7], v[18:19]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v20, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v21, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v25, v1, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, v25, v3, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, s[0:1], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[10:11], s[4:5], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v6, 0, v6, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v7, v25, v7, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v4, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v5, v25, v5, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[8:9], s[2:3], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[12:13], s[6:7], v[20:21]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v2, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s0, v0
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s4, v2
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v1, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v3, s[10:11]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v6, 0, v6, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v4, 0, v4, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v7, s[8:9]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v5, s[12:13]
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s2, v6
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s6, v4
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s1, v0
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s3, v1
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s5, v2
|
|
; GFX900-GISEL-NEXT: v_readfirstlane_b32 s7, v3
|
|
; GFX900-GISEL-NEXT: ; return to shader part epilog
|
|
%result = call <4 x double> @llvm.exp.v4f64(<4 x double> %in)
|
|
%cast = bitcast <4 x double> %result to <8 x i32>
|
|
ret <8 x i32> %cast
|
|
}
|
|
|
|
define double @v_exp_fabs_f64(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_fabs_f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], |v[0:1]|
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_fabs_f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], |v[0:1]|, v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], |v[0:1]|
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_fabs_f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], |v[0:1]|
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_fabs_f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], |v[0:1]|, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], |v[0:1]|
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_fabs_f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], |v[0:1]|
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_fabs_f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], |v[0:1]|, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], |v[0:1]|
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fabs = call double @llvm.fabs.f64(double %in)
|
|
%result = call double @llvm.exp.f64(double %fabs)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_fneg_fabs_f64(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_fneg_fabs_f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], -|v[0:1]|
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xc0900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x4090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_fneg_fabs_f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], -|v[0:1]|, v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], -|v[0:1]|
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, -|v[0:1]|, v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e64 vcc, -|v[0:1]|, v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_fneg_fabs_f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x4090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], -|v[0:1]|
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xc0900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_fneg_fabs_f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], -|v[0:1]|, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], -|v[0:1]|
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], -|v[0:1]|, v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, -|v[0:1]|, v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_fneg_fabs_f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x4090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], -|v[0:1]|
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xc0900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_fneg_fabs_f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], -|v[0:1]|, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], -|v[0:1]|
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], -|v[0:1]|, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, -|v[0:1]|, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fabs = call double @llvm.fabs.f64(double %in)
|
|
%fneg.fabs = fneg double %fabs
|
|
%result = call double @llvm.exp.f64(double %fneg.fabs)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_fneg_f64(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_fneg_f64:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], -v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xc0900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x4090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_fneg_f64:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], -v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], -v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, -v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e64 vcc, -v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_fneg_f64:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0x4090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], -v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xc0900000
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_fneg_f64:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], -v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], -v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], -v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, -v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_fneg_f64:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0x4090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], -v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xc0900000
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_fneg_f64:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], -v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], -v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], -v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, -v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fneg = fneg double %in
|
|
%result = call double @llvm.exp.f64(double %fneg)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_fast(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_fast:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_fast:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_fast:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_fast:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_fast:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_fast:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call fast double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_afn(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_afn:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_afn:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_afn:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_afn:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_afn:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_afn:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call afn double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_ninf(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_ninf:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_ninf:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_ninf:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_ninf:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_ninf:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_ninf:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call ninf double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_nnan(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_nnan:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_nnan:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_nnan:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_nnan:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_nnan:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_nnan:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call nnan double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_fabs_exp_f64_afn(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_fabs_exp_f64_afn:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], |v[0:1]|
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_fabs_exp_f64_afn:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], |v[0:1]|, v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], |v[0:1]|
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_fabs_exp_f64_afn:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], |v[0:1]|
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_fabs_exp_f64_afn:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], |v[0:1]|, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], |v[0:1]|
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_fabs_exp_f64_afn:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], |v[0:1]|, s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], |v[0:1]|
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, s[6:7]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_fabs_exp_f64_afn:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], |v[0:1]|, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], |v[0:1]|
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], |v[0:1]|, v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e64 vcc, |v[0:1]|, v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fabs = call double @llvm.fabs.f64(double %in)
|
|
%result = call afn double @llvm.exp.f64(double %fabs)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_nnan_ninf(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_nnan_ninf:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_nnan_ninf:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_nnan_ninf:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_nnan_ninf:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_nnan_ninf:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_nnan_ninf:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call nnan ninf double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_undef() #0 {
|
|
; GCN-SDAG-LABEL: v_exp_f64_undef:
|
|
; GCN-SDAG: ; %bb.0:
|
|
; GCN-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-SDAG-NEXT: v_mov_b32_e32 v1, 0
|
|
; GCN-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GCN-GISEL-LABEL: v_exp_f64_undef:
|
|
; GCN-GISEL: ; %bb.0:
|
|
; GCN-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call double @llvm.exp.f64(double poison)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_0() #0 {
|
|
; GCN-LABEL: v_exp_f64_0:
|
|
; GCN: ; %bb.0:
|
|
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GCN-NEXT: v_mov_b32_e32 v0, 0
|
|
; GCN-NEXT: v_mov_b32_e32 v1, 0x3ff00000
|
|
; GCN-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call double @llvm.exp.f64(double 0.0)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_from_fpext_f16(half %src) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_from_fpext_f16:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: s_brev_b32 s8, -2
|
|
; SI-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v5, 0x43300000
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, -1
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x432fffff
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s8, v5, v3
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[6:7]
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_from_fpext_f16:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x432fffff
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_mul_f64 v[0:1], v[2:3], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v1
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[12:13], v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, v[6:7]
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[12:13], -v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[8:9], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[10:11], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[14:15], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x555502a1
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 11
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v10, v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x40900000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[0:1], v10
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[2:3], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_from_fpext_f16:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_from_fpext_f16:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_from_fpext_f16:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_from_fpext_f16:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fpext = fpext half %src to double
|
|
%result = call double @llvm.exp.f64(double %fpext)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_from_fpext_f32(float %src) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_from_fpext_f32:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, -1
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_from_fpext_f32:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, -1
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[8:9], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[8:9], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], -v[2:3], v[10:11], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[2:3], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_from_fpext_f32:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_from_fpext_f32:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_from_fpext_f32:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_from_fpext_f32:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fpext = fpext float %src to double
|
|
%result = call double @llvm.exp.f64(double %fpext)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_from_fpext_math_f16(half %src0, half %src1) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_from_fpext_math_f16:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: s_brev_b32 s10, -2
|
|
; SI-SDAG-NEXT: v_add_f32_e32 v0, v0, v1
|
|
; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0x43300000
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0x432fffff
|
|
; SI-SDAG-NEXT: v_cvt_f32_f16_e32 v1, v0
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v0, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s8, 0xfefa39ef
|
|
; SI-SDAG-NEXT: s_mov_b32 s9, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_cvt_f64_f32_e32 v[2:3], v1
|
|
; SI-SDAG-NEXT: v_mul_f64 v[4:5], v[2:3], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: v_bfi_b32 v1, s10, v6, v5
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[4:5]|, s[6:7]
|
|
; SI-SDAG-NEXT: v_add_f64 v[0:1], v[6:7], -v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[0:1], s[8:9], v[2:3]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[0:1], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[0:1]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[0:1], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[0:1], v[0:1], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[2:3]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[2:3]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_from_fpext_math_f16:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f32_e32 v0, v0, v1
|
|
; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v0, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v14, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v15, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mul_f64 v[0:1], v[2:3], v[0:1]
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v1
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[12:13], v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[0:1]|, v[6:7]
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[12:13], -v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[8:9], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[0:1], v[10:11], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[14:15], v[6:7]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 0x555502a1
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v10, 11
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v11, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[10:11]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v10, v[0:1]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x40900000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[0:1], v[0:1], v10
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[2:3], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_from_fpext_math_f16:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
|
|
; VI-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_from_fpext_math_f16:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
|
|
; VI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_from_fpext_math_f16:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
|
|
; GFX900-SDAG-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_from_fpext_math_f16:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
|
|
; GFX900-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%fadd = fadd half %src0, %src1
|
|
%fpext = fpext half %fadd to double
|
|
%result = call double @llvm.exp.f64(double %fpext)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_contract(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_contract:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; SI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; SI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_contract:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x7ff00000
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v6, v3, vcc
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_contract:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; VI-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; VI-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_contract:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; VI-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_contract:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s6, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s7, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x40900000
|
|
; GFX900-SDAG-NEXT: v_cmp_nlt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[6:7], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v4, 0x7ff00000
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_and_b64 vcc, s[4:5], vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v3, s[4:5]
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_contract:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e64 s[4:5], v[0:1], v[6:7]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x40900000
|
|
; GFX900-GISEL-NEXT: v_cmp_ngt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v0, 0x7ff00000
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v3, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s[4:5]
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v1, s[4:5]
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call contract double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
define double @v_exp_f64_contract_nnan_ninf(double %in) #0 {
|
|
; SI-SDAG-LABEL: v_exp_f64_contract_nnan_ninf:
|
|
; SI-SDAG: ; %bb.0:
|
|
; SI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; SI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; SI-SDAG-NEXT: s_brev_b32 s4, -2
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0x43300000
|
|
; SI-SDAG-NEXT: v_bfi_b32 v5, s4, v4, v3
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-SDAG-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, -1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x432fffff
|
|
; SI-SDAG-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-SDAG-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; SI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; SI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; SI-SDAG-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; SI-SDAG-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; SI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; SI-GISEL-LABEL: v_exp_f64_contract_nnan_ninf:
|
|
; SI-GISEL: ; %bb.0:
|
|
; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; SI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_and_b32_e32 v5, 0x80000000, v3
|
|
; SI-GISEL-NEXT: v_or_b32_e32 v5, 0x43300000, v5
|
|
; SI-GISEL-NEXT: v_add_f64 v[6:7], v[2:3], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, -1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x432fffff
|
|
; SI-GISEL-NEXT: v_add_f64 v[4:5], v[6:7], -v[4:5]
|
|
; SI-GISEL-NEXT: v_cmp_gt_f64_e64 vcc, |v[2:3]|, v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0xfefa39ef
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3fe62e42
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[0:1]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; SI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; SI-GISEL-NEXT: v_cvt_i32_f64_e32 v8, v[2:3]
|
|
; SI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_fma_f64 v[2:3], v[4:5], v[6:7], 1.0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v4, 0
|
|
; SI-GISEL-NEXT: v_mov_b32_e32 v5, 0xc090cc00
|
|
; SI-GISEL-NEXT: v_ldexp_f64 v[2:3], v[2:3], v8
|
|
; SI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[4:5]
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; SI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; SI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-SDAG-LABEL: v_exp_f64_contract_nnan_ninf:
|
|
; VI-SDAG: ; %bb.0:
|
|
; VI-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; VI-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; VI-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; VI-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; VI-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; VI-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; VI-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; VI-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; VI-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; VI-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; VI-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; VI-GISEL-LABEL: v_exp_f64_contract_nnan_ninf:
|
|
; VI-GISEL: ; %bb.0:
|
|
; VI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; VI-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; VI-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; VI-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; VI-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; VI-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; VI-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; VI-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; VI-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; VI-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-SDAG-LABEL: v_exp_f64_contract_nnan_ninf:
|
|
; GFX900-SDAG: ; %bb.0:
|
|
; GFX900-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x652b82fe
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ff71547
|
|
; GFX900-SDAG-NEXT: v_mul_f64 v[2:3], v[0:1], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0xfefa39ef
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbfe62e42
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v6, 0xfca7ab0c
|
|
; GFX900-SDAG-NEXT: v_mov_b32_e32 v7, 0x3e928af3
|
|
; GFX900-SDAG-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x3b39803f
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xbc7abc9e
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[2:3], s[4:5], v[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x6a5dcb37
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3e5ade15
|
|
; GFX900-SDAG-NEXT: v_cvt_i32_f64_e32 v2, v[2:3]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], s[4:5], v[6:7]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x623fde64
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3ec71dee
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x7c89e6b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3efa0199
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x14761f6e
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f2a01a0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x1852b7b0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f56c16c
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x11122322
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3f811111
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x555502a1
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fa55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0x55555511
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fc55555
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 11
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0x3fe00000
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], s[4:5]
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s4, 0
|
|
; GFX900-SDAG-NEXT: s_mov_b32 s5, 0xc090cc00
|
|
; GFX900-SDAG-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[0:1]
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-SDAG-NEXT: v_ldexp_f64 v[2:3], v[4:5], v2
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v0, 0, v2, vcc
|
|
; GFX900-SDAG-NEXT: v_cndmask_b32_e32 v1, 0, v3, vcc
|
|
; GFX900-SDAG-NEXT: s_setpc_b64 s[30:31]
|
|
;
|
|
; GFX900-GISEL-LABEL: v_exp_f64_contract_nnan_ninf:
|
|
; GFX900-GISEL: ; %bb.0:
|
|
; GFX900-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0x652b82fe
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0x3ff71547
|
|
; GFX900-GISEL-NEXT: v_mul_f64 v[2:3], v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v4, 0xfefa39ef
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v5, 0x3fe62e42
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x3b39803f
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3c7abc9e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0xfca7ab0c
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3e928af3
|
|
; GFX900-GISEL-NEXT: v_rndne_f64_e32 v[2:3], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[4:5], v[0:1]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], -v[2:3], v[6:7], v[4:5]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v6, 0x6a5dcb37
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v7, 0x3e5ade15
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x623fde64
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3ec71dee
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x7c89e6b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3efa0199
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x14761f6e
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f2a01a0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x1852b7b0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f56c16c
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x11122322
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3f811111
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x555502a1
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fa55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 0x55555511
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fc55555
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v8, 11
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v9, 0x3fe00000
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], v[8:9]
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[6:7], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], 1.0
|
|
; GFX900-GISEL-NEXT: v_cvt_i32_f64_e32 v6, v[2:3]
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v2, 0
|
|
; GFX900-GISEL-NEXT: v_mov_b32_e32 v3, 0xc090cc00
|
|
; GFX900-GISEL-NEXT: v_cmp_nlt_f64_e32 vcc, v[0:1], v[2:3]
|
|
; GFX900-GISEL-NEXT: v_ldexp_f64 v[4:5], v[4:5], v6
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v0, 0, v4, vcc
|
|
; GFX900-GISEL-NEXT: v_cndmask_b32_e32 v1, 0, v5, vcc
|
|
; GFX900-GISEL-NEXT: s_setpc_b64 s[30:31]
|
|
%result = call contract nnan ninf double @llvm.exp.f64(double %in)
|
|
ret double %result
|
|
}
|
|
|
|
attributes #0 = { nounwind }
|
|
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
|
|
; GFX900: {{.*}}
|
|
; SI: {{.*}}
|
|
; VI: {{.*}}
|