shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
Jeffrey Byrnes 6118a254ff
[AMDGPU] Allocate AVRegClass last (#146606) 
This changes the RC priorities such that AVRegClass is the least
prioritized. These registers are less constrained than the VRegClass and
ARegClass as they can be either agpr or vgpr. Thus, assigning them last
removes unnecessary constraints from VRegClass and ARegClass
assignments, and allows the RA to make smarter decisions about whether
to use vgpr / agpr for AVRegClass.

We only have 5 bits for RC priorities, and we still want to prioritize
larger RCs over smaller ones. Since this new prioritization uses the 5th
bit for AVRegClass vs ARegClass / VRegClass, we only have 4 bits to
encode the size priorities. Previously, each RC with a distinct size,
had a distinct priority. However, this PR groups together multiple sizes
to the same priority. Currently, this will have no effect on
prioritization in practice because we only have one actually defined RC
per group per vector register type.

For example, a register class with 15 or 16 32bit registers will have
the same size priority (14). However, we only have VReg_512 (VReg_480
doesn't exist), so only one actual RC in VRegClass has this priority.
Similarly, we give register class with 17-32+ 32 bit registers a size
priority of 15, but we only have VReg_1024.

The effect of this PR is to prioritize first the vector register type
(VReg & Areg have top priority, then AVReg), with the size of the
register class having second priority.

Passes PSDB.

---------

Co-authored-by: Matt Arsenault <Matthew.Arsenault@amd.com>
2025-07-25 13:46:41 -07:00

311 lines
13 KiB
LLVM
Raw Blame History

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -disable-separate-const-offset-from-gep=1 -amdgpu-use-sdag-ptradd=1 < %s | FileCheck --check-prefixes=GFX942,GFX942_PTRADD %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -disable-separate-const-offset-from-gep=1 -amdgpu-use-sdag-ptradd=0 < %s | FileCheck --check-prefixes=GFX942,GFX942_LEGACY %s
; Tests for DAG combines and folds related to the ISD::PTRADD SelectionDAG
; opcode. The RUN lines uses -disable-separate-const-offset-from-gep to disable
; similar transformations in that pass.
; Tests reassociation (ptradd N0:(ptradd p, c1), z) where N0 has only one use.
define i64 @global_load_ZTwoUses(ptr addrspace(1) %base, i64 %voffset) {
; GFX942-LABEL: global_load_ZTwoUses:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
; GFX942-NEXT: global_load_dwordx2 v[0:1], v[0:1], off offset:24
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
; GFX942-NEXT: s_setpc_b64 s[30:31]
%gep0 = getelementptr inbounds i8, ptr addrspace(1) %base, i64 24
%gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 %voffset
%l = load i64, ptr addrspace(1) %gep1, align 8
%r = add i64 %l, %voffset
ret i64 %r
}
define i64 @global_load_gep_add_reassoc(ptr addrspace(1) %base, i64 %voffset) {
; GFX942_PTRADD-LABEL: global_load_gep_add_reassoc:
; GFX942_PTRADD: ; %bb.0:
; GFX942_PTRADD-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_PTRADD-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
; GFX942_PTRADD-NEXT: global_load_dwordx2 v[0:1], v[0:1], off offset:24
; GFX942_PTRADD-NEXT: s_waitcnt vmcnt(0)
; GFX942_PTRADD-NEXT: s_setpc_b64 s[30:31]
;
; GFX942_LEGACY-LABEL: global_load_gep_add_reassoc:
; GFX942_LEGACY: ; %bb.0:
; GFX942_LEGACY-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_LEGACY-NEXT: v_lshl_add_u64 v[0:1], v[2:3], 0, v[0:1]
; GFX942_LEGACY-NEXT: global_load_dwordx2 v[0:1], v[0:1], off offset:24
; GFX942_LEGACY-NEXT: s_waitcnt vmcnt(0)
; GFX942_LEGACY-NEXT: s_setpc_b64 s[30:31]
%add0 = add nuw nsw i64 %voffset, 24
%gep0 = getelementptr nuw inbounds i8, ptr addrspace(1) %base, i64 %add0
%l = load i64, ptr addrspace(1) %gep0, align 8
ret i64 %l
}
; Tests reassociation (ptradd (ptradd p, c1), c2) with two constants. These
; would be folded away in most cases, but the index computation introduced by
; the legalization of wide vector stores can for example introduce them.
define amdgpu_kernel void @store_v16i32(ptr addrspace(1) %out, <16 x i32> %a) {
; GFX942-LABEL: store_v16i32:
; GFX942: ; %bb.0: ; %entry
; GFX942-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
; GFX942-NEXT: v_mov_b32_e32 v0, 0
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
; GFX942-NEXT: v_mov_b32_e32 v2, s20
; GFX942-NEXT: v_mov_b32_e32 v3, s21
; GFX942-NEXT: v_mov_b32_e32 v4, s22
; GFX942-NEXT: v_mov_b32_e32 v5, s23
; GFX942-NEXT: global_store_dwordx4 v0, v[2:5], s[0:1] offset:48
; GFX942-NEXT: s_nop 1
; GFX942-NEXT: v_mov_b32_e32 v2, s16
; GFX942-NEXT: v_mov_b32_e32 v3, s17
; GFX942-NEXT: v_mov_b32_e32 v4, s18
; GFX942-NEXT: v_mov_b32_e32 v5, s19
; GFX942-NEXT: global_store_dwordx4 v0, v[2:5], s[0:1] offset:32
; GFX942-NEXT: s_nop 1
; GFX942-NEXT: v_mov_b32_e32 v2, s12
; GFX942-NEXT: v_mov_b32_e32 v3, s13
; GFX942-NEXT: v_mov_b32_e32 v4, s14
; GFX942-NEXT: v_mov_b32_e32 v5, s15
; GFX942-NEXT: global_store_dwordx4 v0, v[2:5], s[0:1] offset:16
; GFX942-NEXT: s_nop 1
; GFX942-NEXT: v_mov_b32_e32 v2, s8
; GFX942-NEXT: v_mov_b32_e32 v3, s9
; GFX942-NEXT: v_mov_b32_e32 v4, s10
; GFX942-NEXT: v_mov_b32_e32 v5, s11
; GFX942-NEXT: global_store_dwordx4 v0, v[2:5], s[0:1]
; GFX942-NEXT: s_endpgm
entry:
store <16 x i32> %a, ptr addrspace(1) %out
ret void
}
; Tests the (ptradd 0, x) -> x DAG combine.
define void @baseptr_null(i64 %offset, i8 %v) {
; GFX942-LABEL: baseptr_null:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT: flat_store_byte v[0:1], v2
; GFX942-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX942-NEXT: s_setpc_b64 s[30:31]
%gep = getelementptr i8, ptr null, i64 %offset
store i8 %v, ptr %gep, align 1
ret void
}
; Taken from implicit-kernarg-backend-usage.ll, tests the PTRADD handling in the
; assertalign DAG combine.
define amdgpu_kernel void @llvm_amdgcn_queue_ptr(ptr addrspace(1) %ptr) #0 {
; GFX942-LABEL: llvm_amdgcn_queue_ptr:
; GFX942: ; %bb.0:
; GFX942-NEXT: v_mov_b32_e32 v2, 0
; GFX942-NEXT: global_load_ubyte v0, v2, s[2:3] sc0 sc1
; GFX942-NEXT: global_load_ubyte v0, v2, s[4:5] offset:8 sc0 sc1
; GFX942-NEXT: global_load_ubyte v0, v2, s[0:1] sc0 sc1
; GFX942-NEXT: ; kill: killed $sgpr0_sgpr1
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: v_mov_b64_e32 v[0:1], s[6:7]
; GFX942-NEXT: ; kill: killed $sgpr2_sgpr3
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
; GFX942-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] sc0 sc1
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: s_endpgm
%queue.ptr = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()
%implicitarg.ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
%dispatch.ptr = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
%dispatch.id = call i64 @llvm.amdgcn.dispatch.id()
%queue.load = load volatile i8, ptr addrspace(4) %queue.ptr
%implicitarg.load = load volatile i8, ptr addrspace(4) %implicitarg.ptr
%dispatch.load = load volatile i8, ptr addrspace(4) %dispatch.ptr
store volatile i64 %dispatch.id, ptr addrspace(1) %ptr
ret void
}
; Taken from memcpy-param-combinations.ll, tests PTRADD handling in
; SelectionDAGAddressAnalysis.
define void @memcpy_p1_p4_sz16_align_1_1(ptr addrspace(1) align 1 %dst, ptr addrspace(4) align 1 readonly %src) {
; GFX942-LABEL: memcpy_p1_p4_sz16_align_1_1:
; GFX942: ; %bb.0: ; %entry
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT: global_load_dwordx4 v[2:5], v[2:3], off
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: global_store_dwordx4 v[0:1], v[2:5], off
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: s_setpc_b64 s[30:31]
entry:
tail call void @llvm.memcpy.p1.p4.i64(ptr addrspace(1) noundef nonnull align 1 %dst, ptr addrspace(4) noundef nonnull align 1 %src, i64 16, i1 false)
ret void
}
; Test skipping the lower-32-bit addition if it is unnecessary.
define ptr @huge_offset_low_32_unused(ptr %p) {
; GFX942-LABEL: huge_offset_low_32_unused:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT: v_add_u32_e32 v1, 1, v1
; GFX942-NEXT: s_setpc_b64 s[30:31]
%gep = getelementptr inbounds i8, ptr %p, i64 u0x100000000
ret ptr %gep
}
; Reassociate address computation if it leads to more scalar operations.
define amdgpu_kernel void @reassoc_scalar_r(ptr addrspace(1) %out, ptr addrspace(1) %p, i64 %soffset) {
; GFX942-LABEL: reassoc_scalar_r:
; GFX942: ; %bb.0: ; %entry
; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX942-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x10
; GFX942-NEXT: v_mov_b32_e32 v1, 0
; GFX942-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
; GFX942-NEXT: s_add_u32 s2, s2, s6
; GFX942-NEXT: s_addc_u32 s3, s3, s7
; GFX942-NEXT: v_lshl_add_u64 v[2:3], s[2:3], 0, v[0:1]
; GFX942-NEXT: global_store_dwordx2 v1, v[2:3], s[0:1]
; GFX942-NEXT: s_endpgm
entry:
%voffset32 = call i32 @llvm.amdgcn.workitem.id.x()
%voffset = zext i32 %voffset32 to i64
%offset = add nuw nsw i64 %voffset, %soffset
%gep = getelementptr i8, ptr addrspace(1) %p, i64 %offset
store ptr addrspace(1) %gep, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @reassoc_scalar_l(ptr addrspace(1) %out, ptr addrspace(1) %p, i64 %soffset) {
; GFX942-LABEL: reassoc_scalar_l:
; GFX942: ; %bb.0: ; %entry
; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX942-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x10
; GFX942-NEXT: v_mov_b32_e32 v1, 0
; GFX942-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
; GFX942-NEXT: s_add_u32 s2, s2, s6
; GFX942-NEXT: s_addc_u32 s3, s3, s7
; GFX942-NEXT: v_lshl_add_u64 v[2:3], s[2:3], 0, v[0:1]
; GFX942-NEXT: global_store_dwordx2 v1, v[2:3], s[0:1]
; GFX942-NEXT: s_endpgm
entry:
%voffset32 = call i32 @llvm.amdgcn.workitem.id.x()
%voffset = zext i32 %voffset32 to i64
%offset = add nuw nsw i64 %soffset, %voffset
%gep = getelementptr i8, ptr addrspace(1) %p, i64 %offset
store ptr addrspace(1) %gep, ptr addrspace(1) %out
ret void
}
; Tests the target-specific (ptradd x, shl(0 - y, k)) -> sub(x, shl(y, k)) fold
define ptr addrspace(1) @shl_neg_offset(ptr addrspace(1) %p, i64 %noffset, i64 %shift) {
; GFX942-LABEL: shl_neg_offset:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT: v_lshlrev_b64 v[2:3], v4, v[2:3]
; GFX942-NEXT: v_sub_co_u32_e32 v0, vcc, v0, v2
; GFX942-NEXT: s_nop 1
; GFX942-NEXT: v_subb_co_u32_e32 v1, vcc, v1, v3, vcc
; GFX942-NEXT: s_setpc_b64 s[30:31]
%offset = sub i64 0, %noffset
%x = shl i64 %offset, %shift
%gep = getelementptr inbounds i8, ptr addrspace(1) %p, i64 %x
ret ptr addrspace(1) %gep
}
%complextype = type { i64, [10 x i8], float }
@v0 = dso_local addrspace(1) global %complextype zeroinitializer
; Check that offsets are folded into global addresses if possible. For example,
; this is relevant when using --amdgpu-lower-module-lds-strategy=table.
define ptr addrspace(1) @complextype_global_gep(i64 %offset) {
; GFX942_PTRADD-LABEL: complextype_global_gep:
; GFX942_PTRADD: ; %bb.0:
; GFX942_PTRADD-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_PTRADD-NEXT: s_getpc_b64 s[0:1]
; GFX942_PTRADD-NEXT: s_add_u32 s0, s0, v0@rel32@lo+14
; GFX942_PTRADD-NEXT: s_addc_u32 s1, s1, v0@rel32@hi+22
; GFX942_PTRADD-NEXT: v_lshl_add_u64 v[0:1], s[0:1], 0, v[0:1]
; GFX942_PTRADD-NEXT: s_setpc_b64 s[30:31]
;
; GFX942_LEGACY-LABEL: complextype_global_gep:
; GFX942_LEGACY: ; %bb.0:
; GFX942_LEGACY-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_LEGACY-NEXT: s_getpc_b64 s[0:1]
; GFX942_LEGACY-NEXT: s_add_u32 s0, s0, v0@rel32@lo+14
; GFX942_LEGACY-NEXT: s_addc_u32 s1, s1, v0@rel32@hi+22
; GFX942_LEGACY-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
; GFX942_LEGACY-NEXT: s_setpc_b64 s[30:31]
%gep0 = getelementptr inbounds %complextype, ptr addrspace(1) @v0, i64 0, i32 1, i64 %offset
%gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 2
ret ptr addrspace(1) %gep1
}
%S = type <{ float, double }>
; Tests the tryFoldToMad64_32 PTRADD combine.
define amdgpu_kernel void @fold_mad64(ptr addrspace(1) %p) {
; GFX942-LABEL: fold_mad64:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
; GFX942-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX942-NEXT: v_mov_b32_e32 v2, 1.0
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
; GFX942-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v0, 12, s[0:1]
; GFX942-NEXT: global_store_dword v[0:1], v2, off
; GFX942-NEXT: s_endpgm
%voffset32 = call i32 @llvm.amdgcn.workitem.id.x()
%voffset = zext i32 %voffset32 to i64
%p1 = getelementptr inbounds %S, ptr addrspace(1) %p, i64 %voffset, i32 0
store float 1.0, ptr addrspace(1) %p1
ret void
}
; Use non-zero shift amounts in v_lshl_add_u64.
define ptr @select_v_lshl_add_u64(ptr %base, i64 %voffset) {
; GFX942_PTRADD-LABEL: select_v_lshl_add_u64:
; GFX942_PTRADD: ; %bb.0:
; GFX942_PTRADD-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_PTRADD-NEXT: v_lshlrev_b64 v[2:3], 3, v[2:3]
; GFX942_PTRADD-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
; GFX942_PTRADD-NEXT: s_setpc_b64 s[30:31]
;
; GFX942_LEGACY-LABEL: select_v_lshl_add_u64:
; GFX942_LEGACY: ; %bb.0:
; GFX942_LEGACY-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_LEGACY-NEXT: v_lshl_add_u64 v[0:1], v[2:3], 3, v[0:1]
; GFX942_LEGACY-NEXT: s_setpc_b64 s[30:31]
%gep = getelementptr inbounds i64, ptr %base, i64 %voffset
ret ptr %gep
}
; Fold mul and add into v_mad, even if amdgpu-codegenprepare-mul24 turned the
; mul into a mul24.
define ptr @fold_mul24_into_mad(ptr %base, i64 %a, i64 %b) {
; GFX942_PTRADD-LABEL: fold_mul24_into_mad:
; GFX942_PTRADD: ; %bb.0:
; GFX942_PTRADD-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_PTRADD-NEXT: v_and_b32_e32 v2, 0xfffff, v2
; GFX942_PTRADD-NEXT: v_and_b32_e32 v4, 0xfffff, v4
; GFX942_PTRADD-NEXT: v_mul_hi_u32_u24_e32 v3, v2, v4
; GFX942_PTRADD-NEXT: v_mul_u32_u24_e32 v2, v2, v4
; GFX942_PTRADD-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
; GFX942_PTRADD-NEXT: s_setpc_b64 s[30:31]
;
; GFX942_LEGACY-LABEL: fold_mul24_into_mad:
; GFX942_LEGACY: ; %bb.0:
; GFX942_LEGACY-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942_LEGACY-NEXT: v_and_b32_e32 v2, 0xfffff, v2
; GFX942_LEGACY-NEXT: v_and_b32_e32 v3, 0xfffff, v4
; GFX942_LEGACY-NEXT: v_mad_u64_u32 v[0:1], s[0:1], v2, v3, v[0:1]
; GFX942_LEGACY-NEXT: s_setpc_b64 s[30:31]
%a_masked = and i64 %a, u0xfffff
%b_masked = and i64 %b, u0xfffff
%mul = mul i64 %a_masked, %b_masked
%gep = getelementptr inbounds i8, ptr %base, i64 %mul
ret ptr %gep
}
Reference in New Issue View Git Blame Copy Permalink