ac0f64f06d
Allocating wwm-registers and per-thread VGPR operands together imposes many challenges in the way the registers are reused during allocation. There are times when regalloc reuses the registers of regular VGPRs operations for wwm-operations in a small range leading to unwantedly clobbering their inactive lanes causing correctness issues that are hard to trace. This patch splits the VGPR allocation pipeline further to allocate wwm-registers first and the regular VGPR operands in a separate pipeline. The splitting would ensure that the physical registers used for wwm allocations won't take part in the next allocation pipeline to avoid any such clobbering.
80 lines
4.1 KiB
YAML
80 lines
4.1 KiB
YAML
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
|
|
# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-regalloc -stress-regalloc=3 -start-before=greedy,1 -stop-before=virtregrewriter,1 -o - %s | FileCheck %s
|
|
---
|
|
name: split_instruction_subranges
|
|
alignment: 1
|
|
tracksRegLiveness: true
|
|
frameInfo:
|
|
maxAlignment: 1
|
|
hasCalls: true
|
|
machineFunctionInfo:
|
|
maxKernArgAlign: 1
|
|
isEntryFunction: true
|
|
scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
|
|
stackPtrOffsetReg: '$sgpr32'
|
|
argumentInfo:
|
|
privateSegmentBuffer: { reg: '$sgpr0_sgpr1_sgpr2_sgpr3' }
|
|
privateSegmentWaveByteOffset: { reg: '$sgpr17' }
|
|
occupancy: 8
|
|
body: |
|
|
bb.0:
|
|
; CHECK-LABEL: name: split_instruction_subranges
|
|
; CHECK: [[GLOBAL_LOAD_DWORDX2_SADDR:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %1:vgpr_32, 0, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
; CHECK-NEXT: [[GLOBAL_LOAD_DWORDX2_SADDR1:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %3:vgpr_32, 4, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
; CHECK-NEXT: [[GLOBAL_LOAD_DWORDX2_SADDR2:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %5:vgpr_32, 8, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
; CHECK-NEXT: S_NOP 0, implicit [[GLOBAL_LOAD_DWORDX2_SADDR]].sub1
|
|
; CHECK-NEXT: S_NOP 0, implicit [[GLOBAL_LOAD_DWORDX2_SADDR1]].sub0
|
|
; CHECK-NEXT: S_NOP 0, implicit [[GLOBAL_LOAD_DWORDX2_SADDR2]].sub1
|
|
; CHECK-NEXT: S_ENDPGM 0
|
|
%1:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %4:vgpr_32, 0, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
%2:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %5:vgpr_32, 4, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
%3:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %6:vgpr_32, 8, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
S_NOP 0, implicit %1.sub1
|
|
S_NOP 0, implicit %2.sub0
|
|
S_NOP 0, implicit %3.sub1
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
|
|
---
|
|
name: split_instruction_subranges_use_is_subreg_def
|
|
alignment: 1
|
|
tracksRegLiveness: true
|
|
frameInfo:
|
|
maxAlignment: 1
|
|
hasCalls: true
|
|
machineFunctionInfo:
|
|
maxKernArgAlign: 1
|
|
isEntryFunction: true
|
|
scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
|
|
stackPtrOffsetReg: '$sgpr32'
|
|
argumentInfo:
|
|
privateSegmentBuffer: { reg: '$sgpr0_sgpr1_sgpr2_sgpr3' }
|
|
privateSegmentWaveByteOffset: { reg: '$sgpr17' }
|
|
occupancy: 8
|
|
body: |
|
|
bb.0:
|
|
; CHECK-LABEL: name: split_instruction_subranges_use_is_subreg_def
|
|
; CHECK: [[GLOBAL_LOAD_DWORDX2_SADDR:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %1:vgpr_32, 0, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
; CHECK-NEXT: [[GLOBAL_LOAD_DWORDX2_SADDR1:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %3:vgpr_32, 4, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
; CHECK-NEXT: [[GLOBAL_LOAD_DWORDX2_SADDR2:%[0-9]+]]:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %5:vgpr_32, 8, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
; CHECK-NEXT: S_NOP 0, implicit-def [[GLOBAL_LOAD_DWORDX2_SADDR]].sub0
|
|
; CHECK-NEXT: S_NOP 0, implicit-def [[GLOBAL_LOAD_DWORDX2_SADDR1]].sub1
|
|
; CHECK-NEXT: S_NOP 0, implicit-def [[GLOBAL_LOAD_DWORDX2_SADDR2]].sub0
|
|
; CHECK-NEXT: S_NOP 0, implicit [[GLOBAL_LOAD_DWORDX2_SADDR]].sub1
|
|
; CHECK-NEXT: S_NOP 0, implicit [[GLOBAL_LOAD_DWORDX2_SADDR1]].sub0
|
|
; CHECK-NEXT: S_NOP 0, implicit [[GLOBAL_LOAD_DWORDX2_SADDR2]].sub1
|
|
; CHECK-NEXT: S_ENDPGM 0
|
|
%1:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %4:vgpr_32, 0, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
%2:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %5:vgpr_32, 4, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
%3:vreg_64 = GLOBAL_LOAD_DWORDX2_SADDR undef $sgpr0_sgpr1, undef %6:vgpr_32, 8, 0, implicit $exec :: (load (s64), addrspace 1)
|
|
S_NOP 0, implicit-def %1.sub0
|
|
S_NOP 0, implicit-def %2.sub1
|
|
S_NOP 0, implicit-def %3.sub0
|
|
S_NOP 0, implicit %1.sub1
|
|
S_NOP 0, implicit %2.sub0
|
|
S_NOP 0, implicit %3.sub1
|
|
S_ENDPGM 0
|
|
|
|
...
|