shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/av_spill_cross_bb_usage.mir
Ryan Buchner 4fb7d1953d
[AMDGPU] Remove implicit definition of register group when restoring the last sub-register after a spill (#133986) 
Extension of https://reviews.llvm.org/D141101 to remove the define flag
from the last stack memory access. Fixes case where COPY instructions
are used for some of the stack restoration, but the copies get optimized
away during the machine-cp pass.

Prior to this change, was possible to produce the following code:
$agpr16_agpr17_agpr18_agpr19 = SCRATCH_LOAD_DWORDX4_ST 64, 0, implicit
$exec, implicit $flat_scr, implicit-def
$agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
:: (load (s128) from %stack.17, align 4, addrspace 5)
$agpr20_agpr21_agpr22_agpr23 = SCRATCH_LOAD_DWORDX4_ST 80, 0, implicit
$exec, implicit $flat_scr :: (load (s128) from %stack.17 + 16, align 4,
addrspace 5)
$agpr24_agpr25_agpr26_agpr27 = SCRATCH_LOAD_DWORDX4_ST 96, 0, implicit
$exec, implicit $flat_scr :: (load (s128) from %stack.17 + 32, align 4,
addrspace 5)
$agpr31 = COPY $agpr112, implicit
$agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
$agpr30 = COPY $agpr208, implicit
$agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
$agpr28_agpr29 = SCRATCH_LOAD_DWORDX2_ST 112, 0, implicit $exec,
implicit $flat_scr, implicit-def
$agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
:: (load (s64) from %stack.17 + 48, align 4, addrspace 5)

where `$agpr30 = COPY $agpr208` would be optimized away by `machine-cp`
pass. Instead, change to:
$agpr28_agpr29 = SCRATCH_LOAD_DWORDX2_ST 112, 0, implicit $exec,
implicit $flat_scr :: (load (s64) from %stack.17 + 48, align 4,
addrspace 5)

Fixes #131386.

Made the simple fix, but I'm not completely comfortable with this change
since the reason for the previous inclusion of `IsLastSubReg` is unclear
to me.

@krzysz00
2025-05-05 20:17:36 +02:00

145 lines
12 KiB
YAML
Raw Blame History

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -start-before=greedy,1 -stop-after=prologepilog -verify-machineinstrs -verify-regalloc -o - %s | FileCheck --check-prefixes=GCN %s
# The VGPR pair spilled and restored around the callsite is used in the next basic block.
#
# AMDGPU target spill hooks storeRegToStackSlot/loadRegFromStackSlot handle the register spills via
# spill pseudos to insert a single instruction per spill to tackle the limitation during inline spiller
# that incorrectly updates the Liveintervals in case of a spill lowered into multiple instructions.
# AV spills were handled earlier by converting them into equivalent VGPR spills by adding appropriate copies.
# The multiple instructions (a copy + vgpr spill pseudo) introduced an incorrect liverange that caused a
# crash during RA. It is fixed by introducing AV* spill pseudos to ensure a single instruction per spill and
# the test started compiling successfully.
---
name: test_av_spill_cross_bb_usage
tracksRegLiveness: true
frameInfo:
adjustsStack: true
stack:
- { id: 0, name: '', type: spill-slot, offset: 0, size: 4, alignment: 4 }
machineFunctionInfo:
scratchRSrcReg: $sgpr0_sgpr1_sgpr2_sgpr3
stackPtrOffsetReg: '$sgpr32'
body: |
; GCN-LABEL: name: test_av_spill_cross_bb_usage
; GCN: bb.0:
; GCN-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)
; GCN-NEXT: liveins: $sgpr30, $sgpr31, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8, $vgpr9, $vgpr10, $vgpr11, $vgpr12, $vgpr13, $vgpr14, $vgpr15, $vgpr40, $vgpr41, $vgpr42, $vgpr43, $vgpr44, $vgpr45, $vgpr46, $vgpr56, $vgpr57, $vgpr58, $vgpr59, $vgpr60, $vgpr61, $sgpr30_sgpr31
; GCN-NEXT: {{ $}}
; GCN-NEXT: $sgpr4_sgpr5 = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr4_sgpr5
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr41, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr42, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr43, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr44, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr45, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.8, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr46, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.9, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr56, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.10, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr57, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.11, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr58, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.12, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr59, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.13, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr60, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.14, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr61, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.15, addrspace 5)
; GCN-NEXT: renamable $vgpr44 = COPY $vgpr13, implicit $exec
; GCN-NEXT: renamable $vgpr43 = COPY $vgpr12, implicit $exec
; GCN-NEXT: S_CBRANCH_SCC1 %bb.2, implicit undef $scc
; GCN-NEXT: S_BRANCH %bb.1
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.1:
; GCN-NEXT: successors: %bb.2(0x80000000)
; GCN-NEXT: liveins: $exec, $sgpr30, $sgpr31, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8, $vgpr9, $vgpr40, $sgpr30_sgpr31, $vgpr10_vgpr11:0x000000000000000F, $vgpr14_vgpr15:0x000000000000000F, $vgpr41_vgpr42:0x000000000000000F, $vgpr43_vgpr44:0x000000000000000F, $vgpr45_vgpr46:0x000000000000000F, $vgpr56_vgpr57:0x000000000000000F, $vgpr58_vgpr59:0x000000000000000F, $vgpr60_vgpr61:0x000000000000000F
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr57 = COPY $vgpr9, implicit $exec
; GCN-NEXT: renamable $vgpr56 = COPY $vgpr8, implicit $exec
; GCN-NEXT: renamable $vgpr59 = COPY $vgpr7, implicit $exec
; GCN-NEXT: renamable $vgpr58 = COPY $vgpr6, implicit $exec
; GCN-NEXT: renamable $vgpr61 = COPY $vgpr5, implicit $exec
; GCN-NEXT: renamable $vgpr60 = COPY $vgpr4, implicit $exec
; GCN-NEXT: renamable $vgpr42 = COPY $vgpr3, implicit $exec
; GCN-NEXT: renamable $vgpr41 = COPY $vgpr2, implicit $exec
; GCN-NEXT: renamable $vgpr46 = COPY $vgpr1, implicit $exec
; GCN-NEXT: renamable $vgpr45 = COPY $vgpr0, implicit $exec
; GCN-NEXT: renamable $sgpr16_sgpr17 = IMPLICIT_DEF
; GCN-NEXT: $vgpr40 = SI_SPILL_S32_TO_VGPR $sgpr30, 0, $vgpr40, implicit-def $sgpr30_sgpr31, implicit $sgpr30_sgpr31
; GCN-NEXT: $vgpr40 = SI_SPILL_S32_TO_VGPR $sgpr31, 1, $vgpr40, implicit $sgpr30_sgpr31
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $vgpr14_vgpr15, implicit $vgpr14_vgpr15 :: (store (s32) into %stack.1, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit killed $vgpr14_vgpr15 :: (store (s32) into %stack.1 + 4, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $vgpr10_vgpr11, implicit $vgpr10_vgpr11 :: (store (s32) into %stack.2, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit killed $vgpr10_vgpr11 :: (store (s32) into %stack.2 + 4, addrspace 5)
; GCN-NEXT: dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr16_sgpr17, 0, csr_amdgpu, implicit-def dead $vgpr0
; GCN-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $vgpr14_vgpr15 :: (load (s32) from %stack.1, addrspace 5)
; GCN-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.1 + 4, addrspace 5)
; GCN-NEXT: renamable $vgpr0_vgpr1 = nofpexcept V_FMA_F64_e64 0, killed $vgpr45_vgpr46, 0, killed $vgpr41_vgpr42, 0, killed $vgpr60_vgpr61, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: FLAT_STORE_DWORDX2 killed renamable $vgpr58_vgpr59, killed renamable $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
; GCN-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1 :: (load (s32) from %stack.2, addrspace 5)
; GCN-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.2 + 4, addrspace 5)
; GCN-NEXT: FLAT_STORE_DWORDX2 killed renamable $vgpr0_vgpr1, killed renamable $vgpr56_vgpr57, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.2:
; GCN-NEXT: liveins: $vgpr40, $vgpr14_vgpr15:0x000000000000000F, $vgpr43_vgpr44:0x000000000000000F
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr0_vgpr1 = V_MOV_B64_PSEUDO 0, implicit $exec
; GCN-NEXT: FLAT_STORE_DWORDX2 undef renamable $vgpr0_vgpr1, killed renamable $vgpr43_vgpr44, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
; GCN-NEXT: FLAT_STORE_DWORDX2 killed renamable $vgpr0_vgpr1, killed renamable $vgpr14_vgpr15, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
; GCN-NEXT: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.15, addrspace 5)
; GCN-NEXT: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.14, addrspace 5)
; GCN-NEXT: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.13, addrspace 5)
; GCN-NEXT: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.12, addrspace 5)
; GCN-NEXT: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.11, addrspace 5)
; GCN-NEXT: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.10, addrspace 5)
; GCN-NEXT: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.9, addrspace 5)
; GCN-NEXT: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.8, addrspace 5)
; GCN-NEXT: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.7, addrspace 5)
; GCN-NEXT: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.6, addrspace 5)
; GCN-NEXT: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.5, addrspace 5)
; GCN-NEXT: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.4, addrspace 5)
; GCN-NEXT: $sgpr4_sgpr5 = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
; GCN-NEXT: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.3, addrspace 5)
; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr4_sgpr5
; GCN-NEXT: S_SETPC_B64_return undef $sgpr30_sgpr31
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8, $vgpr9, $vgpr10, $vgpr11, $vgpr12, $vgpr13, $vgpr14, $vgpr15, $vgpr40, $sgpr30, $sgpr31, $sgpr30_sgpr31
undef %0.sub1:vreg_64 = COPY $vgpr15
%0.sub0:vreg_64 = COPY $vgpr14
undef %1.sub1:vreg_64 = COPY $vgpr13
%1.sub0:vreg_64 = COPY $vgpr12
undef %2.sub1:vreg_64 = COPY $vgpr11
%2.sub0:vreg_64 = COPY $vgpr10
undef %3.sub1:vreg_64 = COPY $vgpr9
%3.sub0:vreg_64 = COPY $vgpr8
undef %4.sub1:vreg_64 = COPY $vgpr7
%4.sub0:vreg_64 = COPY $vgpr6
undef %5.sub1:vreg_64 = COPY $vgpr5
%5.sub0:vreg_64 = COPY $vgpr4
undef %6.sub1:vreg_64 = COPY $vgpr3
%6.sub0:vreg_64 = COPY $vgpr2
undef %7.sub1:vreg_64 = COPY $vgpr1
%7.sub0:vreg_64 = COPY $vgpr0
S_CBRANCH_SCC1 %bb.2, implicit undef $scc
S_BRANCH %bb.1
bb.1:
liveins: $vgpr40, $sgpr30, $sgpr31, $sgpr30_sgpr31
ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
renamable $sgpr16_sgpr17 = IMPLICIT_DEF
$vgpr40 = SI_SPILL_S32_TO_VGPR $sgpr30, 0, $vgpr40, implicit-def $sgpr30_sgpr31, implicit $sgpr30_sgpr31
$vgpr40 = SI_SPILL_S32_TO_VGPR killed $sgpr31, 1, $vgpr40, implicit killed $sgpr30_sgpr31
dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr16_sgpr17, 0, csr_amdgpu, implicit-def dead $vgpr0
%8:vreg_64 = nofpexcept V_FMA_F64_e64 0, %7, 0, %6, 0, %5, 0, 0, implicit $mode, implicit $exec
ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
FLAT_STORE_DWORDX2 %4, %8, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
FLAT_STORE_DWORDX2 %2, %3, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
bb.2:
liveins: $vgpr40
%9:vreg_64 = V_MOV_B64_PSEUDO 0, implicit $exec
FLAT_STORE_DWORDX2 undef %10:vreg_64, %1, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
FLAT_STORE_DWORDX2 %9, %0, 0, 0, implicit $exec, implicit $flat_scr :: (store (s64))
S_SETPC_B64_return undef $sgpr30_sgpr31
...
Reference in New Issue View Git Blame Copy Permalink