[llvm] [AMDGPU] Improve isBasicBlockPrologue helper function (PR #69924)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 23 05:49:21 PDT 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Christudasan Devadasan (cdevadas)
<details>
<summary>Changes</summary>
This function helps to skip over the basic block prolog
instructions while inserting a new spill or a copy during
liverange split. However, this appears to be incomplete.
It currently skips only the instructions that restores
the exec mask. It should also have skipped over the BB
prolog spills and the liverange split copies inserted
earlier during regalloc while dealing with the BBLiveIns.
This patch partially addresses the limitation by including
only the spills in the query. The live-range-split COPY
handling is pending.
This helper function is also used at multiple places to
skip over the bb prolog. They should continue to produce
functionally correct code even after this patch.
This patch depends on [#<!-- -->69923](https://github.com/llvm/llvm-project/pull/69923)
---
Patch is 413.71 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/69924.diff
47 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIFrameLowering.cpp (+5-7)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+21-4)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.h (+5)
- (modified) llvm/lib/Target/AMDGPU/SIInstructions.td (+19)
- (modified) llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp (+5-5)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll (+3-2)
- (modified) llvm/test/CodeGen/AMDGPU/av_spill_cross_bb_usage.mir (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/collapse-endcf.ll (+41-30)
- (modified) llvm/test/CodeGen/AMDGPU/control-flow-fastregalloc.ll (+8-6)
- (modified) llvm/test/CodeGen/AMDGPU/cross-block-use-is-not-abi-copy.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/csr-sgpr-spill-live-ins.mir (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/extend-wwm-virt-reg-liveness.mir (+8-8)
- (modified) llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands-non-ptr-intrinsics.ll (+30-24)
- (modified) llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll (+32-26)
- (modified) llvm/test/CodeGen/AMDGPU/partial-sgpr-to-vgpr-spills.ll (+19-14)
- (modified) llvm/test/CodeGen/AMDGPU/pei-scavenge-sgpr-carry-out.mir (+2-2)
- (modified) llvm/test/CodeGen/AMDGPU/pei-scavenge-sgpr-gfx9.mir (+5-4)
- (modified) llvm/test/CodeGen/AMDGPU/pei-scavenge-sgpr.mir (+2-2)
- (modified) llvm/test/CodeGen/AMDGPU/pei-scavenge-vgpr-spill.mir (+8-6)
- (modified) llvm/test/CodeGen/AMDGPU/preserve-only-inactive-lane.mir (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/same-slot-agpr-sgpr.mir (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/sgpr-spill-dead-frame-in-dbg-value.mir (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/sgpr-spill-no-vgprs.ll (+4-3)
- (modified) llvm/test/CodeGen/AMDGPU/sgpr-spill-partially-undef.mir (+4-4)
- (modified) llvm/test/CodeGen/AMDGPU/sgpr-spill-to-vmem-scc-clobber.mir (+16-16)
- (modified) llvm/test/CodeGen/AMDGPU/sgpr-spill-vmem-large-frame.mir (+2-2)
- (modified) llvm/test/CodeGen/AMDGPU/sgpr-spill.mir (+445-441)
- (modified) llvm/test/CodeGen/AMDGPU/si-lower-sgpr-spills.mir (+8-8)
- (modified) llvm/test/CodeGen/AMDGPU/snippet-copy-bundle-regression.mir (+24-24)
- (modified) llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir (+8-8)
- (modified) llvm/test/CodeGen/AMDGPU/spill-scavenge-offset.ll (+244-245)
- (modified) llvm/test/CodeGen/AMDGPU/spill-sgpr-csr-live-ins.mir (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/spill-sgpr-to-virtual-vgpr.mir (+106-106)
- (modified) llvm/test/CodeGen/AMDGPU/spill-special-sgpr.mir (+20-18)
- (modified) llvm/test/CodeGen/AMDGPU/spill-writelane-vgprs.ll (-8)
- (modified) llvm/test/CodeGen/AMDGPU/spill192.mir (+14-12)
- (modified) llvm/test/CodeGen/AMDGPU/spill224.mir (+16-14)
- (modified) llvm/test/CodeGen/AMDGPU/spill288.mir (+20-18)
- (modified) llvm/test/CodeGen/AMDGPU/spill320.mir (+22-20)
- (modified) llvm/test/CodeGen/AMDGPU/spill352.mir (+24-22)
- (modified) llvm/test/CodeGen/AMDGPU/spill384.mir (+26-24)
- (modified) llvm/test/CodeGen/AMDGPU/swdev380865.ll (+8-8)
- (modified) llvm/test/CodeGen/AMDGPU/tied-op-for-wwm-scratch-reg-spill-restore.mir (+12-12)
- (modified) llvm/test/CodeGen/AMDGPU/track-spilled-vgpr-liveness.mir (-18)
- (modified) llvm/test/CodeGen/AMDGPU/use_restore_frame_reg.mir (+5-4)
- (modified) llvm/test/CodeGen/AMDGPU/vgpr-spill-placement-issue61083.ll (+2-1)
- (modified) llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll (+3-3)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp b/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp
index 5de993ec3cba1d8..1d855d9bd83157d 100644
--- a/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp
@@ -274,7 +274,7 @@ class PrologEpilogSGPRSpillBuilder {
Register SubReg = NumSubRegs == 1
? SuperReg
: Register(TRI.getSubReg(SuperReg, SplitParts[I]));
- BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_WRITELANE_B32), Spill[I].VGPR)
+ BuildMI(MBB, MI, DL, TII->get(AMDGPU::SI_WRITELANE_PSEUDO), Spill[I].VGPR)
.addReg(SubReg)
.addImm(Spill[I].Lane)
.addReg(Spill[I].VGPR, RegState::Undef);
@@ -319,7 +319,7 @@ class PrologEpilogSGPRSpillBuilder {
Register SubReg = NumSubRegs == 1
? SuperReg
: Register(TRI.getSubReg(SuperReg, SplitParts[I]));
- BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_READLANE_B32), SubReg)
+ BuildMI(MBB, MI, DL, TII->get(AMDGPU::SI_READLANE_PSEUDO), SubReg)
.addReg(Spill[I].VGPR)
.addImm(Spill[I].Lane);
}
@@ -1554,12 +1554,10 @@ void SIFrameLowering::determineCalleeSaves(MachineFunction &MF,
// TODO: Handle this elsewhere at an early point. Walking through all MBBs
// here would be a bad heuristic. A better way should be by calling
// allocateWWMSpill during the regalloc pipeline whenever a physical
- // register is allocated for the intended virtual registers. That will
- // also help excluding the general use of WRITELANE/READLANE intrinsics
- // that won't really need any such special handling.
- if (MI.getOpcode() == AMDGPU::V_WRITELANE_B32)
+ // register is allocated for the intended virtual registers.
+ if (MI.getOpcode() == AMDGPU::SI_WRITELANE_PSEUDO)
MFI->allocateWWMSpill(MF, MI.getOperand(0).getReg());
- else if (MI.getOpcode() == AMDGPU::V_READLANE_B32)
+ else if (MI.getOpcode() == AMDGPU::SI_READLANE_PSEUDO)
MFI->allocateWWMSpill(MF, MI.getOperand(1).getReg());
else if (TII->isWWMRegSpillOpcode(MI.getOpcode()))
NeedExecCopyReservedReg = true;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 4ff7b462f0f3295..f977c2412bbe322 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2090,6 +2090,14 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
MI.setDesc(get(AMDGPU::S_AND_SAVEEXEC_B32));
break;
+ case AMDGPU::SI_WRITELANE_PSEUDO:
+ MI.setDesc(get(AMDGPU::V_WRITELANE_B32));
+ break;
+
+ case AMDGPU::SI_READLANE_PSEUDO:
+ MI.setDesc(get(AMDGPU::V_READLANE_B32));
+ break;
+
case AMDGPU::V_MOV_B64_PSEUDO: {
Register Dst = MI.getOperand(0).getReg();
Register DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
@@ -3907,7 +3915,9 @@ bool SIInstrInfo::hasUnwantedEffectsWhenEXECEmpty(const MachineInstr &MI) const
// However, executing them with EXEC = 0 causes them to operate on undefined
// data, which we avoid by returning true here.
if (Opcode == AMDGPU::V_READFIRSTLANE_B32 ||
- Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32)
+ Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32 ||
+ Opcode == AMDGPU::SI_READLANE_PSEUDO ||
+ Opcode == AMDGPU::SI_WRITELANE_PSEUDO)
return true;
return false;
@@ -4301,7 +4311,9 @@ static bool shouldReadExec(const MachineInstr &MI) {
if (SIInstrInfo::isVALU(MI)) {
switch (MI.getOpcode()) {
case AMDGPU::V_READLANE_B32:
+ case AMDGPU::SI_READLANE_PSEUDO:
case AMDGPU::V_WRITELANE_B32:
+ case AMDGPU::SI_WRITELANE_PSEUDO:
return false;
}
@@ -8326,8 +8338,11 @@ unsigned SIInstrInfo::getLiveRangeSplitOpcode(Register SrcReg,
}
bool SIInstrInfo::isBasicBlockPrologue(const MachineInstr &MI) const {
- return !MI.isTerminator() && MI.getOpcode() != AMDGPU::COPY &&
- MI.modifiesRegister(AMDGPU::EXEC, &RI);
+ uint16_t Opc = MI.getOpcode();
+ // FIXME: Copies inserted in the block prolog for live-range split should also
+ // be included.
+ return (isSpillOpcode(Opc) || (!MI.isTerminator() && Opc != AMDGPU::COPY &&
+ MI.modifiesRegister(AMDGPU::EXEC, &RI)));
}
MachineInstrBuilder
@@ -8970,7 +8985,9 @@ SIInstrInfo::getInstructionUniformity(const MachineInstr &MI) const {
return InstructionUniformity::NeverUniform;
unsigned opcode = MI.getOpcode();
- if (opcode == AMDGPU::V_READLANE_B32 || opcode == AMDGPU::V_READFIRSTLANE_B32)
+ if (opcode == AMDGPU::V_READLANE_B32 ||
+ opcode == AMDGPU::V_READFIRSTLANE_B32 ||
+ opcode == AMDGPU::SI_READLANE_PSEUDO)
return InstructionUniformity::AlwaysUniform;
if (isCopyInstr(MI)) {
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index 5ef17c44f7de389..34453c41109580d 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -671,6 +671,11 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
return get(Opcode).TSFlags & SIInstrFlags::SGPRSpill;
}
+ bool isSpillOpcode(uint16_t Opcode) const {
+ return get(Opcode).TSFlags &
+ (SIInstrFlags::SGPRSpill | SIInstrFlags::VGPRSpill);
+ }
+
static bool isWWMRegSpillOpcode(uint16_t Opcode) {
return Opcode == AMDGPU::SI_SPILL_WWM_V32_SAVE ||
Opcode == AMDGPU::SI_SPILL_WWM_AV32_SAVE ||
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 567f1b812c1808c..6929dc4f4ab91cc 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -875,6 +875,25 @@ defm SI_SPILL_S384 : SI_SPILL_SGPR <SReg_384>;
defm SI_SPILL_S512 : SI_SPILL_SGPR <SReg_512>;
defm SI_SPILL_S1024 : SI_SPILL_SGPR <SReg_1024>;
+let SGPRSpill = 1 in {
+def SI_WRITELANE_PSEUDO : PseudoInstSI <(outs VGPR_32:$vdst),
+ (ins SReg_32:$src0, i32imm:$src1, VGPR_32:$vdst_in)> {
+ let hasSideEffects = 0;
+ let mayLoad = 0;
+ let mayStore = 0;
+ let VALU = 1;
+ let Constraints = "$vdst = $vdst_in";
+}
+
+def SI_READLANE_PSEUDO : PseudoInstSI <(outs SReg_32:$sdst),
+ (ins VGPR_32:$src0, i32imm:$src1)> {
+ let hasSideEffects = 0;
+ let mayLoad = 0;
+ let mayStore = 0;
+ let VALU = 1;
+}
+} // End SGPRSpill = 1
+
// VGPR or AGPR spill instructions. In case of AGPR spilling a temp register
// needs to be used and an extra instruction to move between VGPR and AGPR.
// UsesTmp adds to the total size of an expanded spill in this case.
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
index d0a81673d6528c2..ed81fac6886adca 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -1769,7 +1769,7 @@ bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, int Index,
// Mark the "old value of vgpr" input undef only if this is the first sgpr
// spill to this specific vgpr in the first basic block.
auto MIB = BuildMI(*SB.MBB, MI, SB.DL,
- SB.TII.get(AMDGPU::V_WRITELANE_B32), Spill.VGPR)
+ SB.TII.get(AMDGPU::SI_WRITELANE_PSEUDO), Spill.VGPR)
.addReg(SubReg, getKillRegState(UseKill))
.addImm(Spill.Lane)
.addReg(Spill.VGPR);
@@ -1815,7 +1815,7 @@ bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, int Index,
: Register(getSubReg(SB.SuperReg, SB.SplitParts[i]));
MachineInstrBuilder WriteLane =
- BuildMI(*SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::V_WRITELANE_B32),
+ BuildMI(*SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::SI_WRITELANE_PSEUDO),
SB.TmpVGPR)
.addReg(SubReg, SubKillState)
.addImm(i % PVD.PerVGPR)
@@ -1877,8 +1877,8 @@ bool SIRegisterInfo::restoreSGPR(MachineBasicBlock::iterator MI, int Index,
: Register(getSubReg(SB.SuperReg, SB.SplitParts[i]));
SpilledReg Spill = VGPRSpills[i];
- auto MIB = BuildMI(*SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::V_READLANE_B32),
- SubReg)
+ auto MIB = BuildMI(*SB.MBB, MI, SB.DL,
+ SB.TII.get(AMDGPU::SI_READLANE_PSEUDO), SubReg)
.addReg(Spill.VGPR)
.addImm(Spill.Lane);
if (SB.NumSubRegs > 1 && i == 0)
@@ -1911,7 +1911,7 @@ bool SIRegisterInfo::restoreSGPR(MachineBasicBlock::iterator MI, int Index,
bool LastSubReg = (i + 1 == e);
auto MIB = BuildMI(*SB.MBB, MI, SB.DL,
- SB.TII.get(AMDGPU::V_READLANE_B32), SubReg)
+ SB.TII.get(AMDGPU::SI_READLANE_PSEUDO), SubReg)
.addReg(SB.TmpVGPR, getKillRegState(LastSubReg))
.addImm(i);
if (SB.NumSubRegs > 1 && i == 0)
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll
index b19230c2e876c4f..10cbc56cc5fbea1 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll
@@ -144,8 +144,6 @@ define <4 x float> @waterfall_loop(<8 x i32> %vgpr_srd) {
; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b32 exec_lo, s21
; CHECK-NEXT: ; %bb.2: ; in Loop: Header=BB0_1 Depth=1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; CHECK-NEXT: s_or_saveexec_b32 s21, -1
; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b32 exec_lo, s21
@@ -163,6 +161,9 @@ define <4 x float> @waterfall_loop(<8 x i32> %vgpr_srd) {
; CHECK-NEXT: v_readlane_b32 s17, v2, 1
; CHECK-NEXT: v_readlane_b32 s18, v2, 2
; CHECK-NEXT: v_readlane_b32 s19, v2, 3
+; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: image_sample v0, v[0:1], s[8:15], s[16:19] dmask:0x1 dim:SQ_RSRC_IMG_2D
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
diff --git a/llvm/test/CodeGen/AMDGPU/av_spill_cross_bb_usage.mir b/llvm/test/CodeGen/AMDGPU/av_spill_cross_bb_usage.mir
index 7209d160e6c8a7a..3e56e49bf31d5cb 100644
--- a/llvm/test/CodeGen/AMDGPU/av_spill_cross_bb_usage.mir
+++ b/llvm/test/CodeGen/AMDGPU/av_spill_cross_bb_usage.mir
@@ -60,8 +60,8 @@ body: |
; 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 = V_WRITELANE_B32 $sgpr30, 0, $vgpr40, implicit-def $sgpr30_sgpr31, implicit $sgpr30_sgpr31
- ; GCN-NEXT: $vgpr40 = V_WRITELANE_B32 $sgpr31, 1, $vgpr40, implicit $sgpr30_sgpr31
+ ; GCN-NEXT: $vgpr40 = SI_WRITELANE_PSEUDO $sgpr30, 0, $vgpr40, implicit-def $sgpr30_sgpr31, implicit $sgpr30_sgpr31
+ ; GCN-NEXT: $vgpr40 = SI_WRITELANE_PSEUDO $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)
@@ -124,8 +124,8 @@ body: |
ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
renamable $sgpr16_sgpr17 = IMPLICIT_DEF
- $vgpr40 = V_WRITELANE_B32 $sgpr30, 0, $vgpr40, implicit-def $sgpr30_sgpr31, implicit $sgpr30_sgpr31
- $vgpr40 = V_WRITELANE_B32 killed $sgpr31, 1, $vgpr40, implicit killed $sgpr30_sgpr31
+ $vgpr40 = SI_WRITELANE_PSEUDO $sgpr30, 0, $vgpr40, implicit-def $sgpr30_sgpr31, implicit $sgpr30_sgpr31
+ $vgpr40 = SI_WRITELANE_PSEUDO 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
diff --git a/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll b/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll
index 73d5088141cdb31..dc519df52919add 100644
--- a/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll
+++ b/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll
@@ -75,10 +75,10 @@ define amdgpu_kernel void @simple_nested_if(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_readlane_b32 s4, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s5, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s0, 0
; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
@@ -104,7 +104,6 @@ define amdgpu_kernel void @simple_nested_if(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB0_3
; GCN-O0-NEXT: ; %bb.2: ; %bb.inner.then
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
@@ -112,7 +111,9 @@ define amdgpu_kernel void @simple_nested_if(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
@@ -248,10 +249,10 @@ define amdgpu_kernel void @uncollapsable_nested_if(ptr addrspace(1) nocapture %a
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_readlane_b32 s4, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s5, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s0, 0
; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
@@ -277,7 +278,6 @@ define amdgpu_kernel void @uncollapsable_nested_if(ptr addrspace(1) nocapture %a
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB1_4
; GCN-O0-NEXT: ; %bb.2: ; %bb.inner.then
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
@@ -285,7 +285,9 @@ define amdgpu_kernel void @uncollapsable_nested_if(ptr addrspace(1) nocapture %a
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
@@ -311,7 +313,6 @@ define amdgpu_kernel void @uncollapsable_nested_if(ptr addrspace(1) nocapture %a
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: s_branch .LBB1_5
; GCN-O0-NEXT: .LBB1_4: ; %bb.inner.end
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
@@ -322,7 +323,9 @@ define amdgpu_kernel void @uncollapsable_nested_if(ptr addrspace(1) nocapture %a
; GCN-O0-NEXT: s_or_b64 exec, exec, s[2:3]
; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: v_mov_b32_e32 v0, 2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
@@ -508,7 +511,6 @@ define amdgpu_kernel void @nested_if_if_else(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: s_xor_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB2_5
; GCN-O0-NEXT: ; %bb.3: ; %bb.then
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
@@ -516,7 +518,9 @@ define amdgpu_kernel void @nested_if_if_else(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
@@ -532,7 +536,6 @@ define amdgpu_kernel void @nested_if_if_else(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
; GCN-O0-NEXT: s_branch .LBB2_5
; GCN-O0-NEXT: .LBB2_4: ; %bb.else
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
@@ -540,7 +543,9 @@ define amdgpu_kernel void @nested_if_if_else(ptr addrspace(1) nocapture %arg) {
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: v_mov_b32_e32 v0, 2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
@@ -954,20 +959,21 @@ define amdgpu_kernel void @s_endpgm_unsafe_barrier(ptr addrspace(1) nocapture %a
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB4_2
; GCN-O0-NEXT: ; %bb.1: ; %bb.then
-; GCN-O0-NEXT: s_waitcnt...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/69924
More information about the llvm-commits
mailing list