[llvm] AMDGPU/GlobalISelDivergenceLowering: select divergent i1 phis (PR #76145)
via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 21 04:48:33 PST 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-globalisel
@llvm/pr-subscribers-backend-amdgpu
Author: Petar Avramovic (petar-avramovic)
<details>
<summary>Changes</summary>
Implement PhiLoweringHelper for GlobalISel in DivergenceLoweringHelper. Use machine uniformity analysis to find divergent i1 phis and select them as lane mask phis in same way SILowerI1Copies select VReg_1 phis. Note that divergent i1 phis include phis created by LCSSA and all cases of uses outside of cycle are actually covered by "lowering LCSSA phis". GlobalISel lane masks are registers with sgpr register class and S1 LLT.
TODO: General goal is that instructions created in this pass are fully instruction-selected so that selection of lane mask phis is not split across multiple passes.
patch 3 from: https://github.com/llvm/llvm-project/pull/73337
---
Patch is 125.94 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/76145.diff
15 Files Affected:
- (modified) llvm/include/llvm/CodeGen/MachineUniformityAnalysis.h (+19)
- (modified) llvm/lib/CodeGen/MachineUniformityAnalysis.cpp (-19)
- (modified) llvm/lib/Target/AMDGPU/AMDGPUGlobalISelDivergenceLowering.cpp (+154-1)
- (modified) llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp (+8)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.ll (+1)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.mir (+45-25)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-used-outside-loop.ll (+1)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-used-outside-loop.mir (+249-85)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-structurizer.ll (+1)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-structurizer.mir (+221-65)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-temporal-divergent-i1.ll (+1)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-temporal-divergent-i1.mir (+65-31)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/divergent-control-flow.ll (+1)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-phi.mir (+2-2)
- (modified) llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.div.fmas.ll (+1)
``````````diff
diff --git a/llvm/include/llvm/CodeGen/MachineUniformityAnalysis.h b/llvm/include/llvm/CodeGen/MachineUniformityAnalysis.h
index e6da099751e7ae..1039ac4e5189b3 100644
--- a/llvm/include/llvm/CodeGen/MachineUniformityAnalysis.h
+++ b/llvm/include/llvm/CodeGen/MachineUniformityAnalysis.h
@@ -32,6 +32,25 @@ MachineUniformityInfo computeMachineUniformityInfo(
MachineFunction &F, const MachineCycleInfo &cycleInfo,
const MachineDomTree &domTree, bool HasBranchDivergence);
+/// Legacy analysis pass which computes a \ref MachineUniformityInfo.
+class MachineUniformityAnalysisPass : public MachineFunctionPass {
+ MachineUniformityInfo UI;
+
+public:
+ static char ID;
+
+ MachineUniformityAnalysisPass();
+
+ MachineUniformityInfo &getUniformityInfo() { return UI; }
+ const MachineUniformityInfo &getUniformityInfo() const { return UI; }
+
+ bool runOnMachineFunction(MachineFunction &F) override;
+ void getAnalysisUsage(AnalysisUsage &AU) const override;
+ void print(raw_ostream &OS, const Module *M = nullptr) const override;
+
+ // TODO: verify analysis
+};
+
} // namespace llvm
#endif // LLVM_CODEGEN_MACHINEUNIFORMITYANALYSIS_H
diff --git a/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp b/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp
index 3e0fe2b1ba087f..131138e0649e4c 100644
--- a/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp
+++ b/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp
@@ -165,25 +165,6 @@ MachineUniformityInfo llvm::computeMachineUniformityInfo(
namespace {
-/// Legacy analysis pass which computes a \ref MachineUniformityInfo.
-class MachineUniformityAnalysisPass : public MachineFunctionPass {
- MachineUniformityInfo UI;
-
-public:
- static char ID;
-
- MachineUniformityAnalysisPass();
-
- MachineUniformityInfo &getUniformityInfo() { return UI; }
- const MachineUniformityInfo &getUniformityInfo() const { return UI; }
-
- bool runOnMachineFunction(MachineFunction &F) override;
- void getAnalysisUsage(AnalysisUsage &AU) const override;
- void print(raw_ostream &OS, const Module *M = nullptr) const override;
-
- // TODO: verify analysis
-};
-
class MachineUniformityInfoPrinterPass : public MachineFunctionPass {
public:
static char ID;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUGlobalISelDivergenceLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUGlobalISelDivergenceLowering.cpp
index 4cd8b1ec1051f4..c7ad7f86966e14 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUGlobalISelDivergenceLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUGlobalISelDivergenceLowering.cpp
@@ -16,7 +16,10 @@
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
+#include "SILowerI1Copies.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineUniformityAnalysis.h"
+#include "llvm/InitializePasses.h"
#define DEBUG_TYPE "amdgpu-global-isel-divergence-lowering"
@@ -42,14 +45,155 @@ class AMDGPUGlobalISelDivergenceLowering : public MachineFunctionPass {
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
+ AU.addRequired<MachineDominatorTree>();
+ AU.addRequired<MachinePostDominatorTree>();
+ AU.addRequired<MachineUniformityAnalysisPass>();
MachineFunctionPass::getAnalysisUsage(AU);
}
};
+class DivergenceLoweringHelper : public PhiLoweringHelper {
+public:
+ DivergenceLoweringHelper(MachineFunction *MF, MachineDominatorTree *DT,
+ MachinePostDominatorTree *PDT,
+ MachineUniformityInfo *MUI);
+
+private:
+ MachineUniformityInfo *MUI = nullptr;
+
+public:
+ void markAsLaneMask(Register DstReg) const override;
+ void getCandidatesForLowering(
+ SmallVectorImpl<MachineInstr *> &Vreg1Phis) const override;
+ void collectIncomingValuesFromPhi(
+ const MachineInstr *MI,
+ SmallVectorImpl<Incoming> &Incomings) const override;
+ void replaceDstReg(Register NewReg, Register OldReg,
+ MachineBasicBlock *MBB) override;
+ void buildMergeLaneMasks(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I, const DebugLoc &DL,
+ Register DstReg, Register PrevReg,
+ Register CurReg) override;
+ void constrainIncomingRegisterTakenAsIs(Incoming &In) override;
+};
+
+DivergenceLoweringHelper::DivergenceLoweringHelper(
+ MachineFunction *MF, MachineDominatorTree *DT,
+ MachinePostDominatorTree *PDT, MachineUniformityInfo *MUI)
+ : PhiLoweringHelper(MF, DT, PDT), MUI(MUI) {}
+
+// _(s1) -> SReg_32/64(s1)
+void DivergenceLoweringHelper::markAsLaneMask(Register DstReg) const {
+ assert(MRI->getType(DstReg) == LLT::scalar(1));
+
+ // Can't blindly set a register class on phi, users could have reg class
+ // constraints (e.g. sreg_32/64..._xexec classes for control flow intrinsics).
+ if (MRI->getRegClassOrNull(DstReg))
+ return;
+
+ MRI->setRegClass(DstReg, ST->getBoolRC());
+}
+
+void DivergenceLoweringHelper::getCandidatesForLowering(
+ SmallVectorImpl<MachineInstr *> &Vreg1Phis) const {
+ LLT S1 = LLT::scalar(1);
+
+ // Add divergent i1 phis to the list
+ for (MachineBasicBlock &MBB : *MF) {
+ for (MachineInstr &MI : MBB.phis()) {
+ Register Dst = MI.getOperand(0).getReg();
+ if (MRI->getType(Dst) == S1 && MUI->isDivergent(Dst))
+ Vreg1Phis.push_back(&MI);
+ }
+ }
+}
+
+void DivergenceLoweringHelper::collectIncomingValuesFromPhi(
+ const MachineInstr *MI, SmallVectorImpl<Incoming> &Incomings) const {
+ for (unsigned i = 1; i < MI->getNumOperands(); i += 2) {
+ Incomings.emplace_back(MI->getOperand(i).getReg(),
+ MI->getOperand(i + 1).getMBB(), Register());
+ }
+}
+
+void DivergenceLoweringHelper::replaceDstReg(Register NewReg, Register OldReg,
+ MachineBasicBlock *MBB) {
+ BuildMI(*MBB, MBB->getFirstNonPHI(), {}, TII->get(AMDGPU::COPY), OldReg)
+ .addReg(NewReg);
+}
+
+// Get pointers to build instruction just after MI (skips phis if needed)
+static std::pair<MachineBasicBlock *, MachineBasicBlock::iterator>
+getInsertAfterPtrs(MachineInstr *MI) {
+ MachineBasicBlock *InsertMBB = MI->getParent();
+ return {InsertMBB,
+ InsertMBB->SkipPHIsAndLabels(std::next(MI->getIterator()))};
+}
+
+// bb.previous
+// %PrevReg = ...
+//
+// bb.current
+// %CurReg = ...
+//
+// %DstReg - not defined
+//
+// -> (wave32 example, new registers have sreg_32 reg class and S1 LLT)
+//
+// bb.previous
+// %PrevReg = ...
+// %PrevRegCopy:sreg_32(s1) = COPY %PrevReg
+//
+// bb.current
+// %CurReg = ...
+// %CurRegCopy:sreg_32(s1) = COPY %CurReg
+// ...
+// %PrevMaskedReg:sreg_32(s1) = ANDN2 %PrevRegCopy, ExecReg - active lanes 0
+// %CurMaskedReg:sreg_32(s1) = AND %ExecReg, CurRegCopy - inactive lanes to 0
+// %DstReg:sreg_32(s1) = OR %PrevMaskedReg, CurMaskedReg
+//
+// DstReg = for active lanes rewrite bit in PrevReg with bit from CurReg
+void DivergenceLoweringHelper::buildMergeLaneMasks(
+ MachineBasicBlock &MBB, MachineBasicBlock::iterator I, const DebugLoc &DL,
+ Register DstReg, Register PrevReg, Register CurReg) {
+ // DstReg = (PrevReg & !EXEC) | (CurReg & EXEC)
+ // TODO: check if inputs are constants or results of a compare.
+
+ Register PrevRegCopy = createLaneMaskReg(MRI, LaneMaskRegAttrs);
+ auto [PrevMBB, AfterPrevReg] = getInsertAfterPtrs(MRI->getVRegDef(PrevReg));
+ BuildMI(*PrevMBB, AfterPrevReg, DL, TII->get(AMDGPU::COPY), PrevRegCopy)
+ .addReg(PrevReg);
+ Register PrevMaskedReg = createLaneMaskReg(MRI, LaneMaskRegAttrs);
+ BuildMI(MBB, I, DL, TII->get(AndN2Op), PrevMaskedReg)
+ .addReg(PrevRegCopy)
+ .addReg(ExecReg);
+
+ Register CurRegCopy = createLaneMaskReg(MRI, LaneMaskRegAttrs);
+ auto [CurMBB, AfterCurReg] = getInsertAfterPtrs(MRI->getVRegDef(CurReg));
+ BuildMI(*CurMBB, AfterCurReg, DL, TII->get(AMDGPU::COPY), CurRegCopy)
+ .addReg(CurReg);
+ Register CurMaskedReg = createLaneMaskReg(MRI, LaneMaskRegAttrs);
+ BuildMI(MBB, I, DL, TII->get(AndOp), CurMaskedReg)
+ .addReg(ExecReg)
+ .addReg(CurRegCopy);
+
+ BuildMI(MBB, I, DL, TII->get(OrOp), DstReg)
+ .addReg(PrevMaskedReg)
+ .addReg(CurMaskedReg);
+}
+
+void DivergenceLoweringHelper::constrainIncomingRegisterTakenAsIs(
+ Incoming &In) {
+ return;
+}
+
} // End anonymous namespace.
INITIALIZE_PASS_BEGIN(AMDGPUGlobalISelDivergenceLowering, DEBUG_TYPE,
"AMDGPU GlobalISel divergence lowering", false, false)
+INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
+INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
+INITIALIZE_PASS_DEPENDENCY(MachineUniformityAnalysisPass)
INITIALIZE_PASS_END(AMDGPUGlobalISelDivergenceLowering, DEBUG_TYPE,
"AMDGPU GlobalISel divergence lowering", false, false)
@@ -64,5 +208,14 @@ FunctionPass *llvm::createAMDGPUGlobalISelDivergenceLoweringPass() {
bool AMDGPUGlobalISelDivergenceLowering::runOnMachineFunction(
MachineFunction &MF) {
- return false;
+ MachineDominatorTree &DT = getAnalysis<MachineDominatorTree>();
+ MachinePostDominatorTree &PDT = getAnalysis<MachinePostDominatorTree>();
+ MachineUniformityInfo &MUI =
+ getAnalysis<MachineUniformityAnalysisPass>().getUniformityInfo();
+
+ DivergenceLoweringHelper Helper(&MF, &DT, &PDT, &MUI);
+
+ bool Changed = false;
+ Changed |= Helper.lowerPhis();
+ return Changed;
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 88ef4b5774242d..f2b716c44d47b2 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -210,6 +210,14 @@ bool AMDGPUInstructionSelector::selectCOPY(MachineInstr &I) const {
bool AMDGPUInstructionSelector::selectPHI(MachineInstr &I) const {
const Register DefReg = I.getOperand(0).getReg();
const LLT DefTy = MRI->getType(DefReg);
+ // Lane mask PHIs, PHI where all register operands have sgpr register class
+ // with S1 LLT, are already selected in divergence lowering pass.
+ if (I.getOpcode() == AMDGPU::PHI) {
+ assert(MRI->getType(DefReg) == LLT::scalar(1));
+ assert(TRI.isSGPRClass(MRI->getRegClass(DefReg)));
+ return true;
+ }
+
if (DefTy == LLT::scalar(1)) {
if (!AllowRiskySelect) {
LLVM_DEBUG(dbgs() << "Skipping risky boolean phi\n");
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.ll
index ccf4e84fbbbd16..e573df4e8f76a9 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.ll
@@ -1,5 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
; RUN: llc -global-isel -amdgpu-global-isel-risky-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; REQUIRES: do-not-run-me
; Divergent phis that don't require lowering using lane mask merging
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.mir
index d314ebe355f51d..56f2812b590a8d 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/divergence-divergent-i1-phis-no-lane-mask-merging.mir
@@ -46,7 +46,7 @@ body: |
; GFX10-NEXT: bb.2:
; GFX10-NEXT: successors: %bb.4(0x80000000)
; GFX10-NEXT: {{ $}}
- ; GFX10-NEXT: [[PHI:%[0-9]+]]:_(s1) = G_PHI %14(s1), %bb.3, [[ICMP]](s1), %bb.0
+ ; GFX10-NEXT: [[PHI:%[0-9]+]]:sreg_32(s1) = G_PHI %14(s1), %bb.3, [[ICMP]](s1), %bb.0
; GFX10-NEXT: G_BR %bb.4
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: bb.3:
@@ -126,6 +126,7 @@ body: |
; GFX10-NEXT: [[COPY3:%[0-9]+]]:_(s32) = COPY $sgpr0
; GFX10-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
; GFX10-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[COPY2]](s32), [[C]]
+ ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_32(s1) = COPY [[ICMP]](s1)
; GFX10-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
; GFX10-NEXT: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY3]](s32), [[C1]]
; GFX10-NEXT: G_BRCOND [[ICMP1]](s1), %bb.2
@@ -136,12 +137,17 @@ body: |
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
; GFX10-NEXT: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(ult), [[COPY2]](s32), [[C2]]
+ ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32(s1) = COPY [[ICMP2]](s1)
+ ; GFX10-NEXT: [[S_ANDN2_B32_:%[0-9]+]]:sreg_32(s1) = S_ANDN2_B32 [[COPY4]](s1), $exec_lo, implicit-def $scc
+ ; GFX10-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32(s1) = S_AND_B32 $exec_lo, [[COPY5]](s1), implicit-def $scc
+ ; GFX10-NEXT: [[S_OR_B32_:%[0-9]+]]:sreg_32(s1) = S_OR_B32 [[S_ANDN2_B32_]](s1), [[S_AND_B32_]](s1), implicit-def $scc
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: bb.2:
- ; GFX10-NEXT: [[PHI:%[0-9]+]]:_(s1) = G_PHI [[ICMP]](s1), %bb.0, [[ICMP2]](s1), %bb.1
+ ; GFX10-NEXT: [[PHI:%[0-9]+]]:sreg_32 = PHI [[ICMP]](s1), %bb.0, [[S_OR_B32_]](s1), %bb.1
+ ; GFX10-NEXT: [[COPY6:%[0-9]+]]:sreg_32(s1) = COPY [[PHI]]
; GFX10-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
; GFX10-NEXT: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
- ; GFX10-NEXT: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[PHI]](s1), [[C4]], [[C3]]
+ ; GFX10-NEXT: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[COPY6]](s1), [[C4]], [[C3]]
; GFX10-NEXT: G_STORE [[SELECT]](s32), [[MV]](p1) :: (store (s32), addrspace 1)
; GFX10-NEXT: S_ENDPGM 0
bb.0:
@@ -191,30 +197,37 @@ body: |
; GFX10-NEXT: [[MV:%[0-9]+]]:_(p0) = G_MERGE_VALUES [[COPY1]](s32), [[COPY2]](s32)
; GFX10-NEXT: [[C:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
; GFX10-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+ ; GFX10-NEXT: [[DEF:%[0-9]+]]:sreg_32(s1) = IMPLICIT_DEF
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: bb.1:
; GFX10-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
; GFX10-NEXT: {{ $}}
- ; GFX10-NEXT: [[PHI:%[0-9]+]]:_(s32) = G_PHI %7(s32), %bb.1, [[C1]](s32), %bb.0
- ; GFX10-NEXT: [[PHI1:%[0-9]+]]:_(s32) = G_PHI [[C1]](s32), %bb.0, %9(s32), %bb.1
- ; GFX10-NEXT: [[PHI2:%[0-9]+]]:_(s1) = G_PHI [[C]](s1), %bb.0, %11(s1), %bb.1
+ ; GFX10-NEXT: [[PHI:%[0-9]+]]:sreg_32 = PHI [[DEF]](s1), %bb.0, %22(s1), %bb.1
+ ; GFX10-NEXT: [[PHI1:%[0-9]+]]:_(s32) = G_PHI %7(s32), %bb.1, [[C1]](s32), %bb.0
+ ; GFX10-NEXT: [[PHI2:%[0-9]+]]:_(s32) = G_PHI [[C1]](s32), %bb.0, %9(s32), %bb.1
+ ; GFX10-NEXT: [[PHI3:%[0-9]+]]:_(s1) = G_PHI [[C]](s1), %bb.0, %11(s1), %bb.1
+ ; GFX10-NEXT: [[COPY3:%[0-9]+]]:sreg_32(s1) = COPY [[PHI]]
; GFX10-NEXT: [[C2:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
- ; GFX10-NEXT: [[XOR:%[0-9]+]]:_(s1) = G_XOR [[PHI2]], [[C2]]
- ; GFX10-NEXT: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[PHI1]](s32)
+ ; GFX10-NEXT: [[XOR:%[0-9]+]]:_(s1) = G_XOR [[PHI3]], [[C2]]
+ ; GFX10-NEXT: [[COPY4:%[0-9]+]]:sreg_32(s1) = COPY [[XOR]](s1)
+ ; GFX10-NEXT: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[PHI2]](s32)
; GFX10-NEXT: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(ogt), [[UITOFP]](s32), [[COPY]]
; GFX10-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
- ; GFX10-NEXT: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[PHI1]], [[C3]]
- ; GFX10-NEXT: [[INTRINSIC_CONVERGENT:%[0-9]+]]:sreg_32_xm0_xexec(s32) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.if.break), [[FCMP]](s1), [[PHI]](s32)
+ ; GFX10-NEXT: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[PHI2]], [[C3]]
+ ; GFX10-NEXT: [[INTRINSIC_CONVERGENT:%[0-9]+]]:sreg_32_xm0_xexec(s32) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.if.break), [[FCMP]](s1), [[PHI1]](s32)
+ ; GFX10-NEXT: [[S_ANDN2_B32_:%[0-9]+]]:sreg_32(s1) = S_ANDN2_B32 [[COPY3]](s1), $exec_lo, implicit-def $scc
+ ; GFX10-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32(s1) = S_AND_B32 $exec_lo, [[COPY4]](s1), implicit-def $scc
+ ; GFX10-NEXT: [[S_OR_B32_:%[0-9]+]]:sreg_32(s1) = S_OR_B32 [[S_ANDN2_B32_]](s1), [[S_AND_B32_]](s1), implicit-def $scc
; GFX10-NEXT: SI_LOOP [[INTRINSIC_CONVERGENT]](s32), %bb.1, implicit-def $exec, implicit-def $scc, implicit $exec
; GFX10-NEXT: G_BR %bb.2
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: bb.2:
- ; GFX10-NEXT: [[PHI3:%[0-9]+]]:_(s1) = G_PHI [[XOR]](s1), %bb.1
; GFX10-NEXT: [[PHI4:%[0-9]+]]:_(s32) = G_PHI [[INTRINSIC_CONVERGENT]](s32), %bb.1
+ ; GFX10-NEXT: [[COPY5:%[0-9]+]]:sreg_32(s1) = COPY [[S_OR_B32_]](s1)
; GFX10-NEXT: G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.end.cf), [[PHI4]](s32)
; GFX10-NEXT: [[C4:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
; GFX10-NEXT: [[C5:%[0-9]+]]:_(s32) = G_FCONSTANT float 1.000000e+00
- ; GFX10-NEXT: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[PHI3]](s1), [[C5]], [[C4]]
+ ; GFX10-NEXT: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[COPY5]](s1), [[C5]], [[C4]]
; GFX10-NEXT: G_STORE [[SELECT]](s32), [[MV]](p0) :: (store (s32))
; GFX10-NEXT: SI_RETURN
bb.0:
@@ -279,25 +292,28 @@ body: |
; GFX10-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
; GFX10-NEXT: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 1.000000e+00
; GFX10-NEXT: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(ogt), [[COPY1]](s32), [[C1]]
+ ; GFX10-NEXT: [[DEF:%[0-9]+]]:sreg_32(s1) = IMPLICIT_DEF
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: bb.1:
; GFX10-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000)
; GFX10-NEXT: {{ $}}
- ; GFX10-NEXT: [[PHI:%[0-9]+]]:_(s32) = G_PHI %15(s32), %bb.5, [[C]](s32), %bb.0
- ; GFX10-NEXT: [[PHI1:%[0-9]+]]:_(s32) = G_PHI [[C]](s32), %bb.0, %17(s32), %bb.5
- ; GFX10-NEXT: [[PHI2:%[0-9]+]]:_(s1) = G_PHI [[FCMP]](s1), %bb.0, %19(s1), %bb.5
+ ; GFX10-NEXT: [[PHI:%[0-9]+]]:sreg_32 = PHI [[DEF]](s1), %bb.0, %39(s1), %bb.5
+ ; GFX10-NEXT: [[PHI1:%[0-9]+]]:_(s32) = G_PHI %15(s32), %bb.5, [[C]](s32), %bb.0
+ ; GFX10-NEXT: [[PHI2:%[0-9]+]]:_(s32) = G_PHI [[C]](s32), %bb.0, %17(s32), %bb.5
+ ; GFX10-NEXT: [[PHI3:%[0-9]+]]:sreg_32(s1) = G_PHI [[FCMP]](s1), %bb.0, %19(s1), %bb.5
+ ; GFX10-NEXT: [[COPY8:%[0-9]+]]:sreg_32(s1) = COPY [[PHI]]
; GFX10-NEXT: [[C2:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
; GFX10-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1000
- ; GFX10-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(sle), [[PHI1]](s32), [[C3]]
+ ; GFX10-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(sle), [[PHI2]](s32), [[C3]]
; GFX10-NEXT: G_BRCOND [[ICMP]](s1), %bb.4
; GFX10-NEXT: G_BR %bb.2
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: bb.2:
; GFX10-NEXT: successors: %bb.3(0x40000000), %bb.5(0x40000000)
; GFX10-NEXT: {{ $}}
- ; GFX10-NEXT: [[PHI3:%[0-9]+]]:_(s1) = G_PHI %24(s1), %bb.4, [[C2]](s1), %bb.1
+ ; GFX10-NEXT: [[PHI4:%[0-9]+]]:_(s1) = G_PHI %24(s1), %bb.4, [[C2]](s1), %bb.1
; GFX10-NEXT: [[C4:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
- ; GFX10-NEXT: [[XOR:%[0-9]+]]:_(s1) = G_XOR [[PHI3]], [[C4]]
+ ; GFX10-NEXT: [[XOR:%[0-9]+]]:_(s1) = G_XOR [[PHI4]], [[C4]]
; GFX10-NEXT: G_BRCOND [[XOR]](s1), %bb.5
; GFX10-NEXT: G_BR %bb.3
; GFX10-NEXT: {{ $}}
@@ -320,22 +336,26 @@ body: |
; GFX10-NEXT: successors: %bb.6(0x04000000), %bb.1(0x7c000000)
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: [[C8:%[0-9]+]]:_(s1) = G_CONSTANT i1 true
- ; GFX10-NEXT: [[XOR1:%[0-9]+]]:_(s1) = G_XOR [[PHI2]], [[C8]]
- ; GFX10-NEXT: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[PHI1]](s32)
+ ; GFX10-NEXT: [[XOR1:%[0-9]+]]:_(s1) = G_XOR [[PHI3]], [[C8]]
+ ; GFX10-NEXT: [[COPY9:%[0-9]+]]:sreg_32(s1) = COPY [[XOR1]](s1)
+ ; GFX10-NEXT: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[PHI2]](s32)
; GFX10-NEXT: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(ogt), [[UITOFP]](s32), [[COPY]]
; GFX10-NEXT: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
- ; GFX10-NEXT: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[PHI1]], [[C9]]
- ; GFX10-NEXT: [[INTRINSIC_CONVERGENT:%[0-9]+]]:sreg_32_xm0_xexec(s32) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.if.break), [[FCMP1]](s1), [[PHI]](s32)
+ ; GFX10-NEXT: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[PHI2]], [[C9]]
+ ; GFX10-NEXT: [[INTRINSIC_CONVERGENT:%[0-9]+]]:sreg_32_xm0_xexec(s32) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.if.break), [[FCMP1]](...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/76145
More information about the llvm-commits
mailing list