[llvm] AMDGPU: Fix temporal divergence introduced by machine-sink and performance regression introduced by D155343 (PR #67456)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Sep 26 09:47:44 PDT 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-adt
<details>
<summary>Changes</summary>
D155343 introduced performance regressions by not allowing machine instruction sinking in some cases - SWEDEV-414443.
It originally fixed SWDEV-407790
Revert D155343 to fix performance regression SWEDEV-414443
and introduce temporal divergence lowering pass that also fixes SWDEV-407790
Add a target hook to TargetInstrInfo that can be used to fix temporal divergence that can be introduced by sinking instruction outside cycle with divergent exit. AMDGPU implementation uses machine uniformity analysis to detect temporal divergent uses.
---
Patch is 66.88 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/67456.diff
12 Files Affected:
- (modified) llvm/include/llvm/ADT/GenericUniformityImpl.h (+43)
- (modified) llvm/include/llvm/ADT/GenericUniformityInfo.h (+6)
- (modified) llvm/include/llvm/CodeGen/TargetInstrInfo.h (+8)
- (modified) llvm/lib/Analysis/UniformityAnalysis.cpp (+7-2)
- (modified) llvm/lib/CodeGen/MachineSink.cpp (+10-11)
- (modified) llvm/lib/CodeGen/MachineUniformityAnalysis.cpp (+7-2)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+60)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.h (+3)
- (modified) llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll (+1-1)
- (modified) llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.mir (+1-1)
- (added) llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll (+1094)
- (modified) llvm/test/CodeGen/AMDGPU/sink-after-control-flow.mir (+1-1)
``````````diff
diff --git a/llvm/include/llvm/ADT/GenericUniformityImpl.h b/llvm/include/llvm/ADT/GenericUniformityImpl.h
index ddd0746ccd91632..755e1161b41b521 100644
--- a/llvm/include/llvm/ADT/GenericUniformityImpl.h
+++ b/llvm/include/llvm/ADT/GenericUniformityImpl.h
@@ -341,6 +341,9 @@ template <typename ContextT> class GenericUniformityAnalysisImpl {
using DivergenceDescriptorT =
typename SyncDependenceAnalysisT::DivergenceDescriptor;
using BlockLabelMapT = typename SyncDependenceAnalysisT::BlockLabelMap;
+ using UseOutsideCycleWithDivergentExitInfoT =
+ typename std::tuple<ConstValueRefT, const InstructionT *,
+ SmallVector<BlockT *, 4>>;
GenericUniformityAnalysisImpl(const DominatorTreeT &DT, const CycleInfoT &CI,
const TargetTransformInfo *TTI)
@@ -396,6 +399,9 @@ template <typename ContextT> class GenericUniformityAnalysisImpl {
void print(raw_ostream &out) const;
+ iterator_range<const UseOutsideCycleWithDivergentExitInfoT *>
+ uses_outside_cycles_with_divergent_exit() const;
+
protected:
/// \brief Value/block pair representing a single phi input.
struct PhiInput {
@@ -427,6 +433,8 @@ template <typename ContextT> class GenericUniformityAnalysisImpl {
// Recognized cycles with divergent exits.
SmallPtrSet<const CycleT *, 16> DivergentExitCycles;
+ SmallVector<UseOutsideCycleWithDivergentExitInfoT, 4>
+ UsesOutsideCyclesWithDivergentExit;
// Cycles assumed to be divergent.
//
@@ -470,6 +478,10 @@ template <typename ContextT> class GenericUniformityAnalysisImpl {
/// \brief Whether \p Def is divergent when read in \p ObservingBlock.
bool isTemporalDivergent(const BlockT &ObservingBlock,
const InstructionT &Def) const;
+
+ void recordUseOutsideCycleWithDivergentExit(ConstValueRefT Src,
+ const InstructionT *UserInstr,
+ const CycleT &DefCycle);
};
template <typename ImplT>
@@ -1210,6 +1222,20 @@ void GenericUniformityAnalysisImpl<ContextT>::print(raw_ostream &OS) const {
}
}
+template <typename ContextT>
+using UseOutsideCycleWithDivergentExitInfoT =
+ typename std::tuple<typename ContextT::ConstValueRefT,
+ const typename ContextT::InstructionT *,
+ SmallVector<typename ContextT::BlockT *, 4>>;
+
+template <typename ContextT>
+iterator_range<const UseOutsideCycleWithDivergentExitInfoT<ContextT> *>
+GenericUniformityAnalysisImpl<
+ ContextT>::uses_outside_cycles_with_divergent_exit() const {
+ return make_range(UsesOutsideCyclesWithDivergentExit.begin(),
+ UsesOutsideCyclesWithDivergentExit.end());
+}
+
template <typename ContextT>
bool GenericUniformityInfo<ContextT>::hasDivergence() const {
return DA->hasDivergence();
@@ -1248,6 +1274,13 @@ void GenericUniformityInfo<ContextT>::print(raw_ostream &out) const {
DA->print(out);
}
+template <typename ContextT>
+iterator_range<const UseOutsideCycleWithDivergentExitInfoT<ContextT> *>
+GenericUniformityInfo<ContextT>::uses_outside_cycles_with_divergent_exit()
+ const {
+ return DA->uses_outside_cycles_with_divergent_exit();
+}
+
template <typename ContextT>
void llvm::ModifiedPostOrder<ContextT>::computeStackPO(
SmallVectorImpl<const BlockT *> &Stack, const CycleInfoT &CI,
@@ -1367,6 +1400,16 @@ void llvm::ModifiedPostOrder<ContextT>::compute(const CycleInfoT &CI) {
computeStackPO(Stack, CI, nullptr, Finalized);
}
+template <typename ContextT>
+void GenericUniformityAnalysisImpl<ContextT>::
+ recordUseOutsideCycleWithDivergentExit(ConstValueRefT Src,
+ const InstructionT *UserInstr,
+ const CycleT &DefCycle) {
+ SmallVector<BlockT *, 4> TmpExitBlocks;
+ DefCycle.getExitBlocks(TmpExitBlocks);
+ UsesOutsideCyclesWithDivergentExit.push_back({Src, UserInstr, TmpExitBlocks});
+}
+
} // namespace llvm
#undef DEBUG_TYPE
diff --git a/llvm/include/llvm/ADT/GenericUniformityInfo.h b/llvm/include/llvm/ADT/GenericUniformityInfo.h
index e53afccc020b469..a76813d4bb964a1 100644
--- a/llvm/include/llvm/ADT/GenericUniformityInfo.h
+++ b/llvm/include/llvm/ADT/GenericUniformityInfo.h
@@ -39,6 +39,9 @@ template <typename ContextT> class GenericUniformityInfo {
using CycleInfoT = GenericCycleInfo<ContextT>;
using CycleT = typename CycleInfoT::CycleT;
+ using UseOutsideCycleWithDivergentExitInfoT =
+ typename std::tuple<ConstValueRefT, const InstructionT *,
+ SmallVector<BlockT *, 4>>;
GenericUniformityInfo(const DominatorTreeT &DT, const CycleInfoT &CI,
const TargetTransformInfo *TTI = nullptr);
@@ -78,6 +81,9 @@ template <typename ContextT> class GenericUniformityInfo {
void print(raw_ostream &Out) const;
+ iterator_range<const UseOutsideCycleWithDivergentExitInfoT *>
+ uses_outside_cycles_with_divergent_exit() const;
+
private:
using ImplT = GenericUniformityAnalysisImpl<ContextT>;
diff --git a/llvm/include/llvm/CodeGen/TargetInstrInfo.h b/llvm/include/llvm/CodeGen/TargetInstrInfo.h
index 98679b4dcf3cbfb..2135484448ef4ad 100644
--- a/llvm/include/llvm/CodeGen/TargetInstrInfo.h
+++ b/llvm/include/llvm/CodeGen/TargetInstrInfo.h
@@ -19,6 +19,8 @@
#include "llvm/ADT/Uniformity.h"
#include "llvm/CodeGen/MIRFormatter.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineCycleAnalysis.h"
+#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
@@ -150,6 +152,12 @@ class TargetInstrInfo : public MCInstrInfo {
return false;
}
+ virtual void fixTemporalDivergence(MachineFunction &MF,
+ MachineDominatorTree *DT,
+ MachineCycleInfo *CI) const {
+ return;
+ }
+
protected:
/// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
/// set, this hook lets the target specify whether the instruction is actually
diff --git a/llvm/lib/Analysis/UniformityAnalysis.cpp b/llvm/lib/Analysis/UniformityAnalysis.cpp
index 2d617db431c5888..df1299610469d30 100644
--- a/llvm/lib/Analysis/UniformityAnalysis.cpp
+++ b/llvm/lib/Analysis/UniformityAnalysis.cpp
@@ -80,12 +80,17 @@ template <>
void llvm::GenericUniformityAnalysisImpl<
SSAContext>::propagateTemporalDivergence(const Instruction &I,
const Cycle &DefCycle) {
- if (isDivergent(I))
- return;
for (auto *User : I.users()) {
auto *UserInstr = cast<Instruction>(User);
if (DefCycle.contains(UserInstr->getParent()))
continue;
+
+ recordUseOutsideCycleWithDivergentExit(cast<Value>(&I), UserInstr,
+ DefCycle);
+
+ if (isDivergent(I))
+ continue;
+
markDivergent(*UserInstr);
}
}
diff --git a/llvm/lib/CodeGen/MachineSink.cpp b/llvm/lib/CodeGen/MachineSink.cpp
index 480ac23d43ad879..ef35ec57a56971a 100644
--- a/llvm/lib/CodeGen/MachineSink.cpp
+++ b/llvm/lib/CodeGen/MachineSink.cpp
@@ -300,7 +300,8 @@ static bool blockPrologueInterferes(const MachineBasicBlock *BB,
if (!Reg)
continue;
if (MO.isUse()) {
- if (Reg.isPhysical() && MRI && MRI->isConstantPhysReg(Reg))
+ if (Reg.isPhysical() &&
+ (TII->isIgnorableUse(MO) || (MRI && MRI->isConstantPhysReg(Reg))))
continue;
if (PI->modifiesRegister(Reg, TRI))
return true;
@@ -776,6 +777,12 @@ bool MachineSinking::runOnMachineFunction(MachineFunction &MF) {
MRI->clearKillFlags(I);
RegsToClearKillFlags.clear();
+ // In some tests there is a new baic block, need to recompute
+ DT->calculate(MF);
+ CI->clear();
+ CI->compute(MF);
+ TII->fixTemporalDivergence(MF, DT, CI);
+
return EverMadeChange;
}
@@ -1245,24 +1252,16 @@ MachineSinking::FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB,
if (MBB == SuccToSinkTo)
return nullptr;
- if (!SuccToSinkTo)
- return nullptr;
-
// It's not safe to sink instructions to EH landing pad. Control flow into
// landing pad is implicitly defined.
- if (SuccToSinkTo->isEHPad())
+ if (SuccToSinkTo && SuccToSinkTo->isEHPad())
return nullptr;
// It ought to be okay to sink instructions into an INLINEASM_BR target, but
// only if we make sure that MI occurs _before_ an INLINEASM_BR instruction in
// the source block (which this code does not yet do). So for now, forbid
// doing so.
- if (SuccToSinkTo->isInlineAsmBrIndirectTarget())
- return nullptr;
-
- MachineBasicBlock::const_iterator InsertPos =
- SuccToSinkTo->SkipPHIsAndLabels(SuccToSinkTo->begin());
- if (blockPrologueInterferes(SuccToSinkTo, InsertPos, MI, TRI, TII, MRI))
+ if (SuccToSinkTo && SuccToSinkTo->isInlineAsmBrIndirectTarget())
return nullptr;
return SuccToSinkTo;
diff --git a/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp b/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp
index 3e0fe2b1ba087fe..5bd87d069285857 100644
--- a/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp
+++ b/llvm/lib/CodeGen/MachineUniformityAnalysis.cpp
@@ -117,11 +117,16 @@ void llvm::GenericUniformityAnalysisImpl<MachineSSAContext>::
if (!Op.getReg().isVirtual())
continue;
auto Reg = Op.getReg();
- if (isDivergent(Reg))
- continue;
+
for (MachineInstr &UserInstr : RegInfo.use_instructions(Reg)) {
if (DefCycle.contains(UserInstr.getParent()))
continue;
+
+ recordUseOutsideCycleWithDivergentExit(Reg, &UserInstr, DefCycle);
+
+ if (isDivergent(Reg))
+ continue;
+
markDivergent(UserInstr);
}
}
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 2799a3e78b04d22..5db140bb0fc6c94 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -24,6 +24,7 @@
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineScheduler.h"
+#include "llvm/CodeGen/MachineUniformityAnalysis.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/IR/DiagnosticInfo.h"
@@ -171,6 +172,65 @@ bool SIInstrInfo::isIgnorableUse(const MachineOperand &MO) const {
isVALU(*MO.getParent()) && !resultDependsOnExec(*MO.getParent());
}
+// Get poiners to build instruction after MI (skips phis if needed)
+static std::pair<MachineBasicBlock *, MachineBasicBlock::iterator>
+getInsertAfterPtrs(MachineInstr *MI) {
+ MachineBasicBlock *InsertMBB = MI->getParent();
+ return std::make_pair(
+ InsertMBB, InsertMBB->SkipPHIsAndLabels(std::next(MI->getIterator())));
+}
+
+static void replaceUseRegisterWith(const MachineInstr *MI, Register Reg,
+ Register Newreg) {
+ for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
+ const MachineOperand &Op = MI->getOperand(i);
+ if (Op.isReg() && Op.getReg() == Reg) {
+ const_cast<MachineInstr *>(MI)->getOperand(i).setReg(Newreg);
+ }
+ }
+}
+
+void SIInstrInfo::fixTemporalDivergence(MachineFunction &MF,
+ MachineDominatorTree *DT,
+ MachineCycleInfo *CI) const {
+ MachineUniformityInfo MUI =
+ computeMachineUniformityInfo(MF, *CI, DT->getBase(), true);
+
+ MachineRegisterInfo &MRI = MF.getRegInfo();
+ const GCNSubtarget &Subtarget = MF.getSubtarget<GCNSubtarget>();
+ const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
+ const SIRegisterInfo &TRI = *Subtarget.getRegisterInfo();
+
+ // Temporal divergence lowering is required for uniform UniformSourceReg and
+ // divergent UserInstr. UserInstr is uniform only when cycle has uniform exit.
+ for (auto [SrcReg, UserInstr, CycleExitBlocks] :
+ MUI.uses_outside_cycles_with_divergent_exit()) {
+ if (!MUI.isUniform(SrcReg) || !MUI.isDivergent(UserInstr))
+ continue;
+
+ MachineInstr *UniformSourceInstr = MRI.getVRegDef(SrcReg);
+
+ // FixMe: SrcReg is lane mask in this case. Find a better way to detect it.
+ if (UniformSourceInstr->getOpcode() == AMDGPU::SI_IF_BREAK ||
+ UserInstr->getOpcode() == AMDGPU::SI_IF)
+ continue;
+
+ unsigned Size = TRI.getRegSizeInBits(*MRI.getRegClassOrNull(SrcReg));
+ Register VgprDst =
+ MRI.createVirtualRegister(TRI.getVGPRClassForBitWidth(Size));
+
+ auto [MBB, AfterUniformSourceReg] = getInsertAfterPtrs(UniformSourceInstr);
+ BuildMI(*MBB, AfterUniformSourceReg, {}, TII.get(AMDGPU::COPY))
+ .addDef(VgprDst)
+ .addReg(SrcReg)
+ .addReg(AMDGPU::EXEC, RegState::Implicit);
+
+ replaceUseRegisterWith(UserInstr, SrcReg, VgprDst);
+ }
+
+ return;
+}
+
bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
int64_t &Offset0,
int64_t &Offset1) const {
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index e85917a4c0f3296..4446a962d7cca44 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -222,6 +222,9 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
bool isIgnorableUse(const MachineOperand &MO) const override;
+ void fixTemporalDivergence(MachineFunction &MF, MachineDominatorTree *DT,
+ MachineCycleInfo *CI) const override;
+
bool areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1, int64_t &Offset0,
int64_t &Offset1) const override;
diff --git a/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll b/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll
index e2456b74f7ef1fa..b8e74bc7db09a1a 100644
--- a/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll
+++ b/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll
@@ -21,6 +21,7 @@ define void @machinesink_loop_variable_out_of_divergent_loop(i32 %arg, i1 %cmp49
; CHECK-NEXT: .LBB0_1: ; %Flow
; CHECK-NEXT: ; in Loop: Header=BB0_3 Depth=1
; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; CHECK-NEXT: v_add_nc_u32_e32 v4, -4, v4
; CHECK-NEXT: .LBB0_2: ; %Flow1
; CHECK-NEXT: ; in Loop: Header=BB0_3 Depth=1
; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s7
@@ -53,7 +54,6 @@ define void @machinesink_loop_variable_out_of_divergent_loop(i32 %arg, i1 %cmp49
; CHECK-NEXT: ;;#ASMEND
; CHECK-NEXT: v_add_nc_u32_e32 v4, s9, v2
; CHECK-NEXT: v_cmp_ge_u32_e64 s4, v4, v0
-; CHECK-NEXT: v_add_nc_u32_e32 v4, -4, v4
; CHECK-NEXT: s_or_b32 s8, s4, s8
; CHECK-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
; CHECK-NEXT: s_cbranch_execz .LBB0_1
diff --git a/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.mir b/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.mir
index cc14b4a80d58a7d..037a285794120da 100644
--- a/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.mir
+++ b/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.mir
@@ -42,7 +42,6 @@ body: |
; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.5(0x40000000)
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: INLINEASM &"", 1 /* sideeffect attdialect */
- ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[COPY1]], 0, implicit $exec
; CHECK-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_32 = SI_IF_BREAK killed [[SI_IF1]], [[SI_IF]], implicit-def dead $scc
; CHECK-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.2, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
; CHECK-NEXT: S_BRANCH %bb.5
@@ -52,6 +51,7 @@ body: |
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.4
; CHECK-NEXT: SI_END_CF [[SI_IF_BREAK]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[COPY1]], 0, implicit $exec
; CHECK-NEXT: INLINEASM &"", 1 /* sideeffect attdialect */, implicit [[V_ADD_U32_e64_]]
; CHECK-NEXT: S_BRANCH %bb.2
; CHECK-NEXT: {{ $}}
diff --git a/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll b/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll
new file mode 100644
index 000000000000000..0c631bdcdd374a8
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll
@@ -0,0 +1,1094 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1031 < %s | FileCheck %s
+
+; ModuleID = 'kernel_round1_passing.bc'
+source_filename = "/tmp/comgr-295d04/input/CompileSource"
+target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8"
+target triple = "amdgcn-amd-amdhsa"
+
+ at kernel_round1.first_words_data = external hidden unnamed_addr addrspace(3) global [896 x i8], align 1
+ at kernel_round1.collisionsData = external hidden unnamed_addr addrspace(3) global [3840 x i32], align 4
+ at kernel_round1.collisionsNum = external hidden addrspace(3) global i32, align 4
+
+; Function Attrs: convergent mustprogress nofree nounwind willreturn memory(none)
+declare hidden i64 @_Z13get_global_idj(i32 noundef) local_unnamed_addr #0
+
+; Function Attrs: convergent nounwind
+declare hidden i32 @_Z10atomic_addPU3AS1Vjj(ptr addrspace(1) noundef, i32 noundef) local_unnamed_addr #1
+
+; Function Attrs: convergent nounwind
+declare hidden i32 @_Z10atomic_subPU3AS1Vjj(ptr addrspace(1) noundef, i32 noundef) local_unnamed_addr #1
+
+; Function Attrs: convergent mustprogress nofree nounwind willreturn memory(none)
+declare hidden i64 @_Z12get_local_idj(i32 noundef) local_unnamed_addr #0
+
+; Function Attrs: convergent nounwind
+declare hidden void @_Z7barrierj(i32 noundef) local_unnamed_addr #1
+
+; Function Attrs: convergent mustprogress nofree nounwind willreturn memory(none)
+declare hidden i32 @_Z3minjj(i32 noundef, i32 noundef) local_unnamed_addr #0
+
+; Function Attrs: convergent nounwind
+declare hidden i32 @_Z10atomic_incPU3AS3Vj(ptr addrspace(3) noundef) local_unnamed_addr #1
+
+; Function Attrs: convergent mustprogress nofree nounwind willreturn memory(none)
+declare hidden i64 @_Z14get_local_sizej(i32 noundef) local_unnamed_addr #0
+
+; Function Attrs: convergent norecurse nounwind
+define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture noundef readonly align 1 %0, ptr addrspace(1) nocapture noundef writeonly align 1 %1, ptr addrspace(1) nocapture noundef readonly align 4 %2, ptr addrspace(1) noundef align 4 %3, ptr addrspace(1) nocapture noundef readnone align 4 %4) local_unnamed_addr #2 !kernel_arg_addr_space !5 !kernel_arg_access_qual !6 !kernel_arg_type !7 !kernel_arg_base_type !7 !kernel_arg_type_qual !8 !kernel_arg_name !9 !reqd_work_group_size !10 {
+; CHECK-LABEL: kernel_round1:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_add_u32 s10, s10, s15
+; CHECK-NEXT: s_mov_b32 s32, 0
+; CHECK-NEXT: s_addc_u32 s11, s11, 0
+; CHECK-NEXT: s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s10
+; CHECK-NEXT: s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s11
+; CHECK-NEXT: s_load_dwordx8 s[44:51], s[6:7], 0x0
+; CHECK-NEXT: s_add_u32 s0, s0, s15
+; CHECK-NEXT: s_mov_b64 s[34:35], s[6:7]
+; CHECK-NEXT: s_addc_u32 s1, s1, 0
+; CHECK-NEXT: v_mov_b32_e32 v41, v0
+; CHECK-NEXT: s_add_u32 s42, s34, 40
+; CHECK-NEXT: v_mov_b32_e32 v31, v0
+; CHECK-NEXT: v_mov_b32_e32 v0, 0
+; CHECK-NEXT: s_mov_b64 s[36:37], s[8:9]
+; CHECK-NEXT: s_addc_u32 s43, s35, 0
+; CHECK-NEXT: s_mov_b64 s[10:11], s[36:37]
+; CHECK-NEXT: s_mov_b64 s[8:9], s[42:43]
+; CHECK-NEXT: s_mov_b32 s33, s14
+; CHECK-NEXT: s_mov_b32 s40, s13
+; CHECK-NEXT: s_mov_b32 s41, s12
+; CHECK-NEXT: s_mov_b64 s[38:39], s[4:5]
+; CHECK-NEXT: s_getpc_b64 s[6:7]
+; CHECK-NEXT: s_add_u32 s6, ...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/67456
More information about the llvm-commits
mailing list