[flang] [llvm] [mlir] [mlir][llvm][OpenMP] Hoist __atomic_load alloca (PR #132888)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Apr 7 04:14:33 PDT 2025
https://github.com/NimishMishra updated https://github.com/llvm/llvm-project/pull/132888
>From 629450c05c1fa1573377387d17039c09cd2d0f5f Mon Sep 17 00:00:00 2001
From: Nimish Mishra <neelam.nimish at gmail.com>
Date: Tue, 1 Apr 2025 19:30:53 +0530
Subject: [PATCH] [mlir][llvm][OpenMP] Hoist __atomic_load alloca
---
.../test/Integration/OpenMP/atomic-capture-complex.f90 | 2 +-
llvm/include/llvm/Frontend/Atomic/Atomic.h | 5 +++--
llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h | 10 ++++++----
llvm/lib/Frontend/Atomic/Atomic.cpp | 4 ++++
llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp | 6 +++---
.../Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp | 4 +++-
mlir/test/Target/LLVMIR/openmp-llvm.mlir | 9 +++++----
7 files changed, 25 insertions(+), 15 deletions(-)
diff --git a/flang/test/Integration/OpenMP/atomic-capture-complex.f90 b/flang/test/Integration/OpenMP/atomic-capture-complex.f90
index 4ffd18097d79e..69390427ff1ff 100644
--- a/flang/test/Integration/OpenMP/atomic-capture-complex.f90
+++ b/flang/test/Integration/OpenMP/atomic-capture-complex.f90
@@ -9,6 +9,7 @@
!RUN: %if x86-registered-target %{ %flang_fc1 -triple x86_64-unknown-linux-gnu -emit-llvm -fopenmp %s -o - | FileCheck --check-prefixes=CHECK,X86 %s %}
!RUN: %if aarch64-registerd-target %{ %flang_fc1 -triple aarch64-unknown-linux-gnu -emit-llvm -fopenmp %s -o - | FileCheck --check-prefixes=CHECK,AARCH64 %s %}
+!CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
!CHECK: %[[X_NEW_VAL:.*]] = alloca { float, float }, align 8
!CHECK: %[[VAL_1:.*]] = alloca { float, float }, i64 1, align 8
!CHECK: %[[ORIG_VAL:.*]] = alloca { float, float }, i64 1, align 8
@@ -16,7 +17,6 @@
!CHECK: br label %entry
!CHECK: entry:
-!CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
!CHECK: call void @__atomic_load(i64 8, ptr %[[ORIG_VAL]], ptr %[[ATOMIC_TEMP_LOAD]], i32 0)
!CHECK: %[[PHI_NODE_ENTRY_1:.*]] = load { float, float }, ptr %[[ATOMIC_TEMP_LOAD]], align 8
!CHECK: br label %.atomic.cont
diff --git a/llvm/include/llvm/Frontend/Atomic/Atomic.h b/llvm/include/llvm/Frontend/Atomic/Atomic.h
index 9f46fde6292a9..fb85a3b580b13 100644
--- a/llvm/include/llvm/Frontend/Atomic/Atomic.h
+++ b/llvm/include/llvm/Frontend/Atomic/Atomic.h
@@ -22,14 +22,15 @@ class AtomicInfo {
Align AtomicAlign;
Align ValueAlign;
bool UseLibcall;
+ IRBuilderBase::InsertPoint AllocaIP;
public:
AtomicInfo(IRBuilderBase *Builder, Type *Ty, uint64_t AtomicSizeInBits,
uint64_t ValueSizeInBits, Align AtomicAlign, Align ValueAlign,
- bool UseLibcall)
+ bool UseLibcall, IRBuilderBase::InsertPoint AllocaIP)
: Builder(Builder), Ty(Ty), AtomicSizeInBits(AtomicSizeInBits),
ValueSizeInBits(ValueSizeInBits), AtomicAlign(AtomicAlign),
- ValueAlign(ValueAlign), UseLibcall(UseLibcall) {}
+ ValueAlign(ValueAlign), UseLibcall(UseLibcall), AllocaIP(AllocaIP) {}
virtual ~AtomicInfo() = default;
diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
index ec013d1822439..8e9739cce3e2f 100644
--- a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
@@ -489,9 +489,10 @@ class OpenMPIRBuilder {
public:
AtomicInfo(IRBuilder<> *Builder, llvm::Type *Ty, uint64_t AtomicSizeInBits,
uint64_t ValueSizeInBits, llvm::Align AtomicAlign,
- llvm::Align ValueAlign, bool UseLibcall, llvm::Value *AtomicVar)
+ llvm::Align ValueAlign, bool UseLibcall,
+ IRBuilderBase::InsertPoint AllocaIP, llvm::Value *AtomicVar)
: llvm::AtomicInfo(Builder, Ty, AtomicSizeInBits, ValueSizeInBits,
- AtomicAlign, ValueAlign, UseLibcall),
+ AtomicAlign, ValueAlign, UseLibcall, AllocaIP),
AtomicVar(AtomicVar) {}
llvm::Value *getAtomicPointer() const override { return AtomicVar; }
@@ -3268,11 +3269,12 @@ class OpenMPIRBuilder {
/// value
/// \param AO Atomic ordering of the generated atomic
/// instructions.
- ///
+ /// \param AllocaIP Insert point for allocas
+ //
/// \return Insertion point after generated atomic read IR.
InsertPointTy createAtomicRead(const LocationDescription &Loc,
AtomicOpValue &X, AtomicOpValue &V,
- AtomicOrdering AO);
+ AtomicOrdering AO, InsertPointTy AllocaIP);
/// Emit atomic write for : X = Expr --- Only Scalar data types.
///
diff --git a/llvm/lib/Frontend/Atomic/Atomic.cpp b/llvm/lib/Frontend/Atomic/Atomic.cpp
index c9f9a9dcfb702..37db8fe350053 100644
--- a/llvm/lib/Frontend/Atomic/Atomic.cpp
+++ b/llvm/lib/Frontend/Atomic/Atomic.cpp
@@ -118,8 +118,12 @@ AtomicInfo::EmitAtomicLoadLibcall(AtomicOrdering AO) {
Value *PtrVal = getAtomicPointer();
PtrVal = Builder->CreateAddrSpaceCast(PtrVal, PointerType::getUnqual(Ctx));
Args.push_back(PtrVal);
+
+ auto CurrentIP = Builder->saveIP();
+ Builder->restoreIP(AllocaIP);
AllocaInst *AllocaResult =
CreateAlloca(Ty, getAtomicPointer()->getName() + "atomic.temp.load");
+ Builder->restoreIP(CurrentIP);
const Align AllocaAlignment = DL.getPrefTypeAlign(SizedIntTy);
AllocaResult->setAlignment(AllocaAlignment);
Args.push_back(AllocaResult);
diff --git a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
index 68b1fa42934ad..28e109e1a8ad3 100644
--- a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
+++ b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
@@ -8565,7 +8565,7 @@ bool OpenMPIRBuilder::checkAndEmitFlushAfterAtomic(
OpenMPIRBuilder::InsertPointTy
OpenMPIRBuilder::createAtomicRead(const LocationDescription &Loc,
AtomicOpValue &X, AtomicOpValue &V,
- AtomicOrdering AO) {
+ AtomicOrdering AO, InsertPointTy AllocaIP) {
if (!updateToLocation(Loc))
return Loc.IP;
@@ -8593,7 +8593,7 @@ OpenMPIRBuilder::createAtomicRead(const LocationDescription &Loc,
LoadDL.getTypeStoreSize(OldVal->getPointerOperand()->getType());
OpenMPIRBuilder::AtomicInfo atomicInfo(
&Builder, XElemTy, LoadSize * 8, LoadSize * 8, OldVal->getAlign(),
- OldVal->getAlign(), true /* UseLibcall */, X.Var);
+ OldVal->getAlign(), true /* UseLibcall */, AllocaIP, X.Var);
auto AtomicLoadRes = atomicInfo.EmitAtomicLoadLibcall(AO);
XRead = AtomicLoadRes.first;
OldVal->eraseFromParent();
@@ -8758,7 +8758,7 @@ Expected<std::pair<Value *, Value *>> OpenMPIRBuilder::emitAtomicUpdate(
OpenMPIRBuilder::AtomicInfo atomicInfo(
&Builder, XElemTy, LoadSize * 8, LoadSize * 8, OldVal->getAlign(),
- OldVal->getAlign(), true /* UseLibcall */, X);
+ OldVal->getAlign(), true /* UseLibcall */, AllocaIP, X);
auto AtomicLoadRes = atomicInfo.EmitAtomicLoadLibcall(AO);
BasicBlock *CurBB = Builder.GetInsertBlock();
Instruction *CurBBTI = CurBB->getTerminator();
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 4d610d6e2656d..2dd2185dae31b 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2675,6 +2675,8 @@ convertOmpAtomicRead(Operation &opInst, llvm::IRBuilderBase &builder,
return failure();
llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
+ llvm::OpenMPIRBuilder::InsertPointTy allocaIP =
+ findAllocaInsertPoint(builder, moduleTranslation);
llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
@@ -2687,7 +2689,7 @@ convertOmpAtomicRead(Operation &opInst, llvm::IRBuilderBase &builder,
llvm::OpenMPIRBuilder::AtomicOpValue V = {v, elementType, false, false};
llvm::OpenMPIRBuilder::AtomicOpValue X = {x, elementType, false, false};
- builder.restoreIP(ompBuilder->createAtomicRead(ompLoc, X, V, AO));
+ builder.restoreIP(ompBuilder->createAtomicRead(ompLoc, X, V, AO, allocaIP));
return success();
}
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index d7f4d0a65b24c..15cd429c7cf9a 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -1368,6 +1368,7 @@ llvm.func @omp_atomic_read(%arg0 : !llvm.ptr, %arg1 : !llvm.ptr) -> () {
// CHECK-LABEL: @omp_atomic_read_implicit_cast
llvm.func @omp_atomic_read_implicit_cast () {
+//CHECK: %[[ATOMIC_LOAD_TEMP:.*]] = alloca { float, float }, align 8
//CHECK: %[[Z:.*]] = alloca float, i64 1, align 4
//CHECK: %[[Y:.*]] = alloca double, i64 1, align 8
//CHECK: %[[X:.*]] = alloca [2 x { float, float }], i64 1, align 8
@@ -1392,7 +1393,7 @@ llvm.func @omp_atomic_read_implicit_cast () {
%16 = llvm.mul %10, %9 overflow<nsw> : i64
%17 = llvm.getelementptr %5[%15] : (!llvm.ptr, i64) -> !llvm.ptr, !llvm.struct<(f32, f32)>
-//CHECK: %[[ATOMIC_LOAD_TEMP:.*]] = alloca { float, float }, align 8
+
//CHECK: call void @__atomic_load(i64 8, ptr %[[X_ELEMENT]], ptr %[[ATOMIC_LOAD_TEMP]], i32 0)
//CHECK: %[[LOAD:.*]] = load { float, float }, ptr %[[ATOMIC_LOAD_TEMP]], align 8
//CHECK: %[[EXT:.*]] = extractvalue { float, float } %[[LOAD]], 0
@@ -1480,6 +1481,7 @@ llvm.func @omp_atomic_update(%x:!llvm.ptr, %expr: i32, %xbool: !llvm.ptr, %exprb
// -----
+//CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
//CHECK: %[[X_NEW_VAL:.*]] = alloca { float, float }, align 8
//CHECK: {{.*}} = alloca { float, float }, i64 1, align 8
//CHECK: %[[ORIG_VAL:.*]] = alloca { float, float }, i64 1, align 8
@@ -1487,7 +1489,6 @@ llvm.func @omp_atomic_update(%x:!llvm.ptr, %expr: i32, %xbool: !llvm.ptr, %exprb
//CHECK: br label %entry
//CHECK: entry:
-//CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
//CHECK: call void @__atomic_load(i64 8, ptr %[[ORIG_VAL]], ptr %[[ATOMIC_TEMP_LOAD]], i32 0)
//CHECK: %[[PHI_NODE_ENTRY_1:.*]] = load { float, float }, ptr %[[ATOMIC_TEMP_LOAD]], align 8
//CHECK: br label %.atomic.cont
@@ -1532,6 +1533,7 @@ llvm.func @_QPomp_atomic_update_complex() {
// -----
+//CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
//CHECK: %[[X_NEW_VAL:.*]] = alloca { float, float }, align 8
//CHECK: %[[VAL_1:.*]] = alloca { float, float }, i64 1, align 8
//CHECK: %[[ORIG_VAL:.*]] = alloca { float, float }, i64 1, align 8
@@ -1539,7 +1541,6 @@ llvm.func @_QPomp_atomic_update_complex() {
//CHECK: br label %entry
//CHECK: entry: ; preds = %0
-//CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
//CHECK: call void @__atomic_load(i64 8, ptr %[[ORIG_VAL]], ptr %[[ATOMIC_TEMP_LOAD]], i32 0)
//CHECK: %[[PHI_NODE_ENTRY_1:.*]] = load { float, float }, ptr %[[ATOMIC_TEMP_LOAD]], align 8
//CHECK: br label %.atomic.cont
@@ -1597,9 +1598,9 @@ llvm.func @_QPomp_atomic_capture_complex() {
// CHECK-LABEL: define void @omp_atomic_read_complex() {
llvm.func @omp_atomic_read_complex(){
+// CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
// CHECK: %[[a:.*]] = alloca { float, float }, i64 1, align 8
// CHECK: %[[b:.*]] = alloca { float, float }, i64 1, align 8
-// CHECK: %[[ATOMIC_TEMP_LOAD:.*]] = alloca { float, float }, align 8
// CHECK: call void @__atomic_load(i64 8, ptr %[[b]], ptr %[[ATOMIC_TEMP_LOAD]], i32 0)
// CHECK: %[[LOADED_VAL:.*]] = load { float, float }, ptr %[[ATOMIC_TEMP_LOAD]], align 8
// CHECK: store { float, float } %[[LOADED_VAL]], ptr %[[a]], align 4
More information about the llvm-commits
mailing list