diff options
Diffstat (limited to 'flang/lib/Lower/Support')
| -rw-r--r-- | flang/lib/Lower/Support/PrivateReductionUtils.cpp | 35 |
1 files changed, 22 insertions, 13 deletions
diff --git a/flang/lib/Lower/Support/PrivateReductionUtils.cpp b/flang/lib/Lower/Support/PrivateReductionUtils.cpp index d433ce3..c6c4288 100644 --- a/flang/lib/Lower/Support/PrivateReductionUtils.cpp +++ b/flang/lib/Lower/Support/PrivateReductionUtils.cpp @@ -376,6 +376,8 @@ private: loadedMoldArg = builder.loadIfRef(loc, moldArg); return loadedMoldArg; } + + bool shouldAllocateTempOnStack() const; }; } // namespace @@ -438,8 +440,14 @@ void PopulateInitAndCleanupRegionsHelper::initAndCleanupBoxedScalar( builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front()); } - mlir::Value valAlloc = builder.createHeapTemporary(loc, innerTy, /*name=*/{}, - /*shape=*/{}, lenParams); + bool shouldAllocateOnStack = shouldAllocateTempOnStack(); + mlir::Value valAlloc = + (shouldAllocateOnStack) + ? builder.createTemporary(loc, innerTy, /*name=*/{}, + /*shape=*/{}, lenParams) + : builder.createHeapTemporary(loc, innerTy, /*name=*/{}, + /*shape=*/{}, lenParams); + if (scalarInitValue) builder.createStoreWithConvert(loc, scalarInitValue, valAlloc); mlir::Value box = fir::EmboxOp::create(builder, loc, valType, valAlloc, @@ -451,8 +459,9 @@ void PopulateInitAndCleanupRegionsHelper::initAndCleanupBoxedScalar( fir::StoreOp lastOp = fir::StoreOp::create(builder, loc, box, allocatedPrivVarArg); - createCleanupRegion(converter, loc, argType, cleanupRegion, sym, - isDoConcurrent); + if (!shouldAllocateOnStack) + createCleanupRegion(converter, loc, argType, cleanupRegion, sym, + isDoConcurrent); if (ifUnallocated) builder.setInsertionPointAfter(ifUnallocated); @@ -462,6 +471,14 @@ void PopulateInitAndCleanupRegionsHelper::initAndCleanupBoxedScalar( createYield(allocatedPrivVarArg); } +bool PopulateInitAndCleanupRegionsHelper::shouldAllocateTempOnStack() const { + // On the GPU, always allocate on the stack since heap allocatins are very + // expensive. + auto offloadMod = + llvm::dyn_cast<mlir::omp::OffloadModuleInterface>(*builder.getModule()); + return offloadMod && offloadMod.getIsGPU(); +} + void PopulateInitAndCleanupRegionsHelper::initAndCleanupBoxedArray( fir::BaseBoxType boxTy, bool needsInitialization) { bool isAllocatableOrPointer = @@ -504,15 +521,7 @@ void PopulateInitAndCleanupRegionsHelper::initAndCleanupBoxedArray( // Allocating on the heap in case the whole reduction/privatization is nested // inside of a loop auto temp = [&]() { - bool shouldAllocateOnStack = false; - - // On the GPU, always allocate on the stack since heap allocatins are very - // expensive. - if (auto offloadMod = llvm::dyn_cast<mlir::omp::OffloadModuleInterface>( - *builder.getModule())) - shouldAllocateOnStack = offloadMod.getIsGPU(); - - if (shouldAllocateOnStack) + if (shouldAllocateTempOnStack()) return createStackTempFromMold(loc, builder, source); auto [temp, needsDealloc] = createTempFromMold(loc, builder, source); |