diff options
Diffstat (limited to 'flang/lib/Optimizer')
5 files changed, 69 insertions, 74 deletions
diff --git a/flang/lib/Optimizer/Analysis/AliasAnalysis.cpp b/flang/lib/Optimizer/Analysis/AliasAnalysis.cpp index 73ddd1f..ef98942 100644 --- a/flang/lib/Optimizer/Analysis/AliasAnalysis.cpp +++ b/flang/lib/Optimizer/Analysis/AliasAnalysis.cpp @@ -27,6 +27,26 @@ using namespace mlir; #define DEBUG_TYPE "fir-alias-analysis" +// Inspect for value-scoped Allocate effects and determine whether +// 'candidate' is a new allocation. Returns SourceKind::Allocate if a +// MemAlloc effect is attached +static fir::AliasAnalysis::SourceKind +classifyAllocateFromEffects(mlir::Operation *op, mlir::Value candidate) { + if (!op) + return fir::AliasAnalysis::SourceKind::Unknown; + auto interface = llvm::dyn_cast<mlir::MemoryEffectOpInterface>(op); + if (!interface) + return fir::AliasAnalysis::SourceKind::Unknown; + llvm::SmallVector<mlir::MemoryEffects::EffectInstance, 4> effects; + interface.getEffects(effects); + for (mlir::MemoryEffects::EffectInstance &e : effects) { + if (mlir::isa<mlir::MemoryEffects::Allocate>(e.getEffect()) && + e.getValue() && e.getValue() == candidate) + return fir::AliasAnalysis::SourceKind::Allocate; + } + return fir::AliasAnalysis::SourceKind::Unknown; +} + //===----------------------------------------------------------------------===// // AliasAnalysis: alias //===----------------------------------------------------------------------===// @@ -535,6 +555,11 @@ AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v, mlir::Operation *instantiationPoint{nullptr}; while (defOp && !breakFromLoop) { ty = defOp->getResultTypes()[0]; + // Value-scoped allocation detection via effects. + if (classifyAllocateFromEffects(defOp, v) == SourceKind::Allocate) { + type = SourceKind::Allocate; + break; + } llvm::TypeSwitch<Operation *>(defOp) .Case<hlfir::AsExprOp>([&](auto op) { v = op.getVar(); @@ -554,11 +579,6 @@ AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v, defOp = v.getDefiningOp(); } }) - .Case<fir::AllocaOp, fir::AllocMemOp>([&](auto op) { - // Unique memory allocation. - type = SourceKind::Allocate; - breakFromLoop = true; - }) .Case<fir::ConvertOp>([&](auto op) { // Skip ConvertOp's and track further through the operand. v = op->getOperand(0); @@ -628,16 +648,23 @@ AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v, type = SourceKind::Global; } else { auto def = llvm::cast<mlir::Value>(boxSrc.origin.u); - // TODO: Add support to fir.allocmem - if (auto allocOp = def.template getDefiningOp<fir::AllocaOp>()) { - v = def; - defOp = v.getDefiningOp(); - type = SourceKind::Allocate; - } else if (isDummyArgument(def)) { - defOp = nullptr; - v = def; - } else { - type = SourceKind::Indirect; + bool classified = false; + if (auto defDefOp = def.getDefiningOp()) { + if (classifyAllocateFromEffects(defDefOp, def) == + SourceKind::Allocate) { + v = def; + defOp = defDefOp; + type = SourceKind::Allocate; + classified = true; + } + } + if (!classified) { + if (isDummyArgument(def)) { + defOp = nullptr; + v = def; + } else { + type = SourceKind::Indirect; + } } } breakFromLoop = true; diff --git a/flang/lib/Optimizer/Builder/CUFCommon.cpp b/flang/lib/Optimizer/Builder/CUFCommon.cpp index cf7588f..461deb8 100644 --- a/flang/lib/Optimizer/Builder/CUFCommon.cpp +++ b/flang/lib/Optimizer/Builder/CUFCommon.cpp @@ -9,6 +9,7 @@ #include "flang/Optimizer/Builder/CUFCommon.h" #include "flang/Optimizer/Builder/FIRBuilder.h" #include "flang/Optimizer/Dialect/CUF/CUFOps.h" +#include "flang/Optimizer/Dialect/Support/KindMapping.h" #include "flang/Optimizer/HLFIR/HLFIROps.h" #include "mlir/Dialect/Func/IR/FuncOps.h" #include "mlir/Dialect/LLVMIR/NVVMDialect.h" @@ -91,3 +92,25 @@ void cuf::genPointerSync(const mlir::Value box, fir::FirOpBuilder &builder) { } } } + +int cuf::computeElementByteSize(mlir::Location loc, mlir::Type type, + fir::KindMapping &kindMap, + bool emitErrorOnFailure) { + auto eleTy = fir::unwrapSequenceType(type); + if (auto t{mlir::dyn_cast<mlir::IntegerType>(eleTy)}) + return t.getWidth() / 8; + if (auto t{mlir::dyn_cast<mlir::FloatType>(eleTy)}) + return t.getWidth() / 8; + if (auto t{mlir::dyn_cast<fir::LogicalType>(eleTy)}) + return kindMap.getLogicalBitsize(t.getFKind()) / 8; + if (auto t{mlir::dyn_cast<mlir::ComplexType>(eleTy)}) { + int elemSize = + mlir::cast<mlir::FloatType>(t.getElementType()).getWidth() / 8; + return 2 * elemSize; + } + if (auto t{mlir::dyn_cast<fir::CharacterType>(eleTy)}) + return kindMap.getCharacterBitsize(t.getFKind()) / 8; + if (emitErrorOnFailure) + mlir::emitError(loc, "unsupported type"); + return 0; +} diff --git a/flang/lib/Optimizer/OpenMP/MapInfoFinalization.cpp b/flang/lib/Optimizer/OpenMP/MapInfoFinalization.cpp index d60da89..8382a48 100644 --- a/flang/lib/Optimizer/OpenMP/MapInfoFinalization.cpp +++ b/flang/lib/Optimizer/OpenMP/MapInfoFinalization.cpp @@ -943,38 +943,6 @@ class MapInfoFinalizationPass localBoxAllocas.clear(); deferrableDesc.clear(); - // First, walk `omp.map.info` ops to see if any of them have varPtrs - // with an underlying type of fir.char<k, ?>, i.e a character - // with dynamic length. If so, check if they need bounds added. - func->walk([&](mlir::omp::MapInfoOp op) { - if (!op.getBounds().empty()) - return; - - mlir::Value varPtr = op.getVarPtr(); - mlir::Type underlyingVarType = fir::unwrapRefType(varPtr.getType()); - - if (!fir::characterWithDynamicLen(underlyingVarType)) - return; - - fir::factory::AddrAndBoundsInfo info = - fir::factory::getDataOperandBaseAddr( - builder, varPtr, /*isOptional=*/false, varPtr.getLoc()); - - fir::ExtendedValue extendedValue = - hlfir::translateToExtendedValue(varPtr.getLoc(), builder, - hlfir::Entity{info.addr}, - /*continguousHint=*/true) - .first; - builder.setInsertionPoint(op); - llvm::SmallVector<mlir::Value> boundsOps = - fir::factory::genImplicitBoundsOps<mlir::omp::MapBoundsOp, - mlir::omp::MapBoundsType>( - builder, info, extendedValue, - /*dataExvIsAssumedSize=*/false, varPtr.getLoc()); - - op.getBoundsMutable().append(boundsOps); - }); - // Next, walk `omp.map.info` ops to see if any record members should be // implicitly mapped. func->walk([&](mlir::omp::MapInfoOp op) { diff --git a/flang/lib/Optimizer/Transforms/CUFComputeSharedMemoryOffsetsAndSize.cpp b/flang/lib/Optimizer/Transforms/CUFComputeSharedMemoryOffsetsAndSize.cpp index 09126e0..a644945 100644 --- a/flang/lib/Optimizer/Transforms/CUFComputeSharedMemoryOffsetsAndSize.cpp +++ b/flang/lib/Optimizer/Transforms/CUFComputeSharedMemoryOffsetsAndSize.cpp @@ -41,8 +41,7 @@ namespace { static bool isAssumedSize(mlir::ValueRange shape) { if (shape.size() != 1) return false; - std::optional<std::int64_t> val = fir::getIntIfConstant(shape[0]); - if (val && *val == -1) + if (llvm::isa_and_nonnull<fir::AssumedSizeExtentOp>(shape[0].getDefiningOp())) return true; return false; } diff --git a/flang/lib/Optimizer/Transforms/CUFOpConversion.cpp b/flang/lib/Optimizer/Transforms/CUFOpConversion.cpp index 8d00272..5b1b0a2 100644 --- a/flang/lib/Optimizer/Transforms/CUFOpConversion.cpp +++ b/flang/lib/Optimizer/Transforms/CUFOpConversion.cpp @@ -263,28 +263,6 @@ static bool inDeviceContext(mlir::Operation *op) { return false; } -static int computeWidth(mlir::Location loc, mlir::Type type, - fir::KindMapping &kindMap) { - auto eleTy = fir::unwrapSequenceType(type); - if (auto t{mlir::dyn_cast<mlir::IntegerType>(eleTy)}) - return t.getWidth() / 8; - if (auto t{mlir::dyn_cast<mlir::FloatType>(eleTy)}) - return t.getWidth() / 8; - if (eleTy.isInteger(1)) - return 1; - if (auto t{mlir::dyn_cast<fir::LogicalType>(eleTy)}) - return kindMap.getLogicalBitsize(t.getFKind()) / 8; - if (auto t{mlir::dyn_cast<mlir::ComplexType>(eleTy)}) { - int elemSize = - mlir::cast<mlir::FloatType>(t.getElementType()).getWidth() / 8; - return 2 * elemSize; - } - if (auto t{mlir::dyn_cast_or_null<fir::CharacterType>(eleTy)}) - return kindMap.getCharacterBitsize(t.getFKind()) / 8; - mlir::emitError(loc, "unsupported type"); - return 0; -} - struct CUFAllocOpConversion : public mlir::OpRewritePattern<cuf::AllocOp> { using OpRewritePattern::OpRewritePattern; @@ -320,7 +298,7 @@ struct CUFAllocOpConversion : public mlir::OpRewritePattern<cuf::AllocOp> { mlir::Value bytes; fir::KindMapping kindMap{fir::getKindMapping(mod)}; if (fir::isa_trivial(op.getInType())) { - int width = computeWidth(loc, op.getInType(), kindMap); + int width = cuf::computeElementByteSize(loc, op.getInType(), kindMap); bytes = builder.createIntegerConstant(loc, builder.getIndexType(), width); } else if (auto seqTy = mlir::dyn_cast_or_null<fir::SequenceType>( @@ -330,7 +308,7 @@ struct CUFAllocOpConversion : public mlir::OpRewritePattern<cuf::AllocOp> { mlir::Type structTy = typeConverter->convertType(seqTy.getEleTy()); size = dl->getTypeSizeInBits(structTy) / 8; } else { - size = computeWidth(loc, seqTy.getEleTy(), kindMap); + size = cuf::computeElementByteSize(loc, seqTy.getEleTy(), kindMap); } mlir::Value width = builder.createIntegerConstant(loc, builder.getIndexType(), size); @@ -704,7 +682,7 @@ struct CUFDataTransferOpConversion typeConverter->convertType(fir::unwrapSequenceType(dstTy)); width = dl->getTypeSizeInBits(structTy) / 8; } else { - width = computeWidth(loc, dstTy, kindMap); + width = cuf::computeElementByteSize(loc, dstTy, kindMap); } mlir::Value widthValue = mlir::arith::ConstantOp::create( rewriter, loc, i64Ty, rewriter.getIntegerAttr(i64Ty, width)); |