Skip to content

[flang][cuda] Handle gpu.return in AbstractResult pass #119035

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Dec 10, 2024
Merged

[flang][cuda] Handle gpu.return in AbstractResult pass #119035

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
df36486
[flang][cuda] Change how abstract result pass is scheduled on func.fu…
clementval Dec 6, 2024
72a8929
[flang][cuda] Handle gpu.return in AbstractResult pass
clementval Dec 6, 2024
ddb52ea
Merge branch 'main' into cuf_abstract_result1
clementval Dec 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 9 additions & 2 deletions flang/lib/Optimizer/Passes/Pipelines.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,8 +16,14 @@ namespace fir {
void addNestedPassToAllTopLevelOperations(mlir::PassManager &pm,
PassConstructor ctor) {
addNestedPassToOps<mlir::func::FuncOp, mlir::omp::DeclareReductionOp,
mlir::omp::PrivateClauseOp, fir::GlobalOp,
mlir::gpu::GPUModuleOp>(pm, ctor);
mlir::omp::PrivateClauseOp, fir::GlobalOp>(pm, ctor);
}

template <typename NestOpTy>
void addNestedPassToNest(mlir::PassManager &pm, PassConstructor ctor) {
mlir::OpPassManager &nestPM = pm.nest<NestOpTy>();
nestPM.addNestedPass<mlir::func::FuncOp>(ctor());
nestPM.addNestedPass<mlir::gpu::GPUFuncOp>(ctor());
}

void addNestedPassToAllTopLevelOperationsConditionally(
Expand Down Expand Up @@ -266,6 +272,7 @@ void createDefaultFIRCodeGenPassPipeline(mlir::PassManager &pm,
llvm::StringRef inputFilename) {
fir::addBoxedProcedurePass(pm);
addNestedPassToAllTopLevelOperations(pm, fir::createAbstractResultOpt);
addNestedPassToNest<mlir::gpu::GPUModuleOp>(pm, fir::createAbstractResultOpt);
fir::addCodeGenRewritePass(
pm, (config.DebugInfo != llvm::codegenoptions::NoDebugInfo));
fir::addExternalNameConversionPass(pm, config.Underscoring);
Expand Down
138 changes: 78 additions & 60 deletions flang/lib/Optimizer/Transforms/AbstractResult.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,6 +234,60 @@ class SaveResultOpConversion
}
};

template <typename OpTy>
static mlir::LogicalResult
processReturnLikeOp(OpTy ret, mlir::Value newArg,
mlir::PatternRewriter &rewriter) {
auto loc = ret.getLoc();
rewriter.setInsertionPoint(ret);
mlir::Value resultValue = ret.getOperand(0);
fir::LoadOp resultLoad;
mlir::Value resultStorage;
// Identify result local storage.
if (auto load = resultValue.getDefiningOp<fir::LoadOp>()) {
resultLoad = load;
resultStorage = load.getMemref();
// The result alloca may be behind a fir.declare, if any.
if (auto declare = resultStorage.getDefiningOp<fir::DeclareOp>())
resultStorage = declare.getMemref();
}
// Replace old local storage with new storage argument, unless
// the derived type is C_PTR/C_FUN_PTR, in which case the return
// type is updated to return void* (no new argument is passed).
if (fir::isa_builtin_cptr_type(resultValue.getType())) {
auto module = ret->template getParentOfType<mlir::ModuleOp>();
FirOpBuilder builder(rewriter, module);
mlir::Value cptr = resultValue;
if (resultLoad) {
// Replace whole derived type load by component load.
cptr = resultLoad.getMemref();
rewriter.setInsertionPoint(resultLoad);
}
mlir::Value newResultValue =
fir::factory::genCPtrOrCFunptrValue(builder, loc, cptr);
newResultValue = builder.createConvert(
loc, getVoidPtrType(ret.getContext()), newResultValue);
rewriter.setInsertionPoint(ret);
rewriter.replaceOpWithNewOp<OpTy>(ret, mlir::ValueRange{newResultValue});
} else if (resultStorage) {
resultStorage.replaceAllUsesWith(newArg);
rewriter.replaceOpWithNewOp<OpTy>(ret);
} else {
// The result storage may have been optimized out by a memory to
// register pass, this is possible for fir.box results, or fir.record
// with no length parameters. Simply store the result in the result
// storage. at the return point.
rewriter.create<fir::StoreOp>(loc, resultValue, newArg);
rewriter.replaceOpWithNewOp<OpTy>(ret);
}
// Delete result old local storage if unused.
if (resultStorage)
if (auto alloc = resultStorage.getDefiningOp<fir::AllocaOp>())
if (alloc->use_empty())
rewriter.eraseOp(alloc);
return mlir::success();
}

class ReturnOpConversion : public mlir::OpRewritePattern<mlir::func::ReturnOp> {
public:
using OpRewritePattern::OpRewritePattern;
Expand All @@ -242,55 +296,23 @@ class ReturnOpConversion : public mlir::OpRewritePattern<mlir::func::ReturnOp> {
llvm::LogicalResult
matchAndRewrite(mlir::func::ReturnOp ret,
mlir::PatternRewriter &rewriter) const override {
auto loc = ret.getLoc();
rewriter.setInsertionPoint(ret);
mlir::Value resultValue = ret.getOperand(0);
fir::LoadOp resultLoad;
mlir::Value resultStorage;
// Identify result local storage.
if (auto load = resultValue.getDefiningOp<fir::LoadOp>()) {
resultLoad = load;
resultStorage = load.getMemref();
// The result alloca may be behind a fir.declare, if any.
if (auto declare = resultStorage.getDefiningOp<fir::DeclareOp>())
resultStorage = declare.getMemref();
}
// Replace old local storage with new storage argument, unless
// the derived type is C_PTR/C_FUN_PTR, in which case the return
// type is updated to return void* (no new argument is passed).
if (fir::isa_builtin_cptr_type(resultValue.getType())) {
auto module = ret->getParentOfType<mlir::ModuleOp>();
FirOpBuilder builder(rewriter, module);
mlir::Value cptr = resultValue;
if (resultLoad) {
// Replace whole derived type load by component load.
cptr = resultLoad.getMemref();
rewriter.setInsertionPoint(resultLoad);
}
mlir::Value newResultValue =
fir::factory::genCPtrOrCFunptrValue(builder, loc, cptr);
newResultValue = builder.createConvert(
loc, getVoidPtrType(ret.getContext()), newResultValue);
rewriter.setInsertionPoint(ret);
rewriter.replaceOpWithNewOp<mlir::func::ReturnOp>(
ret, mlir::ValueRange{newResultValue});
} else if (resultStorage) {
resultStorage.replaceAllUsesWith(newArg);
rewriter.replaceOpWithNewOp<mlir::func::ReturnOp>(ret);
} else {
// The result storage may have been optimized out by a memory to
// register pass, this is possible for fir.box results, or fir.record
// with no length parameters. Simply store the result in the result
// storage. at the return point.
rewriter.create<fir::StoreOp>(loc, resultValue, newArg);
rewriter.replaceOpWithNewOp<mlir::func::ReturnOp>(ret);
}
// Delete result old local storage if unused.
if (resultStorage)
if (auto alloc = resultStorage.getDefiningOp<fir::AllocaOp>())
if (alloc->use_empty())
rewriter.eraseOp(alloc);
return mlir::success();
return processReturnLikeOp(ret, newArg, rewriter);
}

private:
mlir::Value newArg;
};

class GPUReturnOpConversion
: public mlir::OpRewritePattern<mlir::gpu::ReturnOp> {
public:
using OpRewritePattern::OpRewritePattern;
GPUReturnOpConversion(mlir::MLIRContext *context, mlir::Value newArg)
: OpRewritePattern(context), newArg{newArg} {}
llvm::LogicalResult
matchAndRewrite(mlir::gpu::ReturnOp ret,
mlir::PatternRewriter &rewriter) const override {
return processReturnLikeOp(ret, newArg, rewriter);
}

private:
Expand Down Expand Up @@ -373,6 +395,9 @@ class AbstractResultOpt
patterns.insert<ReturnOpConversion>(context, newArg);
target.addDynamicallyLegalOp<mlir::func::ReturnOp>(
[](mlir::func::ReturnOp ret) { return ret.getOperands().empty(); });
patterns.insert<GPUReturnOpConversion>(context, newArg);
target.addDynamicallyLegalOp<mlir::gpu::ReturnOp>(
[](mlir::gpu::ReturnOp ret) { return ret.getOperands().empty(); });
assert(func.getFunctionType() ==
getNewFunctionType(funcTy, shouldBoxResult));
} else {
Expand Down Expand Up @@ -460,17 +485,10 @@ class AbstractResultOpt
const bool shouldBoxResult = this->passResultAsBox.getValue();

mlir::TypeSwitch<mlir::Operation *, void>(op)
.Case<mlir::func::FuncOp, fir::GlobalOp>([&](auto op) {
runOnSpecificOperation(op, shouldBoxResult, patterns, target);
})
.Case<mlir::gpu::GPUModuleOp>([&](auto op) {
auto gpuMod = mlir::dyn_cast<mlir::gpu::GPUModuleOp>(*op);
for (auto funcOp : gpuMod.template getOps<mlir::func::FuncOp>())
runOnSpecificOperation(funcOp, shouldBoxResult, patterns, target);
for (auto gpuFuncOp : gpuMod.template getOps<mlir::gpu::GPUFuncOp>())
runOnSpecificOperation(gpuFuncOp, shouldBoxResult, patterns,
target);
});
.Case<mlir::func::FuncOp, fir::GlobalOp, mlir::gpu::GPUFuncOp>(
[&](auto op) {
runOnSpecificOperation(op, shouldBoxResult, patterns, target);
});

// Convert the calls and, if needed, the ReturnOp in the function body.
target.addLegalDialect<fir::FIROpsDialect, mlir::arith::ArithDialect,
Expand Down
9 changes: 2 additions & 7 deletions flang/test/Driver/bbc-mlir-pass-pipeline.f90
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,14 +17,12 @@
! CHECK-NEXT: (S) 0 num-cse'd - Number of operations CSE'd
! CHECK-NEXT: (S) 0 num-dce'd - Number of operations DCE'd

! CHECK-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! CHECK-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! CHECK-NEXT: 'fir.global' Pipeline
! CHECK-NEXT: CharacterConversion
! CHECK-NEXT: 'func.func' Pipeline
! CHECK-NEXT: ArrayValueCopy
! CHECK-NEXT: CharacterConversion
! CHECK-NEXT: 'gpu.module' Pipeline
! CHECK-NEXT: CharacterConversion
! CHECK-NEXT: 'omp.declare_reduction' Pipeline
! CHECK-NEXT: CharacterConversion
! CHECK-NEXT: 'omp.private' Pipeline
Expand All @@ -50,16 +48,13 @@
! CHECK-NEXT: PolymorphicOpConversion
! CHECK-NEXT: AssumedRankOpConversion

! CHECK-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! CHECK-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! CHECK-NEXT: 'fir.global' Pipeline
! CHECK-NEXT: StackReclaim
! CHECK-NEXT: CFGConversion
! CHECK-NEXT: 'func.func' Pipeline
! CHECK-NEXT: StackReclaim
! CHECK-NEXT: CFGConversion
! CHECK-NEXT: 'gpu.module' Pipeline
! CHECK-NEXT: StackReclaim
! CHECK-NEXT: CFGConversion
! CHECK-NEXT: 'omp.declare_reduction' Pipeline
! CHECK-NEXT: StackReclaim
! CHECK-NEXT: CFGConversion
Expand Down
19 changes: 8 additions & 11 deletions flang/test/Driver/mlir-debug-pass-pipeline.f90
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,13 +28,11 @@
! ALL: Pass statistics report

! ALL: Fortran::lower::VerifierPass
! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: 'fir.global' Pipeline
! ALL-NEXT: InlineElementals
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: InlineElementals
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: InlineElementals
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: InlineElementals
! ALL-NEXT: 'omp.private' Pipeline
Expand All @@ -51,14 +49,12 @@
! ALL-NEXT: (S) 0 num-cse'd - Number of operations CSE'd
! ALL-NEXT: (S) 0 num-dce'd - Number of operations DCE'd

! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: 'fir.global' Pipeline
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: ArrayValueCopy
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'omp.private' Pipeline
Expand All @@ -82,16 +78,13 @@
! ALL-NEXT: PolymorphicOpConversion
! ALL-NEXT: AssumedRankOpConversion

! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: 'fir.global' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
Expand All @@ -112,7 +105,11 @@
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: Pipeline Collection : ['func.func', 'gpu.func']
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'gpu.func' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'omp.private' Pipeline
Expand Down
24 changes: 9 additions & 15 deletions flang/test/Driver/mlir-pass-pipeline.f90
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,16 +16,13 @@

! ALL: Fortran::lower::VerifierPass
! O2-NEXT: Canonicalizer
! ALL: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! ALL: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT:'fir.global' Pipeline
! O2-NEXT: SimplifyHLFIRIntrinsics
! ALL: InlineElementals
! ALL-NEXT:'func.func' Pipeline
! O2-NEXT: SimplifyHLFIRIntrinsics
! ALL: InlineElementals
! ALL-NEXT:'gpu.module' Pipeline
! O2-NEXT: SimplifyHLFIRIntrinsics
! ALL: InlineElementals
! ALL-NEXT:'omp.declare_reduction' Pipeline
! O2-NEXT: SimplifyHLFIRIntrinsics
! ALL: InlineElementals
Expand All @@ -36,13 +33,11 @@
! O2-NEXT: CSE
! O2-NEXT: (S) {{.*}} num-cse'd
! O2-NEXT: (S) {{.*}} num-dce'd
! O2-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! O2-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! O2-NEXT: 'fir.global' Pipeline
! O2-NEXT: OptimizedBufferization
! O2-NEXT: 'func.func' Pipeline
! O2-NEXT: OptimizedBufferization
! O2-NEXT: 'gpu.module' Pipeline
! O2-NEXT: OptimizedBufferization
! O2-NEXT: 'omp.declare_reduction' Pipeline
! O2-NEXT: OptimizedBufferization
! O2-NEXT: 'omp.private' Pipeline
Expand All @@ -59,14 +54,12 @@
! ALL-NEXT: (S) 0 num-cse'd - Number of operations CSE'd
! ALL-NEXT: (S) 0 num-dce'd - Number of operations DCE'd

! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: 'fir.global' Pipeline
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: ArrayValueCopy
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: CharacterConversion
! ALL-NEXT: 'omp.private' Pipeline
Expand All @@ -93,16 +86,13 @@
! ALL-NEXT: AssumedRankOpConversion
! O2-NEXT: AddAliasTags

! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'gpu.module', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: Pipeline Collection : ['fir.global', 'func.func', 'omp.declare_reduction', 'omp.private']
! ALL-NEXT: 'fir.global' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: StackReclaim
! ALL-NEXT: CFGConversion
Expand All @@ -124,7 +114,11 @@
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'gpu.module' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: Pipeline Collection : ['func.func', 'gpu.func']
! ALL-NEXT: 'func.func' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'gpu.func' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'omp.declare_reduction' Pipeline
! ALL-NEXT: AbstractResultOpt
! ALL-NEXT: 'omp.private' Pipeline
Expand Down
Loading
Loading