[llvm-branch-commits] [flang] [flang] Lower omp.workshare to other omp constructs (PR #101446)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Aug 22 02:00:02 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-flang-driver
Author: Ivan R. Ivanov (ivanradanov)
<details>
<summary>Changes</summary>
4/4
There are two points which need some discussion in this PR:
1. We need to make a value computed in a omp.single accessible in all threads of the omp.parallel region. This is achieved by allocating temporary memory outside the omp.parallel and atoring that in the omp.single and then reloading it from all threads. However, from reading the standard I dont think we are guaranteed that the workshare is nested in the omp.parallel so there could be a omp.parallel { func.call @<!-- -->contains_workshare }, then we would not be able to access the omp.parallel. So I think adding support in the runtime to be able to yield a value from a omp.single could be the fix to this.
2. For the temporary allocations above not all types are supported by fir.alloca, so I need to use llvm.alloca and unrealized_cast to be able to allocate a temporary for a fir.ref type. This too can be fixed by introducing yielding from omp.single
1/4 https://github.com/llvm/llvm-project/pull/101443
2/4 https://github.com/llvm/llvm-project/pull/101444
3/4 https://github.com/llvm/llvm-project/pull/101445
4/4 https://github.com/llvm/llvm-project/pull/101446
WIP https://github.com/llvm/llvm-project/pull/104748 adds HLFIR lowering that make use of this pipeline
---
Patch is 44.63 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/101446.diff
16 Files Affected:
- (modified) flang/include/flang/Optimizer/OpenMP/Passes.h (+5)
- (modified) flang/include/flang/Optimizer/OpenMP/Passes.td (+5)
- (modified) flang/include/flang/Tools/CLOptions.inc (+4-2)
- (modified) flang/include/flang/Tools/CrossToolHelpers.h (+1)
- (modified) flang/lib/Frontend/FrontendActions.cpp (+9-1)
- (modified) flang/lib/Optimizer/OpenMP/CMakeLists.txt (+1)
- (added) flang/lib/Optimizer/OpenMP/LowerWorkshare.cpp (+446)
- (modified) flang/test/Fir/basic-program.fir (+1)
- (added) flang/test/Transforms/OpenMP/lower-workshare.mlir (+189)
- (added) flang/test/Transforms/OpenMP/lower-workshare2.mlir (+23)
- (added) flang/test/Transforms/OpenMP/lower-workshare3.mlir (+74)
- (added) flang/test/Transforms/OpenMP/lower-workshare4.mlir (+59)
- (added) flang/test/Transforms/OpenMP/lower-workshare5.mlir (+42)
- (added) flang/test/Transforms/OpenMP/lower-workshare6.mlir (+51)
- (modified) flang/tools/bbc/bbc.cpp (+4-1)
- (modified) flang/tools/tco/tco.cpp (+1)
``````````diff
diff --git a/flang/include/flang/Optimizer/OpenMP/Passes.h b/flang/include/flang/Optimizer/OpenMP/Passes.h
index 403d79667bf448..feb395f1a12dbd 100644
--- a/flang/include/flang/Optimizer/OpenMP/Passes.h
+++ b/flang/include/flang/Optimizer/OpenMP/Passes.h
@@ -25,6 +25,11 @@ namespace flangomp {
#define GEN_PASS_REGISTRATION
#include "flang/Optimizer/OpenMP/Passes.h.inc"
+/// Impelements the logic specified in the 2.8.3 workshare Construct section of
+/// the OpenMP standard which specifies what statements or constructs shall be
+/// divided into units of work.
+bool shouldUseWorkshareLowering(mlir::Operation *op);
+
} // namespace flangomp
#endif // FORTRAN_OPTIMIZER_OPENMP_PASSES_H
diff --git a/flang/include/flang/Optimizer/OpenMP/Passes.td b/flang/include/flang/Optimizer/OpenMP/Passes.td
index 395178e26a5762..041240cad12eb3 100644
--- a/flang/include/flang/Optimizer/OpenMP/Passes.td
+++ b/flang/include/flang/Optimizer/OpenMP/Passes.td
@@ -37,4 +37,9 @@ def FunctionFiltering : Pass<"omp-function-filtering"> {
];
}
+// Needs to be scheduled on Module as we create functions in it
+def LowerWorkshare : Pass<"lower-workshare", "::mlir::ModuleOp"> {
+ let summary = "Lower workshare construct";
+}
+
#endif //FORTRAN_OPTIMIZER_OPENMP_PASSES
diff --git a/flang/include/flang/Tools/CLOptions.inc b/flang/include/flang/Tools/CLOptions.inc
index 1881e23b00045a..bb00e079008a0b 100644
--- a/flang/include/flang/Tools/CLOptions.inc
+++ b/flang/include/flang/Tools/CLOptions.inc
@@ -337,7 +337,7 @@ inline void createDefaultFIROptimizerPassPipeline(
/// \param optLevel - optimization level used for creating FIR optimization
/// passes pipeline
inline void createHLFIRToFIRPassPipeline(
- mlir::PassManager &pm, llvm::OptimizationLevel optLevel = defaultOptLevel) {
+ mlir::PassManager &pm, bool enableOpenMP, llvm::OptimizationLevel optLevel = defaultOptLevel) {
if (optLevel.isOptimizingForSpeed()) {
addCanonicalizerPassWithoutRegionSimplification(pm);
addNestedPassToAllTopLevelOperations(
@@ -354,6 +354,8 @@ inline void createHLFIRToFIRPassPipeline(
pm.addPass(hlfir::createLowerHLFIRIntrinsics());
pm.addPass(hlfir::createBufferizeHLFIR());
pm.addPass(hlfir::createConvertHLFIRtoFIR());
+ if (enableOpenMP)
+ pm.addPass(flangomp::createLowerWorkshare());
}
/// Create a pass pipeline for handling certain OpenMP transformations needed
@@ -425,7 +427,7 @@ inline void createDefaultFIRCodeGenPassPipeline(mlir::PassManager &pm,
/// passes pipeline
inline void createMLIRToLLVMPassPipeline(mlir::PassManager &pm,
MLIRToLLVMPassPipelineConfig &config, llvm::StringRef inputFilename = {}) {
- fir::createHLFIRToFIRPassPipeline(pm, config.OptLevel);
+ fir::createHLFIRToFIRPassPipeline(pm, config.EnableOpenMP, config.OptLevel);
// Add default optimizer pass pipeline.
fir::createDefaultFIROptimizerPassPipeline(pm, config);
diff --git a/flang/include/flang/Tools/CrossToolHelpers.h b/flang/include/flang/Tools/CrossToolHelpers.h
index 75fd783af237d0..0911b9bca67332 100644
--- a/flang/include/flang/Tools/CrossToolHelpers.h
+++ b/flang/include/flang/Tools/CrossToolHelpers.h
@@ -123,6 +123,7 @@ struct MLIRToLLVMPassPipelineConfig : public FlangEPCallBacks {
false; ///< Set no-signed-zeros-fp-math attribute for functions.
bool UnsafeFPMath = false; ///< Set unsafe-fp-math attribute for functions.
bool NSWOnLoopVarInc = false; ///< Add nsw flag to loop variable increments.
+ bool EnableOpenMP = false; ///< Enable OpenMP lowering.
};
struct OffloadModuleOpts {
diff --git a/flang/lib/Frontend/FrontendActions.cpp b/flang/lib/Frontend/FrontendActions.cpp
index 5c86bd947ce73f..db5c5649337528 100644
--- a/flang/lib/Frontend/FrontendActions.cpp
+++ b/flang/lib/Frontend/FrontendActions.cpp
@@ -711,7 +711,11 @@ void CodeGenAction::lowerHLFIRToFIR() {
pm.enableVerifier(/*verifyPasses=*/true);
// Create the pass pipeline
- fir::createHLFIRToFIRPassPipeline(pm, level);
+ fir::createHLFIRToFIRPassPipeline(
+ pm,
+ ci.getInvocation().getFrontendOpts().features.IsEnabled(
+ Fortran::common::LanguageFeature::OpenMP),
+ level);
(void)mlir::applyPassManagerCLOptions(pm);
if (!mlir::succeeded(pm.run(*mlirModule))) {
@@ -824,6 +828,10 @@ void CodeGenAction::generateLLVMIR() {
config.VScaleMax = vsr->second;
}
+ if (ci.getInvocation().getFrontendOpts().features.IsEnabled(
+ Fortran::common::LanguageFeature::OpenMP))
+ config.EnableOpenMP = true;
+
if (ci.getInvocation().getLoweringOpts().getNSWOnLoopVarInc())
config.NSWOnLoopVarInc = true;
diff --git a/flang/lib/Optimizer/OpenMP/CMakeLists.txt b/flang/lib/Optimizer/OpenMP/CMakeLists.txt
index 92051634f0378b..39e92d388288d4 100644
--- a/flang/lib/Optimizer/OpenMP/CMakeLists.txt
+++ b/flang/lib/Optimizer/OpenMP/CMakeLists.txt
@@ -4,6 +4,7 @@ add_flang_library(FlangOpenMPTransforms
FunctionFiltering.cpp
MapInfoFinalization.cpp
MarkDeclareTarget.cpp
+ LowerWorkshare.cpp
DEPENDS
FIRDialect
diff --git a/flang/lib/Optimizer/OpenMP/LowerWorkshare.cpp b/flang/lib/Optimizer/OpenMP/LowerWorkshare.cpp
new file mode 100644
index 00000000000000..6e5538b54ba5e0
--- /dev/null
+++ b/flang/lib/Optimizer/OpenMP/LowerWorkshare.cpp
@@ -0,0 +1,446 @@
+//===- LowerWorkshare.cpp - special cases for bufferization -------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the lowering of omp.workshare to other omp constructs.
+//
+// This pass is tasked with parallelizing the loops nested in
+// workshare.loop_wrapper while both the Fortran to mlir lowering and the hlfir
+// to fir lowering pipelines are responsible for emitting the
+// workshare.loop_wrapper ops where appropriate according to the
+// `shouldUseWorkshareLowering` function.
+//
+//===----------------------------------------------------------------------===//
+
+#include <flang/Optimizer/Builder/FIRBuilder.h>
+#include <flang/Optimizer/Dialect/FIROps.h>
+#include <flang/Optimizer/Dialect/FIRType.h>
+#include <flang/Optimizer/HLFIR/HLFIROps.h>
+#include <flang/Optimizer/OpenMP/Passes.h>
+#include <llvm/ADT/BreadthFirstIterator.h>
+#include <llvm/ADT/STLExtras.h>
+#include <llvm/ADT/SmallVectorExtras.h>
+#include <llvm/ADT/iterator_range.h>
+#include <llvm/Support/ErrorHandling.h>
+#include <mlir/Dialect/Arith/IR/Arith.h>
+#include <mlir/Dialect/LLVMIR/LLVMTypes.h>
+#include <mlir/Dialect/OpenMP/OpenMPClauseOperands.h>
+#include <mlir/Dialect/OpenMP/OpenMPDialect.h>
+#include <mlir/Dialect/SCF/IR/SCF.h>
+#include <mlir/IR/BuiltinOps.h>
+#include <mlir/IR/IRMapping.h>
+#include <mlir/IR/OpDefinition.h>
+#include <mlir/IR/PatternMatch.h>
+#include <mlir/IR/Visitors.h>
+#include <mlir/Interfaces/SideEffectInterfaces.h>
+#include <mlir/Support/LLVM.h>
+#include <mlir/Transforms/GreedyPatternRewriteDriver.h>
+
+#include <variant>
+
+namespace flangomp {
+#define GEN_PASS_DEF_LOWERWORKSHARE
+#include "flang/Optimizer/OpenMP/Passes.h.inc"
+} // namespace flangomp
+
+#define DEBUG_TYPE "lower-workshare"
+
+using namespace mlir;
+
+namespace flangomp {
+
+// Checks for nesting pattern below as we need to avoid sharing the work of
+// statements which are nested in some constructs such as omp.critical or
+// another omp.parallel.
+//
+// omp.workshare { // `wsOp`
+// ...
+// omp.T { // `parent`
+// ...
+// `op`
+//
+template <typename T>
+static bool isNestedIn(omp::WorkshareOp wsOp, Operation *op) {
+ T parent = op->getParentOfType<T>();
+ if (!parent)
+ return false;
+ return wsOp->isProperAncestor(parent);
+}
+
+bool shouldUseWorkshareLowering(Operation *op) {
+ auto parentWorkshare = op->getParentOfType<omp::WorkshareOp>();
+
+ if (!parentWorkshare)
+ return false;
+
+ if (isNestedIn<omp::CriticalOp>(parentWorkshare, op))
+ return false;
+
+ // 2.8.3 workshare Construct
+ // For a parallel construct, the construct is a unit of work with respect to
+ // the workshare construct. The statements contained in the parallel construct
+ // are executed by a new thread team.
+ if (isNestedIn<omp::ParallelOp>(parentWorkshare, op))
+ return false;
+
+ // 2.8.2 single Construct
+ // Binding The binding thread set for a single region is the current team. A
+ // single region binds to the innermost enclosing parallel region.
+ // Description Only one of the encountering threads will execute the
+ // structured block associated with the single construct.
+ if (isNestedIn<omp::SingleOp>(parentWorkshare, op))
+ return false;
+
+ return true;
+}
+
+} // namespace flangomp
+
+namespace {
+
+struct SingleRegion {
+ Block::iterator begin, end;
+};
+
+static bool mustParallelizeOp(Operation *op) {
+ return op
+ ->walk([&](Operation *nested) {
+ // We need to be careful not to pick up workshare.loop_wrapper in nested
+ // omp.parallel{omp.workshare} regions, i.e. make sure that `nested`
+ // binds to the workshare region we are currently handling.
+ //
+ // For example:
+ //
+ // omp.parallel {
+ // omp.workshare { // currently handling this
+ // omp.parallel {
+ // omp.workshare { // nested workshare
+ // omp.workshare.loop_wrapper {}
+ //
+ // Therefore, we skip if we encounter a nested omp.workshare.
+ if (isa<omp::WorkshareOp>(op))
+ return WalkResult::skip();
+ if (isa<omp::WorkshareLoopWrapperOp>(op))
+ return WalkResult::interrupt();
+ return WalkResult::advance();
+ })
+ .wasInterrupted();
+}
+
+static bool isSafeToParallelize(Operation *op) {
+ return isa<hlfir::DeclareOp>(op) || isa<fir::DeclareOp>(op) ||
+ isMemoryEffectFree(op);
+}
+
+/// Simple shallow copies suffice for our purposes in this pass, so we implement
+/// this simpler alternative to the full fledged `createCopyFunc` in the
+/// frontend
+static mlir::func::FuncOp createCopyFunc(mlir::Location loc, mlir::Type varType,
+ fir::FirOpBuilder builder) {
+ mlir::ModuleOp module = builder.getModule();
+ auto rt = cast<fir::ReferenceType>(varType);
+ mlir::Type eleTy = rt.getEleTy();
+ std::string copyFuncName =
+ fir::getTypeAsString(eleTy, builder.getKindMap(), "_workshare_copy");
+
+ if (auto decl = module.lookupSymbol<mlir::func::FuncOp>(copyFuncName))
+ return decl;
+ // create function
+ mlir::OpBuilder::InsertionGuard guard(builder);
+ mlir::OpBuilder modBuilder(module.getBodyRegion());
+ llvm::SmallVector<mlir::Type> argsTy = {varType, varType};
+ auto funcType = mlir::FunctionType::get(builder.getContext(), argsTy, {});
+ mlir::func::FuncOp funcOp =
+ modBuilder.create<mlir::func::FuncOp>(loc, copyFuncName, funcType);
+ funcOp.setVisibility(mlir::SymbolTable::Visibility::Private);
+ builder.createBlock(&funcOp.getRegion(), funcOp.getRegion().end(), argsTy,
+ {loc, loc});
+ builder.setInsertionPointToStart(&funcOp.getRegion().back());
+
+ Value loaded = builder.create<fir::LoadOp>(loc, funcOp.getArgument(0));
+ builder.create<fir::StoreOp>(loc, loaded, funcOp.getArgument(1));
+
+ builder.create<mlir::func::ReturnOp>(loc);
+ return funcOp;
+}
+
+static bool isUserOutsideSR(Operation *user, Operation *parentOp,
+ SingleRegion sr) {
+ while (user->getParentOp() != parentOp)
+ user = user->getParentOp();
+ return sr.begin->getBlock() != user->getBlock() ||
+ !(user->isBeforeInBlock(&*sr.end) && sr.begin->isBeforeInBlock(user));
+}
+
+static bool isTransitivelyUsedOutside(Value v, SingleRegion sr) {
+ Block *srBlock = sr.begin->getBlock();
+ Operation *parentOp = srBlock->getParentOp();
+
+ for (auto &use : v.getUses()) {
+ Operation *user = use.getOwner();
+ if (isUserOutsideSR(user, parentOp, sr))
+ return true;
+
+ // Results of nested users cannot be used outside of the SR
+ if (user->getBlock() != srBlock)
+ continue;
+
+ // A non-safe to parallelize operation will be handled separately
+ if (!isSafeToParallelize(user))
+ continue;
+
+ for (auto res : user->getResults())
+ if (isTransitivelyUsedOutside(res, sr))
+ return true;
+ }
+ return false;
+}
+
+/// We clone pure operations in both the parallel and single blocks. this
+/// functions cleans them up if they end up with no uses
+static void cleanupBlock(Block *block) {
+ for (Operation &op : llvm::make_early_inc_range(
+ llvm::make_range(block->rbegin(), block->rend())))
+ if (isOpTriviallyDead(&op))
+ op.erase();
+}
+
+static void parallelizeRegion(Region &sourceRegion, Region &targetRegion,
+ IRMapping &rootMapping, Location loc,
+ mlir::DominanceInfo &di) {
+ OpBuilder rootBuilder(sourceRegion.getContext());
+ ModuleOp m = sourceRegion.getParentOfType<ModuleOp>();
+ OpBuilder copyFuncBuilder(m.getBodyRegion());
+ fir::FirOpBuilder firCopyFuncBuilder(copyFuncBuilder, m);
+
+ auto mapReloadedValue =
+ [&](Value v, OpBuilder allocaBuilder, OpBuilder singleBuilder,
+ OpBuilder parallelBuilder, IRMapping singleMapping) -> Value {
+ if (auto reloaded = rootMapping.lookupOrNull(v))
+ return nullptr;
+ Type ty = v.getType();
+ Value alloc = allocaBuilder.create<fir::AllocaOp>(loc, ty);
+ singleBuilder.create<fir::StoreOp>(loc, singleMapping.lookup(v), alloc);
+ Value reloaded = parallelBuilder.create<fir::LoadOp>(loc, ty, alloc);
+ rootMapping.map(v, reloaded);
+ return alloc;
+ };
+
+ auto moveToSingle = [&](SingleRegion sr, OpBuilder allocaBuilder,
+ OpBuilder singleBuilder,
+ OpBuilder parallelBuilder) -> SmallVector<Value> {
+ IRMapping singleMapping = rootMapping;
+ SmallVector<Value> copyPrivate;
+
+ for (Operation &op : llvm::make_range(sr.begin, sr.end)) {
+ if (isSafeToParallelize(&op)) {
+ singleBuilder.clone(op, singleMapping);
+ parallelBuilder.clone(op, rootMapping);
+ } else if (auto alloca = dyn_cast<fir::AllocaOp>(&op)) {
+ auto hoisted =
+ cast<fir::AllocaOp>(allocaBuilder.clone(*alloca, singleMapping));
+ rootMapping.map(&*alloca, &*hoisted);
+ rootMapping.map(alloca.getResult(), hoisted.getResult());
+ copyPrivate.push_back(hoisted);
+ } else {
+ singleBuilder.clone(op, singleMapping);
+ // Prepare reloaded values for results of operations that cannot be
+ // safely parallelized and which are used after the region `sr`
+ for (auto res : op.getResults()) {
+ if (isTransitivelyUsedOutside(res, sr)) {
+ auto alloc = mapReloadedValue(res, allocaBuilder, singleBuilder,
+ parallelBuilder, singleMapping);
+ if (alloc)
+ copyPrivate.push_back(alloc);
+ }
+ }
+ }
+ }
+ singleBuilder.create<omp::TerminatorOp>(loc);
+ return copyPrivate;
+ };
+
+ for (Block &block : sourceRegion) {
+ Block *targetBlock = rootBuilder.createBlock(
+ &targetRegion, {}, block.getArgumentTypes(),
+ llvm::map_to_vector(block.getArguments(),
+ [](BlockArgument arg) { return arg.getLoc(); }));
+ rootMapping.map(&block, targetBlock);
+ rootMapping.map(block.getArguments(), targetBlock->getArguments());
+ }
+
+ auto handleOneBlock = [&](Block &block) {
+ Block &targetBlock = *rootMapping.lookup(&block);
+ rootBuilder.setInsertionPointToStart(&targetBlock);
+ Operation *terminator = block.getTerminator();
+ SmallVector<std::variant<SingleRegion, Operation *>> regions;
+
+ auto it = block.begin();
+ auto getOneRegion = [&]() {
+ if (&*it == terminator)
+ return false;
+ if (mustParallelizeOp(&*it)) {
+ regions.push_back(&*it);
+ it++;
+ return true;
+ }
+ SingleRegion sr;
+ sr.begin = it;
+ while (&*it != terminator && !mustParallelizeOp(&*it))
+ it++;
+ sr.end = it;
+ assert(sr.begin != sr.end);
+ regions.push_back(sr);
+ return true;
+ };
+ while (getOneRegion())
+ ;
+
+ for (auto [i, opOrSingle] : llvm::enumerate(regions)) {
+ bool isLast = i + 1 == regions.size();
+ if (std::holds_alternative<SingleRegion>(opOrSingle)) {
+ OpBuilder singleBuilder(sourceRegion.getContext());
+ Block *singleBlock = new Block();
+ singleBuilder.setInsertionPointToStart(singleBlock);
+
+ OpBuilder allocaBuilder(sourceRegion.getContext());
+ Block *allocaBlock = new Block();
+ allocaBuilder.setInsertionPointToStart(allocaBlock);
+
+ OpBuilder parallelBuilder(sourceRegion.getContext());
+ Block *parallelBlock = new Block();
+ parallelBuilder.setInsertionPointToStart(parallelBlock);
+
+ omp::SingleOperands singleOperands;
+ if (isLast)
+ singleOperands.nowait = rootBuilder.getUnitAttr();
+ singleOperands.copyprivateVars =
+ moveToSingle(std::get<SingleRegion>(opOrSingle), allocaBuilder,
+ singleBuilder, parallelBuilder);
+ cleanupBlock(singleBlock);
+ for (auto var : singleOperands.copyprivateVars) {
+ mlir::func::FuncOp funcOp =
+ createCopyFunc(loc, var.getType(), firCopyFuncBuilder);
+ singleOperands.copyprivateSyms.push_back(SymbolRefAttr::get(funcOp));
+ }
+ omp::SingleOp singleOp =
+ rootBuilder.create<omp::SingleOp>(loc, singleOperands);
+ singleOp.getRegion().push_back(singleBlock);
+ rootBuilder.getInsertionBlock()->getOperations().splice(
+ rootBuilder.getInsertionPoint(), parallelBlock->getOperations());
+ targetRegion.front().getOperations().splice(
+ singleOp->getIterator(), allocaBlock->getOperations());
+ delete allocaBlock;
+ delete parallelBlock;
+ } else {
+ auto op = std::get<Operation *>(opOrSingle);
+ if (auto wslw = dyn_cast<omp::WorkshareLoopWrapperOp>(op)) {
+ omp::WsloopOperands wsloopOperands;
+ if (isLast)
+ wsloopOperands.nowait = rootBuilder.getUnitAttr();
+ auto wsloop =
+ rootBuilder.create<mlir::omp::WsloopOp>(loc, wsloopOperands);
+ auto clonedWslw = cast<omp::WorkshareLoopWrapperOp>(
+ rootBuilder.clone(*wslw, rootMapping));
+ wsloop.getRegion().takeBody(clonedWslw.getRegion());
+ clonedWslw->erase();
+ } else {
+ assert(mustParallelizeOp(op));
+ Operation *cloned = rootBuilder.cloneWithoutRegions(*op, rootMapping);
+ for (auto [region, clonedRegion] :
+ llvm::zip(op->getRegions(), cloned->getRegions()))
+ parallelizeRegion(region, clonedRegion, rootMapping, loc, di);
+ }
+ }
+ }
+
+ rootBuilder.clone(*block.getTerminator(), rootMapping);
+ };
+
+ if (sourceRegion.hasOneBlock()) {
+ handleOneBlock(sourceRegion.front());
+ } else {
+ auto &domTree = di.getDomTree(&sourceRegion);
+ for (auto node : llvm::breadth_first(domTree.getRootNode())) {
+ handleOneBlock(*node->getBlock());
+ }
+ }
+
+ for (Block &targetBlock : targetRegion)
+ cleanupBlock(&targetBlock);
+}
+
+/// Lowers workshare to a sequence of single-thread regions and parallel loops
+///
+/// For example:
+///
+/// omp.workshare {
+/// %a = fir.allocmem
+/// omp.workshare.loop_wrapper {}
+/// fir.call Assign %b %a
+/// fir.freemem %a
+/// }
+///
+/// becomes
+///
+/// %tmp = fir.alloca
+/// omp.single copyprivate(%tmp) {
+/// %a = fir.allocmem
+/// fir.store %a %tmp
+/// }
+/// %a_reloaded = fir.load %tmp
+/// omp.workshare.loop_wrapper {}
+/// omp.single {
+/// fir.call Assign %b %a_reloaded
+/// fir.freemem %a_reloaded
+/// }
+///
+/// Note that we allocate temporary memory for values in omp.single's which need
+/// to be accessed by all threads and broadcast them using single's copyprivate
+LogicalResult lowerWorkshare(mlir::omp::WorkshareOp wsOp, DominanceInfo &di) {
+ Location loc = wsOp->getLoc();
+ IRMapping ...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/101446
More information about the llvm-branch-commits
mailing list