|
| 1 | +//===- StackToShared.cpp -------------------------------------------===// |
| 2 | +// |
| 3 | +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | +// See https://llvm.org/LICENSE.txt for license information. |
| 5 | +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | +// |
| 7 | +//===----------------------------------------------------------------------===// |
| 8 | +// |
| 9 | +// This file implements transforms to swap stack allocations on the target |
| 10 | +// device with device shared memory where applicable. |
| 11 | +// |
| 12 | +//===----------------------------------------------------------------------===// |
| 13 | + |
| 14 | +#include "flang/Optimizer/Dialect/FIROps.h" |
| 15 | +#include "flang/Optimizer/HLFIR/HLFIROps.h" |
| 16 | +#include "flang/Optimizer/OpenMP/Passes.h" |
| 17 | +#include "mlir/Dialect/Func/IR/FuncOps.h" |
| 18 | +#include "mlir/Dialect/OpenMP/OpenMPDialect.h" |
| 19 | +#include "mlir/Dialect/OpenMP/OpenMPInterfaces.h" |
| 20 | + |
| 21 | +namespace flangomp { |
| 22 | +#define GEN_PASS_DEF_STACKTOSHAREDPASS |
| 23 | +#include "flang/Optimizer/OpenMP/Passes.h.inc" |
| 24 | +} // namespace flangomp |
| 25 | + |
| 26 | +using namespace mlir; |
| 27 | + |
| 28 | +namespace { |
| 29 | +class StackToSharedPass |
| 30 | + : public flangomp::impl::StackToSharedPassBase<StackToSharedPass> { |
| 31 | +public: |
| 32 | + StackToSharedPass() = default; |
| 33 | + |
| 34 | + void runOnOperation() override { |
| 35 | + MLIRContext *context = &getContext(); |
| 36 | + OpBuilder builder(context); |
| 37 | + |
| 38 | + func::FuncOp funcOp = getOperation(); |
| 39 | + auto offloadIface = funcOp->getParentOfType<omp::OffloadModuleInterface>(); |
| 40 | + if (!offloadIface || !offloadIface.getIsTargetDevice()) |
| 41 | + return; |
| 42 | + |
| 43 | + funcOp->walk([&](fir::AllocaOp allocaOp) { |
| 44 | + if (!shouldReplaceAlloca(*allocaOp)) |
| 45 | + return; |
| 46 | + |
| 47 | + // Replace fir.alloca with omp.alloc_shared_mem. |
| 48 | + builder.setInsertionPoint(allocaOp); |
| 49 | + auto sharedAllocOp = omp::AllocSharedMemOp::create( |
| 50 | + builder, allocaOp->getLoc(), allocaOp.getResult().getType(), |
| 51 | + allocaOp.getInType(), allocaOp.getUniqNameAttr(), |
| 52 | + allocaOp.getBindcNameAttr(), allocaOp.getTypeparams(), |
| 53 | + allocaOp.getShape()); |
| 54 | + allocaOp.replaceAllUsesWith(sharedAllocOp.getOperation()); |
| 55 | + allocaOp.erase(); |
| 56 | + |
| 57 | + // Create a new omp.free_shared_mem for the allocated buffer prior to |
| 58 | + // exiting the region. |
| 59 | + Block *allocaBlock = sharedAllocOp->getBlock(); |
| 60 | + DominanceInfo domInfo; |
| 61 | + for (Block &block : sharedAllocOp->getParentRegion()->getBlocks()) { |
| 62 | + Operation *terminator = block.getTerminator(); |
| 63 | + if (!terminator->hasSuccessors() && |
| 64 | + domInfo.dominates(allocaBlock, &block)) { |
| 65 | + builder.setInsertionPoint(terminator); |
| 66 | + omp::FreeSharedMemOp::create(builder, sharedAllocOp.getLoc(), |
| 67 | + sharedAllocOp); |
| 68 | + } |
| 69 | + } |
| 70 | + }); |
| 71 | + } |
| 72 | + |
| 73 | +private: |
| 74 | + // TODO: Refactor the logic in `shouldReplaceAlloca` and `checkAllocaUses` to |
| 75 | + // be reusable by the MLIR to LLVM IR translation stage, as something very |
| 76 | + // similar is also implemented there to choose between allocas and device |
| 77 | + // shared memory allocations when processing OpenMP reductions, mapping and |
| 78 | + // privatization. |
| 79 | + |
| 80 | + // Decide whether to replace a fir.alloca with a pair of device shared memory |
| 81 | + // allocation/deallocation pair based on the location of the allocation and |
| 82 | + // its uses. |
| 83 | + // |
| 84 | + // In summary, it should be done whenever the allocation is placed outside any |
| 85 | + // parallel regions and inside either a target device function or a generic |
| 86 | + // kernel, while being used inside of a parallel region. |
| 87 | + bool shouldReplaceAlloca(Operation &op) { |
| 88 | + auto targetOp = op.getParentOfType<omp::TargetOp>(); |
| 89 | + |
| 90 | + // It must be inside of a generic omp.target or in a target device function, |
| 91 | + // and not inside of omp.parallel. |
| 92 | + if (auto parallelOp = op.getParentOfType<omp::ParallelOp>()) { |
| 93 | + if (!targetOp || !targetOp->isProperAncestor(parallelOp)) |
| 94 | + return false; |
| 95 | + } |
| 96 | + |
| 97 | + if (targetOp) { |
| 98 | + if (targetOp.getKernelExecFlags(targetOp.getInnermostCapturedOmpOp()) != |
| 99 | + mlir::omp::TargetExecMode::generic) |
| 100 | + return false; |
| 101 | + } else { |
| 102 | + auto declTargetIface = dyn_cast<mlir::omp::DeclareTargetInterface>( |
| 103 | + *op.getParentOfType<func::FuncOp>()); |
| 104 | + if (!declTargetIface || !declTargetIface.isDeclareTarget() || |
| 105 | + declTargetIface.getDeclareTargetDeviceType() == |
| 106 | + mlir::omp::DeclareTargetDeviceType::host) |
| 107 | + return false; |
| 108 | + } |
| 109 | + |
| 110 | + return checkAllocaUses(op.getUses()); |
| 111 | + } |
| 112 | + |
| 113 | + // When a use takes place inside an omp.parallel region and it's not as a |
| 114 | + // private clause argument, or when it is a reduction argument passed to |
| 115 | + // omp.parallel, then the defining allocation is eligible for replacement with |
| 116 | + // shared memory. |
| 117 | + // |
| 118 | + // Only one of the uses needs to meet these conditions to return true. |
| 119 | + bool checkAllocaUses(const Operation::use_range &uses) { |
| 120 | + auto checkUse = [&](const OpOperand &use) { |
| 121 | + Operation *owner = use.getOwner(); |
| 122 | + auto moduleOp = owner->getParentOfType<ModuleOp>(); |
| 123 | + if (auto parallelOp = dyn_cast<omp::ParallelOp>(owner)) { |
| 124 | + if (llvm::is_contained(parallelOp.getReductionVars(), use.get())) |
| 125 | + return true; |
| 126 | + } else if (owner->getParentOfType<omp::ParallelOp>()) { |
| 127 | + // If it is used directly inside of a parallel region, it has to be |
| 128 | + // replaced unless the use is a private clause. |
| 129 | + if (auto argIface = dyn_cast<omp::BlockArgOpenMPOpInterface>(owner)) { |
| 130 | + if ( auto privateSyms = llvm::cast_or_null<ArrayAttr>( |
| 131 | + owner->getAttr("private_syms"))) { |
| 132 | + for (auto [var, sym] : |
| 133 | + llvm::zip_equal(argIface.getPrivateVars(), privateSyms)) { |
| 134 | + if (var != use.get()) |
| 135 | + continue; |
| 136 | + |
| 137 | + auto privateOp = cast<omp::PrivateClauseOp>( |
| 138 | + moduleOp.lookupSymbol(cast<SymbolRefAttr>(sym))); |
| 139 | + return privateOp.getDataSharingType() != |
| 140 | + omp::DataSharingClauseType::Private; |
| 141 | + } |
| 142 | + } |
| 143 | + } |
| 144 | + return true; |
| 145 | + } |
| 146 | + return false; |
| 147 | + }; |
| 148 | + |
| 149 | + // Check direct uses and also follow hlfir.declare uses. |
| 150 | + for (const OpOperand &use : uses) { |
| 151 | + if (auto declareOp = dyn_cast<hlfir::DeclareOp>(use.getOwner())) { |
| 152 | + if (checkAllocaUses(declareOp->getUses())) |
| 153 | + return true; |
| 154 | + } else if (checkUse(use)) { |
| 155 | + return true; |
| 156 | + } |
| 157 | + } |
| 158 | + |
| 159 | + return false; |
| 160 | + } |
| 161 | +}; |
| 162 | +} // namespace |
0 commit comments