|
| 1 | +//=== SYCLOptimizeBackToBackBarrier.cpp - SYCL barrier optimization pass ===// |
| 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 pass cleans up back-to-back ControlBarrier calls. |
| 10 | +// |
| 11 | +//===----------------------------------------------------------------------===// |
| 12 | + |
| 13 | +#include "llvm/SYCLLowerIR/SYCLOptimizeBackToBackBarrier.h" |
| 14 | + |
| 15 | +#include "llvm/IR/IRBuilder.h" |
| 16 | + |
| 17 | +using namespace llvm; |
| 18 | + |
| 19 | +namespace { |
| 20 | + |
| 21 | +static constexpr char CONTROL_BARRIER[] = "_Z22__spirv_ControlBarrieriii"; |
| 22 | +static constexpr char ITT_BARRIER[] = "__itt_offload_wg_barrier_wrapper"; |
| 23 | +static constexpr char ITT_RESUME[] = "__itt_offload_wi_resume_wrapper"; |
| 24 | + |
| 25 | +// Known scopes in SPIR-V. |
| 26 | +enum class Scope { |
| 27 | + CrossDevice = 0, |
| 28 | + Device = 1, |
| 29 | + Workgroup = 2, |
| 30 | + Subgroup = 3, |
| 31 | + Invocation = 4 |
| 32 | +}; |
| 33 | + |
| 34 | +enum class CompareRes { BIGGER = 0, SMALLER = 1, EQUAL = 2, UNKNOWN = 3 }; |
| 35 | + |
| 36 | +// This map is added in case of any future scopes are added to SPIR-V and/or |
| 37 | +// SYCL. |
| 38 | +const std::unordered_map<uint64_t, uint64_t> ScopeWeights = { |
| 39 | + {static_cast<uint64_t>(Scope::CrossDevice), 1000}, |
| 40 | + {static_cast<uint64_t>(Scope::Device), 800}, |
| 41 | + {static_cast<uint64_t>(Scope::Workgroup), 600}, |
| 42 | + {static_cast<uint64_t>(Scope::Subgroup), 400}, |
| 43 | + {static_cast<uint64_t>(Scope::Invocation), 10}}; |
| 44 | + |
| 45 | +inline CompareRes compareScopesWithWeights(const uint64_t LHS, |
| 46 | + const uint64_t RHS) { |
| 47 | + auto LHSIt = ScopeWeights.find(LHS); |
| 48 | + auto RHSIt = ScopeWeights.find(RHS); |
| 49 | + |
| 50 | + if (LHSIt == ScopeWeights.end() || RHSIt == ScopeWeights.end()) |
| 51 | + return CompareRes::UNKNOWN; |
| 52 | + |
| 53 | + const uint64_t LHSWeight = LHSIt->second; |
| 54 | + const uint64_t RHSWeight = RHSIt->second; |
| 55 | + |
| 56 | + if (LHSWeight > RHSWeight) |
| 57 | + return CompareRes::BIGGER; |
| 58 | + if (LHSWeight < RHSWeight) |
| 59 | + return CompareRes::SMALLER; |
| 60 | + return CompareRes::EQUAL; |
| 61 | +} |
| 62 | + |
| 63 | +// The function removes back-to-back ControlBarrier calls in case if they |
| 64 | +// have the same memory scope and memory semantics arguments. When two |
| 65 | +// back-to-back ControlBarriers are having different execution scope arguments - |
| 66 | +// pick the one with the 'bigger' scope. |
| 67 | +// It also cleans up ITT annotations surrounding the removed barrier call. |
| 68 | +bool processControlBarrier(Function *F) { |
| 69 | + BasicBlock *PrevBB = nullptr; |
| 70 | + llvm::SmallPtrSet<Instruction *, 8> ToErase; |
| 71 | + for (auto I = F->user_begin(), E = F->user_end(); I != E;) { |
| 72 | + User *U = *I++; |
| 73 | + auto *CI = dyn_cast<CallInst>(U); |
| 74 | + if (!CI) |
| 75 | + continue; |
| 76 | + |
| 77 | + // New basic block - new processing. |
| 78 | + BasicBlock *CurrentBB = CI->getParent(); |
| 79 | + if (CurrentBB != PrevBB) { |
| 80 | + PrevBB = CurrentBB; |
| 81 | + continue; |
| 82 | + } |
| 83 | + |
| 84 | + llvm::SmallPtrSet<Instruction *, 2> ToEraseLocalITT; |
| 85 | + BasicBlock::iterator It(CI); |
| 86 | + // Iterate over the basic block storing back-to-back barriers and their ITT |
| 87 | + // annotations into ToErase container. |
| 88 | + while (It != CurrentBB->begin()) { |
| 89 | + --It; |
| 90 | + auto *Cand = dyn_cast<CallInst>(&*It); |
| 91 | + if (!Cand) |
| 92 | + break; |
| 93 | + CallInst *CIToRemove = Cand; |
| 94 | + StringRef CandName = Cand->getCalledFunction()->getName(); |
| 95 | + if (CandName == ITT_RESUME || CandName == ITT_BARRIER) { |
| 96 | + ToEraseLocalITT.insert(Cand); |
| 97 | + continue; |
| 98 | + } else if (CandName == CONTROL_BARRIER) { |
| 99 | + bool EqualOps = true; |
| 100 | + const auto *ExecutionScopeCI = CI->getOperand(0); |
| 101 | + const auto *ExecutionScopeCand = Cand->getOperand(0); |
| 102 | + if (ExecutionScopeCI != ExecutionScopeCand) { |
| 103 | + if (isa<ConstantInt>(ExecutionScopeCI) && |
| 104 | + isa<ConstantInt>(ExecutionScopeCand)) { |
| 105 | + const auto ConstScopeCI = |
| 106 | + cast<ConstantInt>(ExecutionScopeCI)->getZExtValue(); |
| 107 | + const auto ConstScopeCand = |
| 108 | + cast<ConstantInt>(ExecutionScopeCand)->getZExtValue(); |
| 109 | + // Pick ControlBarrier with the 'bigger' execution scope. |
| 110 | + const auto Compare = |
| 111 | + compareScopesWithWeights(ConstScopeCI, ConstScopeCand); |
| 112 | + if (Compare == CompareRes::SMALLER) |
| 113 | + CIToRemove = CI; |
| 114 | + else if (Compare == CompareRes::UNKNOWN) |
| 115 | + // Unknown scopes = unknown rules. Keep ControlBarrier call. |
| 116 | + EqualOps = false; |
| 117 | + } else |
| 118 | + EqualOps = false; |
| 119 | + } |
| 120 | + // TODO: may be handle a case with not-matching memory scope and |
| 121 | + // memory semantic arguments in a smart way. |
| 122 | + for (unsigned I = 1; I != CI->getNumOperands(); ++I) { |
| 123 | + if (CI->getOperand(I) != Cand->getOperand(I)) { |
| 124 | + EqualOps = false; |
| 125 | + break; |
| 126 | + } |
| 127 | + } |
| 128 | + if (EqualOps) { |
| 129 | + ToErase.insert(CIToRemove); |
| 130 | + for (auto *ITT : ToEraseLocalITT) |
| 131 | + ToErase.insert(ITT); |
| 132 | + ToEraseLocalITT.clear(); |
| 133 | + } |
| 134 | + } |
| 135 | + } |
| 136 | + } |
| 137 | + |
| 138 | + if (ToErase.empty()) |
| 139 | + return false; |
| 140 | + |
| 141 | + for (auto *I : ToErase) { |
| 142 | + I->dropAllReferences(); |
| 143 | + I->eraseFromParent(); |
| 144 | + } |
| 145 | + |
| 146 | + return true; |
| 147 | +} |
| 148 | + |
| 149 | +} // namespace |
| 150 | + |
| 151 | +PreservedAnalyses |
| 152 | +SYCLOptimizeBackToBackBarrierPass::run(Module &M, ModuleAnalysisManager &MAM) { |
| 153 | + bool ModuleChanged = false; |
| 154 | + for (Function &F : M) |
| 155 | + if (F.isDeclaration()) |
| 156 | + if (F.getName() == CONTROL_BARRIER) |
| 157 | + ModuleChanged |= processControlBarrier(&F); |
| 158 | + |
| 159 | + return ModuleChanged ? PreservedAnalyses::none() : PreservedAnalyses::all(); |
| 160 | +} |
0 commit comments