| 1 | //===----------------------------------------------------------------------===// |
| 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 | // Lower LDS global variables with target extension type "amdgpu.named.barrier" |
| 10 | // that require specialized address assignment. It assigns a unique |
| 11 | // barrier identifier to each named-barrier LDS variable and encodes |
| 12 | // this identifier within the !absolute_symbol metadata of that global. |
| 13 | // This encoding ensures that subsequent LDS lowering passes can process these |
| 14 | // barriers correctly without conflicts. |
| 15 | // |
| 16 | //===----------------------------------------------------------------------===// |
| 17 | |
| 18 | #include "AMDGPU.h" |
| 19 | #include "AMDGPUMemoryUtils.h" |
| 20 | #include "AMDGPUTargetMachine.h" |
| 21 | #include "llvm/ADT/DenseMap.h" |
| 22 | #include "llvm/Analysis/CallGraph.h" |
| 23 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 24 | #include "llvm/IR/Constants.h" |
| 25 | #include "llvm/IR/Instructions.h" |
| 26 | #include "llvm/IR/ReplaceConstant.h" |
| 27 | #include "llvm/InitializePasses.h" |
| 28 | #include "llvm/Pass.h" |
| 29 | |
| 30 | #include <algorithm> |
| 31 | |
| 32 | #define DEBUG_TYPE "amdgpu-lower-exec-sync" |
| 33 | |
| 34 | using namespace llvm; |
| 35 | using namespace AMDGPU; |
| 36 | |
| 37 | namespace { |
| 38 | |
| 39 | // If GV is also used directly by other kernels, create a new GV |
| 40 | // used only by this kernel and its function. |
| 41 | static GlobalVariable *uniquifyGVPerKernel(Module &M, GlobalVariable *GV, |
| 42 | Function *KF) { |
| 43 | bool NeedsReplacement = false; |
| 44 | for (Use &U : GV->uses()) { |
| 45 | if (auto *I = dyn_cast<Instruction>(Val: U.getUser())) { |
| 46 | Function *F = I->getFunction(); |
| 47 | if (isKernel(F: *F) && F != KF) { |
| 48 | NeedsReplacement = true; |
| 49 | break; |
| 50 | } |
| 51 | } |
| 52 | } |
| 53 | if (!NeedsReplacement) |
| 54 | return GV; |
| 55 | // Create a new GV used only by this kernel and its function |
| 56 | GlobalVariable *NewGV = new GlobalVariable( |
| 57 | M, GV->getValueType(), GV->isConstant(), GV->getLinkage(), |
| 58 | GV->getInitializer(), GV->getName() + "." + KF->getName(), nullptr, |
| 59 | GV->getThreadLocalMode(), GV->getType()->getAddressSpace()); |
| 60 | NewGV->copyAttributesFrom(Src: GV); |
| 61 | for (Use &U : make_early_inc_range(Range: GV->uses())) { |
| 62 | if (auto *I = dyn_cast<Instruction>(Val: U.getUser())) { |
| 63 | Function *F = I->getFunction(); |
| 64 | if (!isKernel(F: *F) || F == KF) { |
| 65 | U.getUser()->replaceUsesOfWith(From: GV, To: NewGV); |
| 66 | } |
| 67 | } |
| 68 | } |
| 69 | return NewGV; |
| 70 | } |
| 71 | |
| 72 | // Write the specified address into metadata where it can be retrieved by |
| 73 | // the assembler. Format is a half open range, [Address Address+1) |
| 74 | static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV, |
| 75 | uint32_t Address) { |
| 76 | LLVMContext &Ctx = M->getContext(); |
| 77 | auto *IntTy = M->getDataLayout().getIntPtrType(C&: Ctx, AddressSpace: AMDGPUAS::LOCAL_ADDRESS); |
| 78 | auto *MinC = ConstantAsMetadata::get(C: ConstantInt::get(Ty: IntTy, V: Address)); |
| 79 | auto *MaxC = ConstantAsMetadata::get(C: ConstantInt::get(Ty: IntTy, V: Address + 1)); |
| 80 | GV->setMetadata(KindID: LLVMContext::MD_absolute_symbol, |
| 81 | Node: MDNode::get(Context&: Ctx, MDs: {MinC, MaxC})); |
| 82 | } |
| 83 | |
| 84 | template <typename T> SmallVector<T> sortByName(SmallVector<T> &&V) { |
| 85 | sort(V, [](const auto *L, const auto *R) { |
| 86 | return L->getName() < R->getName(); |
| 87 | }); |
| 88 | return {std::move(V)}; |
| 89 | } |
| 90 | |
| 91 | // Main utility function for special LDS variables lowering. |
| 92 | static bool lowerExecSyncGlobalVariables( |
| 93 | Module &M, LDSUsesInfoTy &LDSUsesInfo, |
| 94 | VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly) { |
| 95 | bool Changed = false; |
| 96 | const DataLayout &DL = M.getDataLayout(); |
| 97 | // The 1st round: give module-absolute assignments |
| 98 | int NumAbsolutes = 0; |
| 99 | SmallVector<GlobalVariable *> OrderedGVs; |
| 100 | for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) { |
| 101 | GlobalVariable *GV = K.first; |
| 102 | if (!isNamedBarrier(GV: *GV)) |
| 103 | continue; |
| 104 | // give a module-absolute assignment if it is indirectly accessed by |
| 105 | // multiple kernels. This is not precise, but we don't want to duplicate |
| 106 | // a function when it is called by multiple kernels. |
| 107 | if (LDSToKernelsThatNeedToAccessItIndirectly[GV].size() > 1) { |
| 108 | OrderedGVs.push_back(Elt: GV); |
| 109 | } else { |
| 110 | // leave it to the 2nd round, which will give a kernel-relative |
| 111 | // assignment if it is only indirectly accessed by one kernel |
| 112 | LDSUsesInfo.direct_access[*K.second.begin()].insert(V: GV); |
| 113 | } |
| 114 | LDSToKernelsThatNeedToAccessItIndirectly.erase(Val: GV); |
| 115 | } |
| 116 | OrderedGVs = sortByName(V: std::move(OrderedGVs)); |
| 117 | for (GlobalVariable *GV : OrderedGVs) { |
| 118 | unsigned BarrierScope = AMDGPU::Barrier::BARRIER_SCOPE_WORKGROUP; |
| 119 | unsigned BarId = NumAbsolutes + 1; |
| 120 | unsigned BarCnt = GV->getGlobalSize(DL) / 16; |
| 121 | NumAbsolutes += BarCnt; |
| 122 | |
| 123 | // 4 bits for alignment, 5 bits for the barrier num, |
| 124 | // 3 bits for the barrier scope |
| 125 | unsigned Offset = 0x802000u | BarrierScope << 9 | BarId << 4; |
| 126 | recordLDSAbsoluteAddress(M: &M, GV, Address: Offset); |
| 127 | } |
| 128 | OrderedGVs.clear(); |
| 129 | |
| 130 | // The 2nd round: give a kernel-relative assignment for GV that |
| 131 | // either only indirectly accessed by single kernel or only directly |
| 132 | // accessed by multiple kernels. |
| 133 | SmallVector<Function *> OrderedKernels; |
| 134 | for (auto &K : LDSUsesInfo.direct_access) { |
| 135 | Function *F = K.first; |
| 136 | assert(isKernel(*F)); |
| 137 | OrderedKernels.push_back(Elt: F); |
| 138 | } |
| 139 | OrderedKernels = sortByName(V: std::move(OrderedKernels)); |
| 140 | |
| 141 | DenseMap<Function *, uint32_t> Kernel2BarId; |
| 142 | for (Function *F : OrderedKernels) { |
| 143 | for (GlobalVariable *GV : LDSUsesInfo.direct_access[F]) { |
| 144 | if (!isNamedBarrier(GV: *GV)) |
| 145 | continue; |
| 146 | |
| 147 | LDSUsesInfo.direct_access[F].erase(V: GV); |
| 148 | if (GV->isAbsoluteSymbolRef()) { |
| 149 | // already assigned |
| 150 | continue; |
| 151 | } |
| 152 | OrderedGVs.push_back(Elt: GV); |
| 153 | } |
| 154 | OrderedGVs = sortByName(V: std::move(OrderedGVs)); |
| 155 | for (GlobalVariable *GV : OrderedGVs) { |
| 156 | // GV could also be used directly by other kernels. If so, we need to |
| 157 | // create a new GV used only by this kernel and its function. |
| 158 | auto NewGV = uniquifyGVPerKernel(M, GV, KF: F); |
| 159 | Changed |= (NewGV != GV); |
| 160 | unsigned BarrierScope = AMDGPU::Barrier::BARRIER_SCOPE_WORKGROUP; |
| 161 | unsigned BarId = Kernel2BarId[F]; |
| 162 | BarId += NumAbsolutes + 1; |
| 163 | unsigned BarCnt = GV->getGlobalSize(DL) / 16; |
| 164 | Kernel2BarId[F] += BarCnt; |
| 165 | unsigned Offset = 0x802000u | BarrierScope << 9 | BarId << 4; |
| 166 | recordLDSAbsoluteAddress(M: &M, GV: NewGV, Address: Offset); |
| 167 | } |
| 168 | OrderedGVs.clear(); |
| 169 | } |
| 170 | // Also erase those special LDS variables from indirect_access. |
| 171 | for (auto &K : LDSUsesInfo.indirect_access) { |
| 172 | assert(isKernel(*K.first)); |
| 173 | for (GlobalVariable *GV : K.second) { |
| 174 | if (isNamedBarrier(GV: *GV)) |
| 175 | K.second.erase(V: GV); |
| 176 | } |
| 177 | } |
| 178 | return Changed; |
| 179 | } |
| 180 | |
| 181 | static bool runLowerExecSyncGlobals(Module &M) { |
| 182 | CallGraph CG = CallGraph(M); |
| 183 | bool Changed = false; |
| 184 | Changed |= eliminateConstantExprUsesOfLDSFromAllInstructions(M); |
| 185 | |
| 186 | // For each kernel, what variables does it access directly or through |
| 187 | // callees |
| 188 | LDSUsesInfoTy LDSUsesInfo = getTransitiveUsesOfLDS(CG, M); |
| 189 | |
| 190 | // For each variable accessed through callees, which kernels access it |
| 191 | VariableFunctionMap LDSToKernelsThatNeedToAccessItIndirectly; |
| 192 | for (auto &K : LDSUsesInfo.indirect_access) { |
| 193 | Function *F = K.first; |
| 194 | assert(isKernel(*F)); |
| 195 | for (GlobalVariable *GV : K.second) { |
| 196 | LDSToKernelsThatNeedToAccessItIndirectly[GV].insert(V: F); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | if (LDSUsesInfo.HasSpecialGVs) { |
| 201 | // Special LDS variables need special address assignment |
| 202 | Changed |= lowerExecSyncGlobalVariables( |
| 203 | M, LDSUsesInfo, LDSToKernelsThatNeedToAccessItIndirectly); |
| 204 | } |
| 205 | return Changed; |
| 206 | } |
| 207 | |
| 208 | class AMDGPULowerExecSyncLegacy : public ModulePass { |
| 209 | public: |
| 210 | static char ID; |
| 211 | AMDGPULowerExecSyncLegacy() : ModulePass(ID) {} |
| 212 | bool runOnModule(Module &M) override; |
| 213 | }; |
| 214 | |
| 215 | } // namespace |
| 216 | |
| 217 | char AMDGPULowerExecSyncLegacy::ID = 0; |
| 218 | char &llvm::AMDGPULowerExecSyncLegacyPassID = AMDGPULowerExecSyncLegacy::ID; |
| 219 | |
| 220 | INITIALIZE_PASS_BEGIN(AMDGPULowerExecSyncLegacy, DEBUG_TYPE, |
| 221 | "AMDGPU lowering of execution synchronization" , false, |
| 222 | false) |
| 223 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
| 224 | INITIALIZE_PASS_END(AMDGPULowerExecSyncLegacy, DEBUG_TYPE, |
| 225 | "AMDGPU lowering of execution synchronization" , false, |
| 226 | false) |
| 227 | |
| 228 | bool AMDGPULowerExecSyncLegacy::runOnModule(Module &M) { |
| 229 | return runLowerExecSyncGlobals(M); |
| 230 | } |
| 231 | |
| 232 | ModulePass *llvm::createAMDGPULowerExecSyncLegacyPass() { |
| 233 | return new AMDGPULowerExecSyncLegacy(); |
| 234 | } |
| 235 | |
| 236 | PreservedAnalyses AMDGPULowerExecSyncPass::run(Module &M, |
| 237 | ModuleAnalysisManager &AM) { |
| 238 | return runLowerExecSyncGlobals(M) ? PreservedAnalyses::none() |
| 239 | : PreservedAnalyses::all(); |
| 240 | } |
| 241 | |