| 1 | //===-- AMDGPUMemoryUtils.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 | #include "AMDGPUMemoryUtils.h" |
| 10 | #include "AMDGPU.h" |
| 11 | #include "Utils/AMDGPUBaseInfo.h" |
| 12 | #include "llvm/ADT/SetOperations.h" |
| 13 | #include "llvm/Analysis/AliasAnalysis.h" |
| 14 | #include "llvm/Analysis/CallGraph.h" |
| 15 | #include "llvm/Analysis/MemorySSA.h" |
| 16 | #include "llvm/IR/DataLayout.h" |
| 17 | #include "llvm/IR/Instructions.h" |
| 18 | #include "llvm/IR/IntrinsicInst.h" |
| 19 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 20 | #include "llvm/IR/ReplaceConstant.h" |
| 21 | |
| 22 | #define DEBUG_TYPE "amdgpu-memory-utils" |
| 23 | |
| 24 | using namespace llvm; |
| 25 | |
| 26 | namespace llvm::AMDGPU { |
| 27 | |
| 28 | Align getAlign(const DataLayout &DL, const GlobalVariable *GV) { |
| 29 | return DL.getValueOrABITypeAlignment(Alignment: GV->getPointerAlignment(DL), |
| 30 | Ty: GV->getValueType()); |
| 31 | } |
| 32 | |
| 33 | // Returns the target extension type of a global variable, |
| 34 | // which can only be a TargetExtType, an array or single-element struct of it, |
| 35 | // or their nesting combination. |
| 36 | // TODO: allow struct of multiple TargetExtType elements of the same type. |
| 37 | // TODO: Disallow other uses of target("amdgcn.named.barrier") including: |
| 38 | // - Structs containing barriers in different scope/rank |
| 39 | // - Structs containing a mixture of barriers and other data. |
| 40 | // - Globals in other address spaces. |
| 41 | // - Allocas. |
| 42 | static TargetExtType *getTargetExtType(const GlobalVariable &GV) { |
| 43 | Type *Ty = GV.getValueType(); |
| 44 | while (true) { |
| 45 | if (auto *TTy = dyn_cast<TargetExtType>(Val: Ty)) |
| 46 | return TTy; |
| 47 | if (auto *STy = dyn_cast<StructType>(Val: Ty)) { |
| 48 | if (STy->getNumElements() != 1) |
| 49 | return nullptr; |
| 50 | Ty = STy->getElementType(N: 0); |
| 51 | continue; |
| 52 | } |
| 53 | if (auto *ATy = dyn_cast<ArrayType>(Val: Ty)) { |
| 54 | Ty = ATy->getElementType(); |
| 55 | continue; |
| 56 | } |
| 57 | return nullptr; |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | TargetExtType *isNamedBarrier(const GlobalVariable &GV) { |
| 62 | if (TargetExtType *Ty = getTargetExtType(GV)) |
| 63 | return Ty->getName() == "amdgcn.named.barrier" ? Ty : nullptr; |
| 64 | return nullptr; |
| 65 | } |
| 66 | |
| 67 | bool isDynamicLDS(const GlobalVariable &GV) { |
| 68 | // external zero size addrspace(3) without initializer is dynlds. |
| 69 | const Module *M = GV.getParent(); |
| 70 | const DataLayout &DL = M->getDataLayout(); |
| 71 | if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) |
| 72 | return false; |
| 73 | return GV.getGlobalSize(DL) == 0; |
| 74 | } |
| 75 | |
| 76 | bool isLDSVariableToLower(const GlobalVariable &GV) { |
| 77 | if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) { |
| 78 | return false; |
| 79 | } |
| 80 | if (isDynamicLDS(GV)) { |
| 81 | return true; |
| 82 | } |
| 83 | if (GV.isConstant()) { |
| 84 | // A constant undef variable can't be written to, and any load is |
| 85 | // undef, so it should be eliminated by the optimizer. It could be |
| 86 | // dropped by the back end if not. This pass skips over it. |
| 87 | return false; |
| 88 | } |
| 89 | if (GV.hasInitializer() && !isa<UndefValue>(Val: GV.getInitializer())) { |
| 90 | // Initializers are unimplemented for LDS address space. |
| 91 | // Leave such variables in place for consistent error reporting. |
| 92 | return false; |
| 93 | } |
| 94 | return true; |
| 95 | } |
| 96 | |
| 97 | bool eliminateConstantExprUsesOfLDSFromAllInstructions(Module &M) { |
| 98 | // Constants are uniqued within LLVM. A ConstantExpr referring to a LDS |
| 99 | // global may have uses from multiple different functions as a result. |
| 100 | // This pass specialises LDS variables with respect to the kernel that |
| 101 | // allocates them. |
| 102 | |
| 103 | // This is semantically equivalent to (the unimplemented as slow): |
| 104 | // for (auto &F : M.functions()) |
| 105 | // for (auto &BB : F) |
| 106 | // for (auto &I : BB) |
| 107 | // for (Use &Op : I.operands()) |
| 108 | // if (constantExprUsesLDS(Op)) |
| 109 | // replaceConstantExprInFunction(I, Op); |
| 110 | |
| 111 | SmallVector<Constant *> LDSGlobals; |
| 112 | for (auto &GV : M.globals()) |
| 113 | if (AMDGPU::isLDSVariableToLower(GV)) |
| 114 | LDSGlobals.push_back(Elt: &GV); |
| 115 | return convertUsersOfConstantsToInstructions(Consts: LDSGlobals); |
| 116 | } |
| 117 | |
| 118 | void getUsesOfLDSByFunction(const CallGraph &CG, Module &M, |
| 119 | FunctionVariableMap &kernels, |
| 120 | FunctionVariableMap &Functions) { |
| 121 | // Get uses from the current function, excluding uses by called Functions |
| 122 | // Two output variables to avoid walking the globals list twice |
| 123 | for (auto &GV : M.globals()) { |
| 124 | if (!AMDGPU::isLDSVariableToLower(GV)) |
| 125 | continue; |
| 126 | for (User *V : GV.users()) { |
| 127 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 128 | Function *F = I->getFunction(); |
| 129 | if (isKernel(F: *F)) |
| 130 | kernels[F].insert(V: &GV); |
| 131 | else |
| 132 | Functions[F].insert(V: &GV); |
| 133 | } |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | LDSUsesInfoTy getTransitiveUsesOfLDS(const CallGraph &CG, Module &M) { |
| 139 | |
| 140 | FunctionVariableMap DirectMapKernel; |
| 141 | FunctionVariableMap DirectMapFunction; |
| 142 | getUsesOfLDSByFunction(CG, M, kernels&: DirectMapKernel, Functions&: DirectMapFunction); |
| 143 | |
| 144 | // Collect functions whose address has escaped |
| 145 | DenseSet<Function *> AddressTakenFuncs; |
| 146 | for (Function &F : M.functions()) { |
| 147 | if (!isKernel(F)) |
| 148 | if (F.hasAddressTaken(nullptr, |
| 149 | /* IgnoreCallbackUses */ false, |
| 150 | /* IgnoreAssumeLikeCalls */ false, |
| 151 | /* IgnoreLLVMUsed */ IngoreLLVMUsed: true, |
| 152 | /* IgnoreArcAttachedCall */ IgnoreARCAttachedCall: false)) { |
| 153 | AddressTakenFuncs.insert(V: &F); |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | // Collect variables that are used by functions whose address has escaped |
| 158 | DenseSet<GlobalVariable *> VariablesReachableThroughFunctionPointer; |
| 159 | for (Function *F : AddressTakenFuncs) { |
| 160 | set_union(S1&: VariablesReachableThroughFunctionPointer, S2: DirectMapFunction[F]); |
| 161 | } |
| 162 | |
| 163 | auto FunctionMakesUnknownCall = [&](const Function *F) -> bool { |
| 164 | assert(!F->isDeclaration()); |
| 165 | for (const CallGraphNode::CallRecord &R : *CG[F]) { |
| 166 | if (!R.second->getFunction()) |
| 167 | return true; |
| 168 | } |
| 169 | return false; |
| 170 | }; |
| 171 | |
| 172 | // Work out which variables are reachable through function calls |
| 173 | FunctionVariableMap TransitiveMapFunction = DirectMapFunction; |
| 174 | |
| 175 | // If the function makes any unknown call, assume the worst case that it can |
| 176 | // access all variables accessed by functions whose address escaped |
| 177 | for (Function &F : M.functions()) { |
| 178 | if (!F.isDeclaration() && FunctionMakesUnknownCall(&F)) { |
| 179 | if (!isKernel(F)) { |
| 180 | set_union(S1&: TransitiveMapFunction[&F], |
| 181 | S2: VariablesReachableThroughFunctionPointer); |
| 182 | } |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | // Direct implementation of collecting all variables reachable from each |
| 187 | // function |
| 188 | for (Function &Func : M.functions()) { |
| 189 | if (Func.isDeclaration() || isKernel(F: Func)) |
| 190 | continue; |
| 191 | |
| 192 | DenseSet<Function *> seen; // catches cycles |
| 193 | SmallVector<Function *, 4> wip = {&Func}; |
| 194 | |
| 195 | while (!wip.empty()) { |
| 196 | Function *F = wip.pop_back_val(); |
| 197 | |
| 198 | // Can accelerate this by referring to transitive map for functions that |
| 199 | // have already been computed, with more care than this |
| 200 | set_union(S1&: TransitiveMapFunction[&Func], S2: DirectMapFunction[F]); |
| 201 | |
| 202 | for (const CallGraphNode::CallRecord &R : *CG[F]) { |
| 203 | Function *Ith = R.second->getFunction(); |
| 204 | if (Ith) { |
| 205 | if (!seen.contains(V: Ith)) { |
| 206 | seen.insert(V: Ith); |
| 207 | wip.push_back(Elt: Ith); |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | // Collect variables that are transitively used by functions whose address has |
| 215 | // escaped |
| 216 | for (Function *F : AddressTakenFuncs) { |
| 217 | set_union(S1&: VariablesReachableThroughFunctionPointer, |
| 218 | S2: TransitiveMapFunction[F]); |
| 219 | } |
| 220 | |
| 221 | // DirectMapKernel lists which variables are used by the kernel |
| 222 | // find the variables which are used through a function call |
| 223 | FunctionVariableMap IndirectMapKernel; |
| 224 | |
| 225 | for (Function &Func : M.functions()) { |
| 226 | if (Func.isDeclaration() || !isKernel(F: Func)) |
| 227 | continue; |
| 228 | |
| 229 | for (const CallGraphNode::CallRecord &R : *CG[&Func]) { |
| 230 | Function *Ith = R.second->getFunction(); |
| 231 | if (Ith) { |
| 232 | set_union(S1&: IndirectMapKernel[&Func], S2: TransitiveMapFunction[Ith]); |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | // Check if the kernel encounters unknows calls, wheher directly or |
| 237 | // indirectly. |
| 238 | bool SeesUnknownCalls = [&]() { |
| 239 | SmallVector<Function *> WorkList = {CG[&Func]->getFunction()}; |
| 240 | SmallPtrSet<Function *, 8> Visited; |
| 241 | |
| 242 | while (!WorkList.empty()) { |
| 243 | Function *F = WorkList.pop_back_val(); |
| 244 | |
| 245 | for (const CallGraphNode::CallRecord &CallRecord : *CG[F]) { |
| 246 | if (!CallRecord.second) |
| 247 | continue; |
| 248 | |
| 249 | Function *Callee = CallRecord.second->getFunction(); |
| 250 | if (!Callee) |
| 251 | return true; |
| 252 | |
| 253 | if (Visited.insert(Ptr: Callee).second) |
| 254 | WorkList.push_back(Elt: Callee); |
| 255 | } |
| 256 | } |
| 257 | return false; |
| 258 | }(); |
| 259 | |
| 260 | if (SeesUnknownCalls) { |
| 261 | set_union(S1&: IndirectMapKernel[&Func], |
| 262 | S2: VariablesReachableThroughFunctionPointer); |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | // Verify that we fall into one of 2 cases: |
| 267 | // - All variables are either absolute |
| 268 | // or direct mapped dynamic LDS that is not lowered. |
| 269 | // this is a re-run of the pass |
| 270 | // so we don't have anything to do. |
| 271 | // - No variables are absolute. |
| 272 | // Named-barriers which are absolute symbols are removed |
| 273 | // from the maps. |
| 274 | std::optional<bool> HasAbsoluteGVs; |
| 275 | bool HasSpecialGVs = false; |
| 276 | for (auto &Map : {DirectMapKernel, IndirectMapKernel}) { |
| 277 | for (auto &[Fn, GVs] : Map) { |
| 278 | for (auto *GV : GVs) { |
| 279 | bool IsAbsolute = GV->isAbsoluteSymbolRef(); |
| 280 | bool IsDirectMapDynLDSGV = |
| 281 | AMDGPU::isDynamicLDS(GV: *GV) && DirectMapKernel.contains(Val: Fn); |
| 282 | if (IsDirectMapDynLDSGV) |
| 283 | continue; |
| 284 | if (isNamedBarrier(GV: *GV)) { |
| 285 | if (IsAbsolute) { |
| 286 | DirectMapKernel[Fn].erase(V: GV); |
| 287 | IndirectMapKernel[Fn].erase(V: GV); |
| 288 | } |
| 289 | HasSpecialGVs = true; |
| 290 | continue; |
| 291 | } |
| 292 | if (HasAbsoluteGVs.has_value()) { |
| 293 | if (*HasAbsoluteGVs != IsAbsolute) { |
| 294 | reportFatalUsageError( |
| 295 | reason: "module cannot mix absolute and non-absolute LDS GVs" ); |
| 296 | } |
| 297 | } else |
| 298 | HasAbsoluteGVs = IsAbsolute; |
| 299 | } |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | // If we only had absolute GVs, we have nothing to do, return an empty |
| 304 | // result. |
| 305 | if (HasAbsoluteGVs && *HasAbsoluteGVs) |
| 306 | return {.direct_access: FunctionVariableMap(), .indirect_access: FunctionVariableMap(), .HasSpecialGVs: false}; |
| 307 | |
| 308 | return {.direct_access: std::move(DirectMapKernel), .indirect_access: std::move(IndirectMapKernel), |
| 309 | .HasSpecialGVs: HasSpecialGVs}; |
| 310 | } |
| 311 | |
| 312 | void removeFnAttrFromReachable(CallGraph &CG, Function *KernelRoot, |
| 313 | ArrayRef<StringRef> FnAttrs) { |
| 314 | for (StringRef Attr : FnAttrs) |
| 315 | KernelRoot->removeFnAttr(Kind: Attr); |
| 316 | |
| 317 | SmallVector<Function *> WorkList = {CG[KernelRoot]->getFunction()}; |
| 318 | SmallPtrSet<Function *, 8> Visited; |
| 319 | bool SeenUnknownCall = false; |
| 320 | |
| 321 | while (!WorkList.empty()) { |
| 322 | Function *F = WorkList.pop_back_val(); |
| 323 | |
| 324 | for (auto &CallRecord : *CG[F]) { |
| 325 | if (!CallRecord.second) |
| 326 | continue; |
| 327 | |
| 328 | Function *Callee = CallRecord.second->getFunction(); |
| 329 | if (!Callee) { |
| 330 | if (!SeenUnknownCall) { |
| 331 | SeenUnknownCall = true; |
| 332 | |
| 333 | // If we see any indirect calls, assume nothing about potential |
| 334 | // targets. |
| 335 | // TODO: This could be refined to possible LDS global users. |
| 336 | for (auto &ExternalCallRecord : *CG.getExternalCallingNode()) { |
| 337 | Function *PotentialCallee = |
| 338 | ExternalCallRecord.second->getFunction(); |
| 339 | assert(PotentialCallee); |
| 340 | if (!isKernel(F: *PotentialCallee)) { |
| 341 | for (StringRef Attr : FnAttrs) |
| 342 | PotentialCallee->removeFnAttr(Kind: Attr); |
| 343 | } |
| 344 | } |
| 345 | } |
| 346 | } else { |
| 347 | for (StringRef Attr : FnAttrs) |
| 348 | Callee->removeFnAttr(Kind: Attr); |
| 349 | if (Visited.insert(Ptr: Callee).second) |
| 350 | WorkList.push_back(Elt: Callee); |
| 351 | } |
| 352 | } |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | bool isReallyAClobber(const Value *Ptr, MemoryDef *Def, AAResults *AA) { |
| 357 | Instruction *DefInst = Def->getMemoryInst(); |
| 358 | |
| 359 | if (isa<FenceInst>(Val: DefInst)) |
| 360 | return false; |
| 361 | |
| 362 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: DefInst)) { |
| 363 | switch (II->getIntrinsicID()) { |
| 364 | case Intrinsic::amdgcn_s_barrier: |
| 365 | case Intrinsic::amdgcn_s_cluster_barrier: |
| 366 | case Intrinsic::amdgcn_s_barrier_signal: |
| 367 | case Intrinsic::amdgcn_s_barrier_signal_var: |
| 368 | case Intrinsic::amdgcn_s_barrier_signal_isfirst: |
| 369 | case Intrinsic::amdgcn_s_barrier_init: |
| 370 | case Intrinsic::amdgcn_s_barrier_join: |
| 371 | case Intrinsic::amdgcn_s_barrier_wait: |
| 372 | case Intrinsic::amdgcn_s_barrier_leave: |
| 373 | case Intrinsic::amdgcn_s_get_barrier_state: |
| 374 | case Intrinsic::amdgcn_s_wakeup_barrier: |
| 375 | case Intrinsic::amdgcn_wave_barrier: |
| 376 | case Intrinsic::amdgcn_sched_barrier: |
| 377 | case Intrinsic::amdgcn_sched_group_barrier: |
| 378 | case Intrinsic::amdgcn_iglp_opt: |
| 379 | return false; |
| 380 | default: |
| 381 | break; |
| 382 | } |
| 383 | } |
| 384 | |
| 385 | // Ignore atomics not aliasing with the original load, any atomic is a |
| 386 | // universal MemoryDef from MSSA's point of view too, just like a fence. |
| 387 | const auto checkNoAlias = [AA, Ptr](auto I) -> bool { |
| 388 | return I && AA->isNoAlias(I->getPointerOperand(), Ptr); |
| 389 | }; |
| 390 | |
| 391 | if (checkNoAlias(dyn_cast<AtomicCmpXchgInst>(Val: DefInst)) || |
| 392 | checkNoAlias(dyn_cast<AtomicRMWInst>(Val: DefInst))) |
| 393 | return false; |
| 394 | |
| 395 | return true; |
| 396 | } |
| 397 | |
| 398 | bool isClobberedInFunction(const LoadInst *Load, MemorySSA *MSSA, |
| 399 | AAResults *AA) { |
| 400 | MemorySSAWalker *Walker = MSSA->getWalker(); |
| 401 | SmallVector<MemoryAccess *> WorkList{Walker->getClobberingMemoryAccess(I: Load)}; |
| 402 | SmallPtrSet<MemoryAccess *, 8> Visited; |
| 403 | MemoryLocation Loc(MemoryLocation::get(LI: Load)); |
| 404 | |
| 405 | LLVM_DEBUG(dbgs() << "Checking clobbering of: " << *Load << '\n'); |
| 406 | |
| 407 | // Start with a nearest dominating clobbering access, it will be either |
| 408 | // live on entry (nothing to do, load is not clobbered), MemoryDef, or |
| 409 | // MemoryPhi if several MemoryDefs can define this memory state. In that |
| 410 | // case add all Defs to WorkList and continue going up and checking all |
| 411 | // the definitions of this memory location until the root. When all the |
| 412 | // defs are exhausted and came to the entry state we have no clobber. |
| 413 | // Along the scan ignore barriers and fences which are considered clobbers |
| 414 | // by the MemorySSA, but not really writing anything into the memory. |
| 415 | while (!WorkList.empty()) { |
| 416 | MemoryAccess *MA = WorkList.pop_back_val(); |
| 417 | if (!Visited.insert(Ptr: MA).second) |
| 418 | continue; |
| 419 | |
| 420 | if (MSSA->isLiveOnEntryDef(MA)) |
| 421 | continue; |
| 422 | |
| 423 | if (MemoryDef *Def = dyn_cast<MemoryDef>(Val: MA)) { |
| 424 | LLVM_DEBUG(dbgs() << " Def: " << *Def->getMemoryInst() << '\n'); |
| 425 | |
| 426 | if (isReallyAClobber(Ptr: Load->getPointerOperand(), Def, AA)) { |
| 427 | LLVM_DEBUG(dbgs() << " -> load is clobbered\n" ); |
| 428 | return true; |
| 429 | } |
| 430 | |
| 431 | WorkList.push_back( |
| 432 | Elt: Walker->getClobberingMemoryAccess(MA: Def->getDefiningAccess(), Loc)); |
| 433 | continue; |
| 434 | } |
| 435 | |
| 436 | const MemoryPhi *Phi = cast<MemoryPhi>(Val: MA); |
| 437 | for (const auto &Use : Phi->incoming_values()) |
| 438 | WorkList.push_back(Elt: cast<MemoryAccess>(Val: &Use)); |
| 439 | } |
| 440 | |
| 441 | LLVM_DEBUG(dbgs() << " -> no clobber\n" ); |
| 442 | return false; |
| 443 | } |
| 444 | |
| 445 | } // end namespace llvm::AMDGPU |
| 446 | |