| 1 | //===- AMDGPUTargetTransformInfo.cpp - AMDGPU specific TTI 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 | // \file |
| 10 | // This file implements a TargetTransformInfo analysis pass specific to the |
| 11 | // AMDGPU target machine. It uses the target's detailed information to provide |
| 12 | // more precise answers to certain TTI queries, while letting the target |
| 13 | // independent and default TTI implementations handle the rest. |
| 14 | // |
| 15 | //===----------------------------------------------------------------------===// |
| 16 | |
| 17 | #include "AMDGPUTargetTransformInfo.h" |
| 18 | #include "AMDGPUTargetMachine.h" |
| 19 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
| 20 | #include "SIModeRegisterDefaults.h" |
| 21 | #include "llvm/Analysis/InlineCost.h" |
| 22 | #include "llvm/Analysis/LoopInfo.h" |
| 23 | #include "llvm/Analysis/ValueTracking.h" |
| 24 | #include "llvm/CodeGen/Analysis.h" |
| 25 | #include "llvm/IR/Function.h" |
| 26 | #include "llvm/IR/IRBuilder.h" |
| 27 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 28 | #include "llvm/IR/PatternMatch.h" |
| 29 | #include "llvm/Support/KnownBits.h" |
| 30 | #include <optional> |
| 31 | |
| 32 | using namespace llvm; |
| 33 | |
| 34 | #define DEBUG_TYPE "AMDGPUtti" |
| 35 | |
| 36 | static cl::opt<unsigned> UnrollThresholdPrivate( |
| 37 | "amdgpu-unroll-threshold-private" , |
| 38 | cl::desc("Unroll threshold for AMDGPU if private memory used in a loop" ), |
| 39 | cl::init(Val: 2700), cl::Hidden); |
| 40 | |
| 41 | static cl::opt<unsigned> UnrollThresholdLocal( |
| 42 | "amdgpu-unroll-threshold-local" , |
| 43 | cl::desc("Unroll threshold for AMDGPU if local memory used in a loop" ), |
| 44 | cl::init(Val: 1000), cl::Hidden); |
| 45 | |
| 46 | static cl::opt<unsigned> UnrollThresholdIf( |
| 47 | "amdgpu-unroll-threshold-if" , |
| 48 | cl::desc("Unroll threshold increment for AMDGPU for each if statement inside loop" ), |
| 49 | cl::init(Val: 200), cl::Hidden); |
| 50 | |
| 51 | static cl::opt<bool> UnrollRuntimeLocal( |
| 52 | "amdgpu-unroll-runtime-local" , |
| 53 | cl::desc("Allow runtime unroll for AMDGPU if local memory used in a loop" ), |
| 54 | cl::init(Val: true), cl::Hidden); |
| 55 | |
| 56 | static cl::opt<unsigned> UnrollMaxBlockToAnalyze( |
| 57 | "amdgpu-unroll-max-block-to-analyze" , |
| 58 | cl::desc("Inner loop block size threshold to analyze in unroll for AMDGPU" ), |
| 59 | cl::init(Val: 32), cl::Hidden); |
| 60 | |
| 61 | static cl::opt<unsigned> ArgAllocaCost("amdgpu-inline-arg-alloca-cost" , |
| 62 | cl::Hidden, cl::init(Val: 4000), |
| 63 | cl::desc("Cost of alloca argument" )); |
| 64 | |
| 65 | // If the amount of scratch memory to eliminate exceeds our ability to allocate |
| 66 | // it into registers we gain nothing by aggressively inlining functions for that |
| 67 | // heuristic. |
| 68 | static cl::opt<unsigned> |
| 69 | ArgAllocaCutoff("amdgpu-inline-arg-alloca-cutoff" , cl::Hidden, |
| 70 | cl::init(Val: 256), |
| 71 | cl::desc("Maximum alloca size to use for inline cost" )); |
| 72 | |
| 73 | // Inliner constraint to achieve reasonable compilation time. |
| 74 | static cl::opt<size_t> InlineMaxBB( |
| 75 | "amdgpu-inline-max-bb" , cl::Hidden, cl::init(Val: 1100), |
| 76 | cl::desc("Maximum number of BBs allowed in a function after inlining" |
| 77 | " (compile time constraint)" )); |
| 78 | |
| 79 | // This default unroll factor is based on microbenchmarks on gfx1030. |
| 80 | static cl::opt<unsigned> MemcpyLoopUnroll( |
| 81 | "amdgpu-memcpy-loop-unroll" , |
| 82 | cl::desc("Unroll factor (affecting 4x32-bit operations) to use for memory " |
| 83 | "operations when lowering memcpy as a loop" ), |
| 84 | cl::init(Val: 16), cl::Hidden); |
| 85 | |
| 86 | static bool dependsOnLocalPhi(const Loop *L, const Value *Cond, |
| 87 | unsigned Depth = 0) { |
| 88 | const Instruction *I = dyn_cast<Instruction>(Val: Cond); |
| 89 | if (!I) |
| 90 | return false; |
| 91 | |
| 92 | for (const Value *V : I->operand_values()) { |
| 93 | if (!L->contains(Inst: I)) |
| 94 | continue; |
| 95 | if (const PHINode *PHI = dyn_cast<PHINode>(Val: V)) { |
| 96 | if (llvm::none_of(Range: L->getSubLoops(), P: [PHI](const Loop* SubLoop) { |
| 97 | return SubLoop->contains(Inst: PHI); })) |
| 98 | return true; |
| 99 | } else if (Depth < 10 && dependsOnLocalPhi(L, Cond: V, Depth: Depth+1)) |
| 100 | return true; |
| 101 | } |
| 102 | return false; |
| 103 | } |
| 104 | |
| 105 | AMDGPUTTIImpl::AMDGPUTTIImpl(const AMDGPUTargetMachine *TM, const Function &F) |
| 106 | : BaseT(TM, F.getDataLayout()), |
| 107 | TargetTriple(TM->getTargetTriple()), |
| 108 | ST(static_cast<const GCNSubtarget *>(TM->getSubtargetImpl(F))), |
| 109 | TLI(ST->getTargetLowering()) {} |
| 110 | |
| 111 | void AMDGPUTTIImpl::( |
| 112 | Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP, |
| 113 | OptimizationRemarkEmitter *ORE) const { |
| 114 | const Function &F = *L->getHeader()->getParent(); |
| 115 | UP.Threshold = |
| 116 | F.getFnAttributeAsParsedInteger(Kind: "amdgpu-unroll-threshold" , Default: 300); |
| 117 | UP.MaxCount = std::numeric_limits<unsigned>::max(); |
| 118 | UP.Partial = true; |
| 119 | |
| 120 | // Conditional branch in a loop back edge needs 3 additional exec |
| 121 | // manipulations in average. |
| 122 | UP.BEInsns += 3; |
| 123 | |
| 124 | // We want to run unroll even for the loops which have been vectorized. |
| 125 | UP.UnrollVectorizedLoop = true; |
| 126 | |
| 127 | // TODO: Do we want runtime unrolling? |
| 128 | |
| 129 | // Maximum alloca size than can fit registers. Reserve 16 registers. |
| 130 | const unsigned MaxAlloca = (256 - 16) * 4; |
| 131 | unsigned ThresholdPrivate = UnrollThresholdPrivate; |
| 132 | unsigned ThresholdLocal = UnrollThresholdLocal; |
| 133 | |
| 134 | // If this loop has the amdgpu.loop.unroll.threshold metadata we will use the |
| 135 | // provided threshold value as the default for Threshold |
| 136 | if (MDNode *LoopUnrollThreshold = |
| 137 | findOptionMDForLoop(TheLoop: L, Name: "amdgpu.loop.unroll.threshold" )) { |
| 138 | if (LoopUnrollThreshold->getNumOperands() == 2) { |
| 139 | ConstantInt *MetaThresholdValue = mdconst::extract_or_null<ConstantInt>( |
| 140 | MD: LoopUnrollThreshold->getOperand(I: 1)); |
| 141 | if (MetaThresholdValue) { |
| 142 | // We will also use the supplied value for PartialThreshold for now. |
| 143 | // We may introduce additional metadata if it becomes necessary in the |
| 144 | // future. |
| 145 | UP.Threshold = MetaThresholdValue->getSExtValue(); |
| 146 | UP.PartialThreshold = UP.Threshold; |
| 147 | ThresholdPrivate = std::min(a: ThresholdPrivate, b: UP.Threshold); |
| 148 | ThresholdLocal = std::min(a: ThresholdLocal, b: UP.Threshold); |
| 149 | } |
| 150 | } |
| 151 | } |
| 152 | |
| 153 | unsigned MaxBoost = std::max(a: ThresholdPrivate, b: ThresholdLocal); |
| 154 | for (const BasicBlock *BB : L->getBlocks()) { |
| 155 | const DataLayout &DL = BB->getDataLayout(); |
| 156 | unsigned LocalGEPsSeen = 0; |
| 157 | |
| 158 | if (llvm::any_of(Range: L->getSubLoops(), P: [BB](const Loop* SubLoop) { |
| 159 | return SubLoop->contains(BB); })) |
| 160 | continue; // Block belongs to an inner loop. |
| 161 | |
| 162 | for (const Instruction &I : *BB) { |
| 163 | // Unroll a loop which contains an "if" statement whose condition |
| 164 | // defined by a PHI belonging to the loop. This may help to eliminate |
| 165 | // if region and potentially even PHI itself, saving on both divergence |
| 166 | // and registers used for the PHI. |
| 167 | // Add a small bonus for each of such "if" statements. |
| 168 | if (const BranchInst *Br = dyn_cast<BranchInst>(Val: &I)) { |
| 169 | if (UP.Threshold < MaxBoost && Br->isConditional()) { |
| 170 | BasicBlock *Succ0 = Br->getSuccessor(i: 0); |
| 171 | BasicBlock *Succ1 = Br->getSuccessor(i: 1); |
| 172 | if ((L->contains(BB: Succ0) && L->isLoopExiting(BB: Succ0)) || |
| 173 | (L->contains(BB: Succ1) && L->isLoopExiting(BB: Succ1))) |
| 174 | continue; |
| 175 | if (dependsOnLocalPhi(L, Cond: Br->getCondition())) { |
| 176 | UP.Threshold += UnrollThresholdIf; |
| 177 | LLVM_DEBUG(dbgs() << "Set unroll threshold " << UP.Threshold |
| 178 | << " for loop:\n" |
| 179 | << *L << " due to " << *Br << '\n'); |
| 180 | if (UP.Threshold >= MaxBoost) |
| 181 | return; |
| 182 | } |
| 183 | } |
| 184 | continue; |
| 185 | } |
| 186 | |
| 187 | const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: &I); |
| 188 | if (!GEP) |
| 189 | continue; |
| 190 | |
| 191 | unsigned AS = GEP->getAddressSpace(); |
| 192 | unsigned Threshold = 0; |
| 193 | if (AS == AMDGPUAS::PRIVATE_ADDRESS) |
| 194 | Threshold = ThresholdPrivate; |
| 195 | else if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) |
| 196 | Threshold = ThresholdLocal; |
| 197 | else |
| 198 | continue; |
| 199 | |
| 200 | if (UP.Threshold >= Threshold) |
| 201 | continue; |
| 202 | |
| 203 | if (AS == AMDGPUAS::PRIVATE_ADDRESS) { |
| 204 | const Value *Ptr = GEP->getPointerOperand(); |
| 205 | const AllocaInst *Alloca = |
| 206 | dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Ptr)); |
| 207 | if (!Alloca || !Alloca->isStaticAlloca()) |
| 208 | continue; |
| 209 | auto AllocaSize = Alloca->getAllocationSize(DL); |
| 210 | if (!AllocaSize || AllocaSize->getFixedValue() > MaxAlloca) |
| 211 | continue; |
| 212 | } else if (AS == AMDGPUAS::LOCAL_ADDRESS || |
| 213 | AS == AMDGPUAS::REGION_ADDRESS) { |
| 214 | LocalGEPsSeen++; |
| 215 | // Inhibit unroll for local memory if we have seen addressing not to |
| 216 | // a variable, most likely we will be unable to combine it. |
| 217 | // Do not unroll too deep inner loops for local memory to give a chance |
| 218 | // to unroll an outer loop for a more important reason. |
| 219 | if (LocalGEPsSeen > 1 || L->getLoopDepth() > 2) |
| 220 | continue; |
| 221 | |
| 222 | const Value *V = getUnderlyingObject(V: GEP->getPointerOperand()); |
| 223 | if (!isa<GlobalVariable>(Val: V) && !isa<Argument>(Val: V)) |
| 224 | continue; |
| 225 | |
| 226 | LLVM_DEBUG(dbgs() << "Allow unroll runtime for loop:\n" |
| 227 | << *L << " due to LDS use.\n" ); |
| 228 | UP.Runtime = UnrollRuntimeLocal; |
| 229 | } |
| 230 | |
| 231 | // Check if GEP depends on a value defined by this loop itself. |
| 232 | bool HasLoopDef = false; |
| 233 | for (const Value *Op : GEP->operands()) { |
| 234 | const Instruction *Inst = dyn_cast<Instruction>(Val: Op); |
| 235 | if (!Inst || L->isLoopInvariant(V: Op)) |
| 236 | continue; |
| 237 | |
| 238 | if (llvm::any_of(Range: L->getSubLoops(), P: [Inst](const Loop* SubLoop) { |
| 239 | return SubLoop->contains(Inst); })) |
| 240 | continue; |
| 241 | HasLoopDef = true; |
| 242 | break; |
| 243 | } |
| 244 | if (!HasLoopDef) |
| 245 | continue; |
| 246 | |
| 247 | // We want to do whatever we can to limit the number of alloca |
| 248 | // instructions that make it through to the code generator. allocas |
| 249 | // require us to use indirect addressing, which is slow and prone to |
| 250 | // compiler bugs. If this loop does an address calculation on an |
| 251 | // alloca ptr, then we want to use a higher than normal loop unroll |
| 252 | // threshold. This will give SROA a better chance to eliminate these |
| 253 | // allocas. |
| 254 | // |
| 255 | // We also want to have more unrolling for local memory to let ds |
| 256 | // instructions with different offsets combine. |
| 257 | // |
| 258 | // Don't use the maximum allowed value here as it will make some |
| 259 | // programs way too big. |
| 260 | UP.Threshold = Threshold; |
| 261 | LLVM_DEBUG(dbgs() << "Set unroll threshold " << Threshold |
| 262 | << " for loop:\n" |
| 263 | << *L << " due to " << *GEP << '\n'); |
| 264 | if (UP.Threshold >= MaxBoost) |
| 265 | return; |
| 266 | } |
| 267 | |
| 268 | // If we got a GEP in a small BB from inner loop then increase max trip |
| 269 | // count to analyze for better estimation cost in unroll |
| 270 | if (L->isInnermost() && BB->size() < UnrollMaxBlockToAnalyze) |
| 271 | UP.MaxIterationsCountToAnalyze = 32; |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | void AMDGPUTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE, |
| 276 | TTI::PeelingPreferences &PP) const { |
| 277 | BaseT::getPeelingPreferences(L, SE, PP); |
| 278 | } |
| 279 | |
| 280 | uint64_t AMDGPUTTIImpl::getMaxMemIntrinsicInlineSizeThreshold() const { |
| 281 | return 1024; |
| 282 | } |
| 283 | |
| 284 | const FeatureBitset GCNTTIImpl::InlineFeatureIgnoreList = { |
| 285 | // Codegen control options which don't matter. |
| 286 | AMDGPU::FeatureEnableLoadStoreOpt, AMDGPU::FeatureEnableSIScheduler, |
| 287 | AMDGPU::FeatureEnableUnsafeDSOffsetFolding, AMDGPU::FeatureUseFlatForGlobal, |
| 288 | AMDGPU::FeaturePromoteAlloca, AMDGPU::FeatureUnalignedScratchAccess, |
| 289 | AMDGPU::FeatureUnalignedAccessMode, |
| 290 | |
| 291 | AMDGPU::FeatureAutoWaitcntBeforeBarrier, |
| 292 | |
| 293 | // Property of the kernel/environment which can't actually differ. |
| 294 | AMDGPU::FeatureSGPRInitBug, AMDGPU::FeatureXNACK, |
| 295 | AMDGPU::FeatureTrapHandler, |
| 296 | |
| 297 | // The default assumption needs to be ecc is enabled, but no directly |
| 298 | // exposed operations depend on it, so it can be safely inlined. |
| 299 | AMDGPU::FeatureSRAMECC, |
| 300 | |
| 301 | // Perf-tuning features |
| 302 | AMDGPU::FeatureFastFMAF32, AMDGPU::FeatureHalfRate64Ops}; |
| 303 | |
| 304 | GCNTTIImpl::GCNTTIImpl(const AMDGPUTargetMachine *TM, const Function &F) |
| 305 | : BaseT(TM, F.getDataLayout()), |
| 306 | ST(static_cast<const GCNSubtarget *>(TM->getSubtargetImpl(F))), |
| 307 | TLI(ST->getTargetLowering()), CommonTTI(TM, F), |
| 308 | IsGraphics(AMDGPU::isGraphics(CC: F.getCallingConv())) { |
| 309 | SIModeRegisterDefaults Mode(F, *ST); |
| 310 | HasFP32Denormals = Mode.FP32Denormals != DenormalMode::getPreserveSign(); |
| 311 | HasFP64FP16Denormals = |
| 312 | Mode.FP64FP16Denormals != DenormalMode::getPreserveSign(); |
| 313 | } |
| 314 | |
| 315 | bool GCNTTIImpl::hasBranchDivergence(const Function *F) const { |
| 316 | return !F || !ST->isSingleLaneExecution(Kernel: *F); |
| 317 | } |
| 318 | |
| 319 | unsigned GCNTTIImpl::getNumberOfRegisters(unsigned RCID) const { |
| 320 | // NB: RCID is not an RCID. In fact it is 0 or 1 for scalar or vector |
| 321 | // registers. See getRegisterClassForType for the implementation. |
| 322 | // In this case vector registers are not vector in terms of |
| 323 | // VGPRs, but those which can hold multiple values. |
| 324 | |
| 325 | // This is really the number of registers to fill when vectorizing / |
| 326 | // interleaving loops, so we lie to avoid trying to use all registers. |
| 327 | return 4; |
| 328 | } |
| 329 | |
| 330 | TypeSize |
| 331 | GCNTTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const { |
| 332 | switch (K) { |
| 333 | case TargetTransformInfo::RGK_Scalar: |
| 334 | return TypeSize::getFixed(ExactSize: 32); |
| 335 | case TargetTransformInfo::RGK_FixedWidthVector: |
| 336 | return TypeSize::getFixed(ExactSize: ST->hasPackedFP32Ops() ? 64 : 32); |
| 337 | case TargetTransformInfo::RGK_ScalableVector: |
| 338 | return TypeSize::getScalable(MinimumSize: 0); |
| 339 | } |
| 340 | llvm_unreachable("Unsupported register kind" ); |
| 341 | } |
| 342 | |
| 343 | unsigned GCNTTIImpl::getMinVectorRegisterBitWidth() const { |
| 344 | return 32; |
| 345 | } |
| 346 | |
| 347 | unsigned GCNTTIImpl::getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { |
| 348 | if (Opcode == Instruction::Load || Opcode == Instruction::Store) |
| 349 | return 32 * 4 / ElemWidth; |
| 350 | // For a given width return the max 0number of elements that can be combined |
| 351 | // into a wider bit value: |
| 352 | return (ElemWidth == 8 && ST->has16BitInsts()) ? 4 |
| 353 | : (ElemWidth == 16 && ST->has16BitInsts()) ? 2 |
| 354 | : (ElemWidth == 32 && ST->hasPackedFP32Ops()) ? 2 |
| 355 | : 1; |
| 356 | } |
| 357 | |
| 358 | unsigned GCNTTIImpl::getLoadVectorFactor(unsigned VF, unsigned LoadSize, |
| 359 | unsigned ChainSizeInBytes, |
| 360 | VectorType *VecTy) const { |
| 361 | unsigned VecRegBitWidth = VF * LoadSize; |
| 362 | if (VecRegBitWidth > 128 && VecTy->getScalarSizeInBits() < 32) |
| 363 | // TODO: Support element-size less than 32bit? |
| 364 | return 128 / LoadSize; |
| 365 | |
| 366 | return VF; |
| 367 | } |
| 368 | |
| 369 | unsigned GCNTTIImpl::getStoreVectorFactor(unsigned VF, unsigned StoreSize, |
| 370 | unsigned ChainSizeInBytes, |
| 371 | VectorType *VecTy) const { |
| 372 | unsigned VecRegBitWidth = VF * StoreSize; |
| 373 | if (VecRegBitWidth > 128) |
| 374 | return 128 / StoreSize; |
| 375 | |
| 376 | return VF; |
| 377 | } |
| 378 | |
| 379 | unsigned GCNTTIImpl::getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { |
| 380 | if (AddrSpace == AMDGPUAS::GLOBAL_ADDRESS || |
| 381 | AddrSpace == AMDGPUAS::CONSTANT_ADDRESS || |
| 382 | AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT || |
| 383 | AddrSpace == AMDGPUAS::BUFFER_FAT_POINTER || |
| 384 | AddrSpace == AMDGPUAS::BUFFER_RESOURCE || |
| 385 | AddrSpace == AMDGPUAS::BUFFER_STRIDED_POINTER) { |
| 386 | return 512; |
| 387 | } |
| 388 | |
| 389 | if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) |
| 390 | return 8 * ST->getMaxPrivateElementSize(); |
| 391 | |
| 392 | // Common to flat, global, local and region. Assume for unknown addrspace. |
| 393 | return 128; |
| 394 | } |
| 395 | |
| 396 | bool GCNTTIImpl::isLegalToVectorizeMemChain(unsigned ChainSizeInBytes, |
| 397 | Align Alignment, |
| 398 | unsigned AddrSpace) const { |
| 399 | // We allow vectorization of flat stores, even though we may need to decompose |
| 400 | // them later if they may access private memory. We don't have enough context |
| 401 | // here, and legalization can handle it. |
| 402 | if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS) { |
| 403 | return (Alignment >= 4 || ST->hasUnalignedScratchAccessEnabled()) && |
| 404 | ChainSizeInBytes <= ST->getMaxPrivateElementSize(); |
| 405 | } |
| 406 | return true; |
| 407 | } |
| 408 | |
| 409 | bool GCNTTIImpl::isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, |
| 410 | Align Alignment, |
| 411 | unsigned AddrSpace) const { |
| 412 | return isLegalToVectorizeMemChain(ChainSizeInBytes, Alignment, AddrSpace); |
| 413 | } |
| 414 | |
| 415 | bool GCNTTIImpl::isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, |
| 416 | Align Alignment, |
| 417 | unsigned AddrSpace) const { |
| 418 | return isLegalToVectorizeMemChain(ChainSizeInBytes, Alignment, AddrSpace); |
| 419 | } |
| 420 | |
| 421 | uint64_t GCNTTIImpl::getMaxMemIntrinsicInlineSizeThreshold() const { |
| 422 | return 1024; |
| 423 | } |
| 424 | |
| 425 | Type *GCNTTIImpl::getMemcpyLoopLoweringType( |
| 426 | LLVMContext &Context, Value *Length, unsigned SrcAddrSpace, |
| 427 | unsigned DestAddrSpace, Align SrcAlign, Align DestAlign, |
| 428 | std::optional<uint32_t> AtomicElementSize) const { |
| 429 | |
| 430 | if (AtomicElementSize) |
| 431 | return Type::getIntNTy(C&: Context, N: *AtomicElementSize * 8); |
| 432 | |
| 433 | // 16-byte accesses achieve the highest copy throughput. |
| 434 | // If the operation has a fixed known length that is large enough, it is |
| 435 | // worthwhile to return an even wider type and let legalization lower it into |
| 436 | // multiple accesses, effectively unrolling the memcpy loop. |
| 437 | // We also rely on legalization to decompose into smaller accesses for |
| 438 | // subtargets and address spaces where it is necessary. |
| 439 | // |
| 440 | // Don't unroll if Length is not a constant, since unrolling leads to worse |
| 441 | // performance for length values that are smaller or slightly larger than the |
| 442 | // total size of the type returned here. Mitigating that would require a more |
| 443 | // complex lowering for variable-length memcpy and memmove. |
| 444 | unsigned I32EltsInVector = 4; |
| 445 | if (MemcpyLoopUnroll > 0 && isa<ConstantInt>(Val: Length)) |
| 446 | return FixedVectorType::get(ElementType: Type::getInt32Ty(C&: Context), |
| 447 | NumElts: MemcpyLoopUnroll * I32EltsInVector); |
| 448 | |
| 449 | return FixedVectorType::get(ElementType: Type::getInt32Ty(C&: Context), NumElts: I32EltsInVector); |
| 450 | } |
| 451 | |
| 452 | void GCNTTIImpl::getMemcpyLoopResidualLoweringType( |
| 453 | SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, |
| 454 | unsigned RemainingBytes, unsigned SrcAddrSpace, unsigned DestAddrSpace, |
| 455 | Align SrcAlign, Align DestAlign, |
| 456 | std::optional<uint32_t> AtomicCpySize) const { |
| 457 | |
| 458 | if (AtomicCpySize) |
| 459 | BaseT::getMemcpyLoopResidualLoweringType( |
| 460 | OpsOut, Context, RemainingBytes, SrcAddrSpace, DestAddrSpace, SrcAlign, |
| 461 | DestAlign, AtomicCpySize); |
| 462 | |
| 463 | Type *I32x4Ty = FixedVectorType::get(ElementType: Type::getInt32Ty(C&: Context), NumElts: 4); |
| 464 | while (RemainingBytes >= 16) { |
| 465 | OpsOut.push_back(Elt: I32x4Ty); |
| 466 | RemainingBytes -= 16; |
| 467 | } |
| 468 | |
| 469 | Type *I64Ty = Type::getInt64Ty(C&: Context); |
| 470 | while (RemainingBytes >= 8) { |
| 471 | OpsOut.push_back(Elt: I64Ty); |
| 472 | RemainingBytes -= 8; |
| 473 | } |
| 474 | |
| 475 | Type *I32Ty = Type::getInt32Ty(C&: Context); |
| 476 | while (RemainingBytes >= 4) { |
| 477 | OpsOut.push_back(Elt: I32Ty); |
| 478 | RemainingBytes -= 4; |
| 479 | } |
| 480 | |
| 481 | Type *I16Ty = Type::getInt16Ty(C&: Context); |
| 482 | while (RemainingBytes >= 2) { |
| 483 | OpsOut.push_back(Elt: I16Ty); |
| 484 | RemainingBytes -= 2; |
| 485 | } |
| 486 | |
| 487 | Type *I8Ty = Type::getInt8Ty(C&: Context); |
| 488 | while (RemainingBytes) { |
| 489 | OpsOut.push_back(Elt: I8Ty); |
| 490 | --RemainingBytes; |
| 491 | } |
| 492 | } |
| 493 | |
| 494 | unsigned GCNTTIImpl::getMaxInterleaveFactor(ElementCount VF) const { |
| 495 | // Disable unrolling if the loop is not vectorized. |
| 496 | // TODO: Enable this again. |
| 497 | if (VF.isScalar()) |
| 498 | return 1; |
| 499 | |
| 500 | return 8; |
| 501 | } |
| 502 | |
| 503 | bool GCNTTIImpl::getTgtMemIntrinsic(IntrinsicInst *Inst, |
| 504 | MemIntrinsicInfo &Info) const { |
| 505 | switch (Inst->getIntrinsicID()) { |
| 506 | case Intrinsic::amdgcn_ds_ordered_add: |
| 507 | case Intrinsic::amdgcn_ds_ordered_swap: { |
| 508 | auto *Ordering = dyn_cast<ConstantInt>(Val: Inst->getArgOperand(i: 2)); |
| 509 | auto *Volatile = dyn_cast<ConstantInt>(Val: Inst->getArgOperand(i: 4)); |
| 510 | if (!Ordering || !Volatile) |
| 511 | return false; // Invalid. |
| 512 | |
| 513 | unsigned OrderingVal = Ordering->getZExtValue(); |
| 514 | if (OrderingVal > static_cast<unsigned>(AtomicOrdering::SequentiallyConsistent)) |
| 515 | return false; |
| 516 | |
| 517 | Info.PtrVal = Inst->getArgOperand(i: 0); |
| 518 | Info.Ordering = static_cast<AtomicOrdering>(OrderingVal); |
| 519 | Info.ReadMem = true; |
| 520 | Info.WriteMem = true; |
| 521 | Info.IsVolatile = !Volatile->isZero(); |
| 522 | return true; |
| 523 | } |
| 524 | default: |
| 525 | return false; |
| 526 | } |
| 527 | } |
| 528 | |
| 529 | InstructionCost GCNTTIImpl::getArithmeticInstrCost( |
| 530 | unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, |
| 531 | TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info, |
| 532 | ArrayRef<const Value *> Args, const Instruction *CxtI) const { |
| 533 | |
| 534 | // Legalize the type. |
| 535 | std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty); |
| 536 | int ISD = TLI->InstructionOpcodeToISD(Opcode); |
| 537 | |
| 538 | // Because we don't have any legal vector operations, but the legal types, we |
| 539 | // need to account for split vectors. |
| 540 | unsigned NElts = LT.second.isVector() ? |
| 541 | LT.second.getVectorNumElements() : 1; |
| 542 | |
| 543 | MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy; |
| 544 | |
| 545 | switch (ISD) { |
| 546 | case ISD::SHL: |
| 547 | case ISD::SRL: |
| 548 | case ISD::SRA: |
| 549 | if (SLT == MVT::i64) |
| 550 | return get64BitInstrCost(CostKind) * LT.first * NElts; |
| 551 | |
| 552 | if (ST->has16BitInsts() && SLT == MVT::i16) |
| 553 | NElts = (NElts + 1) / 2; |
| 554 | |
| 555 | // i32 |
| 556 | return getFullRateInstrCost() * LT.first * NElts; |
| 557 | case ISD::ADD: |
| 558 | case ISD::SUB: |
| 559 | case ISD::AND: |
| 560 | case ISD::OR: |
| 561 | case ISD::XOR: |
| 562 | if (SLT == MVT::i64) { |
| 563 | // and, or and xor are typically split into 2 VALU instructions. |
| 564 | return 2 * getFullRateInstrCost() * LT.first * NElts; |
| 565 | } |
| 566 | |
| 567 | if (ST->has16BitInsts() && SLT == MVT::i16) |
| 568 | NElts = (NElts + 1) / 2; |
| 569 | |
| 570 | return LT.first * NElts * getFullRateInstrCost(); |
| 571 | case ISD::MUL: { |
| 572 | const int QuarterRateCost = getQuarterRateInstrCost(CostKind); |
| 573 | if (SLT == MVT::i64) { |
| 574 | const int FullRateCost = getFullRateInstrCost(); |
| 575 | return (4 * QuarterRateCost + (2 * 2) * FullRateCost) * LT.first * NElts; |
| 576 | } |
| 577 | |
| 578 | if (ST->has16BitInsts() && SLT == MVT::i16) |
| 579 | NElts = (NElts + 1) / 2; |
| 580 | |
| 581 | // i32 |
| 582 | return QuarterRateCost * NElts * LT.first; |
| 583 | } |
| 584 | case ISD::FMUL: |
| 585 | // Check possible fuse {fadd|fsub}(a,fmul(b,c)) and return zero cost for |
| 586 | // fmul(b,c) supposing the fadd|fsub will get estimated cost for the whole |
| 587 | // fused operation. |
| 588 | if (CxtI && CxtI->hasOneUse()) |
| 589 | if (const auto *FAdd = dyn_cast<BinaryOperator>(Val: *CxtI->user_begin())) { |
| 590 | const int OPC = TLI->InstructionOpcodeToISD(Opcode: FAdd->getOpcode()); |
| 591 | if (OPC == ISD::FADD || OPC == ISD::FSUB) { |
| 592 | if (ST->hasMadMacF32Insts() && SLT == MVT::f32 && !HasFP32Denormals) |
| 593 | return TargetTransformInfo::TCC_Free; |
| 594 | if (ST->has16BitInsts() && SLT == MVT::f16 && !HasFP64FP16Denormals) |
| 595 | return TargetTransformInfo::TCC_Free; |
| 596 | |
| 597 | // Estimate all types may be fused with contract/unsafe flags |
| 598 | const TargetOptions &Options = TLI->getTargetMachine().Options; |
| 599 | if (Options.AllowFPOpFusion == FPOpFusion::Fast || |
| 600 | (FAdd->hasAllowContract() && CxtI->hasAllowContract())) |
| 601 | return TargetTransformInfo::TCC_Free; |
| 602 | } |
| 603 | } |
| 604 | [[fallthrough]]; |
| 605 | case ISD::FADD: |
| 606 | case ISD::FSUB: |
| 607 | if (ST->hasPackedFP32Ops() && SLT == MVT::f32) |
| 608 | NElts = (NElts + 1) / 2; |
| 609 | if (ST->hasBF16PackedInsts() && SLT == MVT::bf16) |
| 610 | NElts = (NElts + 1) / 2; |
| 611 | if (SLT == MVT::f64) |
| 612 | return LT.first * NElts * get64BitInstrCost(CostKind); |
| 613 | |
| 614 | if (ST->has16BitInsts() && SLT == MVT::f16) |
| 615 | NElts = (NElts + 1) / 2; |
| 616 | |
| 617 | if (SLT == MVT::f32 || SLT == MVT::f16 || SLT == MVT::bf16) |
| 618 | return LT.first * NElts * getFullRateInstrCost(); |
| 619 | break; |
| 620 | case ISD::FDIV: |
| 621 | case ISD::FREM: |
| 622 | // FIXME: frem should be handled separately. The fdiv in it is most of it, |
| 623 | // but the current lowering is also not entirely correct. |
| 624 | if (SLT == MVT::f64) { |
| 625 | int Cost = 7 * get64BitInstrCost(CostKind) + |
| 626 | getQuarterRateInstrCost(CostKind) + |
| 627 | 3 * getHalfRateInstrCost(CostKind); |
| 628 | // Add cost of workaround. |
| 629 | if (!ST->hasUsableDivScaleConditionOutput()) |
| 630 | Cost += 3 * getFullRateInstrCost(); |
| 631 | |
| 632 | return LT.first * Cost * NElts; |
| 633 | } |
| 634 | |
| 635 | if (!Args.empty() && match(V: Args[0], P: PatternMatch::m_FPOne())) { |
| 636 | // TODO: This is more complicated, unsafe flags etc. |
| 637 | if ((SLT == MVT::f32 && !HasFP32Denormals) || |
| 638 | (SLT == MVT::f16 && ST->has16BitInsts())) { |
| 639 | return LT.first * getQuarterRateInstrCost(CostKind) * NElts; |
| 640 | } |
| 641 | } |
| 642 | |
| 643 | if (SLT == MVT::f16 && ST->has16BitInsts()) { |
| 644 | // 2 x v_cvt_f32_f16 |
| 645 | // f32 rcp |
| 646 | // f32 fmul |
| 647 | // v_cvt_f16_f32 |
| 648 | // f16 div_fixup |
| 649 | int Cost = |
| 650 | 4 * getFullRateInstrCost() + 2 * getQuarterRateInstrCost(CostKind); |
| 651 | return LT.first * Cost * NElts; |
| 652 | } |
| 653 | |
| 654 | if (SLT == MVT::f32 && (CxtI && CxtI->hasApproxFunc())) { |
| 655 | // Fast unsafe fdiv lowering: |
| 656 | // f32 rcp |
| 657 | // f32 fmul |
| 658 | int Cost = getQuarterRateInstrCost(CostKind) + getFullRateInstrCost(); |
| 659 | return LT.first * Cost * NElts; |
| 660 | } |
| 661 | |
| 662 | if (SLT == MVT::f32 || SLT == MVT::f16) { |
| 663 | // 4 more v_cvt_* insts without f16 insts support |
| 664 | int Cost = (SLT == MVT::f16 ? 14 : 10) * getFullRateInstrCost() + |
| 665 | 1 * getQuarterRateInstrCost(CostKind); |
| 666 | |
| 667 | if (!HasFP32Denormals) { |
| 668 | // FP mode switches. |
| 669 | Cost += 2 * getFullRateInstrCost(); |
| 670 | } |
| 671 | |
| 672 | return LT.first * NElts * Cost; |
| 673 | } |
| 674 | break; |
| 675 | case ISD::FNEG: |
| 676 | // Use the backend' estimation. If fneg is not free each element will cost |
| 677 | // one additional instruction. |
| 678 | return TLI->isFNegFree(VT: SLT) ? 0 : NElts; |
| 679 | default: |
| 680 | break; |
| 681 | } |
| 682 | |
| 683 | return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info, |
| 684 | Args, CxtI); |
| 685 | } |
| 686 | |
| 687 | // Return true if there's a potential benefit from using v2f16/v2i16 |
| 688 | // instructions for an intrinsic, even if it requires nontrivial legalization. |
| 689 | static bool intrinsicHasPackedVectorBenefit(Intrinsic::ID ID) { |
| 690 | switch (ID) { |
| 691 | case Intrinsic::fma: |
| 692 | case Intrinsic::fmuladd: |
| 693 | case Intrinsic::copysign: |
| 694 | case Intrinsic::minimumnum: |
| 695 | case Intrinsic::maximumnum: |
| 696 | case Intrinsic::canonicalize: |
| 697 | // There's a small benefit to using vector ops in the legalized code. |
| 698 | case Intrinsic::round: |
| 699 | case Intrinsic::uadd_sat: |
| 700 | case Intrinsic::usub_sat: |
| 701 | case Intrinsic::sadd_sat: |
| 702 | case Intrinsic::ssub_sat: |
| 703 | case Intrinsic::abs: |
| 704 | return true; |
| 705 | default: |
| 706 | return false; |
| 707 | } |
| 708 | } |
| 709 | |
| 710 | InstructionCost |
| 711 | GCNTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, |
| 712 | TTI::TargetCostKind CostKind) const { |
| 713 | switch (ICA.getID()) { |
| 714 | case Intrinsic::fabs: |
| 715 | // Free source modifier in the common case. |
| 716 | return 0; |
| 717 | case Intrinsic::amdgcn_workitem_id_x: |
| 718 | case Intrinsic::amdgcn_workitem_id_y: |
| 719 | case Intrinsic::amdgcn_workitem_id_z: |
| 720 | // TODO: If hasPackedTID, or if the calling context is not an entry point |
| 721 | // there may be a bit instruction. |
| 722 | return 0; |
| 723 | case Intrinsic::amdgcn_workgroup_id_x: |
| 724 | case Intrinsic::amdgcn_workgroup_id_y: |
| 725 | case Intrinsic::amdgcn_workgroup_id_z: |
| 726 | case Intrinsic::amdgcn_lds_kernel_id: |
| 727 | case Intrinsic::amdgcn_dispatch_ptr: |
| 728 | case Intrinsic::amdgcn_dispatch_id: |
| 729 | case Intrinsic::amdgcn_implicitarg_ptr: |
| 730 | case Intrinsic::amdgcn_queue_ptr: |
| 731 | // Read from an argument register. |
| 732 | return 0; |
| 733 | default: |
| 734 | break; |
| 735 | } |
| 736 | |
| 737 | if (!intrinsicHasPackedVectorBenefit(ID: ICA.getID())) |
| 738 | return BaseT::getIntrinsicInstrCost(ICA, CostKind); |
| 739 | |
| 740 | Type *RetTy = ICA.getReturnType(); |
| 741 | |
| 742 | // Legalize the type. |
| 743 | std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: RetTy); |
| 744 | |
| 745 | unsigned NElts = LT.second.isVector() ? |
| 746 | LT.second.getVectorNumElements() : 1; |
| 747 | |
| 748 | MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy; |
| 749 | |
| 750 | if ((ST->hasVOP3PInsts() && |
| 751 | (SLT == MVT::f16 || SLT == MVT::i16 || |
| 752 | (SLT == MVT::bf16 && ST->hasBF16PackedInsts()))) || |
| 753 | (ST->hasPackedFP32Ops() && SLT == MVT::f32)) |
| 754 | NElts = (NElts + 1) / 2; |
| 755 | |
| 756 | // TODO: Get more refined intrinsic costs? |
| 757 | unsigned InstRate = getQuarterRateInstrCost(CostKind); |
| 758 | |
| 759 | switch (ICA.getID()) { |
| 760 | case Intrinsic::fma: |
| 761 | case Intrinsic::fmuladd: |
| 762 | if (SLT == MVT::f64) { |
| 763 | InstRate = get64BitInstrCost(CostKind); |
| 764 | break; |
| 765 | } |
| 766 | |
| 767 | if ((SLT == MVT::f32 && ST->hasFastFMAF32()) || SLT == MVT::f16) |
| 768 | InstRate = getFullRateInstrCost(); |
| 769 | else { |
| 770 | InstRate = ST->hasFastFMAF32() ? getHalfRateInstrCost(CostKind) |
| 771 | : getQuarterRateInstrCost(CostKind); |
| 772 | } |
| 773 | break; |
| 774 | case Intrinsic::copysign: |
| 775 | return NElts * getFullRateInstrCost(); |
| 776 | case Intrinsic::minimumnum: |
| 777 | case Intrinsic::maximumnum: { |
| 778 | // Instruction + 2 canonicalizes. For cases that need type promotion, we the |
| 779 | // promotion takes the place of the canonicalize. |
| 780 | unsigned NumOps = 3; |
| 781 | if (const IntrinsicInst *II = ICA.getInst()) { |
| 782 | // Directly legal with ieee=0 |
| 783 | // TODO: Not directly legal with strictfp |
| 784 | if (fpenvIEEEMode(I: *II) == KnownIEEEMode::Off) |
| 785 | NumOps = 1; |
| 786 | } |
| 787 | |
| 788 | unsigned BaseRate = |
| 789 | SLT == MVT::f64 ? get64BitInstrCost(CostKind) : getFullRateInstrCost(); |
| 790 | InstRate = BaseRate * NumOps; |
| 791 | break; |
| 792 | } |
| 793 | case Intrinsic::canonicalize: { |
| 794 | InstRate = |
| 795 | SLT == MVT::f64 ? get64BitInstrCost(CostKind) : getFullRateInstrCost(); |
| 796 | break; |
| 797 | } |
| 798 | case Intrinsic::uadd_sat: |
| 799 | case Intrinsic::usub_sat: |
| 800 | case Intrinsic::sadd_sat: |
| 801 | case Intrinsic::ssub_sat: { |
| 802 | if (SLT == MVT::i16 || SLT == MVT::i32) |
| 803 | InstRate = getFullRateInstrCost(); |
| 804 | |
| 805 | static const auto ValidSatTys = {MVT::v2i16, MVT::v4i16}; |
| 806 | if (any_of(Range: ValidSatTys, P: equal_to(Arg&: LT.second))) |
| 807 | NElts = 1; |
| 808 | break; |
| 809 | } |
| 810 | case Intrinsic::abs: |
| 811 | // Expansion takes 2 instructions for VALU |
| 812 | if (SLT == MVT::i16 || SLT == MVT::i32) |
| 813 | InstRate = 2 * getFullRateInstrCost(); |
| 814 | break; |
| 815 | default: |
| 816 | break; |
| 817 | } |
| 818 | |
| 819 | return LT.first * NElts * InstRate; |
| 820 | } |
| 821 | |
| 822 | InstructionCost GCNTTIImpl::getCFInstrCost(unsigned Opcode, |
| 823 | TTI::TargetCostKind CostKind, |
| 824 | const Instruction *I) const { |
| 825 | assert((I == nullptr || I->getOpcode() == Opcode) && |
| 826 | "Opcode should reflect passed instruction." ); |
| 827 | const bool SCost = |
| 828 | (CostKind == TTI::TCK_CodeSize || CostKind == TTI::TCK_SizeAndLatency); |
| 829 | const int CBrCost = SCost ? 5 : 7; |
| 830 | switch (Opcode) { |
| 831 | case Instruction::Br: { |
| 832 | // Branch instruction takes about 4 slots on gfx900. |
| 833 | const auto *BI = dyn_cast_or_null<BranchInst>(Val: I); |
| 834 | if (BI && BI->isUnconditional()) |
| 835 | return SCost ? 1 : 4; |
| 836 | // Suppose conditional branch takes additional 3 exec manipulations |
| 837 | // instructions in average. |
| 838 | return CBrCost; |
| 839 | } |
| 840 | case Instruction::Switch: { |
| 841 | const auto *SI = dyn_cast_or_null<SwitchInst>(Val: I); |
| 842 | // Each case (including default) takes 1 cmp + 1 cbr instructions in |
| 843 | // average. |
| 844 | return (SI ? (SI->getNumCases() + 1) : 4) * (CBrCost + 1); |
| 845 | } |
| 846 | case Instruction::Ret: |
| 847 | return SCost ? 1 : 10; |
| 848 | } |
| 849 | return BaseT::getCFInstrCost(Opcode, CostKind, I); |
| 850 | } |
| 851 | |
| 852 | InstructionCost |
| 853 | GCNTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *Ty, |
| 854 | std::optional<FastMathFlags> FMF, |
| 855 | TTI::TargetCostKind CostKind) const { |
| 856 | if (TTI::requiresOrderedReduction(FMF)) |
| 857 | return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind); |
| 858 | |
| 859 | EVT OrigTy = TLI->getValueType(DL, Ty); |
| 860 | |
| 861 | // Computes cost on targets that have packed math instructions(which support |
| 862 | // 16-bit types only). |
| 863 | if (!ST->hasVOP3PInsts() || OrigTy.getScalarSizeInBits() != 16) |
| 864 | return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind); |
| 865 | |
| 866 | std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty); |
| 867 | return LT.first * getFullRateInstrCost(); |
| 868 | } |
| 869 | |
| 870 | InstructionCost |
| 871 | GCNTTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, |
| 872 | FastMathFlags FMF, |
| 873 | TTI::TargetCostKind CostKind) const { |
| 874 | EVT OrigTy = TLI->getValueType(DL, Ty); |
| 875 | |
| 876 | // Computes cost on targets that have packed math instructions(which support |
| 877 | // 16-bit types only). |
| 878 | if (!ST->hasVOP3PInsts() || OrigTy.getScalarSizeInBits() != 16) |
| 879 | return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind); |
| 880 | |
| 881 | std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty); |
| 882 | return LT.first * getHalfRateInstrCost(CostKind); |
| 883 | } |
| 884 | |
| 885 | InstructionCost GCNTTIImpl::getVectorInstrCost( |
| 886 | unsigned Opcode, Type *ValTy, TTI::TargetCostKind CostKind, unsigned Index, |
| 887 | const Value *Op0, const Value *Op1, TTI::VectorInstrContext VIC) const { |
| 888 | switch (Opcode) { |
| 889 | case Instruction::ExtractElement: |
| 890 | case Instruction::InsertElement: { |
| 891 | unsigned EltSize |
| 892 | = DL.getTypeSizeInBits(Ty: cast<VectorType>(Val: ValTy)->getElementType()); |
| 893 | if (EltSize < 32) { |
| 894 | if (EltSize == 16 && Index == 0 && ST->has16BitInsts()) |
| 895 | return 0; |
| 896 | return BaseT::getVectorInstrCost(Opcode, Val: ValTy, CostKind, Index, Op0, Op1, |
| 897 | VIC); |
| 898 | } |
| 899 | |
| 900 | // Extracts are just reads of a subregister, so are free. Inserts are |
| 901 | // considered free because we don't want to have any cost for scalarizing |
| 902 | // operations, and we don't have to copy into a different register class. |
| 903 | |
| 904 | // Dynamic indexing isn't free and is best avoided. |
| 905 | return Index == ~0u ? 2 : 0; |
| 906 | } |
| 907 | default: |
| 908 | return BaseT::getVectorInstrCost(Opcode, Val: ValTy, CostKind, Index, Op0, Op1, |
| 909 | VIC); |
| 910 | } |
| 911 | } |
| 912 | |
| 913 | /// Analyze if the results of inline asm are divergent. If \p Indices is empty, |
| 914 | /// this is analyzing the collective result of all output registers. Otherwise, |
| 915 | /// this is only querying a specific result index if this returns multiple |
| 916 | /// registers in a struct. |
| 917 | bool GCNTTIImpl::isInlineAsmSourceOfDivergence( |
| 918 | const CallInst *CI, ArrayRef<unsigned> Indices) const { |
| 919 | // TODO: Handle complex extract indices |
| 920 | if (Indices.size() > 1) |
| 921 | return true; |
| 922 | |
| 923 | const DataLayout &DL = CI->getDataLayout(); |
| 924 | const SIRegisterInfo *TRI = ST->getRegisterInfo(); |
| 925 | TargetLowering::AsmOperandInfoVector TargetConstraints = |
| 926 | TLI->ParseConstraints(DL, TRI: ST->getRegisterInfo(), Call: *CI); |
| 927 | |
| 928 | const int TargetOutputIdx = Indices.empty() ? -1 : Indices[0]; |
| 929 | |
| 930 | int OutputIdx = 0; |
| 931 | for (auto &TC : TargetConstraints) { |
| 932 | if (TC.Type != InlineAsm::isOutput) |
| 933 | continue; |
| 934 | |
| 935 | // Skip outputs we don't care about. |
| 936 | if (TargetOutputIdx != -1 && TargetOutputIdx != OutputIdx++) |
| 937 | continue; |
| 938 | |
| 939 | TLI->ComputeConstraintToUse(OpInfo&: TC, Op: SDValue()); |
| 940 | |
| 941 | const TargetRegisterClass *RC = TLI->getRegForInlineAsmConstraint( |
| 942 | TRI, Constraint: TC.ConstraintCode, VT: TC.ConstraintVT).second; |
| 943 | |
| 944 | // For AGPR constraints null is returned on subtargets without AGPRs, so |
| 945 | // assume divergent for null. |
| 946 | if (!RC || !TRI->isSGPRClass(RC)) |
| 947 | return true; |
| 948 | } |
| 949 | |
| 950 | return false; |
| 951 | } |
| 952 | |
| 953 | bool GCNTTIImpl::isReadRegisterSourceOfDivergence( |
| 954 | const IntrinsicInst *ReadReg) const { |
| 955 | Metadata *MD = |
| 956 | cast<MetadataAsValue>(Val: ReadReg->getArgOperand(i: 0))->getMetadata(); |
| 957 | StringRef RegName = |
| 958 | cast<MDString>(Val: cast<MDNode>(Val: MD)->getOperand(I: 0))->getString(); |
| 959 | |
| 960 | // Special case registers that look like VCC. |
| 961 | MVT VT = MVT::getVT(Ty: ReadReg->getType()); |
| 962 | if (VT == MVT::i1) |
| 963 | return true; |
| 964 | |
| 965 | // Special case scalar registers that start with 'v'. |
| 966 | if (RegName.starts_with(Prefix: "vcc" ) || RegName.empty()) |
| 967 | return false; |
| 968 | |
| 969 | // VGPR or AGPR is divergent. There aren't any specially named vector |
| 970 | // registers. |
| 971 | return RegName[0] == 'v' || RegName[0] == 'a'; |
| 972 | } |
| 973 | |
| 974 | /// \returns true if the result of the value could potentially be |
| 975 | /// different across workitems in a wavefront. |
| 976 | bool GCNTTIImpl::isSourceOfDivergence(const Value *V) const { |
| 977 | if (const Argument *A = dyn_cast<Argument>(Val: V)) |
| 978 | return !AMDGPU::isArgPassedInSGPR(Arg: A); |
| 979 | |
| 980 | // Loads from the private and flat address spaces are divergent, because |
| 981 | // threads can execute the load instruction with the same inputs and get |
| 982 | // different results. |
| 983 | // |
| 984 | // All other loads are not divergent, because if threads issue loads with the |
| 985 | // same arguments, they will always get the same result. |
| 986 | if (const LoadInst *Load = dyn_cast<LoadInst>(Val: V)) |
| 987 | return Load->getPointerAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS || |
| 988 | Load->getPointerAddressSpace() == AMDGPUAS::FLAT_ADDRESS; |
| 989 | |
| 990 | // Atomics are divergent because they are executed sequentially: when an |
| 991 | // atomic operation refers to the same address in each thread, then each |
| 992 | // thread after the first sees the value written by the previous thread as |
| 993 | // original value. |
| 994 | if (isa<AtomicRMWInst, AtomicCmpXchgInst>(Val: V)) |
| 995 | return true; |
| 996 | |
| 997 | if (const IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(Val: V)) { |
| 998 | Intrinsic::ID IID = Intrinsic->getIntrinsicID(); |
| 999 | switch (IID) { |
| 1000 | case Intrinsic::read_register: |
| 1001 | return isReadRegisterSourceOfDivergence(ReadReg: Intrinsic); |
| 1002 | case Intrinsic::amdgcn_addrspacecast_nonnull: { |
| 1003 | unsigned SrcAS = |
| 1004 | Intrinsic->getOperand(i_nocapture: 0)->getType()->getPointerAddressSpace(); |
| 1005 | unsigned DstAS = Intrinsic->getType()->getPointerAddressSpace(); |
| 1006 | return SrcAS == AMDGPUAS::PRIVATE_ADDRESS && |
| 1007 | DstAS == AMDGPUAS::FLAT_ADDRESS && |
| 1008 | ST->hasGloballyAddressableScratch(); |
| 1009 | } |
| 1010 | case Intrinsic::amdgcn_workitem_id_y: |
| 1011 | case Intrinsic::amdgcn_workitem_id_z: { |
| 1012 | const Function *F = Intrinsic->getFunction(); |
| 1013 | bool HasUniformYZ = |
| 1014 | ST->hasWavefrontsEvenlySplittingXDim(F: *F, /*RequitezUniformYZ=*/REquiresUniformYZ: true); |
| 1015 | std::optional<unsigned> ThisDimSize = ST->getReqdWorkGroupSize( |
| 1016 | F: *F, Dim: IID == Intrinsic::amdgcn_workitem_id_y ? 1 : 2); |
| 1017 | return !HasUniformYZ && (!ThisDimSize || *ThisDimSize != 1); |
| 1018 | } |
| 1019 | default: |
| 1020 | return AMDGPU::isIntrinsicSourceOfDivergence(IntrID: IID); |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | // Assume all function calls are a source of divergence. |
| 1025 | if (const CallInst *CI = dyn_cast<CallInst>(Val: V)) { |
| 1026 | if (CI->isInlineAsm()) |
| 1027 | return isInlineAsmSourceOfDivergence(CI); |
| 1028 | return true; |
| 1029 | } |
| 1030 | |
| 1031 | // Assume all function calls are a source of divergence. |
| 1032 | if (isa<InvokeInst>(Val: V)) |
| 1033 | return true; |
| 1034 | |
| 1035 | // If the target supports globally addressable scratch, the mapping from |
| 1036 | // scratch memory to the flat aperture changes therefore an address space cast |
| 1037 | // is no longer uniform. |
| 1038 | if (auto *CastI = dyn_cast<AddrSpaceCastInst>(Val: V)) { |
| 1039 | return CastI->getSrcAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS && |
| 1040 | CastI->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS && |
| 1041 | ST->hasGloballyAddressableScratch(); |
| 1042 | } |
| 1043 | |
| 1044 | return false; |
| 1045 | } |
| 1046 | |
| 1047 | bool GCNTTIImpl::isAlwaysUniform(const Value *V) const { |
| 1048 | if (const IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(Val: V)) |
| 1049 | return AMDGPU::isIntrinsicAlwaysUniform(IntrID: Intrinsic->getIntrinsicID()); |
| 1050 | |
| 1051 | if (const CallInst *CI = dyn_cast<CallInst>(Val: V)) { |
| 1052 | if (CI->isInlineAsm()) |
| 1053 | return !isInlineAsmSourceOfDivergence(CI); |
| 1054 | return false; |
| 1055 | } |
| 1056 | |
| 1057 | // In most cases TID / wavefrontsize is uniform. |
| 1058 | // |
| 1059 | // However, if a kernel has uneven dimesions we can have a value of |
| 1060 | // workitem-id-x divided by the wavefrontsize non-uniform. For example |
| 1061 | // dimensions (65, 2) will have workitems with address (64, 0) and (0, 1) |
| 1062 | // packed into a same wave which gives 1 and 0 after the division by 64 |
| 1063 | // respectively. |
| 1064 | // |
| 1065 | // The X dimension doesn't reset within a wave if either both the Y |
| 1066 | // and Z dimensions are of length 1, or if the X dimension's required |
| 1067 | // size is a power of 2. Note, however, if the X dimension's maximum |
| 1068 | // size is a power of 2 < the wavefront size, division by the wavefront |
| 1069 | // size is guaranteed to yield 0, so this is also a no-reset case. |
| 1070 | bool XDimDoesntResetWithinWaves = false; |
| 1071 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 1072 | const Function *F = I->getFunction(); |
| 1073 | XDimDoesntResetWithinWaves = ST->hasWavefrontsEvenlySplittingXDim(F: *F); |
| 1074 | } |
| 1075 | using namespace llvm::PatternMatch; |
| 1076 | uint64_t C; |
| 1077 | if (match(V, P: m_LShr(L: m_Intrinsic<Intrinsic::amdgcn_workitem_id_x>(), |
| 1078 | R: m_ConstantInt(V&: C))) || |
| 1079 | match(V, P: m_AShr(L: m_Intrinsic<Intrinsic::amdgcn_workitem_id_x>(), |
| 1080 | R: m_ConstantInt(V&: C)))) { |
| 1081 | return C >= ST->getWavefrontSizeLog2() && XDimDoesntResetWithinWaves; |
| 1082 | } |
| 1083 | |
| 1084 | Value *Mask; |
| 1085 | if (match(V, P: m_c_And(L: m_Intrinsic<Intrinsic::amdgcn_workitem_id_x>(), |
| 1086 | R: m_Value(V&: Mask)))) { |
| 1087 | return computeKnownBits(V: Mask, DL).countMinTrailingZeros() >= |
| 1088 | ST->getWavefrontSizeLog2() && |
| 1089 | XDimDoesntResetWithinWaves; |
| 1090 | } |
| 1091 | |
| 1092 | const ExtractValueInst *ExtValue = dyn_cast<ExtractValueInst>(Val: V); |
| 1093 | if (!ExtValue) |
| 1094 | return false; |
| 1095 | |
| 1096 | const CallInst *CI = dyn_cast<CallInst>(Val: ExtValue->getOperand(i_nocapture: 0)); |
| 1097 | if (!CI) |
| 1098 | return false; |
| 1099 | |
| 1100 | if (const IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(Val: CI)) { |
| 1101 | switch (Intrinsic->getIntrinsicID()) { |
| 1102 | default: |
| 1103 | return false; |
| 1104 | case Intrinsic::amdgcn_if: |
| 1105 | case Intrinsic::amdgcn_else: { |
| 1106 | ArrayRef<unsigned> Indices = ExtValue->getIndices(); |
| 1107 | return Indices.size() == 1 && Indices[0] == 1; |
| 1108 | } |
| 1109 | } |
| 1110 | } |
| 1111 | |
| 1112 | // If we have inline asm returning mixed SGPR and VGPR results, we inferred |
| 1113 | // divergent for the overall struct return. We need to override it in the |
| 1114 | // case we're extracting an SGPR component here. |
| 1115 | if (CI->isInlineAsm()) |
| 1116 | return !isInlineAsmSourceOfDivergence(CI, Indices: ExtValue->getIndices()); |
| 1117 | |
| 1118 | return false; |
| 1119 | } |
| 1120 | |
| 1121 | bool GCNTTIImpl::collectFlatAddressOperands(SmallVectorImpl<int> &OpIndexes, |
| 1122 | Intrinsic::ID IID) const { |
| 1123 | switch (IID) { |
| 1124 | case Intrinsic::amdgcn_is_shared: |
| 1125 | case Intrinsic::amdgcn_is_private: |
| 1126 | case Intrinsic::amdgcn_flat_atomic_fmax_num: |
| 1127 | case Intrinsic::amdgcn_flat_atomic_fmin_num: |
| 1128 | case Intrinsic::amdgcn_load_to_lds: |
| 1129 | case Intrinsic::amdgcn_make_buffer_rsrc: |
| 1130 | OpIndexes.push_back(Elt: 0); |
| 1131 | return true; |
| 1132 | default: |
| 1133 | return false; |
| 1134 | } |
| 1135 | } |
| 1136 | |
| 1137 | Value *GCNTTIImpl::rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, |
| 1138 | Value *OldV, |
| 1139 | Value *NewV) const { |
| 1140 | auto IntrID = II->getIntrinsicID(); |
| 1141 | switch (IntrID) { |
| 1142 | case Intrinsic::amdgcn_is_shared: |
| 1143 | case Intrinsic::amdgcn_is_private: { |
| 1144 | unsigned TrueAS = IntrID == Intrinsic::amdgcn_is_shared ? |
| 1145 | AMDGPUAS::LOCAL_ADDRESS : AMDGPUAS::PRIVATE_ADDRESS; |
| 1146 | unsigned NewAS = NewV->getType()->getPointerAddressSpace(); |
| 1147 | LLVMContext &Ctx = NewV->getType()->getContext(); |
| 1148 | ConstantInt *NewVal = (TrueAS == NewAS) ? |
| 1149 | ConstantInt::getTrue(Context&: Ctx) : ConstantInt::getFalse(Context&: Ctx); |
| 1150 | return NewVal; |
| 1151 | } |
| 1152 | case Intrinsic::amdgcn_flat_atomic_fmax_num: |
| 1153 | case Intrinsic::amdgcn_flat_atomic_fmin_num: { |
| 1154 | Type *DestTy = II->getType(); |
| 1155 | Type *SrcTy = NewV->getType(); |
| 1156 | unsigned NewAS = SrcTy->getPointerAddressSpace(); |
| 1157 | if (!AMDGPU::isExtendedGlobalAddrSpace(AS: NewAS)) |
| 1158 | return nullptr; |
| 1159 | Module *M = II->getModule(); |
| 1160 | Function *NewDecl = Intrinsic::getOrInsertDeclaration( |
| 1161 | M, id: II->getIntrinsicID(), Tys: {DestTy, SrcTy, DestTy}); |
| 1162 | II->setArgOperand(i: 0, v: NewV); |
| 1163 | II->setCalledFunction(NewDecl); |
| 1164 | return II; |
| 1165 | } |
| 1166 | case Intrinsic::amdgcn_load_to_lds: { |
| 1167 | Type *SrcTy = NewV->getType(); |
| 1168 | Module *M = II->getModule(); |
| 1169 | Function *NewDecl = |
| 1170 | Intrinsic::getOrInsertDeclaration(M, id: II->getIntrinsicID(), Tys: {SrcTy}); |
| 1171 | II->setArgOperand(i: 0, v: NewV); |
| 1172 | II->setCalledFunction(NewDecl); |
| 1173 | return II; |
| 1174 | } |
| 1175 | case Intrinsic::amdgcn_make_buffer_rsrc: { |
| 1176 | Type *SrcTy = NewV->getType(); |
| 1177 | Type *DstTy = II->getType(); |
| 1178 | Module *M = II->getModule(); |
| 1179 | Function *NewDecl = Intrinsic::getOrInsertDeclaration( |
| 1180 | M, id: II->getIntrinsicID(), Tys: {DstTy, SrcTy}); |
| 1181 | II->setArgOperand(i: 0, v: NewV); |
| 1182 | II->setCalledFunction(NewDecl); |
| 1183 | return II; |
| 1184 | } |
| 1185 | default: |
| 1186 | return nullptr; |
| 1187 | } |
| 1188 | } |
| 1189 | |
| 1190 | InstructionCost GCNTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, |
| 1191 | VectorType *DstTy, VectorType *SrcTy, |
| 1192 | ArrayRef<int> Mask, |
| 1193 | TTI::TargetCostKind CostKind, |
| 1194 | int Index, VectorType *SubTp, |
| 1195 | ArrayRef<const Value *> Args, |
| 1196 | const Instruction *CxtI) const { |
| 1197 | if (!isa<FixedVectorType>(Val: SrcTy)) |
| 1198 | return BaseT::getShuffleCost(Kind, DstTy, SrcTy, Mask, CostKind, Index, |
| 1199 | SubTp); |
| 1200 | |
| 1201 | Kind = improveShuffleKindFromMask(Kind, Mask, SrcTy, Index, SubTy&: SubTp); |
| 1202 | |
| 1203 | unsigned ScalarSize = DL.getTypeSizeInBits(Ty: SrcTy->getElementType()); |
| 1204 | if (ST->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS && |
| 1205 | (ScalarSize == 16 || ScalarSize == 8)) { |
| 1206 | // Larger vector widths may require additional instructions, but are |
| 1207 | // typically cheaper than scalarized versions. |
| 1208 | // |
| 1209 | // We assume that shuffling at a register granularity can be done for free. |
| 1210 | // This is not true for vectors fed into memory instructions, but it is |
| 1211 | // effectively true for all other shuffling. The emphasis of the logic here |
| 1212 | // is to assist generic transform in cleaning up / canonicalizing those |
| 1213 | // shuffles. |
| 1214 | |
| 1215 | // With op_sel VOP3P instructions freely can access the low half or high |
| 1216 | // half of a register, so any swizzle of two elements is free. |
| 1217 | if (auto *SrcVecTy = dyn_cast<FixedVectorType>(Val: SrcTy)) { |
| 1218 | unsigned NumSrcElts = SrcVecTy->getNumElements(); |
| 1219 | if (ST->hasVOP3PInsts() && ScalarSize == 16 && NumSrcElts == 2 && |
| 1220 | (Kind == TTI::SK_Broadcast || Kind == TTI::SK_Reverse || |
| 1221 | Kind == TTI::SK_PermuteSingleSrc)) |
| 1222 | return 0; |
| 1223 | } |
| 1224 | |
| 1225 | unsigned EltsPerReg = 32 / ScalarSize; |
| 1226 | switch (Kind) { |
| 1227 | case TTI::SK_Broadcast: |
| 1228 | // A single v_perm_b32 can be re-used for all destination registers. |
| 1229 | return 1; |
| 1230 | case TTI::SK_Reverse: |
| 1231 | // One instruction per register. |
| 1232 | if (auto *DstVecTy = dyn_cast<FixedVectorType>(Val: DstTy)) |
| 1233 | return divideCeil(Numerator: DstVecTy->getNumElements(), Denominator: EltsPerReg); |
| 1234 | return InstructionCost::getInvalid(); |
| 1235 | case TTI::SK_ExtractSubvector: |
| 1236 | if (Index % EltsPerReg == 0) |
| 1237 | return 0; // Shuffling at register granularity |
| 1238 | if (auto *DstVecTy = dyn_cast<FixedVectorType>(Val: DstTy)) |
| 1239 | return divideCeil(Numerator: DstVecTy->getNumElements(), Denominator: EltsPerReg); |
| 1240 | return InstructionCost::getInvalid(); |
| 1241 | case TTI::SK_InsertSubvector: { |
| 1242 | auto *DstVecTy = dyn_cast<FixedVectorType>(Val: DstTy); |
| 1243 | if (!DstVecTy) |
| 1244 | return InstructionCost::getInvalid(); |
| 1245 | unsigned NumDstElts = DstVecTy->getNumElements(); |
| 1246 | unsigned NumInsertElts = cast<FixedVectorType>(Val: SubTp)->getNumElements(); |
| 1247 | unsigned EndIndex = Index + NumInsertElts; |
| 1248 | unsigned BeginSubIdx = Index % EltsPerReg; |
| 1249 | unsigned EndSubIdx = EndIndex % EltsPerReg; |
| 1250 | unsigned Cost = 0; |
| 1251 | |
| 1252 | if (BeginSubIdx != 0) { |
| 1253 | // Need to shift the inserted vector into place. The cost is the number |
| 1254 | // of destination registers overlapped by the inserted vector. |
| 1255 | Cost = divideCeil(Numerator: EndIndex, Denominator: EltsPerReg) - (Index / EltsPerReg); |
| 1256 | } |
| 1257 | |
| 1258 | // If the last register overlap is partial, there may be three source |
| 1259 | // registers feeding into it; that takes an extra instruction. |
| 1260 | if (EndIndex < NumDstElts && BeginSubIdx < EndSubIdx) |
| 1261 | Cost += 1; |
| 1262 | |
| 1263 | return Cost; |
| 1264 | } |
| 1265 | case TTI::SK_Splice: { |
| 1266 | auto *DstVecTy = dyn_cast<FixedVectorType>(Val: DstTy); |
| 1267 | if (!DstVecTy) |
| 1268 | return InstructionCost::getInvalid(); |
| 1269 | unsigned NumElts = DstVecTy->getNumElements(); |
| 1270 | assert(NumElts == cast<FixedVectorType>(SrcTy)->getNumElements()); |
| 1271 | // Determine the sub-region of the result vector that requires |
| 1272 | // sub-register shuffles / mixing. |
| 1273 | unsigned EltsFromLHS = NumElts - Index; |
| 1274 | bool LHSIsAligned = (Index % EltsPerReg) == 0; |
| 1275 | bool RHSIsAligned = (EltsFromLHS % EltsPerReg) == 0; |
| 1276 | if (LHSIsAligned && RHSIsAligned) |
| 1277 | return 0; |
| 1278 | if (LHSIsAligned && !RHSIsAligned) |
| 1279 | return divideCeil(Numerator: NumElts, Denominator: EltsPerReg) - (EltsFromLHS / EltsPerReg); |
| 1280 | if (!LHSIsAligned && RHSIsAligned) |
| 1281 | return divideCeil(Numerator: EltsFromLHS, Denominator: EltsPerReg); |
| 1282 | return divideCeil(Numerator: NumElts, Denominator: EltsPerReg); |
| 1283 | } |
| 1284 | default: |
| 1285 | break; |
| 1286 | } |
| 1287 | |
| 1288 | if (!Mask.empty()) { |
| 1289 | unsigned NumSrcElts = cast<FixedVectorType>(Val: SrcTy)->getNumElements(); |
| 1290 | |
| 1291 | // Generically estimate the cost by assuming that each destination |
| 1292 | // register is derived from sources via v_perm_b32 instructions if it |
| 1293 | // can't be copied as-is. |
| 1294 | // |
| 1295 | // For each destination register, derive the cost of obtaining it based |
| 1296 | // on the number of source registers that feed into it. |
| 1297 | unsigned Cost = 0; |
| 1298 | for (unsigned DstIdx = 0; DstIdx < Mask.size(); DstIdx += EltsPerReg) { |
| 1299 | SmallVector<int, 4> Regs; |
| 1300 | bool Aligned = true; |
| 1301 | for (unsigned I = 0; I < EltsPerReg && DstIdx + I < Mask.size(); ++I) { |
| 1302 | int SrcIdx = Mask[DstIdx + I]; |
| 1303 | if (SrcIdx == -1) |
| 1304 | continue; |
| 1305 | int Reg; |
| 1306 | if (SrcIdx < (int)NumSrcElts) { |
| 1307 | Reg = SrcIdx / EltsPerReg; |
| 1308 | if (SrcIdx % EltsPerReg != I) |
| 1309 | Aligned = false; |
| 1310 | } else { |
| 1311 | Reg = NumSrcElts + (SrcIdx - NumSrcElts) / EltsPerReg; |
| 1312 | if ((SrcIdx - NumSrcElts) % EltsPerReg != I) |
| 1313 | Aligned = false; |
| 1314 | } |
| 1315 | if (!llvm::is_contained(Range&: Regs, Element: Reg)) |
| 1316 | Regs.push_back(Elt: Reg); |
| 1317 | } |
| 1318 | if (Regs.size() >= 2) |
| 1319 | Cost += Regs.size() - 1; |
| 1320 | else if (!Aligned) |
| 1321 | Cost += 1; |
| 1322 | } |
| 1323 | return Cost; |
| 1324 | } |
| 1325 | } |
| 1326 | |
| 1327 | return BaseT::getShuffleCost(Kind, DstTy, SrcTy, Mask, CostKind, Index, |
| 1328 | SubTp); |
| 1329 | } |
| 1330 | |
| 1331 | /// Whether it is profitable to sink the operands of an |
| 1332 | /// Instruction I to the basic block of I. |
| 1333 | /// This helps using several modifiers (like abs and neg) more often. |
| 1334 | bool GCNTTIImpl::isProfitableToSinkOperands(Instruction *I, |
| 1335 | SmallVectorImpl<Use *> &Ops) const { |
| 1336 | using namespace PatternMatch; |
| 1337 | |
| 1338 | for (auto &Op : I->operands()) { |
| 1339 | // Ensure we are not already sinking this operand. |
| 1340 | if (any_of(Range&: Ops, P: [&](Use *U) { return U->get() == Op.get(); })) |
| 1341 | continue; |
| 1342 | |
| 1343 | if (match(V: &Op, P: m_FAbs(Op0: m_Value())) || match(V: &Op, P: m_FNeg(X: m_Value()))) |
| 1344 | Ops.push_back(Elt: &Op); |
| 1345 | } |
| 1346 | |
| 1347 | return !Ops.empty(); |
| 1348 | } |
| 1349 | |
| 1350 | bool GCNTTIImpl::areInlineCompatible(const Function *Caller, |
| 1351 | const Function *Callee) const { |
| 1352 | const TargetMachine &TM = getTLI()->getTargetMachine(); |
| 1353 | const GCNSubtarget *CallerST |
| 1354 | = static_cast<const GCNSubtarget *>(TM.getSubtargetImpl(*Caller)); |
| 1355 | const GCNSubtarget *CalleeST |
| 1356 | = static_cast<const GCNSubtarget *>(TM.getSubtargetImpl(*Callee)); |
| 1357 | |
| 1358 | const FeatureBitset &CallerBits = CallerST->getFeatureBits(); |
| 1359 | const FeatureBitset &CalleeBits = CalleeST->getFeatureBits(); |
| 1360 | |
| 1361 | FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList; |
| 1362 | FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList; |
| 1363 | if ((RealCallerBits & RealCalleeBits) != RealCalleeBits) |
| 1364 | return false; |
| 1365 | |
| 1366 | // FIXME: dx10_clamp can just take the caller setting, but there seems to be |
| 1367 | // no way to support merge for backend defined attributes. |
| 1368 | SIModeRegisterDefaults CallerMode(*Caller, *CallerST); |
| 1369 | SIModeRegisterDefaults CalleeMode(*Callee, *CalleeST); |
| 1370 | if (!CallerMode.isInlineCompatible(CalleeMode)) |
| 1371 | return false; |
| 1372 | |
| 1373 | if (Callee->hasFnAttribute(Kind: Attribute::AlwaysInline) || |
| 1374 | Callee->hasFnAttribute(Kind: Attribute::InlineHint)) |
| 1375 | return true; |
| 1376 | |
| 1377 | // Hack to make compile times reasonable. |
| 1378 | if (InlineMaxBB) { |
| 1379 | // Single BB does not increase total BB amount. |
| 1380 | if (Callee->size() == 1) |
| 1381 | return true; |
| 1382 | size_t BBSize = Caller->size() + Callee->size() - 1; |
| 1383 | return BBSize <= InlineMaxBB; |
| 1384 | } |
| 1385 | |
| 1386 | return true; |
| 1387 | } |
| 1388 | |
| 1389 | static unsigned adjustInliningThresholdUsingCallee(const CallBase *CB, |
| 1390 | const SITargetLowering *TLI, |
| 1391 | const GCNTTIImpl *TTIImpl) { |
| 1392 | const int NrOfSGPRUntilSpill = 26; |
| 1393 | const int NrOfVGPRUntilSpill = 32; |
| 1394 | |
| 1395 | const DataLayout &DL = TTIImpl->getDataLayout(); |
| 1396 | |
| 1397 | unsigned adjustThreshold = 0; |
| 1398 | int SGPRsInUse = 0; |
| 1399 | int VGPRsInUse = 0; |
| 1400 | for (const Use &A : CB->args()) { |
| 1401 | SmallVector<EVT, 4> ValueVTs; |
| 1402 | ComputeValueVTs(TLI: *TLI, DL, Ty: A.get()->getType(), ValueVTs); |
| 1403 | for (auto ArgVT : ValueVTs) { |
| 1404 | unsigned CCRegNum = TLI->getNumRegistersForCallingConv( |
| 1405 | Context&: CB->getContext(), CC: CB->getCallingConv(), VT: ArgVT); |
| 1406 | if (AMDGPU::isArgPassedInSGPR(CB, ArgNo: CB->getArgOperandNo(U: &A))) |
| 1407 | SGPRsInUse += CCRegNum; |
| 1408 | else |
| 1409 | VGPRsInUse += CCRegNum; |
| 1410 | } |
| 1411 | } |
| 1412 | |
| 1413 | // The cost of passing function arguments through the stack: |
| 1414 | // 1 instruction to put a function argument on the stack in the caller. |
| 1415 | // 1 instruction to take a function argument from the stack in callee. |
| 1416 | // 1 instruction is explicitly take care of data dependencies in callee |
| 1417 | // function. |
| 1418 | InstructionCost ArgStackCost(1); |
| 1419 | ArgStackCost += const_cast<GCNTTIImpl *>(TTIImpl)->getMemoryOpCost( |
| 1420 | Opcode: Instruction::Store, Src: Type::getInt32Ty(C&: CB->getContext()), Alignment: Align(4), |
| 1421 | AddressSpace: AMDGPUAS::PRIVATE_ADDRESS, CostKind: TTI::TCK_SizeAndLatency); |
| 1422 | ArgStackCost += const_cast<GCNTTIImpl *>(TTIImpl)->getMemoryOpCost( |
| 1423 | Opcode: Instruction::Load, Src: Type::getInt32Ty(C&: CB->getContext()), Alignment: Align(4), |
| 1424 | AddressSpace: AMDGPUAS::PRIVATE_ADDRESS, CostKind: TTI::TCK_SizeAndLatency); |
| 1425 | |
| 1426 | // The penalty cost is computed relative to the cost of instructions and does |
| 1427 | // not model any storage costs. |
| 1428 | adjustThreshold += std::max(a: 0, b: SGPRsInUse - NrOfSGPRUntilSpill) * |
| 1429 | ArgStackCost.getValue() * InlineConstants::getInstrCost(); |
| 1430 | adjustThreshold += std::max(a: 0, b: VGPRsInUse - NrOfVGPRUntilSpill) * |
| 1431 | ArgStackCost.getValue() * InlineConstants::getInstrCost(); |
| 1432 | return adjustThreshold; |
| 1433 | } |
| 1434 | |
| 1435 | static unsigned getCallArgsTotalAllocaSize(const CallBase *CB, |
| 1436 | const DataLayout &DL) { |
| 1437 | // If we have a pointer to a private array passed into a function |
| 1438 | // it will not be optimized out, leaving scratch usage. |
| 1439 | // This function calculates the total size in bytes of the memory that would |
| 1440 | // end in scratch if the call was not inlined. |
| 1441 | unsigned AllocaSize = 0; |
| 1442 | SmallPtrSet<const AllocaInst *, 8> AIVisited; |
| 1443 | for (Value *PtrArg : CB->args()) { |
| 1444 | PointerType *Ty = dyn_cast<PointerType>(Val: PtrArg->getType()); |
| 1445 | if (!Ty) |
| 1446 | continue; |
| 1447 | |
| 1448 | unsigned AddrSpace = Ty->getAddressSpace(); |
| 1449 | if (AddrSpace != AMDGPUAS::FLAT_ADDRESS && |
| 1450 | AddrSpace != AMDGPUAS::PRIVATE_ADDRESS) |
| 1451 | continue; |
| 1452 | |
| 1453 | const AllocaInst *AI = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: PtrArg)); |
| 1454 | if (!AI || !AI->isStaticAlloca() || !AIVisited.insert(Ptr: AI).second) |
| 1455 | continue; |
| 1456 | |
| 1457 | if (auto Size = AI->getAllocationSize(DL)) |
| 1458 | AllocaSize += Size->getFixedValue(); |
| 1459 | } |
| 1460 | return AllocaSize; |
| 1461 | } |
| 1462 | |
| 1463 | int GCNTTIImpl::getInliningLastCallToStaticBonus() const { |
| 1464 | return BaseT::getInliningLastCallToStaticBonus() * |
| 1465 | getInliningThresholdMultiplier(); |
| 1466 | } |
| 1467 | |
| 1468 | unsigned GCNTTIImpl::adjustInliningThreshold(const CallBase *CB) const { |
| 1469 | unsigned Threshold = adjustInliningThresholdUsingCallee(CB, TLI, TTIImpl: this); |
| 1470 | |
| 1471 | // Private object passed as arguments may end up in scratch usage if the call |
| 1472 | // is not inlined. Increase the inline threshold to promote inlining. |
| 1473 | unsigned AllocaSize = getCallArgsTotalAllocaSize(CB, DL); |
| 1474 | if (AllocaSize > 0) |
| 1475 | Threshold += ArgAllocaCost; |
| 1476 | return Threshold; |
| 1477 | } |
| 1478 | |
| 1479 | unsigned GCNTTIImpl::getCallerAllocaCost(const CallBase *CB, |
| 1480 | const AllocaInst *AI) const { |
| 1481 | |
| 1482 | // Below the cutoff, assume that the private memory objects would be |
| 1483 | // optimized |
| 1484 | auto AllocaSize = getCallArgsTotalAllocaSize(CB, DL); |
| 1485 | if (AllocaSize <= ArgAllocaCutoff) |
| 1486 | return 0; |
| 1487 | |
| 1488 | // Above the cutoff, we give a cost to each private memory object |
| 1489 | // depending its size. If the array can be optimized by SROA this cost is not |
| 1490 | // added to the total-cost in the inliner cost analysis. |
| 1491 | // |
| 1492 | // We choose the total cost of the alloca such that their sum cancels the |
| 1493 | // bonus given in the threshold (ArgAllocaCost). |
| 1494 | // |
| 1495 | // Cost_Alloca_0 + ... + Cost_Alloca_N == ArgAllocaCost |
| 1496 | // |
| 1497 | // Awkwardly, the ArgAllocaCost bonus is multiplied by threshold-multiplier, |
| 1498 | // the single-bb bonus and the vector-bonus. |
| 1499 | // |
| 1500 | // We compensate the first two multipliers, by repeating logic from the |
| 1501 | // inliner-cost in here. The vector-bonus is 0 on AMDGPU. |
| 1502 | static_assert(InlinerVectorBonusPercent == 0, "vector bonus assumed to be 0" ); |
| 1503 | unsigned Threshold = ArgAllocaCost * getInliningThresholdMultiplier(); |
| 1504 | |
| 1505 | bool SingleBB = none_of(Range&: *CB->getCalledFunction(), P: [](const BasicBlock &BB) { |
| 1506 | return BB.getTerminator()->getNumSuccessors() > 1; |
| 1507 | }); |
| 1508 | if (SingleBB) { |
| 1509 | Threshold += Threshold / 2; |
| 1510 | } |
| 1511 | |
| 1512 | auto ArgAllocaSize = AI->getAllocationSize(DL); |
| 1513 | if (!ArgAllocaSize) |
| 1514 | return 0; |
| 1515 | |
| 1516 | // Attribute the bonus proportionally to the alloca size |
| 1517 | unsigned AllocaThresholdBonus = |
| 1518 | (Threshold * ArgAllocaSize->getFixedValue()) / AllocaSize; |
| 1519 | |
| 1520 | return AllocaThresholdBonus; |
| 1521 | } |
| 1522 | |
| 1523 | void GCNTTIImpl::(Loop *L, ScalarEvolution &SE, |
| 1524 | TTI::UnrollingPreferences &UP, |
| 1525 | OptimizationRemarkEmitter *ORE) const { |
| 1526 | CommonTTI.getUnrollingPreferences(L, SE, UP, ORE); |
| 1527 | } |
| 1528 | |
| 1529 | void GCNTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE, |
| 1530 | TTI::PeelingPreferences &PP) const { |
| 1531 | CommonTTI.getPeelingPreferences(L, SE, PP); |
| 1532 | } |
| 1533 | |
| 1534 | int GCNTTIImpl::get64BitInstrCost(TTI::TargetCostKind CostKind) const { |
| 1535 | return ST->hasFullRate64Ops() |
| 1536 | ? getFullRateInstrCost() |
| 1537 | : ST->hasHalfRate64Ops() ? getHalfRateInstrCost(CostKind) |
| 1538 | : getQuarterRateInstrCost(CostKind); |
| 1539 | } |
| 1540 | |
| 1541 | std::pair<InstructionCost, MVT> |
| 1542 | GCNTTIImpl::getTypeLegalizationCost(Type *Ty) const { |
| 1543 | std::pair<InstructionCost, MVT> Cost = BaseT::getTypeLegalizationCost(Ty); |
| 1544 | auto Size = DL.getTypeSizeInBits(Ty); |
| 1545 | // Maximum load or store can handle 8 dwords for scalar and 4 for |
| 1546 | // vector ALU. Let's assume anything above 8 dwords is expensive |
| 1547 | // even if legal. |
| 1548 | if (Size <= 256) |
| 1549 | return Cost; |
| 1550 | |
| 1551 | Cost.first += (Size + 255) / 256; |
| 1552 | return Cost; |
| 1553 | } |
| 1554 | |
| 1555 | unsigned GCNTTIImpl::getPrefetchDistance() const { |
| 1556 | return ST->hasPrefetch() ? 128 : 0; |
| 1557 | } |
| 1558 | |
| 1559 | bool GCNTTIImpl::shouldPrefetchAddressSpace(unsigned AS) const { |
| 1560 | return AMDGPU::isFlatGlobalAddrSpace(AS); |
| 1561 | } |
| 1562 | |
| 1563 | void GCNTTIImpl::collectKernelLaunchBounds( |
| 1564 | const Function &F, |
| 1565 | SmallVectorImpl<std::pair<StringRef, int64_t>> &LB) const { |
| 1566 | SmallVector<unsigned> MaxNumWorkgroups = ST->getMaxNumWorkGroups(F); |
| 1567 | LB.push_back(Elt: {"amdgpu-max-num-workgroups[0]" , MaxNumWorkgroups[0]}); |
| 1568 | LB.push_back(Elt: {"amdgpu-max-num-workgroups[1]" , MaxNumWorkgroups[1]}); |
| 1569 | LB.push_back(Elt: {"amdgpu-max-num-workgroups[2]" , MaxNumWorkgroups[2]}); |
| 1570 | std::pair<unsigned, unsigned> FlatWorkGroupSize = |
| 1571 | ST->getFlatWorkGroupSizes(F); |
| 1572 | LB.push_back(Elt: {"amdgpu-flat-work-group-size[0]" , FlatWorkGroupSize.first}); |
| 1573 | LB.push_back(Elt: {"amdgpu-flat-work-group-size[1]" , FlatWorkGroupSize.second}); |
| 1574 | std::pair<unsigned, unsigned> WavesPerEU = ST->getWavesPerEU(F); |
| 1575 | LB.push_back(Elt: {"amdgpu-waves-per-eu[0]" , WavesPerEU.first}); |
| 1576 | LB.push_back(Elt: {"amdgpu-waves-per-eu[1]" , WavesPerEU.second}); |
| 1577 | } |
| 1578 | |
| 1579 | GCNTTIImpl::KnownIEEEMode |
| 1580 | GCNTTIImpl::fpenvIEEEMode(const Instruction &I) const { |
| 1581 | if (!ST->hasIEEEMode()) // Only mode on gfx12 |
| 1582 | return KnownIEEEMode::On; |
| 1583 | |
| 1584 | const Function *F = I.getFunction(); |
| 1585 | if (!F) |
| 1586 | return KnownIEEEMode::Unknown; |
| 1587 | |
| 1588 | Attribute IEEEAttr = F->getFnAttribute(Kind: "amdgpu-ieee" ); |
| 1589 | if (IEEEAttr.isValid()) |
| 1590 | return IEEEAttr.getValueAsBool() ? KnownIEEEMode::On : KnownIEEEMode::Off; |
| 1591 | |
| 1592 | return AMDGPU::isShader(CC: F->getCallingConv()) ? KnownIEEEMode::Off |
| 1593 | : KnownIEEEMode::On; |
| 1594 | } |
| 1595 | |
| 1596 | InstructionCost GCNTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, |
| 1597 | Align Alignment, |
| 1598 | unsigned AddressSpace, |
| 1599 | TTI::TargetCostKind CostKind, |
| 1600 | TTI::OperandValueInfo OpInfo, |
| 1601 | const Instruction *I) const { |
| 1602 | if (VectorType *VecTy = dyn_cast<VectorType>(Val: Src)) { |
| 1603 | if ((Opcode == Instruction::Load || Opcode == Instruction::Store) && |
| 1604 | VecTy->getElementType()->isIntegerTy(Bitwidth: 8)) { |
| 1605 | return divideCeil(Numerator: DL.getTypeSizeInBits(Ty: VecTy) - 1, |
| 1606 | Denominator: getLoadStoreVecRegBitWidth(AddrSpace: AddressSpace)); |
| 1607 | } |
| 1608 | } |
| 1609 | return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind, |
| 1610 | OpInfo, I); |
| 1611 | } |
| 1612 | |
| 1613 | unsigned GCNTTIImpl::getNumberOfParts(Type *Tp) const { |
| 1614 | if (VectorType *VecTy = dyn_cast<VectorType>(Val: Tp)) { |
| 1615 | if (VecTy->getElementType()->isIntegerTy(Bitwidth: 8)) { |
| 1616 | unsigned ElementCount = VecTy->getElementCount().getFixedValue(); |
| 1617 | return divideCeil(Numerator: ElementCount - 1, Denominator: 4); |
| 1618 | } |
| 1619 | } |
| 1620 | return BaseT::getNumberOfParts(Tp); |
| 1621 | } |
| 1622 | |
| 1623 | InstructionUniformity |
| 1624 | GCNTTIImpl::getInstructionUniformity(const Value *V) const { |
| 1625 | if (isAlwaysUniform(V)) |
| 1626 | return InstructionUniformity::AlwaysUniform; |
| 1627 | |
| 1628 | if (isSourceOfDivergence(V)) |
| 1629 | return InstructionUniformity::NeverUniform; |
| 1630 | |
| 1631 | return InstructionUniformity::Default; |
| 1632 | } |
| 1633 | |