1 //===-- PPCTargetTransformInfo.cpp - PPC specific TTI ---------------------===// 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 "PPCTargetTransformInfo.h" 10 #include "llvm/Analysis/CodeMetrics.h" 11 #include "llvm/Analysis/TargetLibraryInfo.h" 12 #include "llvm/Analysis/TargetTransformInfo.h" 13 #include "llvm/CodeGen/BasicTTIImpl.h" 14 #include "llvm/CodeGen/CostTable.h" 15 #include "llvm/CodeGen/TargetLowering.h" 16 #include "llvm/CodeGen/TargetSchedule.h" 17 #include "llvm/IR/IntrinsicsPowerPC.h" 18 #include "llvm/Support/CommandLine.h" 19 #include "llvm/Support/Debug.h" 20 #include "llvm/Transforms/InstCombine/InstCombiner.h" 21 #include "llvm/Transforms/Utils/Local.h" 22 23 using namespace llvm; 24 25 #define DEBUG_TYPE "ppctti" 26 27 static cl::opt<bool> DisablePPCConstHoist("disable-ppc-constant-hoisting", 28 cl::desc("disable constant hoisting on PPC"), cl::init(false), cl::Hidden); 29 30 // This is currently only used for the data prefetch pass 31 static cl::opt<unsigned> 32 CacheLineSize("ppc-loop-prefetch-cache-line", cl::Hidden, cl::init(64), 33 cl::desc("The loop prefetch cache line size")); 34 35 static cl::opt<bool> 36 EnablePPCColdCC("ppc-enable-coldcc", cl::Hidden, cl::init(false), 37 cl::desc("Enable using coldcc calling conv for cold " 38 "internal functions")); 39 40 static cl::opt<bool> 41 LsrNoInsnsCost("ppc-lsr-no-insns-cost", cl::Hidden, cl::init(false), 42 cl::desc("Do not add instruction count to lsr cost model")); 43 44 // The latency of mtctr is only justified if there are more than 4 45 // comparisons that will be removed as a result. 46 static cl::opt<unsigned> 47 SmallCTRLoopThreshold("min-ctr-loop-threshold", cl::init(4), cl::Hidden, 48 cl::desc("Loops with a constant trip count smaller than " 49 "this value will not use the count register.")); 50 51 //===----------------------------------------------------------------------===// 52 // 53 // PPC cost model. 54 // 55 //===----------------------------------------------------------------------===// 56 57 TargetTransformInfo::PopcntSupportKind 58 PPCTTIImpl::getPopcntSupport(unsigned TyWidth) { 59 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 60 if (ST->hasPOPCNTD() != PPCSubtarget::POPCNTD_Unavailable && TyWidth <= 64) 61 return ST->hasPOPCNTD() == PPCSubtarget::POPCNTD_Slow ? 62 TTI::PSK_SlowHardware : TTI::PSK_FastHardware; 63 return TTI::PSK_Software; 64 } 65 66 Optional<Instruction *> 67 PPCTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const { 68 Intrinsic::ID IID = II.getIntrinsicID(); 69 switch (IID) { 70 default: 71 break; 72 case Intrinsic::ppc_altivec_lvx: 73 case Intrinsic::ppc_altivec_lvxl: 74 // Turn PPC lvx -> load if the pointer is known aligned. 75 if (getOrEnforceKnownAlignment( 76 II.getArgOperand(0), Align(16), IC.getDataLayout(), &II, 77 &IC.getAssumptionCache(), &IC.getDominatorTree()) >= 16) { 78 Value *Ptr = IC.Builder.CreateBitCast( 79 II.getArgOperand(0), PointerType::getUnqual(II.getType())); 80 return new LoadInst(II.getType(), Ptr, "", false, Align(16)); 81 } 82 break; 83 case Intrinsic::ppc_vsx_lxvw4x: 84 case Intrinsic::ppc_vsx_lxvd2x: { 85 // Turn PPC VSX loads into normal loads. 86 Value *Ptr = IC.Builder.CreateBitCast(II.getArgOperand(0), 87 PointerType::getUnqual(II.getType())); 88 return new LoadInst(II.getType(), Ptr, Twine(""), false, Align(1)); 89 } 90 case Intrinsic::ppc_altivec_stvx: 91 case Intrinsic::ppc_altivec_stvxl: 92 // Turn stvx -> store if the pointer is known aligned. 93 if (getOrEnforceKnownAlignment( 94 II.getArgOperand(1), Align(16), IC.getDataLayout(), &II, 95 &IC.getAssumptionCache(), &IC.getDominatorTree()) >= 16) { 96 Type *OpPtrTy = PointerType::getUnqual(II.getArgOperand(0)->getType()); 97 Value *Ptr = IC.Builder.CreateBitCast(II.getArgOperand(1), OpPtrTy); 98 return new StoreInst(II.getArgOperand(0), Ptr, false, Align(16)); 99 } 100 break; 101 case Intrinsic::ppc_vsx_stxvw4x: 102 case Intrinsic::ppc_vsx_stxvd2x: { 103 // Turn PPC VSX stores into normal stores. 104 Type *OpPtrTy = PointerType::getUnqual(II.getArgOperand(0)->getType()); 105 Value *Ptr = IC.Builder.CreateBitCast(II.getArgOperand(1), OpPtrTy); 106 return new StoreInst(II.getArgOperand(0), Ptr, false, Align(1)); 107 } 108 case Intrinsic::ppc_altivec_vperm: 109 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 110 // Note that ppc_altivec_vperm has a big-endian bias, so when creating 111 // a vectorshuffle for little endian, we must undo the transformation 112 // performed on vec_perm in altivec.h. That is, we must complement 113 // the permutation mask with respect to 31 and reverse the order of 114 // V1 and V2. 115 if (Constant *Mask = dyn_cast<Constant>(II.getArgOperand(2))) { 116 assert(cast<FixedVectorType>(Mask->getType())->getNumElements() == 16 && 117 "Bad type for intrinsic!"); 118 119 // Check that all of the elements are integer constants or undefs. 120 bool AllEltsOk = true; 121 for (unsigned i = 0; i != 16; ++i) { 122 Constant *Elt = Mask->getAggregateElement(i); 123 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 124 AllEltsOk = false; 125 break; 126 } 127 } 128 129 if (AllEltsOk) { 130 // Cast the input vectors to byte vectors. 131 Value *Op0 = 132 IC.Builder.CreateBitCast(II.getArgOperand(0), Mask->getType()); 133 Value *Op1 = 134 IC.Builder.CreateBitCast(II.getArgOperand(1), Mask->getType()); 135 Value *Result = UndefValue::get(Op0->getType()); 136 137 // Only extract each element once. 138 Value *ExtractedElts[32]; 139 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 140 141 for (unsigned i = 0; i != 16; ++i) { 142 if (isa<UndefValue>(Mask->getAggregateElement(i))) 143 continue; 144 unsigned Idx = 145 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 146 Idx &= 31; // Match the hardware behavior. 147 if (DL.isLittleEndian()) 148 Idx = 31 - Idx; 149 150 if (!ExtractedElts[Idx]) { 151 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0; 152 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1; 153 ExtractedElts[Idx] = IC.Builder.CreateExtractElement( 154 Idx < 16 ? Op0ToUse : Op1ToUse, IC.Builder.getInt32(Idx & 15)); 155 } 156 157 // Insert this value into the result vector. 158 Result = IC.Builder.CreateInsertElement(Result, ExtractedElts[Idx], 159 IC.Builder.getInt32(i)); 160 } 161 return CastInst::Create(Instruction::BitCast, Result, II.getType()); 162 } 163 } 164 break; 165 } 166 return None; 167 } 168 169 int PPCTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty, 170 TTI::TargetCostKind CostKind) { 171 if (DisablePPCConstHoist) 172 return BaseT::getIntImmCost(Imm, Ty, CostKind); 173 174 assert(Ty->isIntegerTy()); 175 176 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 177 if (BitSize == 0) 178 return ~0U; 179 180 if (Imm == 0) 181 return TTI::TCC_Free; 182 183 if (Imm.getBitWidth() <= 64) { 184 if (isInt<16>(Imm.getSExtValue())) 185 return TTI::TCC_Basic; 186 187 if (isInt<32>(Imm.getSExtValue())) { 188 // A constant that can be materialized using lis. 189 if ((Imm.getZExtValue() & 0xFFFF) == 0) 190 return TTI::TCC_Basic; 191 192 return 2 * TTI::TCC_Basic; 193 } 194 } 195 196 return 4 * TTI::TCC_Basic; 197 } 198 199 int PPCTTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 200 const APInt &Imm, Type *Ty, 201 TTI::TargetCostKind CostKind) { 202 if (DisablePPCConstHoist) 203 return BaseT::getIntImmCostIntrin(IID, Idx, Imm, Ty, CostKind); 204 205 assert(Ty->isIntegerTy()); 206 207 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 208 if (BitSize == 0) 209 return ~0U; 210 211 switch (IID) { 212 default: 213 return TTI::TCC_Free; 214 case Intrinsic::sadd_with_overflow: 215 case Intrinsic::uadd_with_overflow: 216 case Intrinsic::ssub_with_overflow: 217 case Intrinsic::usub_with_overflow: 218 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<16>(Imm.getSExtValue())) 219 return TTI::TCC_Free; 220 break; 221 case Intrinsic::experimental_stackmap: 222 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 223 return TTI::TCC_Free; 224 break; 225 case Intrinsic::experimental_patchpoint_void: 226 case Intrinsic::experimental_patchpoint_i64: 227 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))) 228 return TTI::TCC_Free; 229 break; 230 } 231 return PPCTTIImpl::getIntImmCost(Imm, Ty, CostKind); 232 } 233 234 int PPCTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx, 235 const APInt &Imm, Type *Ty, 236 TTI::TargetCostKind CostKind) { 237 if (DisablePPCConstHoist) 238 return BaseT::getIntImmCostInst(Opcode, Idx, Imm, Ty, CostKind); 239 240 assert(Ty->isIntegerTy()); 241 242 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 243 if (BitSize == 0) 244 return ~0U; 245 246 unsigned ImmIdx = ~0U; 247 bool ShiftedFree = false, RunFree = false, UnsignedFree = false, 248 ZeroFree = false; 249 switch (Opcode) { 250 default: 251 return TTI::TCC_Free; 252 case Instruction::GetElementPtr: 253 // Always hoist the base address of a GetElementPtr. This prevents the 254 // creation of new constants for every base constant that gets constant 255 // folded with the offset. 256 if (Idx == 0) 257 return 2 * TTI::TCC_Basic; 258 return TTI::TCC_Free; 259 case Instruction::And: 260 RunFree = true; // (for the rotate-and-mask instructions) 261 LLVM_FALLTHROUGH; 262 case Instruction::Add: 263 case Instruction::Or: 264 case Instruction::Xor: 265 ShiftedFree = true; 266 LLVM_FALLTHROUGH; 267 case Instruction::Sub: 268 case Instruction::Mul: 269 case Instruction::Shl: 270 case Instruction::LShr: 271 case Instruction::AShr: 272 ImmIdx = 1; 273 break; 274 case Instruction::ICmp: 275 UnsignedFree = true; 276 ImmIdx = 1; 277 // Zero comparisons can use record-form instructions. 278 LLVM_FALLTHROUGH; 279 case Instruction::Select: 280 ZeroFree = true; 281 break; 282 case Instruction::PHI: 283 case Instruction::Call: 284 case Instruction::Ret: 285 case Instruction::Load: 286 case Instruction::Store: 287 break; 288 } 289 290 if (ZeroFree && Imm == 0) 291 return TTI::TCC_Free; 292 293 if (Idx == ImmIdx && Imm.getBitWidth() <= 64) { 294 if (isInt<16>(Imm.getSExtValue())) 295 return TTI::TCC_Free; 296 297 if (RunFree) { 298 if (Imm.getBitWidth() <= 32 && 299 (isShiftedMask_32(Imm.getZExtValue()) || 300 isShiftedMask_32(~Imm.getZExtValue()))) 301 return TTI::TCC_Free; 302 303 if (ST->isPPC64() && 304 (isShiftedMask_64(Imm.getZExtValue()) || 305 isShiftedMask_64(~Imm.getZExtValue()))) 306 return TTI::TCC_Free; 307 } 308 309 if (UnsignedFree && isUInt<16>(Imm.getZExtValue())) 310 return TTI::TCC_Free; 311 312 if (ShiftedFree && (Imm.getZExtValue() & 0xFFFF) == 0) 313 return TTI::TCC_Free; 314 } 315 316 return PPCTTIImpl::getIntImmCost(Imm, Ty, CostKind); 317 } 318 319 unsigned 320 PPCTTIImpl::getUserCost(const User *U, ArrayRef<const Value *> Operands, 321 TTI::TargetCostKind CostKind) { 322 // We already implement getCastInstrCost and getMemoryOpCost where we perform 323 // the vector adjustment there. 324 if (isa<CastInst>(U) || isa<LoadInst>(U) || isa<StoreInst>(U)) 325 return BaseT::getUserCost(U, Operands, CostKind); 326 327 if (U->getType()->isVectorTy()) { 328 // Instructions that need to be split should cost more. 329 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, U->getType()); 330 return LT.first * BaseT::getUserCost(U, Operands, CostKind); 331 } 332 333 return BaseT::getUserCost(U, Operands, CostKind); 334 } 335 336 bool PPCTTIImpl::mightUseCTR(BasicBlock *BB, TargetLibraryInfo *LibInfo, 337 SmallPtrSetImpl<const Value *> &Visited) { 338 const PPCTargetMachine &TM = ST->getTargetMachine(); 339 340 // Loop through the inline asm constraints and look for something that 341 // clobbers ctr. 342 auto asmClobbersCTR = [](InlineAsm *IA) { 343 InlineAsm::ConstraintInfoVector CIV = IA->ParseConstraints(); 344 for (unsigned i = 0, ie = CIV.size(); i < ie; ++i) { 345 InlineAsm::ConstraintInfo &C = CIV[i]; 346 if (C.Type != InlineAsm::isInput) 347 for (unsigned j = 0, je = C.Codes.size(); j < je; ++j) 348 if (StringRef(C.Codes[j]).equals_lower("{ctr}")) 349 return true; 350 } 351 return false; 352 }; 353 354 // Determining the address of a TLS variable results in a function call in 355 // certain TLS models. 356 std::function<bool(const Value *)> memAddrUsesCTR = 357 [&memAddrUsesCTR, &TM, &Visited](const Value *MemAddr) -> bool { 358 // No need to traverse again if we already checked this operand. 359 if (!Visited.insert(MemAddr).second) 360 return false; 361 const auto *GV = dyn_cast<GlobalValue>(MemAddr); 362 if (!GV) { 363 // Recurse to check for constants that refer to TLS global variables. 364 if (const auto *CV = dyn_cast<Constant>(MemAddr)) 365 for (const auto &CO : CV->operands()) 366 if (memAddrUsesCTR(CO)) 367 return true; 368 369 return false; 370 } 371 372 if (!GV->isThreadLocal()) 373 return false; 374 TLSModel::Model Model = TM.getTLSModel(GV); 375 return Model == TLSModel::GeneralDynamic || 376 Model == TLSModel::LocalDynamic; 377 }; 378 379 auto isLargeIntegerTy = [](bool Is32Bit, Type *Ty) { 380 if (IntegerType *ITy = dyn_cast<IntegerType>(Ty)) 381 return ITy->getBitWidth() > (Is32Bit ? 32U : 64U); 382 383 return false; 384 }; 385 386 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); 387 J != JE; ++J) { 388 if (CallInst *CI = dyn_cast<CallInst>(J)) { 389 // Inline ASM is okay, unless it clobbers the ctr register. 390 if (InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledOperand())) { 391 if (asmClobbersCTR(IA)) 392 return true; 393 continue; 394 } 395 396 if (Function *F = CI->getCalledFunction()) { 397 // Most intrinsics don't become function calls, but some might. 398 // sin, cos, exp and log are always calls. 399 unsigned Opcode = 0; 400 if (F->getIntrinsicID() != Intrinsic::not_intrinsic) { 401 switch (F->getIntrinsicID()) { 402 default: continue; 403 // If we have a call to loop_decrement or set_loop_iterations, 404 // we're definitely using CTR. 405 case Intrinsic::set_loop_iterations: 406 case Intrinsic::loop_decrement: 407 return true; 408 409 // Binary operations on 128-bit value will use CTR. 410 case Intrinsic::experimental_constrained_fadd: 411 case Intrinsic::experimental_constrained_fsub: 412 case Intrinsic::experimental_constrained_fmul: 413 case Intrinsic::experimental_constrained_fdiv: 414 case Intrinsic::experimental_constrained_frem: 415 if (F->getType()->getScalarType()->isFP128Ty() || 416 F->getType()->getScalarType()->isPPC_FP128Ty()) 417 return true; 418 break; 419 420 case Intrinsic::experimental_constrained_fptosi: 421 case Intrinsic::experimental_constrained_fptoui: 422 case Intrinsic::experimental_constrained_sitofp: 423 case Intrinsic::experimental_constrained_uitofp: { 424 Type *SrcType = CI->getArgOperand(0)->getType()->getScalarType(); 425 Type *DstType = CI->getType()->getScalarType(); 426 if (SrcType->isPPC_FP128Ty() || DstType->isPPC_FP128Ty() || 427 isLargeIntegerTy(!TM.isPPC64(), SrcType) || 428 isLargeIntegerTy(!TM.isPPC64(), DstType)) 429 return true; 430 break; 431 } 432 433 // Exclude eh_sjlj_setjmp; we don't need to exclude eh_sjlj_longjmp 434 // because, although it does clobber the counter register, the 435 // control can't then return to inside the loop unless there is also 436 // an eh_sjlj_setjmp. 437 case Intrinsic::eh_sjlj_setjmp: 438 439 case Intrinsic::memcpy: 440 case Intrinsic::memmove: 441 case Intrinsic::memset: 442 case Intrinsic::powi: 443 case Intrinsic::log: 444 case Intrinsic::log2: 445 case Intrinsic::log10: 446 case Intrinsic::exp: 447 case Intrinsic::exp2: 448 case Intrinsic::pow: 449 case Intrinsic::sin: 450 case Intrinsic::cos: 451 case Intrinsic::experimental_constrained_powi: 452 case Intrinsic::experimental_constrained_log: 453 case Intrinsic::experimental_constrained_log2: 454 case Intrinsic::experimental_constrained_log10: 455 case Intrinsic::experimental_constrained_exp: 456 case Intrinsic::experimental_constrained_exp2: 457 case Intrinsic::experimental_constrained_pow: 458 case Intrinsic::experimental_constrained_sin: 459 case Intrinsic::experimental_constrained_cos: 460 return true; 461 case Intrinsic::copysign: 462 if (CI->getArgOperand(0)->getType()->getScalarType()-> 463 isPPC_FP128Ty()) 464 return true; 465 else 466 continue; // ISD::FCOPYSIGN is never a library call. 467 case Intrinsic::fma: Opcode = ISD::FMA; break; 468 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break; 469 case Intrinsic::floor: Opcode = ISD::FFLOOR; break; 470 case Intrinsic::ceil: Opcode = ISD::FCEIL; break; 471 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break; 472 case Intrinsic::rint: Opcode = ISD::FRINT; break; 473 case Intrinsic::lrint: Opcode = ISD::LRINT; break; 474 case Intrinsic::llrint: Opcode = ISD::LLRINT; break; 475 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break; 476 case Intrinsic::round: Opcode = ISD::FROUND; break; 477 case Intrinsic::lround: Opcode = ISD::LROUND; break; 478 case Intrinsic::llround: Opcode = ISD::LLROUND; break; 479 case Intrinsic::minnum: Opcode = ISD::FMINNUM; break; 480 case Intrinsic::maxnum: Opcode = ISD::FMAXNUM; break; 481 case Intrinsic::experimental_constrained_fcmp: 482 Opcode = ISD::STRICT_FSETCC; 483 break; 484 case Intrinsic::experimental_constrained_fcmps: 485 Opcode = ISD::STRICT_FSETCCS; 486 break; 487 case Intrinsic::experimental_constrained_fma: 488 Opcode = ISD::STRICT_FMA; 489 break; 490 case Intrinsic::experimental_constrained_sqrt: 491 Opcode = ISD::STRICT_FSQRT; 492 break; 493 case Intrinsic::experimental_constrained_floor: 494 Opcode = ISD::STRICT_FFLOOR; 495 break; 496 case Intrinsic::experimental_constrained_ceil: 497 Opcode = ISD::STRICT_FCEIL; 498 break; 499 case Intrinsic::experimental_constrained_trunc: 500 Opcode = ISD::STRICT_FTRUNC; 501 break; 502 case Intrinsic::experimental_constrained_rint: 503 Opcode = ISD::STRICT_FRINT; 504 break; 505 case Intrinsic::experimental_constrained_lrint: 506 Opcode = ISD::STRICT_LRINT; 507 break; 508 case Intrinsic::experimental_constrained_llrint: 509 Opcode = ISD::STRICT_LLRINT; 510 break; 511 case Intrinsic::experimental_constrained_nearbyint: 512 Opcode = ISD::STRICT_FNEARBYINT; 513 break; 514 case Intrinsic::experimental_constrained_round: 515 Opcode = ISD::STRICT_FROUND; 516 break; 517 case Intrinsic::experimental_constrained_lround: 518 Opcode = ISD::STRICT_LROUND; 519 break; 520 case Intrinsic::experimental_constrained_llround: 521 Opcode = ISD::STRICT_LLROUND; 522 break; 523 case Intrinsic::experimental_constrained_minnum: 524 Opcode = ISD::STRICT_FMINNUM; 525 break; 526 case Intrinsic::experimental_constrained_maxnum: 527 Opcode = ISD::STRICT_FMAXNUM; 528 break; 529 case Intrinsic::umul_with_overflow: Opcode = ISD::UMULO; break; 530 case Intrinsic::smul_with_overflow: Opcode = ISD::SMULO; break; 531 } 532 } 533 534 // PowerPC does not use [US]DIVREM or other library calls for 535 // operations on regular types which are not otherwise library calls 536 // (i.e. soft float or atomics). If adapting for targets that do, 537 // additional care is required here. 538 539 LibFunc Func; 540 if (!F->hasLocalLinkage() && F->hasName() && LibInfo && 541 LibInfo->getLibFunc(F->getName(), Func) && 542 LibInfo->hasOptimizedCodeGen(Func)) { 543 // Non-read-only functions are never treated as intrinsics. 544 if (!CI->onlyReadsMemory()) 545 return true; 546 547 // Conversion happens only for FP calls. 548 if (!CI->getArgOperand(0)->getType()->isFloatingPointTy()) 549 return true; 550 551 switch (Func) { 552 default: return true; 553 case LibFunc_copysign: 554 case LibFunc_copysignf: 555 continue; // ISD::FCOPYSIGN is never a library call. 556 case LibFunc_copysignl: 557 return true; 558 case LibFunc_fabs: 559 case LibFunc_fabsf: 560 case LibFunc_fabsl: 561 continue; // ISD::FABS is never a library call. 562 case LibFunc_sqrt: 563 case LibFunc_sqrtf: 564 case LibFunc_sqrtl: 565 Opcode = ISD::FSQRT; break; 566 case LibFunc_floor: 567 case LibFunc_floorf: 568 case LibFunc_floorl: 569 Opcode = ISD::FFLOOR; break; 570 case LibFunc_nearbyint: 571 case LibFunc_nearbyintf: 572 case LibFunc_nearbyintl: 573 Opcode = ISD::FNEARBYINT; break; 574 case LibFunc_ceil: 575 case LibFunc_ceilf: 576 case LibFunc_ceill: 577 Opcode = ISD::FCEIL; break; 578 case LibFunc_rint: 579 case LibFunc_rintf: 580 case LibFunc_rintl: 581 Opcode = ISD::FRINT; break; 582 case LibFunc_round: 583 case LibFunc_roundf: 584 case LibFunc_roundl: 585 Opcode = ISD::FROUND; break; 586 case LibFunc_trunc: 587 case LibFunc_truncf: 588 case LibFunc_truncl: 589 Opcode = ISD::FTRUNC; break; 590 case LibFunc_fmin: 591 case LibFunc_fminf: 592 case LibFunc_fminl: 593 Opcode = ISD::FMINNUM; break; 594 case LibFunc_fmax: 595 case LibFunc_fmaxf: 596 case LibFunc_fmaxl: 597 Opcode = ISD::FMAXNUM; break; 598 } 599 } 600 601 if (Opcode) { 602 EVT EVTy = 603 TLI->getValueType(DL, CI->getArgOperand(0)->getType(), true); 604 605 if (EVTy == MVT::Other) 606 return true; 607 608 if (TLI->isOperationLegalOrCustom(Opcode, EVTy)) 609 continue; 610 else if (EVTy.isVector() && 611 TLI->isOperationLegalOrCustom(Opcode, EVTy.getScalarType())) 612 continue; 613 614 return true; 615 } 616 } 617 618 return true; 619 } else if (isa<BinaryOperator>(J) && 620 (J->getType()->getScalarType()->isFP128Ty() || 621 J->getType()->getScalarType()->isPPC_FP128Ty())) { 622 // Most operations on f128 or ppc_f128 values become calls. 623 return true; 624 } else if (isa<UIToFPInst>(J) || isa<SIToFPInst>(J) || 625 isa<FPToUIInst>(J) || isa<FPToSIInst>(J)) { 626 CastInst *CI = cast<CastInst>(J); 627 if (CI->getSrcTy()->getScalarType()->isPPC_FP128Ty() || 628 CI->getDestTy()->getScalarType()->isPPC_FP128Ty() || 629 isLargeIntegerTy(!TM.isPPC64(), CI->getSrcTy()->getScalarType()) || 630 isLargeIntegerTy(!TM.isPPC64(), CI->getDestTy()->getScalarType())) 631 return true; 632 } else if (isLargeIntegerTy(!TM.isPPC64(), 633 J->getType()->getScalarType()) && 634 (J->getOpcode() == Instruction::UDiv || 635 J->getOpcode() == Instruction::SDiv || 636 J->getOpcode() == Instruction::URem || 637 J->getOpcode() == Instruction::SRem)) { 638 return true; 639 } else if (!TM.isPPC64() && 640 isLargeIntegerTy(false, J->getType()->getScalarType()) && 641 (J->getOpcode() == Instruction::Shl || 642 J->getOpcode() == Instruction::AShr || 643 J->getOpcode() == Instruction::LShr)) { 644 // Only on PPC32, for 128-bit integers (specifically not 64-bit 645 // integers), these might be runtime calls. 646 return true; 647 } else if (isa<IndirectBrInst>(J) || isa<InvokeInst>(J)) { 648 // On PowerPC, indirect jumps use the counter register. 649 return true; 650 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(J)) { 651 if (SI->getNumCases() + 1 >= (unsigned)TLI->getMinimumJumpTableEntries()) 652 return true; 653 } 654 655 // FREM is always a call. 656 if (J->getOpcode() == Instruction::FRem) 657 return true; 658 659 if (ST->useSoftFloat()) { 660 switch(J->getOpcode()) { 661 case Instruction::FAdd: 662 case Instruction::FSub: 663 case Instruction::FMul: 664 case Instruction::FDiv: 665 case Instruction::FPTrunc: 666 case Instruction::FPExt: 667 case Instruction::FPToUI: 668 case Instruction::FPToSI: 669 case Instruction::UIToFP: 670 case Instruction::SIToFP: 671 case Instruction::FCmp: 672 return true; 673 } 674 } 675 676 for (Value *Operand : J->operands()) 677 if (memAddrUsesCTR(Operand)) 678 return true; 679 } 680 681 return false; 682 } 683 684 bool PPCTTIImpl::isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE, 685 AssumptionCache &AC, 686 TargetLibraryInfo *LibInfo, 687 HardwareLoopInfo &HWLoopInfo) { 688 const PPCTargetMachine &TM = ST->getTargetMachine(); 689 TargetSchedModel SchedModel; 690 SchedModel.init(ST); 691 692 // Do not convert small short loops to CTR loop. 693 unsigned ConstTripCount = SE.getSmallConstantTripCount(L); 694 if (ConstTripCount && ConstTripCount < SmallCTRLoopThreshold) { 695 SmallPtrSet<const Value *, 32> EphValues; 696 CodeMetrics::collectEphemeralValues(L, &AC, EphValues); 697 CodeMetrics Metrics; 698 for (BasicBlock *BB : L->blocks()) 699 Metrics.analyzeBasicBlock(BB, *this, EphValues); 700 // 6 is an approximate latency for the mtctr instruction. 701 if (Metrics.NumInsts <= (6 * SchedModel.getIssueWidth())) 702 return false; 703 } 704 705 // We don't want to spill/restore the counter register, and so we don't 706 // want to use the counter register if the loop contains calls. 707 SmallPtrSet<const Value *, 4> Visited; 708 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end(); 709 I != IE; ++I) 710 if (mightUseCTR(*I, LibInfo, Visited)) 711 return false; 712 713 SmallVector<BasicBlock*, 4> ExitingBlocks; 714 L->getExitingBlocks(ExitingBlocks); 715 716 // If there is an exit edge known to be frequently taken, 717 // we should not transform this loop. 718 for (auto &BB : ExitingBlocks) { 719 Instruction *TI = BB->getTerminator(); 720 if (!TI) continue; 721 722 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 723 uint64_t TrueWeight = 0, FalseWeight = 0; 724 if (!BI->isConditional() || 725 !BI->extractProfMetadata(TrueWeight, FalseWeight)) 726 continue; 727 728 // If the exit path is more frequent than the loop path, 729 // we return here without further analysis for this loop. 730 bool TrueIsExit = !L->contains(BI->getSuccessor(0)); 731 if (( TrueIsExit && FalseWeight < TrueWeight) || 732 (!TrueIsExit && FalseWeight > TrueWeight)) 733 return false; 734 } 735 } 736 737 LLVMContext &C = L->getHeader()->getContext(); 738 HWLoopInfo.CountType = TM.isPPC64() ? 739 Type::getInt64Ty(C) : Type::getInt32Ty(C); 740 HWLoopInfo.LoopDecrement = ConstantInt::get(HWLoopInfo.CountType, 1); 741 return true; 742 } 743 744 void PPCTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 745 TTI::UnrollingPreferences &UP) { 746 if (ST->getCPUDirective() == PPC::DIR_A2) { 747 // The A2 is in-order with a deep pipeline, and concatenation unrolling 748 // helps expose latency-hiding opportunities to the instruction scheduler. 749 UP.Partial = UP.Runtime = true; 750 751 // We unroll a lot on the A2 (hundreds of instructions), and the benefits 752 // often outweigh the cost of a division to compute the trip count. 753 UP.AllowExpensiveTripCount = true; 754 } 755 756 BaseT::getUnrollingPreferences(L, SE, UP); 757 } 758 759 void PPCTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE, 760 TTI::PeelingPreferences &PP) { 761 BaseT::getPeelingPreferences(L, SE, PP); 762 } 763 // This function returns true to allow using coldcc calling convention. 764 // Returning true results in coldcc being used for functions which are cold at 765 // all call sites when the callers of the functions are not calling any other 766 // non coldcc functions. 767 bool PPCTTIImpl::useColdCCForColdCall(Function &F) { 768 return EnablePPCColdCC; 769 } 770 771 bool PPCTTIImpl::enableAggressiveInterleaving(bool LoopHasReductions) { 772 // On the A2, always unroll aggressively. 773 if (ST->getCPUDirective() == PPC::DIR_A2) 774 return true; 775 776 return LoopHasReductions; 777 } 778 779 PPCTTIImpl::TTI::MemCmpExpansionOptions 780 PPCTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 781 TTI::MemCmpExpansionOptions Options; 782 Options.LoadSizes = {8, 4, 2, 1}; 783 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); 784 return Options; 785 } 786 787 bool PPCTTIImpl::enableInterleavedAccessVectorization() { 788 return true; 789 } 790 791 unsigned PPCTTIImpl::getNumberOfRegisters(unsigned ClassID) const { 792 assert(ClassID == GPRRC || ClassID == FPRRC || 793 ClassID == VRRC || ClassID == VSXRC); 794 if (ST->hasVSX()) { 795 assert(ClassID == GPRRC || ClassID == VSXRC || ClassID == VRRC); 796 return ClassID == VSXRC ? 64 : 32; 797 } 798 assert(ClassID == GPRRC || ClassID == FPRRC || ClassID == VRRC); 799 return 32; 800 } 801 802 unsigned PPCTTIImpl::getRegisterClassForType(bool Vector, Type *Ty) const { 803 if (Vector) 804 return ST->hasVSX() ? VSXRC : VRRC; 805 else if (Ty && (Ty->getScalarType()->isFloatTy() || 806 Ty->getScalarType()->isDoubleTy())) 807 return ST->hasVSX() ? VSXRC : FPRRC; 808 else if (Ty && (Ty->getScalarType()->isFP128Ty() || 809 Ty->getScalarType()->isPPC_FP128Ty())) 810 return VRRC; 811 else if (Ty && Ty->getScalarType()->isHalfTy()) 812 return VSXRC; 813 else 814 return GPRRC; 815 } 816 817 const char* PPCTTIImpl::getRegisterClassName(unsigned ClassID) const { 818 819 switch (ClassID) { 820 default: 821 llvm_unreachable("unknown register class"); 822 return "PPC::unknown register class"; 823 case GPRRC: return "PPC::GPRRC"; 824 case FPRRC: return "PPC::FPRRC"; 825 case VRRC: return "PPC::VRRC"; 826 case VSXRC: return "PPC::VSXRC"; 827 } 828 } 829 830 unsigned PPCTTIImpl::getRegisterBitWidth(bool Vector) const { 831 if (Vector) { 832 if (ST->hasAltivec()) return 128; 833 return 0; 834 } 835 836 if (ST->isPPC64()) 837 return 64; 838 return 32; 839 840 } 841 842 unsigned PPCTTIImpl::getCacheLineSize() const { 843 // Check first if the user specified a custom line size. 844 if (CacheLineSize.getNumOccurrences() > 0) 845 return CacheLineSize; 846 847 // Starting with P7 we have a cache line size of 128. 848 unsigned Directive = ST->getCPUDirective(); 849 // Assume that Future CPU has the same cache line size as the others. 850 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8 || 851 Directive == PPC::DIR_PWR9 || Directive == PPC::DIR_PWR10 || 852 Directive == PPC::DIR_PWR_FUTURE) 853 return 128; 854 855 // On other processors return a default of 64 bytes. 856 return 64; 857 } 858 859 unsigned PPCTTIImpl::getPrefetchDistance() const { 860 return 300; 861 } 862 863 unsigned PPCTTIImpl::getMaxInterleaveFactor(unsigned VF) { 864 unsigned Directive = ST->getCPUDirective(); 865 // The 440 has no SIMD support, but floating-point instructions 866 // have a 5-cycle latency, so unroll by 5x for latency hiding. 867 if (Directive == PPC::DIR_440) 868 return 5; 869 870 // The A2 has no SIMD support, but floating-point instructions 871 // have a 6-cycle latency, so unroll by 6x for latency hiding. 872 if (Directive == PPC::DIR_A2) 873 return 6; 874 875 // FIXME: For lack of any better information, do no harm... 876 if (Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) 877 return 1; 878 879 // For P7 and P8, floating-point instructions have a 6-cycle latency and 880 // there are two execution units, so unroll by 12x for latency hiding. 881 // FIXME: the same for P9 as previous gen until POWER9 scheduling is ready 882 // FIXME: the same for P10 as previous gen until POWER10 scheduling is ready 883 // Assume that future is the same as the others. 884 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8 || 885 Directive == PPC::DIR_PWR9 || Directive == PPC::DIR_PWR10 || 886 Directive == PPC::DIR_PWR_FUTURE) 887 return 12; 888 889 // For most things, modern systems have two execution units (and 890 // out-of-order execution). 891 return 2; 892 } 893 894 // Adjust the cost of vector instructions on targets which there is overlap 895 // between the vector and scalar units, thereby reducing the overall throughput 896 // of vector code wrt. scalar code. 897 int PPCTTIImpl::vectorCostAdjustment(int Cost, unsigned Opcode, Type *Ty1, 898 Type *Ty2) { 899 if (!ST->vectorsUseTwoUnits() || !Ty1->isVectorTy()) 900 return Cost; 901 902 std::pair<int, MVT> LT1 = TLI->getTypeLegalizationCost(DL, Ty1); 903 // If type legalization involves splitting the vector, we don't want to 904 // double the cost at every step - only the last step. 905 if (LT1.first != 1 || !LT1.second.isVector()) 906 return Cost; 907 908 int ISD = TLI->InstructionOpcodeToISD(Opcode); 909 if (TLI->isOperationExpand(ISD, LT1.second)) 910 return Cost; 911 912 if (Ty2) { 913 std::pair<int, MVT> LT2 = TLI->getTypeLegalizationCost(DL, Ty2); 914 if (LT2.first != 1 || !LT2.second.isVector()) 915 return Cost; 916 } 917 918 return Cost * 2; 919 } 920 921 int PPCTTIImpl::getArithmeticInstrCost(unsigned Opcode, Type *Ty, 922 TTI::TargetCostKind CostKind, 923 TTI::OperandValueKind Op1Info, 924 TTI::OperandValueKind Op2Info, 925 TTI::OperandValueProperties Opd1PropInfo, 926 TTI::OperandValueProperties Opd2PropInfo, 927 ArrayRef<const Value *> Args, 928 const Instruction *CxtI) { 929 assert(TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode"); 930 // TODO: Handle more cost kinds. 931 if (CostKind != TTI::TCK_RecipThroughput) 932 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, 933 Op2Info, Opd1PropInfo, 934 Opd2PropInfo, Args, CxtI); 935 936 // Fallback to the default implementation. 937 int Cost = BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, 938 Op2Info, 939 Opd1PropInfo, Opd2PropInfo); 940 return vectorCostAdjustment(Cost, Opcode, Ty, nullptr); 941 } 942 943 int PPCTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, 944 Type *SubTp) { 945 // Legalize the type. 946 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp); 947 948 // PPC, for both Altivec/VSX, support cheap arbitrary permutations 949 // (at least in the sense that there need only be one non-loop-invariant 950 // instruction). We need one such shuffle instruction for each actual 951 // register (this is not true for arbitrary shuffles, but is true for the 952 // structured types of shuffles covered by TTI::ShuffleKind). 953 return vectorCostAdjustment(LT.first, Instruction::ShuffleVector, Tp, 954 nullptr); 955 } 956 957 int PPCTTIImpl::getCFInstrCost(unsigned Opcode, TTI::TargetCostKind CostKind) { 958 if (CostKind != TTI::TCK_RecipThroughput) 959 return Opcode == Instruction::PHI ? 0 : 1; 960 // Branches are assumed to be predicted. 961 return CostKind == TTI::TCK_RecipThroughput ? 0 : 1; 962 } 963 964 int PPCTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 965 TTI::CastContextHint CCH, 966 TTI::TargetCostKind CostKind, 967 const Instruction *I) { 968 assert(TLI->InstructionOpcodeToISD(Opcode) && "Invalid opcode"); 969 970 int Cost = BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I); 971 Cost = vectorCostAdjustment(Cost, Opcode, Dst, Src); 972 // TODO: Allow non-throughput costs that aren't binary. 973 if (CostKind != TTI::TCK_RecipThroughput) 974 return Cost == 0 ? 0 : 1; 975 return Cost; 976 } 977 978 int PPCTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, 979 TTI::TargetCostKind CostKind, 980 const Instruction *I) { 981 int Cost = BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, CostKind, I); 982 // TODO: Handle other cost kinds. 983 if (CostKind != TTI::TCK_RecipThroughput) 984 return Cost; 985 return vectorCostAdjustment(Cost, Opcode, ValTy, nullptr); 986 } 987 988 int PPCTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 989 assert(Val->isVectorTy() && "This must be a vector type"); 990 991 int ISD = TLI->InstructionOpcodeToISD(Opcode); 992 assert(ISD && "Invalid opcode"); 993 994 int Cost = BaseT::getVectorInstrCost(Opcode, Val, Index); 995 Cost = vectorCostAdjustment(Cost, Opcode, Val, nullptr); 996 997 if (ST->hasVSX() && Val->getScalarType()->isDoubleTy()) { 998 // Double-precision scalars are already located in index #0 (or #1 if LE). 999 if (ISD == ISD::EXTRACT_VECTOR_ELT && 1000 Index == (ST->isLittleEndian() ? 1 : 0)) 1001 return 0; 1002 1003 return Cost; 1004 1005 } else if (Val->getScalarType()->isIntegerTy() && Index != -1U) { 1006 if (ST->hasP9Altivec()) { 1007 if (ISD == ISD::INSERT_VECTOR_ELT) 1008 // A move-to VSR and a permute/insert. Assume vector operation cost 1009 // for both (cost will be 2x on P9). 1010 return vectorCostAdjustment(2, Opcode, Val, nullptr); 1011 1012 // It's an extract. Maybe we can do a cheap move-from VSR. 1013 unsigned EltSize = Val->getScalarSizeInBits(); 1014 if (EltSize == 64) { 1015 unsigned MfvsrdIndex = ST->isLittleEndian() ? 1 : 0; 1016 if (Index == MfvsrdIndex) 1017 return 1; 1018 } else if (EltSize == 32) { 1019 unsigned MfvsrwzIndex = ST->isLittleEndian() ? 2 : 1; 1020 if (Index == MfvsrwzIndex) 1021 return 1; 1022 } 1023 1024 // We need a vector extract (or mfvsrld). Assume vector operation cost. 1025 // The cost of the load constant for a vector extract is disregarded 1026 // (invariant, easily schedulable). 1027 return vectorCostAdjustment(1, Opcode, Val, nullptr); 1028 1029 } else if (ST->hasDirectMove()) 1030 // Assume permute has standard cost. 1031 // Assume move-to/move-from VSR have 2x standard cost. 1032 return 3; 1033 } 1034 1035 // Estimated cost of a load-hit-store delay. This was obtained 1036 // experimentally as a minimum needed to prevent unprofitable 1037 // vectorization for the paq8p benchmark. It may need to be 1038 // raised further if other unprofitable cases remain. 1039 unsigned LHSPenalty = 2; 1040 if (ISD == ISD::INSERT_VECTOR_ELT) 1041 LHSPenalty += 7; 1042 1043 // Vector element insert/extract with Altivec is very expensive, 1044 // because they require store and reload with the attendant 1045 // processor stall for load-hit-store. Until VSX is available, 1046 // these need to be estimated as very costly. 1047 if (ISD == ISD::EXTRACT_VECTOR_ELT || 1048 ISD == ISD::INSERT_VECTOR_ELT) 1049 return LHSPenalty + Cost; 1050 1051 return Cost; 1052 } 1053 1054 int PPCTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src, 1055 MaybeAlign Alignment, unsigned AddressSpace, 1056 TTI::TargetCostKind CostKind, 1057 const Instruction *I) { 1058 if (TLI->getValueType(DL, Src, true) == MVT::Other) 1059 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 1060 CostKind); 1061 // Legalize the type. 1062 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, Src); 1063 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 1064 "Invalid Opcode"); 1065 1066 int Cost = BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, 1067 CostKind); 1068 // TODO: Handle other cost kinds. 1069 if (CostKind != TTI::TCK_RecipThroughput) 1070 return Cost; 1071 1072 Cost = vectorCostAdjustment(Cost, Opcode, Src, nullptr); 1073 1074 bool IsAltivecType = ST->hasAltivec() && 1075 (LT.second == MVT::v16i8 || LT.second == MVT::v8i16 || 1076 LT.second == MVT::v4i32 || LT.second == MVT::v4f32); 1077 bool IsVSXType = ST->hasVSX() && 1078 (LT.second == MVT::v2f64 || LT.second == MVT::v2i64); 1079 1080 // VSX has 32b/64b load instructions. Legalization can handle loading of 1081 // 32b/64b to VSR correctly and cheaply. But BaseT::getMemoryOpCost and 1082 // PPCTargetLowering can't compute the cost appropriately. So here we 1083 // explicitly check this case. 1084 unsigned MemBytes = Src->getPrimitiveSizeInBits(); 1085 if (Opcode == Instruction::Load && ST->hasVSX() && IsAltivecType && 1086 (MemBytes == 64 || (ST->hasP8Vector() && MemBytes == 32))) 1087 return 1; 1088 1089 // Aligned loads and stores are easy. 1090 unsigned SrcBytes = LT.second.getStoreSize(); 1091 if (!SrcBytes || !Alignment || *Alignment >= SrcBytes) 1092 return Cost; 1093 1094 // If we can use the permutation-based load sequence, then this is also 1095 // relatively cheap (not counting loop-invariant instructions): one load plus 1096 // one permute (the last load in a series has extra cost, but we're 1097 // neglecting that here). Note that on the P7, we could do unaligned loads 1098 // for Altivec types using the VSX instructions, but that's more expensive 1099 // than using the permutation-based load sequence. On the P8, that's no 1100 // longer true. 1101 if (Opcode == Instruction::Load && (!ST->hasP8Vector() && IsAltivecType) && 1102 *Alignment >= LT.second.getScalarType().getStoreSize()) 1103 return Cost + LT.first; // Add the cost of the permutations. 1104 1105 // For VSX, we can do unaligned loads and stores on Altivec/VSX types. On the 1106 // P7, unaligned vector loads are more expensive than the permutation-based 1107 // load sequence, so that might be used instead, but regardless, the net cost 1108 // is about the same (not counting loop-invariant instructions). 1109 if (IsVSXType || (ST->hasVSX() && IsAltivecType)) 1110 return Cost; 1111 1112 // Newer PPC supports unaligned memory access. 1113 if (TLI->allowsMisalignedMemoryAccesses(LT.second, 0)) 1114 return Cost; 1115 1116 // PPC in general does not support unaligned loads and stores. They'll need 1117 // to be decomposed based on the alignment factor. 1118 1119 // Add the cost of each scalar load or store. 1120 assert(Alignment); 1121 Cost += LT.first * ((SrcBytes / Alignment->value()) - 1); 1122 1123 // For a vector type, there is also scalarization overhead (only for 1124 // stores, loads are expanded using the vector-load + permutation sequence, 1125 // which is much less expensive). 1126 if (Src->isVectorTy() && Opcode == Instruction::Store) 1127 for (int i = 0, e = cast<FixedVectorType>(Src)->getNumElements(); i < e; 1128 ++i) 1129 Cost += getVectorInstrCost(Instruction::ExtractElement, Src, i); 1130 1131 return Cost; 1132 } 1133 1134 int PPCTTIImpl::getInterleavedMemoryOpCost( 1135 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 1136 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, 1137 bool UseMaskForCond, bool UseMaskForGaps) { 1138 if (UseMaskForCond || UseMaskForGaps) 1139 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 1140 Alignment, AddressSpace, CostKind, 1141 UseMaskForCond, UseMaskForGaps); 1142 1143 assert(isa<VectorType>(VecTy) && 1144 "Expect a vector type for interleaved memory op"); 1145 1146 // Legalize the type. 1147 std::pair<int, MVT> LT = TLI->getTypeLegalizationCost(DL, VecTy); 1148 1149 // Firstly, the cost of load/store operation. 1150 int Cost = 1151 getMemoryOpCost(Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, 1152 CostKind); 1153 1154 // PPC, for both Altivec/VSX, support cheap arbitrary permutations 1155 // (at least in the sense that there need only be one non-loop-invariant 1156 // instruction). For each result vector, we need one shuffle per incoming 1157 // vector (except that the first shuffle can take two incoming vectors 1158 // because it does not need to take itself). 1159 Cost += Factor*(LT.first-1); 1160 1161 return Cost; 1162 } 1163 1164 unsigned PPCTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, 1165 TTI::TargetCostKind CostKind) { 1166 return BaseT::getIntrinsicInstrCost(ICA, CostKind); 1167 } 1168 1169 bool PPCTTIImpl::canSaveCmp(Loop *L, BranchInst **BI, ScalarEvolution *SE, 1170 LoopInfo *LI, DominatorTree *DT, 1171 AssumptionCache *AC, TargetLibraryInfo *LibInfo) { 1172 // Process nested loops first. 1173 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 1174 if (canSaveCmp(*I, BI, SE, LI, DT, AC, LibInfo)) 1175 return false; // Stop search. 1176 1177 HardwareLoopInfo HWLoopInfo(L); 1178 1179 if (!HWLoopInfo.canAnalyze(*LI)) 1180 return false; 1181 1182 if (!isHardwareLoopProfitable(L, *SE, *AC, LibInfo, HWLoopInfo)) 1183 return false; 1184 1185 if (!HWLoopInfo.isHardwareLoopCandidate(*SE, *LI, *DT)) 1186 return false; 1187 1188 *BI = HWLoopInfo.ExitBranch; 1189 return true; 1190 } 1191 1192 bool PPCTTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1, 1193 TargetTransformInfo::LSRCost &C2) { 1194 // PowerPC default behaviour here is "instruction number 1st priority". 1195 // If LsrNoInsnsCost is set, call default implementation. 1196 if (!LsrNoInsnsCost) 1197 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, 1198 C1.NumBaseAdds, C1.ScaleCost, C1.ImmCost, C1.SetupCost) < 1199 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, 1200 C2.NumBaseAdds, C2.ScaleCost, C2.ImmCost, C2.SetupCost); 1201 else 1202 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2); 1203 } 1204