1 //===- AMDGPULibCalls.cpp -------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// This file does AMD library function optimizations. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "AMDGPU.h" 15 #include "AMDGPULibFunc.h" 16 #include "GCNSubtarget.h" 17 #include "llvm/Analysis/AssumptionCache.h" 18 #include "llvm/Analysis/TargetLibraryInfo.h" 19 #include "llvm/Analysis/ValueTracking.h" 20 #include "llvm/IR/AttributeMask.h" 21 #include "llvm/IR/Dominators.h" 22 #include "llvm/IR/IRBuilder.h" 23 #include "llvm/IR/IntrinsicInst.h" 24 #include "llvm/IR/IntrinsicsAMDGPU.h" 25 #include "llvm/IR/PatternMatch.h" 26 #include "llvm/InitializePasses.h" 27 #include <cmath> 28 29 #define DEBUG_TYPE "amdgpu-simplifylib" 30 31 using namespace llvm; 32 using namespace llvm::PatternMatch; 33 34 static cl::opt<bool> EnablePreLink("amdgpu-prelink", 35 cl::desc("Enable pre-link mode optimizations"), 36 cl::init(false), 37 cl::Hidden); 38 39 static cl::list<std::string> UseNative("amdgpu-use-native", 40 cl::desc("Comma separated list of functions to replace with native, or all"), 41 cl::CommaSeparated, cl::ValueOptional, 42 cl::Hidden); 43 44 #define MATH_PI numbers::pi 45 #define MATH_E numbers::e 46 #define MATH_SQRT2 numbers::sqrt2 47 #define MATH_SQRT1_2 numbers::inv_sqrt2 48 49 namespace llvm { 50 51 class AMDGPULibCalls { 52 private: 53 const TargetLibraryInfo *TLInfo = nullptr; 54 AssumptionCache *AC = nullptr; 55 DominatorTree *DT = nullptr; 56 57 typedef llvm::AMDGPULibFunc FuncInfo; 58 59 bool UnsafeFPMath = false; 60 61 // -fuse-native. 62 bool AllNative = false; 63 64 bool useNativeFunc(const StringRef F) const; 65 66 // Return a pointer (pointer expr) to the function if function definition with 67 // "FuncName" exists. It may create a new function prototype in pre-link mode. 68 FunctionCallee getFunction(Module *M, const FuncInfo &fInfo); 69 70 bool parseFunctionName(const StringRef &FMangledName, FuncInfo &FInfo); 71 72 bool TDOFold(CallInst *CI, const FuncInfo &FInfo); 73 74 /* Specialized optimizations */ 75 76 // pow/powr/pown 77 bool fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo); 78 79 // rootn 80 bool fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo); 81 82 // -fuse-native for sincos 83 bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo); 84 85 // evaluate calls if calls' arguments are constants. 86 bool evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0, double &Res1, 87 Constant *copr0, Constant *copr1); 88 bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo); 89 90 /// Insert a value to sincos function \p Fsincos. Returns (value of sin, value 91 /// of cos, sincos call). 92 std::tuple<Value *, Value *, Value *> insertSinCos(Value *Arg, 93 FastMathFlags FMF, 94 IRBuilder<> &B, 95 FunctionCallee Fsincos); 96 97 // sin/cos 98 bool fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo); 99 100 // __read_pipe/__write_pipe 101 bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 102 const FuncInfo &FInfo); 103 104 // Get a scalar native builtin single argument FP function 105 FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo); 106 107 /// Substitute a call to a known libcall with an intrinsic call. If \p 108 /// AllowMinSize is true, allow the replacement in a minsize function. 109 bool shouldReplaceLibcallWithIntrinsic(const CallInst *CI, 110 bool AllowMinSizeF32 = false, 111 bool AllowF64 = false, 112 bool AllowStrictFP = false); 113 void replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI, 114 Intrinsic::ID IntrID); 115 116 bool tryReplaceLibcallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI, 117 Intrinsic::ID IntrID, 118 bool AllowMinSizeF32 = false, 119 bool AllowF64 = false, 120 bool AllowStrictFP = false); 121 122 protected: 123 bool isUnsafeMath(const FPMathOperator *FPOp) const; 124 bool isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const; 125 126 bool canIncreasePrecisionOfConstantFold(const FPMathOperator *FPOp) const; 127 128 static void replaceCall(Instruction *I, Value *With) { 129 I->replaceAllUsesWith(With); 130 I->eraseFromParent(); 131 } 132 133 static void replaceCall(FPMathOperator *I, Value *With) { 134 replaceCall(cast<Instruction>(I), With); 135 } 136 137 public: 138 AMDGPULibCalls() {} 139 140 bool fold(CallInst *CI); 141 142 void initFunction(Function &F, FunctionAnalysisManager &FAM); 143 void initNativeFuncs(); 144 145 // Replace a normal math function call with that native version 146 bool useNative(CallInst *CI); 147 }; 148 149 } // end llvm namespace 150 151 template <typename IRB> 152 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg, 153 const Twine &Name = "") { 154 CallInst *R = B.CreateCall(Callee, Arg, Name); 155 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 156 R->setCallingConv(F->getCallingConv()); 157 return R; 158 } 159 160 template <typename IRB> 161 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1, 162 Value *Arg2, const Twine &Name = "") { 163 CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name); 164 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 165 R->setCallingConv(F->getCallingConv()); 166 return R; 167 } 168 169 static FunctionType *getPownType(FunctionType *FT) { 170 Type *PowNExpTy = Type::getInt32Ty(FT->getContext()); 171 if (VectorType *VecTy = dyn_cast<VectorType>(FT->getReturnType())) 172 PowNExpTy = VectorType::get(PowNExpTy, VecTy->getElementCount()); 173 174 return FunctionType::get(FT->getReturnType(), 175 {FT->getParamType(0), PowNExpTy}, false); 176 } 177 178 // Data structures for table-driven optimizations. 179 // FuncTbl works for both f32 and f64 functions with 1 input argument 180 181 struct TableEntry { 182 double result; 183 double input; 184 }; 185 186 /* a list of {result, input} */ 187 static const TableEntry tbl_acos[] = { 188 {MATH_PI / 2.0, 0.0}, 189 {MATH_PI / 2.0, -0.0}, 190 {0.0, 1.0}, 191 {MATH_PI, -1.0} 192 }; 193 static const TableEntry tbl_acosh[] = { 194 {0.0, 1.0} 195 }; 196 static const TableEntry tbl_acospi[] = { 197 {0.5, 0.0}, 198 {0.5, -0.0}, 199 {0.0, 1.0}, 200 {1.0, -1.0} 201 }; 202 static const TableEntry tbl_asin[] = { 203 {0.0, 0.0}, 204 {-0.0, -0.0}, 205 {MATH_PI / 2.0, 1.0}, 206 {-MATH_PI / 2.0, -1.0} 207 }; 208 static const TableEntry tbl_asinh[] = { 209 {0.0, 0.0}, 210 {-0.0, -0.0} 211 }; 212 static const TableEntry tbl_asinpi[] = { 213 {0.0, 0.0}, 214 {-0.0, -0.0}, 215 {0.5, 1.0}, 216 {-0.5, -1.0} 217 }; 218 static const TableEntry tbl_atan[] = { 219 {0.0, 0.0}, 220 {-0.0, -0.0}, 221 {MATH_PI / 4.0, 1.0}, 222 {-MATH_PI / 4.0, -1.0} 223 }; 224 static const TableEntry tbl_atanh[] = { 225 {0.0, 0.0}, 226 {-0.0, -0.0} 227 }; 228 static const TableEntry tbl_atanpi[] = { 229 {0.0, 0.0}, 230 {-0.0, -0.0}, 231 {0.25, 1.0}, 232 {-0.25, -1.0} 233 }; 234 static const TableEntry tbl_cbrt[] = { 235 {0.0, 0.0}, 236 {-0.0, -0.0}, 237 {1.0, 1.0}, 238 {-1.0, -1.0}, 239 }; 240 static const TableEntry tbl_cos[] = { 241 {1.0, 0.0}, 242 {1.0, -0.0} 243 }; 244 static const TableEntry tbl_cosh[] = { 245 {1.0, 0.0}, 246 {1.0, -0.0} 247 }; 248 static const TableEntry tbl_cospi[] = { 249 {1.0, 0.0}, 250 {1.0, -0.0} 251 }; 252 static const TableEntry tbl_erfc[] = { 253 {1.0, 0.0}, 254 {1.0, -0.0} 255 }; 256 static const TableEntry tbl_erf[] = { 257 {0.0, 0.0}, 258 {-0.0, -0.0} 259 }; 260 static const TableEntry tbl_exp[] = { 261 {1.0, 0.0}, 262 {1.0, -0.0}, 263 {MATH_E, 1.0} 264 }; 265 static const TableEntry tbl_exp2[] = { 266 {1.0, 0.0}, 267 {1.0, -0.0}, 268 {2.0, 1.0} 269 }; 270 static const TableEntry tbl_exp10[] = { 271 {1.0, 0.0}, 272 {1.0, -0.0}, 273 {10.0, 1.0} 274 }; 275 static const TableEntry tbl_expm1[] = { 276 {0.0, 0.0}, 277 {-0.0, -0.0} 278 }; 279 static const TableEntry tbl_log[] = { 280 {0.0, 1.0}, 281 {1.0, MATH_E} 282 }; 283 static const TableEntry tbl_log2[] = { 284 {0.0, 1.0}, 285 {1.0, 2.0} 286 }; 287 static const TableEntry tbl_log10[] = { 288 {0.0, 1.0}, 289 {1.0, 10.0} 290 }; 291 static const TableEntry tbl_rsqrt[] = { 292 {1.0, 1.0}, 293 {MATH_SQRT1_2, 2.0} 294 }; 295 static const TableEntry tbl_sin[] = { 296 {0.0, 0.0}, 297 {-0.0, -0.0} 298 }; 299 static const TableEntry tbl_sinh[] = { 300 {0.0, 0.0}, 301 {-0.0, -0.0} 302 }; 303 static const TableEntry tbl_sinpi[] = { 304 {0.0, 0.0}, 305 {-0.0, -0.0} 306 }; 307 static const TableEntry tbl_sqrt[] = { 308 {0.0, 0.0}, 309 {1.0, 1.0}, 310 {MATH_SQRT2, 2.0} 311 }; 312 static const TableEntry tbl_tan[] = { 313 {0.0, 0.0}, 314 {-0.0, -0.0} 315 }; 316 static const TableEntry tbl_tanh[] = { 317 {0.0, 0.0}, 318 {-0.0, -0.0} 319 }; 320 static const TableEntry tbl_tanpi[] = { 321 {0.0, 0.0}, 322 {-0.0, -0.0} 323 }; 324 static const TableEntry tbl_tgamma[] = { 325 {1.0, 1.0}, 326 {1.0, 2.0}, 327 {2.0, 3.0}, 328 {6.0, 4.0} 329 }; 330 331 static bool HasNative(AMDGPULibFunc::EFuncId id) { 332 switch(id) { 333 case AMDGPULibFunc::EI_DIVIDE: 334 case AMDGPULibFunc::EI_COS: 335 case AMDGPULibFunc::EI_EXP: 336 case AMDGPULibFunc::EI_EXP2: 337 case AMDGPULibFunc::EI_EXP10: 338 case AMDGPULibFunc::EI_LOG: 339 case AMDGPULibFunc::EI_LOG2: 340 case AMDGPULibFunc::EI_LOG10: 341 case AMDGPULibFunc::EI_POWR: 342 case AMDGPULibFunc::EI_RECIP: 343 case AMDGPULibFunc::EI_RSQRT: 344 case AMDGPULibFunc::EI_SIN: 345 case AMDGPULibFunc::EI_SINCOS: 346 case AMDGPULibFunc::EI_SQRT: 347 case AMDGPULibFunc::EI_TAN: 348 return true; 349 default:; 350 } 351 return false; 352 } 353 354 using TableRef = ArrayRef<TableEntry>; 355 356 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) { 357 switch(id) { 358 case AMDGPULibFunc::EI_ACOS: return TableRef(tbl_acos); 359 case AMDGPULibFunc::EI_ACOSH: return TableRef(tbl_acosh); 360 case AMDGPULibFunc::EI_ACOSPI: return TableRef(tbl_acospi); 361 case AMDGPULibFunc::EI_ASIN: return TableRef(tbl_asin); 362 case AMDGPULibFunc::EI_ASINH: return TableRef(tbl_asinh); 363 case AMDGPULibFunc::EI_ASINPI: return TableRef(tbl_asinpi); 364 case AMDGPULibFunc::EI_ATAN: return TableRef(tbl_atan); 365 case AMDGPULibFunc::EI_ATANH: return TableRef(tbl_atanh); 366 case AMDGPULibFunc::EI_ATANPI: return TableRef(tbl_atanpi); 367 case AMDGPULibFunc::EI_CBRT: return TableRef(tbl_cbrt); 368 case AMDGPULibFunc::EI_NCOS: 369 case AMDGPULibFunc::EI_COS: return TableRef(tbl_cos); 370 case AMDGPULibFunc::EI_COSH: return TableRef(tbl_cosh); 371 case AMDGPULibFunc::EI_COSPI: return TableRef(tbl_cospi); 372 case AMDGPULibFunc::EI_ERFC: return TableRef(tbl_erfc); 373 case AMDGPULibFunc::EI_ERF: return TableRef(tbl_erf); 374 case AMDGPULibFunc::EI_EXP: return TableRef(tbl_exp); 375 case AMDGPULibFunc::EI_NEXP2: 376 case AMDGPULibFunc::EI_EXP2: return TableRef(tbl_exp2); 377 case AMDGPULibFunc::EI_EXP10: return TableRef(tbl_exp10); 378 case AMDGPULibFunc::EI_EXPM1: return TableRef(tbl_expm1); 379 case AMDGPULibFunc::EI_LOG: return TableRef(tbl_log); 380 case AMDGPULibFunc::EI_NLOG2: 381 case AMDGPULibFunc::EI_LOG2: return TableRef(tbl_log2); 382 case AMDGPULibFunc::EI_LOG10: return TableRef(tbl_log10); 383 case AMDGPULibFunc::EI_NRSQRT: 384 case AMDGPULibFunc::EI_RSQRT: return TableRef(tbl_rsqrt); 385 case AMDGPULibFunc::EI_NSIN: 386 case AMDGPULibFunc::EI_SIN: return TableRef(tbl_sin); 387 case AMDGPULibFunc::EI_SINH: return TableRef(tbl_sinh); 388 case AMDGPULibFunc::EI_SINPI: return TableRef(tbl_sinpi); 389 case AMDGPULibFunc::EI_NSQRT: 390 case AMDGPULibFunc::EI_SQRT: return TableRef(tbl_sqrt); 391 case AMDGPULibFunc::EI_TAN: return TableRef(tbl_tan); 392 case AMDGPULibFunc::EI_TANH: return TableRef(tbl_tanh); 393 case AMDGPULibFunc::EI_TANPI: return TableRef(tbl_tanpi); 394 case AMDGPULibFunc::EI_TGAMMA: return TableRef(tbl_tgamma); 395 default:; 396 } 397 return TableRef(); 398 } 399 400 static inline int getVecSize(const AMDGPULibFunc& FInfo) { 401 return FInfo.getLeads()[0].VectorSize; 402 } 403 404 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) { 405 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType; 406 } 407 408 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) { 409 // If we are doing PreLinkOpt, the function is external. So it is safe to 410 // use getOrInsertFunction() at this stage. 411 412 return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo) 413 : AMDGPULibFunc::getFunction(M, fInfo); 414 } 415 416 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName, 417 FuncInfo &FInfo) { 418 return AMDGPULibFunc::parse(FMangledName, FInfo); 419 } 420 421 bool AMDGPULibCalls::isUnsafeMath(const FPMathOperator *FPOp) const { 422 return UnsafeFPMath || FPOp->isFast(); 423 } 424 425 bool AMDGPULibCalls::isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const { 426 return UnsafeFPMath || 427 (FPOp->hasApproxFunc() && FPOp->hasNoNaNs() && FPOp->hasNoInfs()); 428 } 429 430 bool AMDGPULibCalls::canIncreasePrecisionOfConstantFold( 431 const FPMathOperator *FPOp) const { 432 // TODO: Refine to approxFunc or contract 433 return isUnsafeMath(FPOp); 434 } 435 436 void AMDGPULibCalls::initFunction(Function &F, FunctionAnalysisManager &FAM) { 437 UnsafeFPMath = F.getFnAttribute("unsafe-fp-math").getValueAsBool(); 438 AC = &FAM.getResult<AssumptionAnalysis>(F); 439 TLInfo = &FAM.getResult<TargetLibraryAnalysis>(F); 440 DT = FAM.getCachedResult<DominatorTreeAnalysis>(F); 441 } 442 443 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const { 444 return AllNative || llvm::is_contained(UseNative, F); 445 } 446 447 void AMDGPULibCalls::initNativeFuncs() { 448 AllNative = useNativeFunc("all") || 449 (UseNative.getNumOccurrences() && UseNative.size() == 1 && 450 UseNative.begin()->empty()); 451 } 452 453 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) { 454 bool native_sin = useNativeFunc("sin"); 455 bool native_cos = useNativeFunc("cos"); 456 457 if (native_sin && native_cos) { 458 Module *M = aCI->getModule(); 459 Value *opr0 = aCI->getArgOperand(0); 460 461 AMDGPULibFunc nf; 462 nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType; 463 nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize; 464 465 nf.setPrefix(AMDGPULibFunc::NATIVE); 466 nf.setId(AMDGPULibFunc::EI_SIN); 467 FunctionCallee sinExpr = getFunction(M, nf); 468 469 nf.setPrefix(AMDGPULibFunc::NATIVE); 470 nf.setId(AMDGPULibFunc::EI_COS); 471 FunctionCallee cosExpr = getFunction(M, nf); 472 if (sinExpr && cosExpr) { 473 Value *sinval = 474 CallInst::Create(sinExpr, opr0, "splitsin", aCI->getIterator()); 475 Value *cosval = 476 CallInst::Create(cosExpr, opr0, "splitcos", aCI->getIterator()); 477 new StoreInst(cosval, aCI->getArgOperand(1), aCI->getIterator()); 478 479 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 480 << " with native version of sin/cos"); 481 482 replaceCall(aCI, sinval); 483 return true; 484 } 485 } 486 return false; 487 } 488 489 bool AMDGPULibCalls::useNative(CallInst *aCI) { 490 Function *Callee = aCI->getCalledFunction(); 491 if (!Callee || aCI->isNoBuiltin()) 492 return false; 493 494 FuncInfo FInfo; 495 if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() || 496 FInfo.getPrefix() != AMDGPULibFunc::NOPFX || 497 getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) || 498 !(AllNative || useNativeFunc(FInfo.getName()))) { 499 return false; 500 } 501 502 if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS) 503 return sincosUseNative(aCI, FInfo); 504 505 FInfo.setPrefix(AMDGPULibFunc::NATIVE); 506 FunctionCallee F = getFunction(aCI->getModule(), FInfo); 507 if (!F) 508 return false; 509 510 aCI->setCalledFunction(F); 511 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 512 << " with native version"); 513 return true; 514 } 515 516 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe 517 // builtin, with appended type size and alignment arguments, where 2 or 4 518 // indicates the original number of arguments. The library has optimized version 519 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same 520 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N 521 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ..., 522 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4. 523 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 524 const FuncInfo &FInfo) { 525 auto *Callee = CI->getCalledFunction(); 526 if (!Callee->isDeclaration()) 527 return false; 528 529 assert(Callee->hasName() && "Invalid read_pipe/write_pipe function"); 530 auto *M = Callee->getParent(); 531 std::string Name = std::string(Callee->getName()); 532 auto NumArg = CI->arg_size(); 533 if (NumArg != 4 && NumArg != 6) 534 return false; 535 ConstantInt *PacketSize = 536 dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 2)); 537 ConstantInt *PacketAlign = 538 dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 1)); 539 if (!PacketSize || !PacketAlign) 540 return false; 541 542 unsigned Size = PacketSize->getZExtValue(); 543 Align Alignment = PacketAlign->getAlignValue(); 544 if (Alignment != Size) 545 return false; 546 547 unsigned PtrArgLoc = CI->arg_size() - 3; 548 Value *PtrArg = CI->getArgOperand(PtrArgLoc); 549 Type *PtrTy = PtrArg->getType(); 550 551 SmallVector<llvm::Type *, 6> ArgTys; 552 for (unsigned I = 0; I != PtrArgLoc; ++I) 553 ArgTys.push_back(CI->getArgOperand(I)->getType()); 554 ArgTys.push_back(PtrTy); 555 556 Name = Name + "_" + std::to_string(Size); 557 auto *FTy = FunctionType::get(Callee->getReturnType(), 558 ArrayRef<Type *>(ArgTys), false); 559 AMDGPULibFunc NewLibFunc(Name, FTy); 560 FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc); 561 if (!F) 562 return false; 563 564 SmallVector<Value *, 6> Args; 565 for (unsigned I = 0; I != PtrArgLoc; ++I) 566 Args.push_back(CI->getArgOperand(I)); 567 Args.push_back(PtrArg); 568 569 auto *NCI = B.CreateCall(F, Args); 570 NCI->setAttributes(CI->getAttributes()); 571 CI->replaceAllUsesWith(NCI); 572 CI->dropAllReferences(); 573 CI->eraseFromParent(); 574 575 return true; 576 } 577 578 static bool isKnownIntegral(const Value *V, const DataLayout &DL, 579 FastMathFlags FMF) { 580 if (isa<PoisonValue>(V)) 581 return true; 582 if (isa<UndefValue>(V)) 583 return false; 584 585 if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) 586 return CF->getValueAPF().isInteger(); 587 588 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); 589 const Constant *CV = dyn_cast<Constant>(V); 590 if (VFVTy && CV) { 591 unsigned NumElts = VFVTy->getNumElements(); 592 for (unsigned i = 0; i != NumElts; ++i) { 593 Constant *Elt = CV->getAggregateElement(i); 594 if (!Elt) 595 return false; 596 if (isa<PoisonValue>(Elt)) 597 continue; 598 599 const ConstantFP *CFP = dyn_cast<ConstantFP>(Elt); 600 if (!CFP || !CFP->getValue().isInteger()) 601 return false; 602 } 603 604 return true; 605 } 606 607 const Instruction *I = dyn_cast<Instruction>(V); 608 if (!I) 609 return false; 610 611 switch (I->getOpcode()) { 612 case Instruction::SIToFP: 613 case Instruction::UIToFP: 614 // TODO: Could check nofpclass(inf) on incoming argument 615 if (FMF.noInfs()) 616 return true; 617 618 // Need to check int size cannot produce infinity, which computeKnownFPClass 619 // knows how to do already. 620 return isKnownNeverInfinity(I, /*Depth=*/0, SimplifyQuery(DL)); 621 case Instruction::Call: { 622 const CallInst *CI = cast<CallInst>(I); 623 switch (CI->getIntrinsicID()) { 624 case Intrinsic::trunc: 625 case Intrinsic::floor: 626 case Intrinsic::ceil: 627 case Intrinsic::rint: 628 case Intrinsic::nearbyint: 629 case Intrinsic::round: 630 case Intrinsic::roundeven: 631 return (FMF.noInfs() && FMF.noNaNs()) || 632 isKnownNeverInfOrNaN(I, /*Depth=*/0, SimplifyQuery(DL)); 633 default: 634 break; 635 } 636 637 break; 638 } 639 default: 640 break; 641 } 642 643 return false; 644 } 645 646 // This function returns false if no change; return true otherwise. 647 bool AMDGPULibCalls::fold(CallInst *CI) { 648 Function *Callee = CI->getCalledFunction(); 649 // Ignore indirect calls. 650 if (!Callee || Callee->isIntrinsic() || CI->isNoBuiltin()) 651 return false; 652 653 FuncInfo FInfo; 654 if (!parseFunctionName(Callee->getName(), FInfo)) 655 return false; 656 657 // Further check the number of arguments to see if they match. 658 // TODO: Check calling convention matches too 659 if (!FInfo.isCompatibleSignature(CI->getFunctionType())) 660 return false; 661 662 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << '\n'); 663 664 if (TDOFold(CI, FInfo)) 665 return true; 666 667 IRBuilder<> B(CI); 668 if (CI->isStrictFP()) 669 B.setIsFPConstrained(true); 670 671 if (FPMathOperator *FPOp = dyn_cast<FPMathOperator>(CI)) { 672 // Under unsafe-math, evaluate calls if possible. 673 // According to Brian Sumner, we can do this for all f32 function calls 674 // using host's double function calls. 675 if (canIncreasePrecisionOfConstantFold(FPOp) && evaluateCall(CI, FInfo)) 676 return true; 677 678 // Copy fast flags from the original call. 679 FastMathFlags FMF = FPOp->getFastMathFlags(); 680 B.setFastMathFlags(FMF); 681 682 // Specialized optimizations for each function call. 683 // 684 // TODO: Handle native functions 685 switch (FInfo.getId()) { 686 case AMDGPULibFunc::EI_EXP: 687 if (FMF.none()) 688 return false; 689 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp, 690 FMF.approxFunc()); 691 case AMDGPULibFunc::EI_EXP2: 692 if (FMF.none()) 693 return false; 694 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp2, 695 FMF.approxFunc()); 696 case AMDGPULibFunc::EI_LOG: 697 if (FMF.none()) 698 return false; 699 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log, 700 FMF.approxFunc()); 701 case AMDGPULibFunc::EI_LOG2: 702 if (FMF.none()) 703 return false; 704 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log2, 705 FMF.approxFunc()); 706 case AMDGPULibFunc::EI_LOG10: 707 if (FMF.none()) 708 return false; 709 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log10, 710 FMF.approxFunc()); 711 case AMDGPULibFunc::EI_FMIN: 712 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::minnum, 713 true, true); 714 case AMDGPULibFunc::EI_FMAX: 715 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::maxnum, 716 true, true); 717 case AMDGPULibFunc::EI_FMA: 718 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fma, true, 719 true); 720 case AMDGPULibFunc::EI_MAD: 721 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fmuladd, 722 true, true); 723 case AMDGPULibFunc::EI_FABS: 724 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fabs, true, 725 true, true); 726 case AMDGPULibFunc::EI_COPYSIGN: 727 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::copysign, 728 true, true, true); 729 case AMDGPULibFunc::EI_FLOOR: 730 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::floor, true, 731 true); 732 case AMDGPULibFunc::EI_CEIL: 733 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::ceil, true, 734 true); 735 case AMDGPULibFunc::EI_TRUNC: 736 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::trunc, true, 737 true); 738 case AMDGPULibFunc::EI_RINT: 739 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::rint, true, 740 true); 741 case AMDGPULibFunc::EI_ROUND: 742 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::round, true, 743 true); 744 case AMDGPULibFunc::EI_LDEXP: { 745 if (!shouldReplaceLibcallWithIntrinsic(CI, true, true)) 746 return false; 747 748 Value *Arg1 = CI->getArgOperand(1); 749 if (VectorType *VecTy = dyn_cast<VectorType>(CI->getType()); 750 VecTy && !isa<VectorType>(Arg1->getType())) { 751 Value *SplatArg1 = B.CreateVectorSplat(VecTy->getElementCount(), Arg1); 752 CI->setArgOperand(1, SplatArg1); 753 } 754 755 CI->setCalledFunction(Intrinsic::getDeclaration( 756 CI->getModule(), Intrinsic::ldexp, 757 {CI->getType(), CI->getArgOperand(1)->getType()})); 758 return true; 759 } 760 case AMDGPULibFunc::EI_POW: { 761 Module *M = Callee->getParent(); 762 AMDGPULibFunc PowrInfo(AMDGPULibFunc::EI_POWR, FInfo); 763 FunctionCallee PowrFunc = getFunction(M, PowrInfo); 764 CallInst *Call = cast<CallInst>(FPOp); 765 766 // pow(x, y) -> powr(x, y) for x >= -0.0 767 // TODO: Account for flags on current call 768 if (PowrFunc && 769 cannotBeOrderedLessThanZero( 770 FPOp->getOperand(0), /*Depth=*/0, 771 SimplifyQuery(M->getDataLayout(), TLInfo, DT, AC, Call))) { 772 Call->setCalledFunction(PowrFunc); 773 return fold_pow(FPOp, B, PowrInfo) || true; 774 } 775 776 // pow(x, y) -> pown(x, y) for known integral y 777 if (isKnownIntegral(FPOp->getOperand(1), M->getDataLayout(), 778 FPOp->getFastMathFlags())) { 779 FunctionType *PownType = getPownType(CI->getFunctionType()); 780 AMDGPULibFunc PownInfo(AMDGPULibFunc::EI_POWN, PownType, true); 781 FunctionCallee PownFunc = getFunction(M, PownInfo); 782 if (PownFunc) { 783 // TODO: If the incoming integral value is an sitofp/uitofp, it won't 784 // fold out without a known range. We can probably take the source 785 // value directly. 786 Value *CastedArg = 787 B.CreateFPToSI(FPOp->getOperand(1), PownType->getParamType(1)); 788 // Have to drop any nofpclass attributes on the original call site. 789 Call->removeParamAttrs( 790 1, AttributeFuncs::typeIncompatible(CastedArg->getType())); 791 Call->setCalledFunction(PownFunc); 792 Call->setArgOperand(1, CastedArg); 793 return fold_pow(FPOp, B, PownInfo) || true; 794 } 795 } 796 797 return fold_pow(FPOp, B, FInfo); 798 } 799 case AMDGPULibFunc::EI_POWR: 800 case AMDGPULibFunc::EI_POWN: 801 return fold_pow(FPOp, B, FInfo); 802 case AMDGPULibFunc::EI_ROOTN: 803 return fold_rootn(FPOp, B, FInfo); 804 case AMDGPULibFunc::EI_SQRT: 805 // TODO: Allow with strictfp + constrained intrinsic 806 return tryReplaceLibcallWithSimpleIntrinsic( 807 B, CI, Intrinsic::sqrt, true, true, /*AllowStrictFP=*/false); 808 case AMDGPULibFunc::EI_COS: 809 case AMDGPULibFunc::EI_SIN: 810 return fold_sincos(FPOp, B, FInfo); 811 default: 812 break; 813 } 814 } else { 815 // Specialized optimizations for each function call 816 switch (FInfo.getId()) { 817 case AMDGPULibFunc::EI_READ_PIPE_2: 818 case AMDGPULibFunc::EI_READ_PIPE_4: 819 case AMDGPULibFunc::EI_WRITE_PIPE_2: 820 case AMDGPULibFunc::EI_WRITE_PIPE_4: 821 return fold_read_write_pipe(CI, B, FInfo); 822 default: 823 break; 824 } 825 } 826 827 return false; 828 } 829 830 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) { 831 // Table-Driven optimization 832 const TableRef tr = getOptTable(FInfo.getId()); 833 if (tr.empty()) 834 return false; 835 836 int const sz = (int)tr.size(); 837 Value *opr0 = CI->getArgOperand(0); 838 839 if (getVecSize(FInfo) > 1) { 840 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) { 841 SmallVector<double, 0> DVal; 842 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) { 843 ConstantFP *eltval = dyn_cast<ConstantFP>( 844 CV->getElementAsConstant((unsigned)eltNo)); 845 assert(eltval && "Non-FP arguments in math function!"); 846 bool found = false; 847 for (int i=0; i < sz; ++i) { 848 if (eltval->isExactlyValue(tr[i].input)) { 849 DVal.push_back(tr[i].result); 850 found = true; 851 break; 852 } 853 } 854 if (!found) { 855 // This vector constants not handled yet. 856 return false; 857 } 858 } 859 LLVMContext &context = CI->getParent()->getParent()->getContext(); 860 Constant *nval; 861 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 862 SmallVector<float, 0> FVal; 863 for (unsigned i = 0; i < DVal.size(); ++i) { 864 FVal.push_back((float)DVal[i]); 865 } 866 ArrayRef<float> tmp(FVal); 867 nval = ConstantDataVector::get(context, tmp); 868 } else { // F64 869 ArrayRef<double> tmp(DVal); 870 nval = ConstantDataVector::get(context, tmp); 871 } 872 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 873 replaceCall(CI, nval); 874 return true; 875 } 876 } else { 877 // Scalar version 878 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) { 879 for (int i = 0; i < sz; ++i) { 880 if (CF->isExactlyValue(tr[i].input)) { 881 Value *nval = ConstantFP::get(CF->getType(), tr[i].result); 882 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 883 replaceCall(CI, nval); 884 return true; 885 } 886 } 887 } 888 } 889 890 return false; 891 } 892 893 namespace llvm { 894 static double log2(double V) { 895 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L 896 return ::log2(V); 897 #else 898 return log(V) / numbers::ln2; 899 #endif 900 } 901 } 902 903 bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, 904 const FuncInfo &FInfo) { 905 assert((FInfo.getId() == AMDGPULibFunc::EI_POW || 906 FInfo.getId() == AMDGPULibFunc::EI_POWR || 907 FInfo.getId() == AMDGPULibFunc::EI_POWN) && 908 "fold_pow: encounter a wrong function call"); 909 910 Module *M = B.GetInsertBlock()->getModule(); 911 Type *eltType = FPOp->getType()->getScalarType(); 912 Value *opr0 = FPOp->getOperand(0); 913 Value *opr1 = FPOp->getOperand(1); 914 915 const APFloat *CF = nullptr; 916 const APInt *CINT = nullptr; 917 if (!match(opr1, m_APFloatAllowPoison(CF))) 918 match(opr1, m_APIntAllowPoison(CINT)); 919 920 // 0x1111111 means that we don't do anything for this call. 921 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111); 922 923 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0)) { 924 // pow/powr/pown(x, 0) == 1 925 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1\n"); 926 Constant *cnval = ConstantFP::get(eltType, 1.0); 927 if (getVecSize(FInfo) > 1) { 928 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 929 } 930 replaceCall(FPOp, cnval); 931 return true; 932 } 933 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) { 934 // pow/powr/pown(x, 1.0) = x 935 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n"); 936 replaceCall(FPOp, opr0); 937 return true; 938 } 939 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) { 940 // pow/powr/pown(x, 2.0) = x*x 941 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << " * " 942 << *opr0 << "\n"); 943 Value *nval = B.CreateFMul(opr0, opr0, "__pow2"); 944 replaceCall(FPOp, nval); 945 return true; 946 } 947 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) { 948 // pow/powr/pown(x, -1.0) = 1.0/x 949 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1 / " << *opr0 << "\n"); 950 Constant *cnval = ConstantFP::get(eltType, 1.0); 951 if (getVecSize(FInfo) > 1) { 952 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 953 } 954 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip"); 955 replaceCall(FPOp, nval); 956 return true; 957 } 958 959 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) { 960 // pow[r](x, [-]0.5) = sqrt(x) 961 bool issqrt = CF->isExactlyValue(0.5); 962 if (FunctionCallee FPExpr = 963 getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT 964 : AMDGPULibFunc::EI_RSQRT, 965 FInfo))) { 966 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << FInfo.getName() 967 << '(' << *opr0 << ")\n"); 968 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt" 969 : "__pow2rsqrt"); 970 replaceCall(FPOp, nval); 971 return true; 972 } 973 } 974 975 if (!isUnsafeFiniteOnlyMath(FPOp)) 976 return false; 977 978 // Unsafe Math optimization 979 980 // Remember that ci_opr1 is set if opr1 is integral 981 if (CF) { 982 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32) 983 ? (double)CF->convertToFloat() 984 : CF->convertToDouble(); 985 int ival = (int)dval; 986 if ((double)ival == dval) { 987 ci_opr1 = ival; 988 } else 989 ci_opr1 = 0x11111111; 990 } 991 992 // pow/powr/pown(x, c) = [1/](x*x*..x); where 993 // trunc(c) == c && the number of x == c && |c| <= 12 994 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1; 995 if (abs_opr1 <= 12) { 996 Constant *cnval; 997 Value *nval; 998 if (abs_opr1 == 0) { 999 cnval = ConstantFP::get(eltType, 1.0); 1000 if (getVecSize(FInfo) > 1) { 1001 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 1002 } 1003 nval = cnval; 1004 } else { 1005 Value *valx2 = nullptr; 1006 nval = nullptr; 1007 while (abs_opr1 > 0) { 1008 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0; 1009 if (abs_opr1 & 1) { 1010 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2; 1011 } 1012 abs_opr1 >>= 1; 1013 } 1014 } 1015 1016 if (ci_opr1 < 0) { 1017 cnval = ConstantFP::get(eltType, 1.0); 1018 if (getVecSize(FInfo) > 1) { 1019 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 1020 } 1021 nval = B.CreateFDiv(cnval, nval, "__1powprod"); 1022 } 1023 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " 1024 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0 1025 << ")\n"); 1026 replaceCall(FPOp, nval); 1027 return true; 1028 } 1029 1030 // If we should use the generic intrinsic instead of emitting a libcall 1031 const bool ShouldUseIntrinsic = eltType->isFloatTy() || eltType->isHalfTy(); 1032 1033 // powr ---> exp2(y * log2(x)) 1034 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31)) 1035 FunctionCallee ExpExpr; 1036 if (ShouldUseIntrinsic) 1037 ExpExpr = Intrinsic::getDeclaration(M, Intrinsic::exp2, {FPOp->getType()}); 1038 else { 1039 ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo)); 1040 if (!ExpExpr) 1041 return false; 1042 } 1043 1044 bool needlog = false; 1045 bool needabs = false; 1046 bool needcopysign = false; 1047 Constant *cnval = nullptr; 1048 if (getVecSize(FInfo) == 1) { 1049 CF = nullptr; 1050 match(opr0, m_APFloatAllowPoison(CF)); 1051 1052 if (CF) { 1053 double V = (getArgType(FInfo) == AMDGPULibFunc::F32) 1054 ? (double)CF->convertToFloat() 1055 : CF->convertToDouble(); 1056 1057 V = log2(std::abs(V)); 1058 cnval = ConstantFP::get(eltType, V); 1059 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) && 1060 CF->isNegative(); 1061 } else { 1062 needlog = true; 1063 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR; 1064 } 1065 } else { 1066 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0); 1067 1068 if (!CDV) { 1069 needlog = true; 1070 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR; 1071 } else { 1072 assert ((int)CDV->getNumElements() == getVecSize(FInfo) && 1073 "Wrong vector size detected"); 1074 1075 SmallVector<double, 0> DVal; 1076 for (int i=0; i < getVecSize(FInfo); ++i) { 1077 double V = CDV->getElementAsAPFloat(i).convertToDouble(); 1078 if (V < 0.0) needcopysign = true; 1079 V = log2(std::abs(V)); 1080 DVal.push_back(V); 1081 } 1082 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1083 SmallVector<float, 0> FVal; 1084 for (unsigned i=0; i < DVal.size(); ++i) { 1085 FVal.push_back((float)DVal[i]); 1086 } 1087 ArrayRef<float> tmp(FVal); 1088 cnval = ConstantDataVector::get(M->getContext(), tmp); 1089 } else { 1090 ArrayRef<double> tmp(DVal); 1091 cnval = ConstantDataVector::get(M->getContext(), tmp); 1092 } 1093 } 1094 } 1095 1096 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) { 1097 // We cannot handle corner cases for a general pow() function, give up 1098 // unless y is a constant integral value. Then proceed as if it were pown. 1099 if (!isKnownIntegral(opr1, M->getDataLayout(), FPOp->getFastMathFlags())) 1100 return false; 1101 } 1102 1103 Value *nval; 1104 if (needabs) { 1105 nval = B.CreateUnaryIntrinsic(Intrinsic::fabs, opr0, nullptr, "__fabs"); 1106 } else { 1107 nval = cnval ? cnval : opr0; 1108 } 1109 if (needlog) { 1110 FunctionCallee LogExpr; 1111 if (ShouldUseIntrinsic) { 1112 LogExpr = 1113 Intrinsic::getDeclaration(M, Intrinsic::log2, {FPOp->getType()}); 1114 } else { 1115 LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo)); 1116 if (!LogExpr) 1117 return false; 1118 } 1119 1120 nval = CreateCallEx(B,LogExpr, nval, "__log2"); 1121 } 1122 1123 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) { 1124 // convert int(32) to fp(f32 or f64) 1125 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F"); 1126 } 1127 nval = B.CreateFMul(opr1, nval, "__ylogx"); 1128 nval = CreateCallEx(B,ExpExpr, nval, "__exp2"); 1129 1130 if (needcopysign) { 1131 Value *opr_n; 1132 Type* rTy = opr0->getType(); 1133 Type* nTyS = B.getIntNTy(eltType->getPrimitiveSizeInBits()); 1134 Type *nTy = nTyS; 1135 if (const auto *vTy = dyn_cast<FixedVectorType>(rTy)) 1136 nTy = FixedVectorType::get(nTyS, vTy); 1137 unsigned size = nTy->getScalarSizeInBits(); 1138 opr_n = FPOp->getOperand(1); 1139 if (opr_n->getType()->isIntegerTy()) 1140 opr_n = B.CreateZExtOrTrunc(opr_n, nTy, "__ytou"); 1141 else 1142 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou"); 1143 1144 Value *sign = B.CreateShl(opr_n, size-1, "__yeven"); 1145 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign"); 1146 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign); 1147 nval = B.CreateBitCast(nval, opr0->getType()); 1148 } 1149 1150 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " 1151 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n"); 1152 replaceCall(FPOp, nval); 1153 1154 return true; 1155 } 1156 1157 bool AMDGPULibCalls::fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B, 1158 const FuncInfo &FInfo) { 1159 Value *opr0 = FPOp->getOperand(0); 1160 Value *opr1 = FPOp->getOperand(1); 1161 1162 const APInt *CINT = nullptr; 1163 if (!match(opr1, m_APIntAllowPoison(CINT))) 1164 return false; 1165 1166 int ci_opr1 = (int)CINT->getSExtValue(); 1167 if (ci_opr1 == 1) { // rootn(x, 1) = x 1168 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n"); 1169 replaceCall(FPOp, opr0); 1170 return true; 1171 } 1172 1173 Module *M = B.GetInsertBlock()->getModule(); 1174 if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x) 1175 if (FunctionCallee FPExpr = 1176 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) { 1177 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> sqrt(" << *opr0 1178 << ")\n"); 1179 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt"); 1180 replaceCall(FPOp, nval); 1181 return true; 1182 } 1183 } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x) 1184 if (FunctionCallee FPExpr = 1185 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) { 1186 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> cbrt(" << *opr0 1187 << ")\n"); 1188 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt"); 1189 replaceCall(FPOp, nval); 1190 return true; 1191 } 1192 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x 1193 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1.0 / " << *opr0 << "\n"); 1194 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0), 1195 opr0, 1196 "__rootn2div"); 1197 replaceCall(FPOp, nval); 1198 return true; 1199 } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x) 1200 if (FunctionCallee FPExpr = 1201 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) { 1202 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> rsqrt(" << *opr0 1203 << ")\n"); 1204 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt"); 1205 replaceCall(FPOp, nval); 1206 return true; 1207 } 1208 } 1209 return false; 1210 } 1211 1212 // Get a scalar native builtin single argument FP function 1213 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M, 1214 const FuncInfo &FInfo) { 1215 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId())) 1216 return nullptr; 1217 FuncInfo nf = FInfo; 1218 nf.setPrefix(AMDGPULibFunc::NATIVE); 1219 return getFunction(M, nf); 1220 } 1221 1222 // Some library calls are just wrappers around llvm intrinsics, but compiled 1223 // conservatively. Preserve the flags from the original call site by 1224 // substituting them with direct calls with all the flags. 1225 bool AMDGPULibCalls::shouldReplaceLibcallWithIntrinsic(const CallInst *CI, 1226 bool AllowMinSizeF32, 1227 bool AllowF64, 1228 bool AllowStrictFP) { 1229 Type *FltTy = CI->getType()->getScalarType(); 1230 const bool IsF32 = FltTy->isFloatTy(); 1231 1232 // f64 intrinsics aren't implemented for most operations. 1233 if (!IsF32 && !FltTy->isHalfTy() && (!AllowF64 || !FltTy->isDoubleTy())) 1234 return false; 1235 1236 // We're implicitly inlining by replacing the libcall with the intrinsic, so 1237 // don't do it for noinline call sites. 1238 if (CI->isNoInline()) 1239 return false; 1240 1241 const Function *ParentF = CI->getFunction(); 1242 // TODO: Handle strictfp 1243 if (!AllowStrictFP && ParentF->hasFnAttribute(Attribute::StrictFP)) 1244 return false; 1245 1246 if (IsF32 && !AllowMinSizeF32 && ParentF->hasMinSize()) 1247 return false; 1248 return true; 1249 } 1250 1251 void AMDGPULibCalls::replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, 1252 CallInst *CI, 1253 Intrinsic::ID IntrID) { 1254 if (CI->arg_size() == 2) { 1255 Value *Arg0 = CI->getArgOperand(0); 1256 Value *Arg1 = CI->getArgOperand(1); 1257 VectorType *Arg0VecTy = dyn_cast<VectorType>(Arg0->getType()); 1258 VectorType *Arg1VecTy = dyn_cast<VectorType>(Arg1->getType()); 1259 if (Arg0VecTy && !Arg1VecTy) { 1260 Value *SplatRHS = B.CreateVectorSplat(Arg0VecTy->getElementCount(), Arg1); 1261 CI->setArgOperand(1, SplatRHS); 1262 } else if (!Arg0VecTy && Arg1VecTy) { 1263 Value *SplatLHS = B.CreateVectorSplat(Arg1VecTy->getElementCount(), Arg0); 1264 CI->setArgOperand(0, SplatLHS); 1265 } 1266 } 1267 1268 CI->setCalledFunction( 1269 Intrinsic::getDeclaration(CI->getModule(), IntrID, {CI->getType()})); 1270 } 1271 1272 bool AMDGPULibCalls::tryReplaceLibcallWithSimpleIntrinsic( 1273 IRBuilder<> &B, CallInst *CI, Intrinsic::ID IntrID, bool AllowMinSizeF32, 1274 bool AllowF64, bool AllowStrictFP) { 1275 if (!shouldReplaceLibcallWithIntrinsic(CI, AllowMinSizeF32, AllowF64, 1276 AllowStrictFP)) 1277 return false; 1278 replaceLibCallWithSimpleIntrinsic(B, CI, IntrID); 1279 return true; 1280 } 1281 1282 std::tuple<Value *, Value *, Value *> 1283 AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B, 1284 FunctionCallee Fsincos) { 1285 DebugLoc DL = B.getCurrentDebugLocation(); 1286 Function *F = B.GetInsertBlock()->getParent(); 1287 B.SetInsertPointPastAllocas(F); 1288 1289 AllocaInst *Alloc = B.CreateAlloca(Arg->getType(), nullptr, "__sincos_"); 1290 1291 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) { 1292 // If the argument is an instruction, it must dominate all uses so put our 1293 // sincos call there. Otherwise, right after the allocas works well enough 1294 // if it's an argument or constant. 1295 1296 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator()); 1297 1298 // SetInsertPoint unwelcomely always tries to set the debug loc. 1299 B.SetCurrentDebugLocation(DL); 1300 } 1301 1302 Type *CosPtrTy = Fsincos.getFunctionType()->getParamType(1); 1303 1304 // The allocaInst allocates the memory in private address space. This need 1305 // to be addrspacecasted to point to the address space of cos pointer type. 1306 // In OpenCL 2.0 this is generic, while in 1.2 that is private. 1307 Value *CastAlloc = B.CreateAddrSpaceCast(Alloc, CosPtrTy); 1308 1309 CallInst *SinCos = CreateCallEx2(B, Fsincos, Arg, CastAlloc); 1310 1311 // TODO: Is it worth trying to preserve the location for the cos calls for the 1312 // load? 1313 1314 LoadInst *LoadCos = B.CreateLoad(Alloc->getAllocatedType(), Alloc); 1315 return {SinCos, LoadCos, SinCos}; 1316 } 1317 1318 // fold sin, cos -> sincos. 1319 bool AMDGPULibCalls::fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, 1320 const FuncInfo &fInfo) { 1321 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN || 1322 fInfo.getId() == AMDGPULibFunc::EI_COS); 1323 1324 if ((getArgType(fInfo) != AMDGPULibFunc::F32 && 1325 getArgType(fInfo) != AMDGPULibFunc::F64) || 1326 fInfo.getPrefix() != AMDGPULibFunc::NOPFX) 1327 return false; 1328 1329 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN; 1330 1331 Value *CArgVal = FPOp->getOperand(0); 1332 CallInst *CI = cast<CallInst>(FPOp); 1333 1334 Function *F = B.GetInsertBlock()->getParent(); 1335 Module *M = F->getParent(); 1336 1337 // Merge the sin and cos. For OpenCL 2.0, there may only be a generic pointer 1338 // implementation. Prefer the private form if available. 1339 AMDGPULibFunc SinCosLibFuncPrivate(AMDGPULibFunc::EI_SINCOS, fInfo); 1340 SinCosLibFuncPrivate.getLeads()[0].PtrKind = 1341 AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::PRIVATE_ADDRESS); 1342 1343 AMDGPULibFunc SinCosLibFuncGeneric(AMDGPULibFunc::EI_SINCOS, fInfo); 1344 SinCosLibFuncGeneric.getLeads()[0].PtrKind = 1345 AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS); 1346 1347 FunctionCallee FSinCosPrivate = getFunction(M, SinCosLibFuncPrivate); 1348 FunctionCallee FSinCosGeneric = getFunction(M, SinCosLibFuncGeneric); 1349 FunctionCallee FSinCos = FSinCosPrivate ? FSinCosPrivate : FSinCosGeneric; 1350 if (!FSinCos) 1351 return false; 1352 1353 SmallVector<CallInst *> SinCalls; 1354 SmallVector<CallInst *> CosCalls; 1355 SmallVector<CallInst *> SinCosCalls; 1356 FuncInfo PartnerInfo(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN, 1357 fInfo); 1358 const std::string PairName = PartnerInfo.mangle(); 1359 1360 StringRef SinName = isSin ? CI->getCalledFunction()->getName() : PairName; 1361 StringRef CosName = isSin ? PairName : CI->getCalledFunction()->getName(); 1362 const std::string SinCosPrivateName = SinCosLibFuncPrivate.mangle(); 1363 const std::string SinCosGenericName = SinCosLibFuncGeneric.mangle(); 1364 1365 // Intersect the two sets of flags. 1366 FastMathFlags FMF = FPOp->getFastMathFlags(); 1367 MDNode *FPMath = CI->getMetadata(LLVMContext::MD_fpmath); 1368 1369 SmallVector<DILocation *> MergeDbgLocs = {CI->getDebugLoc()}; 1370 1371 for (User* U : CArgVal->users()) { 1372 CallInst *XI = dyn_cast<CallInst>(U); 1373 if (!XI || XI->getFunction() != F || XI->isNoBuiltin()) 1374 continue; 1375 1376 Function *UCallee = XI->getCalledFunction(); 1377 if (!UCallee) 1378 continue; 1379 1380 bool Handled = true; 1381 1382 if (UCallee->getName() == SinName) 1383 SinCalls.push_back(XI); 1384 else if (UCallee->getName() == CosName) 1385 CosCalls.push_back(XI); 1386 else if (UCallee->getName() == SinCosPrivateName || 1387 UCallee->getName() == SinCosGenericName) 1388 SinCosCalls.push_back(XI); 1389 else 1390 Handled = false; 1391 1392 if (Handled) { 1393 MergeDbgLocs.push_back(XI->getDebugLoc()); 1394 auto *OtherOp = cast<FPMathOperator>(XI); 1395 FMF &= OtherOp->getFastMathFlags(); 1396 FPMath = MDNode::getMostGenericFPMath( 1397 FPMath, XI->getMetadata(LLVMContext::MD_fpmath)); 1398 } 1399 } 1400 1401 if (SinCalls.empty() || CosCalls.empty()) 1402 return false; 1403 1404 B.setFastMathFlags(FMF); 1405 B.setDefaultFPMathTag(FPMath); 1406 DILocation *DbgLoc = DILocation::getMergedLocations(MergeDbgLocs); 1407 B.SetCurrentDebugLocation(DbgLoc); 1408 1409 auto [Sin, Cos, SinCos] = insertSinCos(CArgVal, FMF, B, FSinCos); 1410 1411 auto replaceTrigInsts = [](ArrayRef<CallInst *> Calls, Value *Res) { 1412 for (CallInst *C : Calls) 1413 C->replaceAllUsesWith(Res); 1414 1415 // Leave the other dead instructions to avoid clobbering iterators. 1416 }; 1417 1418 replaceTrigInsts(SinCalls, Sin); 1419 replaceTrigInsts(CosCalls, Cos); 1420 replaceTrigInsts(SinCosCalls, SinCos); 1421 1422 // It's safe to delete the original now. 1423 CI->eraseFromParent(); 1424 return true; 1425 } 1426 1427 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0, 1428 double &Res1, Constant *copr0, 1429 Constant *copr1) { 1430 // By default, opr0/opr1/opr3 holds values of float/double type. 1431 // If they are not float/double, each function has to its 1432 // operand separately. 1433 double opr0 = 0.0, opr1 = 0.0; 1434 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0); 1435 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1); 1436 if (fpopr0) { 1437 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1438 ? fpopr0->getValueAPF().convertToDouble() 1439 : (double)fpopr0->getValueAPF().convertToFloat(); 1440 } 1441 1442 if (fpopr1) { 1443 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1444 ? fpopr1->getValueAPF().convertToDouble() 1445 : (double)fpopr1->getValueAPF().convertToFloat(); 1446 } 1447 1448 switch (FInfo.getId()) { 1449 default : return false; 1450 1451 case AMDGPULibFunc::EI_ACOS: 1452 Res0 = acos(opr0); 1453 return true; 1454 1455 case AMDGPULibFunc::EI_ACOSH: 1456 // acosh(x) == log(x + sqrt(x*x - 1)) 1457 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0)); 1458 return true; 1459 1460 case AMDGPULibFunc::EI_ACOSPI: 1461 Res0 = acos(opr0) / MATH_PI; 1462 return true; 1463 1464 case AMDGPULibFunc::EI_ASIN: 1465 Res0 = asin(opr0); 1466 return true; 1467 1468 case AMDGPULibFunc::EI_ASINH: 1469 // asinh(x) == log(x + sqrt(x*x + 1)) 1470 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0)); 1471 return true; 1472 1473 case AMDGPULibFunc::EI_ASINPI: 1474 Res0 = asin(opr0) / MATH_PI; 1475 return true; 1476 1477 case AMDGPULibFunc::EI_ATAN: 1478 Res0 = atan(opr0); 1479 return true; 1480 1481 case AMDGPULibFunc::EI_ATANH: 1482 // atanh(x) == (log(x+1) - log(x-1))/2; 1483 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0; 1484 return true; 1485 1486 case AMDGPULibFunc::EI_ATANPI: 1487 Res0 = atan(opr0) / MATH_PI; 1488 return true; 1489 1490 case AMDGPULibFunc::EI_CBRT: 1491 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0); 1492 return true; 1493 1494 case AMDGPULibFunc::EI_COS: 1495 Res0 = cos(opr0); 1496 return true; 1497 1498 case AMDGPULibFunc::EI_COSH: 1499 Res0 = cosh(opr0); 1500 return true; 1501 1502 case AMDGPULibFunc::EI_COSPI: 1503 Res0 = cos(MATH_PI * opr0); 1504 return true; 1505 1506 case AMDGPULibFunc::EI_EXP: 1507 Res0 = exp(opr0); 1508 return true; 1509 1510 case AMDGPULibFunc::EI_EXP2: 1511 Res0 = pow(2.0, opr0); 1512 return true; 1513 1514 case AMDGPULibFunc::EI_EXP10: 1515 Res0 = pow(10.0, opr0); 1516 return true; 1517 1518 case AMDGPULibFunc::EI_LOG: 1519 Res0 = log(opr0); 1520 return true; 1521 1522 case AMDGPULibFunc::EI_LOG2: 1523 Res0 = log(opr0) / log(2.0); 1524 return true; 1525 1526 case AMDGPULibFunc::EI_LOG10: 1527 Res0 = log(opr0) / log(10.0); 1528 return true; 1529 1530 case AMDGPULibFunc::EI_RSQRT: 1531 Res0 = 1.0 / sqrt(opr0); 1532 return true; 1533 1534 case AMDGPULibFunc::EI_SIN: 1535 Res0 = sin(opr0); 1536 return true; 1537 1538 case AMDGPULibFunc::EI_SINH: 1539 Res0 = sinh(opr0); 1540 return true; 1541 1542 case AMDGPULibFunc::EI_SINPI: 1543 Res0 = sin(MATH_PI * opr0); 1544 return true; 1545 1546 case AMDGPULibFunc::EI_TAN: 1547 Res0 = tan(opr0); 1548 return true; 1549 1550 case AMDGPULibFunc::EI_TANH: 1551 Res0 = tanh(opr0); 1552 return true; 1553 1554 case AMDGPULibFunc::EI_TANPI: 1555 Res0 = tan(MATH_PI * opr0); 1556 return true; 1557 1558 // two-arg functions 1559 case AMDGPULibFunc::EI_POW: 1560 case AMDGPULibFunc::EI_POWR: 1561 Res0 = pow(opr0, opr1); 1562 return true; 1563 1564 case AMDGPULibFunc::EI_POWN: { 1565 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1566 double val = (double)iopr1->getSExtValue(); 1567 Res0 = pow(opr0, val); 1568 return true; 1569 } 1570 return false; 1571 } 1572 1573 case AMDGPULibFunc::EI_ROOTN: { 1574 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1575 double val = (double)iopr1->getSExtValue(); 1576 Res0 = pow(opr0, 1.0 / val); 1577 return true; 1578 } 1579 return false; 1580 } 1581 1582 // with ptr arg 1583 case AMDGPULibFunc::EI_SINCOS: 1584 Res0 = sin(opr0); 1585 Res1 = cos(opr0); 1586 return true; 1587 } 1588 1589 return false; 1590 } 1591 1592 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) { 1593 int numArgs = (int)aCI->arg_size(); 1594 if (numArgs > 3) 1595 return false; 1596 1597 Constant *copr0 = nullptr; 1598 Constant *copr1 = nullptr; 1599 if (numArgs > 0) { 1600 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr) 1601 return false; 1602 } 1603 1604 if (numArgs > 1) { 1605 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) { 1606 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS) 1607 return false; 1608 } 1609 } 1610 1611 // At this point, all arguments to aCI are constants. 1612 1613 // max vector size is 16, and sincos will generate two results. 1614 double DVal0[16], DVal1[16]; 1615 int FuncVecSize = getVecSize(FInfo); 1616 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS); 1617 if (FuncVecSize == 1) { 1618 if (!evaluateScalarMathFunc(FInfo, DVal0[0], DVal1[0], copr0, copr1)) { 1619 return false; 1620 } 1621 } else { 1622 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0); 1623 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1); 1624 for (int i = 0; i < FuncVecSize; ++i) { 1625 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr; 1626 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr; 1627 if (!evaluateScalarMathFunc(FInfo, DVal0[i], DVal1[i], celt0, celt1)) { 1628 return false; 1629 } 1630 } 1631 } 1632 1633 LLVMContext &context = aCI->getContext(); 1634 Constant *nval0, *nval1; 1635 if (FuncVecSize == 1) { 1636 nval0 = ConstantFP::get(aCI->getType(), DVal0[0]); 1637 if (hasTwoResults) 1638 nval1 = ConstantFP::get(aCI->getType(), DVal1[0]); 1639 } else { 1640 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1641 SmallVector <float, 0> FVal0, FVal1; 1642 for (int i = 0; i < FuncVecSize; ++i) 1643 FVal0.push_back((float)DVal0[i]); 1644 ArrayRef<float> tmp0(FVal0); 1645 nval0 = ConstantDataVector::get(context, tmp0); 1646 if (hasTwoResults) { 1647 for (int i = 0; i < FuncVecSize; ++i) 1648 FVal1.push_back((float)DVal1[i]); 1649 ArrayRef<float> tmp1(FVal1); 1650 nval1 = ConstantDataVector::get(context, tmp1); 1651 } 1652 } else { 1653 ArrayRef<double> tmp0(DVal0); 1654 nval0 = ConstantDataVector::get(context, tmp0); 1655 if (hasTwoResults) { 1656 ArrayRef<double> tmp1(DVal1); 1657 nval1 = ConstantDataVector::get(context, tmp1); 1658 } 1659 } 1660 } 1661 1662 if (hasTwoResults) { 1663 // sincos 1664 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS && 1665 "math function with ptr arg not supported yet"); 1666 new StoreInst(nval1, aCI->getArgOperand(1), aCI->getIterator()); 1667 } 1668 1669 replaceCall(aCI, nval0); 1670 return true; 1671 } 1672 1673 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F, 1674 FunctionAnalysisManager &AM) { 1675 AMDGPULibCalls Simplifier; 1676 Simplifier.initNativeFuncs(); 1677 Simplifier.initFunction(F, AM); 1678 1679 bool Changed = false; 1680 1681 LLVM_DEBUG(dbgs() << "AMDIC: process function "; 1682 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';); 1683 1684 for (auto &BB : F) { 1685 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) { 1686 // Ignore non-calls. 1687 CallInst *CI = dyn_cast<CallInst>(I); 1688 ++I; 1689 1690 if (CI) { 1691 if (Simplifier.fold(CI)) 1692 Changed = true; 1693 } 1694 } 1695 } 1696 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 1697 } 1698 1699 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F, 1700 FunctionAnalysisManager &AM) { 1701 if (UseNative.empty()) 1702 return PreservedAnalyses::all(); 1703 1704 AMDGPULibCalls Simplifier; 1705 Simplifier.initNativeFuncs(); 1706 Simplifier.initFunction(F, AM); 1707 1708 bool Changed = false; 1709 for (auto &BB : F) { 1710 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) { 1711 // Ignore non-calls. 1712 CallInst *CI = dyn_cast<CallInst>(I); 1713 ++I; 1714 if (CI && Simplifier.useNative(CI)) 1715 Changed = true; 1716 } 1717 } 1718 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 1719 } 1720