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/MDBuilder.h" 26 #include "llvm/IR/PatternMatch.h" 27 #include "llvm/InitializePasses.h" 28 #include <cmath> 29 30 #define DEBUG_TYPE "amdgpu-simplifylib" 31 32 using namespace llvm; 33 using namespace llvm::PatternMatch; 34 35 static cl::opt<bool> EnablePreLink("amdgpu-prelink", 36 cl::desc("Enable pre-link mode optimizations"), 37 cl::init(false), 38 cl::Hidden); 39 40 static cl::list<std::string> UseNative("amdgpu-use-native", 41 cl::desc("Comma separated list of functions to replace with native, or all"), 42 cl::CommaSeparated, cl::ValueOptional, 43 cl::Hidden); 44 45 #define MATH_PI numbers::pi 46 #define MATH_E numbers::e 47 #define MATH_SQRT2 numbers::sqrt2 48 #define MATH_SQRT1_2 numbers::inv_sqrt2 49 50 namespace llvm { 51 52 class AMDGPULibCalls { 53 private: 54 const TargetLibraryInfo *TLInfo = nullptr; 55 AssumptionCache *AC = nullptr; 56 DominatorTree *DT = nullptr; 57 58 using FuncInfo = llvm::AMDGPULibFunc; 59 60 bool UnsafeFPMath = false; 61 62 // -fuse-native. 63 bool AllNative = false; 64 65 bool useNativeFunc(const StringRef F) const; 66 67 // Return a pointer (pointer expr) to the function if function definition with 68 // "FuncName" exists. It may create a new function prototype in pre-link mode. 69 FunctionCallee getFunction(Module *M, const FuncInfo &fInfo); 70 71 bool parseFunctionName(const StringRef &FMangledName, FuncInfo &FInfo); 72 73 bool TDOFold(CallInst *CI, const FuncInfo &FInfo); 74 75 /* Specialized optimizations */ 76 77 // pow/powr/pown 78 bool fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo); 79 80 // rootn 81 bool fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo); 82 83 // -fuse-native for sincos 84 bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo); 85 86 // evaluate calls if calls' arguments are constants. 87 bool evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0, double &Res1, 88 Constant *copr0, Constant *copr1); 89 bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo); 90 91 /// Insert a value to sincos function \p Fsincos. Returns (value of sin, value 92 /// of cos, sincos call). 93 std::tuple<Value *, Value *, Value *> insertSinCos(Value *Arg, 94 FastMathFlags FMF, 95 IRBuilder<> &B, 96 FunctionCallee Fsincos); 97 98 // sin/cos 99 bool fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo); 100 101 // __read_pipe/__write_pipe 102 bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 103 const FuncInfo &FInfo); 104 105 // Get a scalar native builtin single argument FP function 106 FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo); 107 108 /// Substitute a call to a known libcall with an intrinsic call. If \p 109 /// AllowMinSize is true, allow the replacement in a minsize function. 110 bool shouldReplaceLibcallWithIntrinsic(const CallInst *CI, 111 bool AllowMinSizeF32 = false, 112 bool AllowF64 = false, 113 bool AllowStrictFP = false); 114 void replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI, 115 Intrinsic::ID IntrID); 116 117 bool tryReplaceLibcallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI, 118 Intrinsic::ID IntrID, 119 bool AllowMinSizeF32 = false, 120 bool AllowF64 = false, 121 bool AllowStrictFP = false); 122 123 protected: 124 bool isUnsafeMath(const FPMathOperator *FPOp) const; 125 bool isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const; 126 127 bool canIncreasePrecisionOfConstantFold(const FPMathOperator *FPOp) const; 128 129 static void replaceCall(Instruction *I, Value *With) { 130 I->replaceAllUsesWith(With); 131 I->eraseFromParent(); 132 } 133 134 static void replaceCall(FPMathOperator *I, Value *With) { 135 replaceCall(cast<Instruction>(I), With); 136 } 137 138 public: 139 AMDGPULibCalls() = default; 140 141 bool fold(CallInst *CI); 142 143 void initFunction(Function &F, FunctionAnalysisManager &FAM); 144 void initNativeFuncs(); 145 146 // Replace a normal math function call with that native version 147 bool useNative(CallInst *CI); 148 }; 149 150 } // end namespace llvm 151 152 template <typename IRB> 153 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg, 154 const Twine &Name = "") { 155 CallInst *R = B.CreateCall(Callee, Arg, Name); 156 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 157 R->setCallingConv(F->getCallingConv()); 158 return R; 159 } 160 161 template <typename IRB> 162 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1, 163 Value *Arg2, const Twine &Name = "") { 164 CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name); 165 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 166 R->setCallingConv(F->getCallingConv()); 167 return R; 168 } 169 170 static FunctionType *getPownType(FunctionType *FT) { 171 Type *PowNExpTy = Type::getInt32Ty(FT->getContext()); 172 if (VectorType *VecTy = dyn_cast<VectorType>(FT->getReturnType())) 173 PowNExpTy = VectorType::get(PowNExpTy, VecTy->getElementCount()); 174 175 return FunctionType::get(FT->getReturnType(), 176 {FT->getParamType(0), PowNExpTy}, false); 177 } 178 179 // Data structures for table-driven optimizations. 180 // FuncTbl works for both f32 and f64 functions with 1 input argument 181 182 struct TableEntry { 183 double result; 184 double input; 185 }; 186 187 /* a list of {result, input} */ 188 static const TableEntry tbl_acos[] = { 189 {MATH_PI / 2.0, 0.0}, 190 {MATH_PI / 2.0, -0.0}, 191 {0.0, 1.0}, 192 {MATH_PI, -1.0} 193 }; 194 static const TableEntry tbl_acosh[] = { 195 {0.0, 1.0} 196 }; 197 static const TableEntry tbl_acospi[] = { 198 {0.5, 0.0}, 199 {0.5, -0.0}, 200 {0.0, 1.0}, 201 {1.0, -1.0} 202 }; 203 static const TableEntry tbl_asin[] = { 204 {0.0, 0.0}, 205 {-0.0, -0.0}, 206 {MATH_PI / 2.0, 1.0}, 207 {-MATH_PI / 2.0, -1.0} 208 }; 209 static const TableEntry tbl_asinh[] = { 210 {0.0, 0.0}, 211 {-0.0, -0.0} 212 }; 213 static const TableEntry tbl_asinpi[] = { 214 {0.0, 0.0}, 215 {-0.0, -0.0}, 216 {0.5, 1.0}, 217 {-0.5, -1.0} 218 }; 219 static const TableEntry tbl_atan[] = { 220 {0.0, 0.0}, 221 {-0.0, -0.0}, 222 {MATH_PI / 4.0, 1.0}, 223 {-MATH_PI / 4.0, -1.0} 224 }; 225 static const TableEntry tbl_atanh[] = { 226 {0.0, 0.0}, 227 {-0.0, -0.0} 228 }; 229 static const TableEntry tbl_atanpi[] = { 230 {0.0, 0.0}, 231 {-0.0, -0.0}, 232 {0.25, 1.0}, 233 {-0.25, -1.0} 234 }; 235 static const TableEntry tbl_cbrt[] = { 236 {0.0, 0.0}, 237 {-0.0, -0.0}, 238 {1.0, 1.0}, 239 {-1.0, -1.0}, 240 }; 241 static const TableEntry tbl_cos[] = { 242 {1.0, 0.0}, 243 {1.0, -0.0} 244 }; 245 static const TableEntry tbl_cosh[] = { 246 {1.0, 0.0}, 247 {1.0, -0.0} 248 }; 249 static const TableEntry tbl_cospi[] = { 250 {1.0, 0.0}, 251 {1.0, -0.0} 252 }; 253 static const TableEntry tbl_erfc[] = { 254 {1.0, 0.0}, 255 {1.0, -0.0} 256 }; 257 static const TableEntry tbl_erf[] = { 258 {0.0, 0.0}, 259 {-0.0, -0.0} 260 }; 261 static const TableEntry tbl_exp[] = { 262 {1.0, 0.0}, 263 {1.0, -0.0}, 264 {MATH_E, 1.0} 265 }; 266 static const TableEntry tbl_exp2[] = { 267 {1.0, 0.0}, 268 {1.0, -0.0}, 269 {2.0, 1.0} 270 }; 271 static const TableEntry tbl_exp10[] = { 272 {1.0, 0.0}, 273 {1.0, -0.0}, 274 {10.0, 1.0} 275 }; 276 static const TableEntry tbl_expm1[] = { 277 {0.0, 0.0}, 278 {-0.0, -0.0} 279 }; 280 static const TableEntry tbl_log[] = { 281 {0.0, 1.0}, 282 {1.0, MATH_E} 283 }; 284 static const TableEntry tbl_log2[] = { 285 {0.0, 1.0}, 286 {1.0, 2.0} 287 }; 288 static const TableEntry tbl_log10[] = { 289 {0.0, 1.0}, 290 {1.0, 10.0} 291 }; 292 static const TableEntry tbl_rsqrt[] = { 293 {1.0, 1.0}, 294 {MATH_SQRT1_2, 2.0} 295 }; 296 static const TableEntry tbl_sin[] = { 297 {0.0, 0.0}, 298 {-0.0, -0.0} 299 }; 300 static const TableEntry tbl_sinh[] = { 301 {0.0, 0.0}, 302 {-0.0, -0.0} 303 }; 304 static const TableEntry tbl_sinpi[] = { 305 {0.0, 0.0}, 306 {-0.0, -0.0} 307 }; 308 static const TableEntry tbl_sqrt[] = { 309 {0.0, 0.0}, 310 {1.0, 1.0}, 311 {MATH_SQRT2, 2.0} 312 }; 313 static const TableEntry tbl_tan[] = { 314 {0.0, 0.0}, 315 {-0.0, -0.0} 316 }; 317 static const TableEntry tbl_tanh[] = { 318 {0.0, 0.0}, 319 {-0.0, -0.0} 320 }; 321 static const TableEntry tbl_tanpi[] = { 322 {0.0, 0.0}, 323 {-0.0, -0.0} 324 }; 325 static const TableEntry tbl_tgamma[] = { 326 {1.0, 1.0}, 327 {1.0, 2.0}, 328 {2.0, 3.0}, 329 {6.0, 4.0} 330 }; 331 332 static bool HasNative(AMDGPULibFunc::EFuncId id) { 333 switch(id) { 334 case AMDGPULibFunc::EI_DIVIDE: 335 case AMDGPULibFunc::EI_COS: 336 case AMDGPULibFunc::EI_EXP: 337 case AMDGPULibFunc::EI_EXP2: 338 case AMDGPULibFunc::EI_EXP10: 339 case AMDGPULibFunc::EI_LOG: 340 case AMDGPULibFunc::EI_LOG2: 341 case AMDGPULibFunc::EI_LOG10: 342 case AMDGPULibFunc::EI_POWR: 343 case AMDGPULibFunc::EI_RECIP: 344 case AMDGPULibFunc::EI_RSQRT: 345 case AMDGPULibFunc::EI_SIN: 346 case AMDGPULibFunc::EI_SINCOS: 347 case AMDGPULibFunc::EI_SQRT: 348 case AMDGPULibFunc::EI_TAN: 349 return true; 350 default:; 351 } 352 return false; 353 } 354 355 using TableRef = ArrayRef<TableEntry>; 356 357 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) { 358 switch(id) { 359 case AMDGPULibFunc::EI_ACOS: return TableRef(tbl_acos); 360 case AMDGPULibFunc::EI_ACOSH: return TableRef(tbl_acosh); 361 case AMDGPULibFunc::EI_ACOSPI: return TableRef(tbl_acospi); 362 case AMDGPULibFunc::EI_ASIN: return TableRef(tbl_asin); 363 case AMDGPULibFunc::EI_ASINH: return TableRef(tbl_asinh); 364 case AMDGPULibFunc::EI_ASINPI: return TableRef(tbl_asinpi); 365 case AMDGPULibFunc::EI_ATAN: return TableRef(tbl_atan); 366 case AMDGPULibFunc::EI_ATANH: return TableRef(tbl_atanh); 367 case AMDGPULibFunc::EI_ATANPI: return TableRef(tbl_atanpi); 368 case AMDGPULibFunc::EI_CBRT: return TableRef(tbl_cbrt); 369 case AMDGPULibFunc::EI_NCOS: 370 case AMDGPULibFunc::EI_COS: return TableRef(tbl_cos); 371 case AMDGPULibFunc::EI_COSH: return TableRef(tbl_cosh); 372 case AMDGPULibFunc::EI_COSPI: return TableRef(tbl_cospi); 373 case AMDGPULibFunc::EI_ERFC: return TableRef(tbl_erfc); 374 case AMDGPULibFunc::EI_ERF: return TableRef(tbl_erf); 375 case AMDGPULibFunc::EI_EXP: return TableRef(tbl_exp); 376 case AMDGPULibFunc::EI_NEXP2: 377 case AMDGPULibFunc::EI_EXP2: return TableRef(tbl_exp2); 378 case AMDGPULibFunc::EI_EXP10: return TableRef(tbl_exp10); 379 case AMDGPULibFunc::EI_EXPM1: return TableRef(tbl_expm1); 380 case AMDGPULibFunc::EI_LOG: return TableRef(tbl_log); 381 case AMDGPULibFunc::EI_NLOG2: 382 case AMDGPULibFunc::EI_LOG2: return TableRef(tbl_log2); 383 case AMDGPULibFunc::EI_LOG10: return TableRef(tbl_log10); 384 case AMDGPULibFunc::EI_NRSQRT: 385 case AMDGPULibFunc::EI_RSQRT: return TableRef(tbl_rsqrt); 386 case AMDGPULibFunc::EI_NSIN: 387 case AMDGPULibFunc::EI_SIN: return TableRef(tbl_sin); 388 case AMDGPULibFunc::EI_SINH: return TableRef(tbl_sinh); 389 case AMDGPULibFunc::EI_SINPI: return TableRef(tbl_sinpi); 390 case AMDGPULibFunc::EI_NSQRT: 391 case AMDGPULibFunc::EI_SQRT: return TableRef(tbl_sqrt); 392 case AMDGPULibFunc::EI_TAN: return TableRef(tbl_tan); 393 case AMDGPULibFunc::EI_TANH: return TableRef(tbl_tanh); 394 case AMDGPULibFunc::EI_TANPI: return TableRef(tbl_tanpi); 395 case AMDGPULibFunc::EI_TGAMMA: return TableRef(tbl_tgamma); 396 default:; 397 } 398 return TableRef(); 399 } 400 401 static inline int getVecSize(const AMDGPULibFunc& FInfo) { 402 return FInfo.getLeads()[0].VectorSize; 403 } 404 405 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) { 406 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType; 407 } 408 409 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) { 410 // If we are doing PreLinkOpt, the function is external. So it is safe to 411 // use getOrInsertFunction() at this stage. 412 413 return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo) 414 : AMDGPULibFunc::getFunction(M, fInfo); 415 } 416 417 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName, 418 FuncInfo &FInfo) { 419 return AMDGPULibFunc::parse(FMangledName, FInfo); 420 } 421 422 bool AMDGPULibCalls::isUnsafeMath(const FPMathOperator *FPOp) const { 423 return UnsafeFPMath || FPOp->isFast(); 424 } 425 426 bool AMDGPULibCalls::isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const { 427 return UnsafeFPMath || 428 (FPOp->hasApproxFunc() && FPOp->hasNoNaNs() && FPOp->hasNoInfs()); 429 } 430 431 bool AMDGPULibCalls::canIncreasePrecisionOfConstantFold( 432 const FPMathOperator *FPOp) const { 433 // TODO: Refine to approxFunc or contract 434 return isUnsafeMath(FPOp); 435 } 436 437 void AMDGPULibCalls::initFunction(Function &F, FunctionAnalysisManager &FAM) { 438 UnsafeFPMath = F.getFnAttribute("unsafe-fp-math").getValueAsBool(); 439 AC = &FAM.getResult<AssumptionAnalysis>(F); 440 TLInfo = &FAM.getResult<TargetLibraryAnalysis>(F); 441 DT = FAM.getCachedResult<DominatorTreeAnalysis>(F); 442 } 443 444 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const { 445 return AllNative || llvm::is_contained(UseNative, F); 446 } 447 448 void AMDGPULibCalls::initNativeFuncs() { 449 AllNative = useNativeFunc("all") || 450 (UseNative.getNumOccurrences() && UseNative.size() == 1 && 451 UseNative.begin()->empty()); 452 } 453 454 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) { 455 bool native_sin = useNativeFunc("sin"); 456 bool native_cos = useNativeFunc("cos"); 457 458 if (native_sin && native_cos) { 459 Module *M = aCI->getModule(); 460 Value *opr0 = aCI->getArgOperand(0); 461 462 AMDGPULibFunc nf; 463 nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType; 464 nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize; 465 466 nf.setPrefix(AMDGPULibFunc::NATIVE); 467 nf.setId(AMDGPULibFunc::EI_SIN); 468 FunctionCallee sinExpr = getFunction(M, nf); 469 470 nf.setPrefix(AMDGPULibFunc::NATIVE); 471 nf.setId(AMDGPULibFunc::EI_COS); 472 FunctionCallee cosExpr = getFunction(M, nf); 473 if (sinExpr && cosExpr) { 474 Value *sinval = 475 CallInst::Create(sinExpr, opr0, "splitsin", aCI->getIterator()); 476 Value *cosval = 477 CallInst::Create(cosExpr, opr0, "splitcos", aCI->getIterator()); 478 new StoreInst(cosval, aCI->getArgOperand(1), aCI->getIterator()); 479 480 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 481 << " with native version of sin/cos"); 482 483 replaceCall(aCI, sinval); 484 return true; 485 } 486 } 487 return false; 488 } 489 490 bool AMDGPULibCalls::useNative(CallInst *aCI) { 491 Function *Callee = aCI->getCalledFunction(); 492 if (!Callee || aCI->isNoBuiltin()) 493 return false; 494 495 FuncInfo FInfo; 496 if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() || 497 FInfo.getPrefix() != AMDGPULibFunc::NOPFX || 498 getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) || 499 !(AllNative || useNativeFunc(FInfo.getName()))) { 500 return false; 501 } 502 503 if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS) 504 return sincosUseNative(aCI, FInfo); 505 506 FInfo.setPrefix(AMDGPULibFunc::NATIVE); 507 FunctionCallee F = getFunction(aCI->getModule(), FInfo); 508 if (!F) 509 return false; 510 511 aCI->setCalledFunction(F); 512 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 513 << " with native version"); 514 return true; 515 } 516 517 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe 518 // builtin, with appended type size and alignment arguments, where 2 or 4 519 // indicates the original number of arguments. The library has optimized version 520 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same 521 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N 522 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ..., 523 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4. 524 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 525 const FuncInfo &FInfo) { 526 auto *Callee = CI->getCalledFunction(); 527 if (!Callee->isDeclaration()) 528 return false; 529 530 assert(Callee->hasName() && "Invalid read_pipe/write_pipe function"); 531 auto *M = Callee->getParent(); 532 std::string Name = std::string(Callee->getName()); 533 auto NumArg = CI->arg_size(); 534 if (NumArg != 4 && NumArg != 6) 535 return false; 536 ConstantInt *PacketSize = 537 dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 2)); 538 ConstantInt *PacketAlign = 539 dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 1)); 540 if (!PacketSize || !PacketAlign) 541 return false; 542 543 unsigned Size = PacketSize->getZExtValue(); 544 Align Alignment = PacketAlign->getAlignValue(); 545 if (Alignment != Size) 546 return false; 547 548 unsigned PtrArgLoc = CI->arg_size() - 3; 549 Value *PtrArg = CI->getArgOperand(PtrArgLoc); 550 Type *PtrTy = PtrArg->getType(); 551 552 SmallVector<llvm::Type *, 6> ArgTys; 553 for (unsigned I = 0; I != PtrArgLoc; ++I) 554 ArgTys.push_back(CI->getArgOperand(I)->getType()); 555 ArgTys.push_back(PtrTy); 556 557 Name = Name + "_" + std::to_string(Size); 558 auto *FTy = FunctionType::get(Callee->getReturnType(), 559 ArrayRef<Type *>(ArgTys), false); 560 AMDGPULibFunc NewLibFunc(Name, FTy); 561 FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc); 562 if (!F) 563 return false; 564 565 SmallVector<Value *, 6> Args; 566 for (unsigned I = 0; I != PtrArgLoc; ++I) 567 Args.push_back(CI->getArgOperand(I)); 568 Args.push_back(PtrArg); 569 570 auto *NCI = B.CreateCall(F, Args); 571 NCI->setAttributes(CI->getAttributes()); 572 CI->replaceAllUsesWith(NCI); 573 CI->dropAllReferences(); 574 CI->eraseFromParent(); 575 576 return true; 577 } 578 579 static bool isKnownIntegral(const Value *V, const DataLayout &DL, 580 FastMathFlags FMF) { 581 if (isa<PoisonValue>(V)) 582 return true; 583 if (isa<UndefValue>(V)) 584 return false; 585 586 if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) 587 return CF->getValueAPF().isInteger(); 588 589 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); 590 const Constant *CV = dyn_cast<Constant>(V); 591 if (VFVTy && CV) { 592 unsigned NumElts = VFVTy->getNumElements(); 593 for (unsigned i = 0; i != NumElts; ++i) { 594 Constant *Elt = CV->getAggregateElement(i); 595 if (!Elt) 596 return false; 597 if (isa<PoisonValue>(Elt)) 598 continue; 599 600 const ConstantFP *CFP = dyn_cast<ConstantFP>(Elt); 601 if (!CFP || !CFP->getValue().isInteger()) 602 return false; 603 } 604 605 return true; 606 } 607 608 const Instruction *I = dyn_cast<Instruction>(V); 609 if (!I) 610 return false; 611 612 switch (I->getOpcode()) { 613 case Instruction::SIToFP: 614 case Instruction::UIToFP: 615 // TODO: Could check nofpclass(inf) on incoming argument 616 if (FMF.noInfs()) 617 return true; 618 619 // Need to check int size cannot produce infinity, which computeKnownFPClass 620 // knows how to do already. 621 return isKnownNeverInfinity(I, /*Depth=*/0, SimplifyQuery(DL)); 622 case Instruction::Call: { 623 const CallInst *CI = cast<CallInst>(I); 624 switch (CI->getIntrinsicID()) { 625 case Intrinsic::trunc: 626 case Intrinsic::floor: 627 case Intrinsic::ceil: 628 case Intrinsic::rint: 629 case Intrinsic::nearbyint: 630 case Intrinsic::round: 631 case Intrinsic::roundeven: 632 return (FMF.noInfs() && FMF.noNaNs()) || 633 isKnownNeverInfOrNaN(I, /*Depth=*/0, SimplifyQuery(DL)); 634 default: 635 break; 636 } 637 638 break; 639 } 640 default: 641 break; 642 } 643 644 return false; 645 } 646 647 // This function returns false if no change; return true otherwise. 648 bool AMDGPULibCalls::fold(CallInst *CI) { 649 Function *Callee = CI->getCalledFunction(); 650 // Ignore indirect calls. 651 if (!Callee || Callee->isIntrinsic() || CI->isNoBuiltin()) 652 return false; 653 654 FuncInfo FInfo; 655 if (!parseFunctionName(Callee->getName(), FInfo)) 656 return false; 657 658 // Further check the number of arguments to see if they match. 659 // TODO: Check calling convention matches too 660 if (!FInfo.isCompatibleSignature(CI->getFunctionType())) 661 return false; 662 663 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << '\n'); 664 665 if (TDOFold(CI, FInfo)) 666 return true; 667 668 IRBuilder<> B(CI); 669 if (CI->isStrictFP()) 670 B.setIsFPConstrained(true); 671 672 if (FPMathOperator *FPOp = dyn_cast<FPMathOperator>(CI)) { 673 // Under unsafe-math, evaluate calls if possible. 674 // According to Brian Sumner, we can do this for all f32 function calls 675 // using host's double function calls. 676 if (canIncreasePrecisionOfConstantFold(FPOp) && evaluateCall(CI, FInfo)) 677 return true; 678 679 // Copy fast flags from the original call. 680 FastMathFlags FMF = FPOp->getFastMathFlags(); 681 B.setFastMathFlags(FMF); 682 683 // Specialized optimizations for each function call. 684 // 685 // TODO: Handle native functions 686 switch (FInfo.getId()) { 687 case AMDGPULibFunc::EI_EXP: 688 if (FMF.none()) 689 return false; 690 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp, 691 FMF.approxFunc()); 692 case AMDGPULibFunc::EI_EXP2: 693 if (FMF.none()) 694 return false; 695 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp2, 696 FMF.approxFunc()); 697 case AMDGPULibFunc::EI_LOG: 698 if (FMF.none()) 699 return false; 700 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log, 701 FMF.approxFunc()); 702 case AMDGPULibFunc::EI_LOG2: 703 if (FMF.none()) 704 return false; 705 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log2, 706 FMF.approxFunc()); 707 case AMDGPULibFunc::EI_LOG10: 708 if (FMF.none()) 709 return false; 710 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log10, 711 FMF.approxFunc()); 712 case AMDGPULibFunc::EI_FMIN: 713 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::minnum, 714 true, true); 715 case AMDGPULibFunc::EI_FMAX: 716 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::maxnum, 717 true, true); 718 case AMDGPULibFunc::EI_FMA: 719 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fma, true, 720 true); 721 case AMDGPULibFunc::EI_MAD: 722 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fmuladd, 723 true, true); 724 case AMDGPULibFunc::EI_FABS: 725 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fabs, true, 726 true, true); 727 case AMDGPULibFunc::EI_COPYSIGN: 728 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::copysign, 729 true, true, true); 730 case AMDGPULibFunc::EI_FLOOR: 731 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::floor, true, 732 true); 733 case AMDGPULibFunc::EI_CEIL: 734 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::ceil, true, 735 true); 736 case AMDGPULibFunc::EI_TRUNC: 737 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::trunc, true, 738 true); 739 case AMDGPULibFunc::EI_RINT: 740 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::rint, true, 741 true); 742 case AMDGPULibFunc::EI_ROUND: 743 return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::round, true, 744 true); 745 case AMDGPULibFunc::EI_LDEXP: { 746 if (!shouldReplaceLibcallWithIntrinsic(CI, true, true)) 747 return false; 748 749 Value *Arg1 = CI->getArgOperand(1); 750 if (VectorType *VecTy = dyn_cast<VectorType>(CI->getType()); 751 VecTy && !isa<VectorType>(Arg1->getType())) { 752 Value *SplatArg1 = B.CreateVectorSplat(VecTy->getElementCount(), Arg1); 753 CI->setArgOperand(1, SplatArg1); 754 } 755 756 CI->setCalledFunction(Intrinsic::getDeclaration( 757 CI->getModule(), Intrinsic::ldexp, 758 {CI->getType(), CI->getArgOperand(1)->getType()})); 759 return true; 760 } 761 case AMDGPULibFunc::EI_POW: { 762 Module *M = Callee->getParent(); 763 AMDGPULibFunc PowrInfo(AMDGPULibFunc::EI_POWR, FInfo); 764 FunctionCallee PowrFunc = getFunction(M, PowrInfo); 765 CallInst *Call = cast<CallInst>(FPOp); 766 767 // pow(x, y) -> powr(x, y) for x >= -0.0 768 // TODO: Account for flags on current call 769 if (PowrFunc && 770 cannotBeOrderedLessThanZero( 771 FPOp->getOperand(0), /*Depth=*/0, 772 SimplifyQuery(M->getDataLayout(), TLInfo, DT, AC, Call))) { 773 Call->setCalledFunction(PowrFunc); 774 return fold_pow(FPOp, B, PowrInfo) || true; 775 } 776 777 // pow(x, y) -> pown(x, y) for known integral y 778 if (isKnownIntegral(FPOp->getOperand(1), M->getDataLayout(), 779 FPOp->getFastMathFlags())) { 780 FunctionType *PownType = getPownType(CI->getFunctionType()); 781 AMDGPULibFunc PownInfo(AMDGPULibFunc::EI_POWN, PownType, true); 782 FunctionCallee PownFunc = getFunction(M, PownInfo); 783 if (PownFunc) { 784 // TODO: If the incoming integral value is an sitofp/uitofp, it won't 785 // fold out without a known range. We can probably take the source 786 // value directly. 787 Value *CastedArg = 788 B.CreateFPToSI(FPOp->getOperand(1), PownType->getParamType(1)); 789 // Have to drop any nofpclass attributes on the original call site. 790 Call->removeParamAttrs( 791 1, AttributeFuncs::typeIncompatible(CastedArg->getType())); 792 Call->setCalledFunction(PownFunc); 793 Call->setArgOperand(1, CastedArg); 794 return fold_pow(FPOp, B, PownInfo) || true; 795 } 796 } 797 798 return fold_pow(FPOp, B, FInfo); 799 } 800 case AMDGPULibFunc::EI_POWR: 801 case AMDGPULibFunc::EI_POWN: 802 return fold_pow(FPOp, B, FInfo); 803 case AMDGPULibFunc::EI_ROOTN: 804 return fold_rootn(FPOp, B, FInfo); 805 case AMDGPULibFunc::EI_SQRT: 806 // TODO: Allow with strictfp + constrained intrinsic 807 return tryReplaceLibcallWithSimpleIntrinsic( 808 B, CI, Intrinsic::sqrt, true, true, /*AllowStrictFP=*/false); 809 case AMDGPULibFunc::EI_COS: 810 case AMDGPULibFunc::EI_SIN: 811 return fold_sincos(FPOp, B, FInfo); 812 default: 813 break; 814 } 815 } else { 816 // Specialized optimizations for each function call 817 switch (FInfo.getId()) { 818 case AMDGPULibFunc::EI_READ_PIPE_2: 819 case AMDGPULibFunc::EI_READ_PIPE_4: 820 case AMDGPULibFunc::EI_WRITE_PIPE_2: 821 case AMDGPULibFunc::EI_WRITE_PIPE_4: 822 return fold_read_write_pipe(CI, B, FInfo); 823 default: 824 break; 825 } 826 } 827 828 return false; 829 } 830 831 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) { 832 // Table-Driven optimization 833 const TableRef tr = getOptTable(FInfo.getId()); 834 if (tr.empty()) 835 return false; 836 837 int const sz = (int)tr.size(); 838 Value *opr0 = CI->getArgOperand(0); 839 840 if (getVecSize(FInfo) > 1) { 841 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) { 842 SmallVector<double, 0> DVal; 843 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) { 844 ConstantFP *eltval = dyn_cast<ConstantFP>( 845 CV->getElementAsConstant((unsigned)eltNo)); 846 assert(eltval && "Non-FP arguments in math function!"); 847 bool found = false; 848 for (int i=0; i < sz; ++i) { 849 if (eltval->isExactlyValue(tr[i].input)) { 850 DVal.push_back(tr[i].result); 851 found = true; 852 break; 853 } 854 } 855 if (!found) { 856 // This vector constants not handled yet. 857 return false; 858 } 859 } 860 LLVMContext &context = CI->getParent()->getParent()->getContext(); 861 Constant *nval; 862 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 863 SmallVector<float, 0> FVal; 864 for (unsigned i = 0; i < DVal.size(); ++i) { 865 FVal.push_back((float)DVal[i]); 866 } 867 ArrayRef<float> tmp(FVal); 868 nval = ConstantDataVector::get(context, tmp); 869 } else { // F64 870 ArrayRef<double> tmp(DVal); 871 nval = ConstantDataVector::get(context, tmp); 872 } 873 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 874 replaceCall(CI, nval); 875 return true; 876 } 877 } else { 878 // Scalar version 879 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) { 880 for (int i = 0; i < sz; ++i) { 881 if (CF->isExactlyValue(tr[i].input)) { 882 Value *nval = ConstantFP::get(CF->getType(), tr[i].result); 883 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 884 replaceCall(CI, nval); 885 return true; 886 } 887 } 888 } 889 } 890 891 return false; 892 } 893 894 namespace llvm { 895 static double log2(double V) { 896 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L 897 return ::log2(V); 898 #else 899 return log(V) / numbers::ln2; 900 #endif 901 } 902 } // namespace llvm 903 904 bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, 905 const FuncInfo &FInfo) { 906 assert((FInfo.getId() == AMDGPULibFunc::EI_POW || 907 FInfo.getId() == AMDGPULibFunc::EI_POWR || 908 FInfo.getId() == AMDGPULibFunc::EI_POWN) && 909 "fold_pow: encounter a wrong function call"); 910 911 Module *M = B.GetInsertBlock()->getModule(); 912 Type *eltType = FPOp->getType()->getScalarType(); 913 Value *opr0 = FPOp->getOperand(0); 914 Value *opr1 = FPOp->getOperand(1); 915 916 const APFloat *CF = nullptr; 917 const APInt *CINT = nullptr; 918 if (!match(opr1, m_APFloatAllowPoison(CF))) 919 match(opr1, m_APIntAllowPoison(CINT)); 920 921 // 0x1111111 means that we don't do anything for this call. 922 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111); 923 924 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0)) { 925 // pow/powr/pown(x, 0) == 1 926 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1\n"); 927 Constant *cnval = ConstantFP::get(eltType, 1.0); 928 if (getVecSize(FInfo) > 1) { 929 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 930 } 931 replaceCall(FPOp, cnval); 932 return true; 933 } 934 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) { 935 // pow/powr/pown(x, 1.0) = x 936 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n"); 937 replaceCall(FPOp, opr0); 938 return true; 939 } 940 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) { 941 // pow/powr/pown(x, 2.0) = x*x 942 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << " * " 943 << *opr0 << "\n"); 944 Value *nval = B.CreateFMul(opr0, opr0, "__pow2"); 945 replaceCall(FPOp, nval); 946 return true; 947 } 948 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) { 949 // pow/powr/pown(x, -1.0) = 1.0/x 950 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1 / " << *opr0 << "\n"); 951 Constant *cnval = ConstantFP::get(eltType, 1.0); 952 if (getVecSize(FInfo) > 1) { 953 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 954 } 955 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip"); 956 replaceCall(FPOp, nval); 957 return true; 958 } 959 960 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) { 961 // pow[r](x, [-]0.5) = sqrt(x) 962 bool issqrt = CF->isExactlyValue(0.5); 963 if (FunctionCallee FPExpr = 964 getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT 965 : AMDGPULibFunc::EI_RSQRT, 966 FInfo))) { 967 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << FInfo.getName() 968 << '(' << *opr0 << ")\n"); 969 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt" 970 : "__pow2rsqrt"); 971 replaceCall(FPOp, nval); 972 return true; 973 } 974 } 975 976 if (!isUnsafeFiniteOnlyMath(FPOp)) 977 return false; 978 979 // Unsafe Math optimization 980 981 // Remember that ci_opr1 is set if opr1 is integral 982 if (CF) { 983 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32) 984 ? (double)CF->convertToFloat() 985 : CF->convertToDouble(); 986 int ival = (int)dval; 987 if ((double)ival == dval) { 988 ci_opr1 = ival; 989 } else 990 ci_opr1 = 0x11111111; 991 } 992 993 // pow/powr/pown(x, c) = [1/](x*x*..x); where 994 // trunc(c) == c && the number of x == c && |c| <= 12 995 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1; 996 if (abs_opr1 <= 12) { 997 Constant *cnval; 998 Value *nval; 999 if (abs_opr1 == 0) { 1000 cnval = ConstantFP::get(eltType, 1.0); 1001 if (getVecSize(FInfo) > 1) { 1002 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 1003 } 1004 nval = cnval; 1005 } else { 1006 Value *valx2 = nullptr; 1007 nval = nullptr; 1008 while (abs_opr1 > 0) { 1009 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0; 1010 if (abs_opr1 & 1) { 1011 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2; 1012 } 1013 abs_opr1 >>= 1; 1014 } 1015 } 1016 1017 if (ci_opr1 < 0) { 1018 cnval = ConstantFP::get(eltType, 1.0); 1019 if (getVecSize(FInfo) > 1) { 1020 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 1021 } 1022 nval = B.CreateFDiv(cnval, nval, "__1powprod"); 1023 } 1024 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " 1025 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0 1026 << ")\n"); 1027 replaceCall(FPOp, nval); 1028 return true; 1029 } 1030 1031 // If we should use the generic intrinsic instead of emitting a libcall 1032 const bool ShouldUseIntrinsic = eltType->isFloatTy() || eltType->isHalfTy(); 1033 1034 // powr ---> exp2(y * log2(x)) 1035 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31)) 1036 FunctionCallee ExpExpr; 1037 if (ShouldUseIntrinsic) 1038 ExpExpr = Intrinsic::getDeclaration(M, Intrinsic::exp2, {FPOp->getType()}); 1039 else { 1040 ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo)); 1041 if (!ExpExpr) 1042 return false; 1043 } 1044 1045 bool needlog = false; 1046 bool needabs = false; 1047 bool needcopysign = false; 1048 Constant *cnval = nullptr; 1049 if (getVecSize(FInfo) == 1) { 1050 CF = nullptr; 1051 match(opr0, m_APFloatAllowPoison(CF)); 1052 1053 if (CF) { 1054 double V = (getArgType(FInfo) == AMDGPULibFunc::F32) 1055 ? (double)CF->convertToFloat() 1056 : CF->convertToDouble(); 1057 1058 V = log2(std::abs(V)); 1059 cnval = ConstantFP::get(eltType, V); 1060 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) && 1061 CF->isNegative(); 1062 } else { 1063 needlog = true; 1064 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR; 1065 } 1066 } else { 1067 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0); 1068 1069 if (!CDV) { 1070 needlog = true; 1071 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR; 1072 } else { 1073 assert ((int)CDV->getNumElements() == getVecSize(FInfo) && 1074 "Wrong vector size detected"); 1075 1076 SmallVector<double, 0> DVal; 1077 for (int i=0; i < getVecSize(FInfo); ++i) { 1078 double V = CDV->getElementAsAPFloat(i).convertToDouble(); 1079 if (V < 0.0) needcopysign = true; 1080 V = log2(std::abs(V)); 1081 DVal.push_back(V); 1082 } 1083 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1084 SmallVector<float, 0> FVal; 1085 for (unsigned i=0; i < DVal.size(); ++i) { 1086 FVal.push_back((float)DVal[i]); 1087 } 1088 ArrayRef<float> tmp(FVal); 1089 cnval = ConstantDataVector::get(M->getContext(), tmp); 1090 } else { 1091 ArrayRef<double> tmp(DVal); 1092 cnval = ConstantDataVector::get(M->getContext(), tmp); 1093 } 1094 } 1095 } 1096 1097 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) { 1098 // We cannot handle corner cases for a general pow() function, give up 1099 // unless y is a constant integral value. Then proceed as if it were pown. 1100 if (!isKnownIntegral(opr1, M->getDataLayout(), FPOp->getFastMathFlags())) 1101 return false; 1102 } 1103 1104 Value *nval; 1105 if (needabs) { 1106 nval = B.CreateUnaryIntrinsic(Intrinsic::fabs, opr0, nullptr, "__fabs"); 1107 } else { 1108 nval = cnval ? cnval : opr0; 1109 } 1110 if (needlog) { 1111 FunctionCallee LogExpr; 1112 if (ShouldUseIntrinsic) { 1113 LogExpr = 1114 Intrinsic::getDeclaration(M, Intrinsic::log2, {FPOp->getType()}); 1115 } else { 1116 LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo)); 1117 if (!LogExpr) 1118 return false; 1119 } 1120 1121 nval = CreateCallEx(B,LogExpr, nval, "__log2"); 1122 } 1123 1124 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) { 1125 // convert int(32) to fp(f32 or f64) 1126 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F"); 1127 } 1128 nval = B.CreateFMul(opr1, nval, "__ylogx"); 1129 nval = CreateCallEx(B,ExpExpr, nval, "__exp2"); 1130 1131 if (needcopysign) { 1132 Type* nTyS = B.getIntNTy(eltType->getPrimitiveSizeInBits()); 1133 Type *nTy = FPOp->getType()->getWithNewType(nTyS); 1134 unsigned size = nTy->getScalarSizeInBits(); 1135 Value *opr_n = FPOp->getOperand(1); 1136 if (opr_n->getType()->getScalarType()->isIntegerTy()) 1137 opr_n = B.CreateZExtOrTrunc(opr_n, nTy, "__ytou"); 1138 else 1139 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou"); 1140 1141 Value *sign = B.CreateShl(opr_n, size-1, "__yeven"); 1142 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign"); 1143 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign); 1144 nval = B.CreateBitCast(nval, opr0->getType()); 1145 } 1146 1147 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " 1148 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n"); 1149 replaceCall(FPOp, nval); 1150 1151 return true; 1152 } 1153 1154 bool AMDGPULibCalls::fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B, 1155 const FuncInfo &FInfo) { 1156 Value *opr0 = FPOp->getOperand(0); 1157 Value *opr1 = FPOp->getOperand(1); 1158 1159 const APInt *CINT = nullptr; 1160 if (!match(opr1, m_APIntAllowPoison(CINT))) 1161 return false; 1162 1163 Function *Parent = B.GetInsertBlock()->getParent(); 1164 1165 int ci_opr1 = (int)CINT->getSExtValue(); 1166 if (ci_opr1 == 1 && !Parent->hasFnAttribute(Attribute::StrictFP)) { 1167 // rootn(x, 1) = x 1168 // 1169 // TODO: Insert constrained canonicalize for strictfp case. 1170 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << '\n'); 1171 replaceCall(FPOp, opr0); 1172 return true; 1173 } 1174 1175 Module *M = B.GetInsertBlock()->getModule(); 1176 1177 CallInst *CI = cast<CallInst>(FPOp); 1178 if (ci_opr1 == 2 && 1179 shouldReplaceLibcallWithIntrinsic(CI, 1180 /*AllowMinSizeF32=*/true, 1181 /*AllowF64=*/true)) { 1182 // rootn(x, 2) = sqrt(x) 1183 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> sqrt(" << *opr0 << ")\n"); 1184 1185 CallInst *NewCall = B.CreateUnaryIntrinsic(Intrinsic::sqrt, opr0, CI); 1186 NewCall->takeName(CI); 1187 1188 // OpenCL rootn has a looser ulp of 2 requirement than sqrt, so add some 1189 // metadata. 1190 MDBuilder MDHelper(M->getContext()); 1191 MDNode *FPMD = MDHelper.createFPMath(std::max(FPOp->getFPAccuracy(), 2.0f)); 1192 NewCall->setMetadata(LLVMContext::MD_fpmath, FPMD); 1193 1194 replaceCall(CI, NewCall); 1195 return true; 1196 } 1197 1198 if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x) 1199 if (FunctionCallee FPExpr = 1200 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) { 1201 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> cbrt(" << *opr0 1202 << ")\n"); 1203 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt"); 1204 replaceCall(FPOp, nval); 1205 return true; 1206 } 1207 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x 1208 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1.0 / " << *opr0 << "\n"); 1209 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0), 1210 opr0, 1211 "__rootn2div"); 1212 replaceCall(FPOp, nval); 1213 return true; 1214 } 1215 1216 if (ci_opr1 == -2 && 1217 shouldReplaceLibcallWithIntrinsic(CI, 1218 /*AllowMinSizeF32=*/true, 1219 /*AllowF64=*/true)) { 1220 // rootn(x, -2) = rsqrt(x) 1221 1222 // The original rootn had looser ulp requirements than the resultant sqrt 1223 // and fdiv. 1224 MDBuilder MDHelper(M->getContext()); 1225 MDNode *FPMD = MDHelper.createFPMath(std::max(FPOp->getFPAccuracy(), 2.0f)); 1226 1227 // TODO: Could handle strictfp but need to fix strict sqrt emission 1228 FastMathFlags FMF = FPOp->getFastMathFlags(); 1229 FMF.setAllowContract(true); 1230 1231 CallInst *Sqrt = B.CreateUnaryIntrinsic(Intrinsic::sqrt, opr0, CI); 1232 Instruction *RSqrt = cast<Instruction>( 1233 B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0), Sqrt)); 1234 Sqrt->setFastMathFlags(FMF); 1235 RSqrt->setFastMathFlags(FMF); 1236 RSqrt->setMetadata(LLVMContext::MD_fpmath, FPMD); 1237 1238 LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> rsqrt(" << *opr0 1239 << ")\n"); 1240 replaceCall(CI, RSqrt); 1241 return true; 1242 } 1243 1244 return false; 1245 } 1246 1247 // Get a scalar native builtin single argument FP function 1248 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M, 1249 const FuncInfo &FInfo) { 1250 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId())) 1251 return nullptr; 1252 FuncInfo nf = FInfo; 1253 nf.setPrefix(AMDGPULibFunc::NATIVE); 1254 return getFunction(M, nf); 1255 } 1256 1257 // Some library calls are just wrappers around llvm intrinsics, but compiled 1258 // conservatively. Preserve the flags from the original call site by 1259 // substituting them with direct calls with all the flags. 1260 bool AMDGPULibCalls::shouldReplaceLibcallWithIntrinsic(const CallInst *CI, 1261 bool AllowMinSizeF32, 1262 bool AllowF64, 1263 bool AllowStrictFP) { 1264 Type *FltTy = CI->getType()->getScalarType(); 1265 const bool IsF32 = FltTy->isFloatTy(); 1266 1267 // f64 intrinsics aren't implemented for most operations. 1268 if (!IsF32 && !FltTy->isHalfTy() && (!AllowF64 || !FltTy->isDoubleTy())) 1269 return false; 1270 1271 // We're implicitly inlining by replacing the libcall with the intrinsic, so 1272 // don't do it for noinline call sites. 1273 if (CI->isNoInline()) 1274 return false; 1275 1276 const Function *ParentF = CI->getFunction(); 1277 // TODO: Handle strictfp 1278 if (!AllowStrictFP && ParentF->hasFnAttribute(Attribute::StrictFP)) 1279 return false; 1280 1281 if (IsF32 && !AllowMinSizeF32 && ParentF->hasMinSize()) 1282 return false; 1283 return true; 1284 } 1285 1286 void AMDGPULibCalls::replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, 1287 CallInst *CI, 1288 Intrinsic::ID IntrID) { 1289 if (CI->arg_size() == 2) { 1290 Value *Arg0 = CI->getArgOperand(0); 1291 Value *Arg1 = CI->getArgOperand(1); 1292 VectorType *Arg0VecTy = dyn_cast<VectorType>(Arg0->getType()); 1293 VectorType *Arg1VecTy = dyn_cast<VectorType>(Arg1->getType()); 1294 if (Arg0VecTy && !Arg1VecTy) { 1295 Value *SplatRHS = B.CreateVectorSplat(Arg0VecTy->getElementCount(), Arg1); 1296 CI->setArgOperand(1, SplatRHS); 1297 } else if (!Arg0VecTy && Arg1VecTy) { 1298 Value *SplatLHS = B.CreateVectorSplat(Arg1VecTy->getElementCount(), Arg0); 1299 CI->setArgOperand(0, SplatLHS); 1300 } 1301 } 1302 1303 CI->setCalledFunction( 1304 Intrinsic::getDeclaration(CI->getModule(), IntrID, {CI->getType()})); 1305 } 1306 1307 bool AMDGPULibCalls::tryReplaceLibcallWithSimpleIntrinsic( 1308 IRBuilder<> &B, CallInst *CI, Intrinsic::ID IntrID, bool AllowMinSizeF32, 1309 bool AllowF64, bool AllowStrictFP) { 1310 if (!shouldReplaceLibcallWithIntrinsic(CI, AllowMinSizeF32, AllowF64, 1311 AllowStrictFP)) 1312 return false; 1313 replaceLibCallWithSimpleIntrinsic(B, CI, IntrID); 1314 return true; 1315 } 1316 1317 std::tuple<Value *, Value *, Value *> 1318 AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B, 1319 FunctionCallee Fsincos) { 1320 DebugLoc DL = B.getCurrentDebugLocation(); 1321 Function *F = B.GetInsertBlock()->getParent(); 1322 B.SetInsertPointPastAllocas(F); 1323 1324 AllocaInst *Alloc = B.CreateAlloca(Arg->getType(), nullptr, "__sincos_"); 1325 1326 if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) { 1327 // If the argument is an instruction, it must dominate all uses so put our 1328 // sincos call there. Otherwise, right after the allocas works well enough 1329 // if it's an argument or constant. 1330 1331 B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator()); 1332 1333 // SetInsertPoint unwelcomely always tries to set the debug loc. 1334 B.SetCurrentDebugLocation(DL); 1335 } 1336 1337 Type *CosPtrTy = Fsincos.getFunctionType()->getParamType(1); 1338 1339 // The allocaInst allocates the memory in private address space. This need 1340 // to be addrspacecasted to point to the address space of cos pointer type. 1341 // In OpenCL 2.0 this is generic, while in 1.2 that is private. 1342 Value *CastAlloc = B.CreateAddrSpaceCast(Alloc, CosPtrTy); 1343 1344 CallInst *SinCos = CreateCallEx2(B, Fsincos, Arg, CastAlloc); 1345 1346 // TODO: Is it worth trying to preserve the location for the cos calls for the 1347 // load? 1348 1349 LoadInst *LoadCos = B.CreateLoad(Alloc->getAllocatedType(), Alloc); 1350 return {SinCos, LoadCos, SinCos}; 1351 } 1352 1353 // fold sin, cos -> sincos. 1354 bool AMDGPULibCalls::fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, 1355 const FuncInfo &fInfo) { 1356 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN || 1357 fInfo.getId() == AMDGPULibFunc::EI_COS); 1358 1359 if ((getArgType(fInfo) != AMDGPULibFunc::F32 && 1360 getArgType(fInfo) != AMDGPULibFunc::F64) || 1361 fInfo.getPrefix() != AMDGPULibFunc::NOPFX) 1362 return false; 1363 1364 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN; 1365 1366 Value *CArgVal = FPOp->getOperand(0); 1367 CallInst *CI = cast<CallInst>(FPOp); 1368 1369 Function *F = B.GetInsertBlock()->getParent(); 1370 Module *M = F->getParent(); 1371 1372 // Merge the sin and cos. For OpenCL 2.0, there may only be a generic pointer 1373 // implementation. Prefer the private form if available. 1374 AMDGPULibFunc SinCosLibFuncPrivate(AMDGPULibFunc::EI_SINCOS, fInfo); 1375 SinCosLibFuncPrivate.getLeads()[0].PtrKind = 1376 AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::PRIVATE_ADDRESS); 1377 1378 AMDGPULibFunc SinCosLibFuncGeneric(AMDGPULibFunc::EI_SINCOS, fInfo); 1379 SinCosLibFuncGeneric.getLeads()[0].PtrKind = 1380 AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS); 1381 1382 FunctionCallee FSinCosPrivate = getFunction(M, SinCosLibFuncPrivate); 1383 FunctionCallee FSinCosGeneric = getFunction(M, SinCosLibFuncGeneric); 1384 FunctionCallee FSinCos = FSinCosPrivate ? FSinCosPrivate : FSinCosGeneric; 1385 if (!FSinCos) 1386 return false; 1387 1388 SmallVector<CallInst *> SinCalls; 1389 SmallVector<CallInst *> CosCalls; 1390 SmallVector<CallInst *> SinCosCalls; 1391 FuncInfo PartnerInfo(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN, 1392 fInfo); 1393 const std::string PairName = PartnerInfo.mangle(); 1394 1395 StringRef SinName = isSin ? CI->getCalledFunction()->getName() : PairName; 1396 StringRef CosName = isSin ? PairName : CI->getCalledFunction()->getName(); 1397 const std::string SinCosPrivateName = SinCosLibFuncPrivate.mangle(); 1398 const std::string SinCosGenericName = SinCosLibFuncGeneric.mangle(); 1399 1400 // Intersect the two sets of flags. 1401 FastMathFlags FMF = FPOp->getFastMathFlags(); 1402 MDNode *FPMath = CI->getMetadata(LLVMContext::MD_fpmath); 1403 1404 SmallVector<DILocation *> MergeDbgLocs = {CI->getDebugLoc()}; 1405 1406 for (User* U : CArgVal->users()) { 1407 CallInst *XI = dyn_cast<CallInst>(U); 1408 if (!XI || XI->getFunction() != F || XI->isNoBuiltin()) 1409 continue; 1410 1411 Function *UCallee = XI->getCalledFunction(); 1412 if (!UCallee) 1413 continue; 1414 1415 bool Handled = true; 1416 1417 if (UCallee->getName() == SinName) 1418 SinCalls.push_back(XI); 1419 else if (UCallee->getName() == CosName) 1420 CosCalls.push_back(XI); 1421 else if (UCallee->getName() == SinCosPrivateName || 1422 UCallee->getName() == SinCosGenericName) 1423 SinCosCalls.push_back(XI); 1424 else 1425 Handled = false; 1426 1427 if (Handled) { 1428 MergeDbgLocs.push_back(XI->getDebugLoc()); 1429 auto *OtherOp = cast<FPMathOperator>(XI); 1430 FMF &= OtherOp->getFastMathFlags(); 1431 FPMath = MDNode::getMostGenericFPMath( 1432 FPMath, XI->getMetadata(LLVMContext::MD_fpmath)); 1433 } 1434 } 1435 1436 if (SinCalls.empty() || CosCalls.empty()) 1437 return false; 1438 1439 B.setFastMathFlags(FMF); 1440 B.setDefaultFPMathTag(FPMath); 1441 DILocation *DbgLoc = DILocation::getMergedLocations(MergeDbgLocs); 1442 B.SetCurrentDebugLocation(DbgLoc); 1443 1444 auto [Sin, Cos, SinCos] = insertSinCos(CArgVal, FMF, B, FSinCos); 1445 1446 auto replaceTrigInsts = [](ArrayRef<CallInst *> Calls, Value *Res) { 1447 for (CallInst *C : Calls) 1448 C->replaceAllUsesWith(Res); 1449 1450 // Leave the other dead instructions to avoid clobbering iterators. 1451 }; 1452 1453 replaceTrigInsts(SinCalls, Sin); 1454 replaceTrigInsts(CosCalls, Cos); 1455 replaceTrigInsts(SinCosCalls, SinCos); 1456 1457 // It's safe to delete the original now. 1458 CI->eraseFromParent(); 1459 return true; 1460 } 1461 1462 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0, 1463 double &Res1, Constant *copr0, 1464 Constant *copr1) { 1465 // By default, opr0/opr1/opr3 holds values of float/double type. 1466 // If they are not float/double, each function has to its 1467 // operand separately. 1468 double opr0 = 0.0, opr1 = 0.0; 1469 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0); 1470 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1); 1471 if (fpopr0) { 1472 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1473 ? fpopr0->getValueAPF().convertToDouble() 1474 : (double)fpopr0->getValueAPF().convertToFloat(); 1475 } 1476 1477 if (fpopr1) { 1478 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1479 ? fpopr1->getValueAPF().convertToDouble() 1480 : (double)fpopr1->getValueAPF().convertToFloat(); 1481 } 1482 1483 switch (FInfo.getId()) { 1484 default : return false; 1485 1486 case AMDGPULibFunc::EI_ACOS: 1487 Res0 = acos(opr0); 1488 return true; 1489 1490 case AMDGPULibFunc::EI_ACOSH: 1491 // acosh(x) == log(x + sqrt(x*x - 1)) 1492 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0)); 1493 return true; 1494 1495 case AMDGPULibFunc::EI_ACOSPI: 1496 Res0 = acos(opr0) / MATH_PI; 1497 return true; 1498 1499 case AMDGPULibFunc::EI_ASIN: 1500 Res0 = asin(opr0); 1501 return true; 1502 1503 case AMDGPULibFunc::EI_ASINH: 1504 // asinh(x) == log(x + sqrt(x*x + 1)) 1505 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0)); 1506 return true; 1507 1508 case AMDGPULibFunc::EI_ASINPI: 1509 Res0 = asin(opr0) / MATH_PI; 1510 return true; 1511 1512 case AMDGPULibFunc::EI_ATAN: 1513 Res0 = atan(opr0); 1514 return true; 1515 1516 case AMDGPULibFunc::EI_ATANH: 1517 // atanh(x) == (log(x+1) - log(x-1))/2; 1518 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0; 1519 return true; 1520 1521 case AMDGPULibFunc::EI_ATANPI: 1522 Res0 = atan(opr0) / MATH_PI; 1523 return true; 1524 1525 case AMDGPULibFunc::EI_CBRT: 1526 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0); 1527 return true; 1528 1529 case AMDGPULibFunc::EI_COS: 1530 Res0 = cos(opr0); 1531 return true; 1532 1533 case AMDGPULibFunc::EI_COSH: 1534 Res0 = cosh(opr0); 1535 return true; 1536 1537 case AMDGPULibFunc::EI_COSPI: 1538 Res0 = cos(MATH_PI * opr0); 1539 return true; 1540 1541 case AMDGPULibFunc::EI_EXP: 1542 Res0 = exp(opr0); 1543 return true; 1544 1545 case AMDGPULibFunc::EI_EXP2: 1546 Res0 = pow(2.0, opr0); 1547 return true; 1548 1549 case AMDGPULibFunc::EI_EXP10: 1550 Res0 = pow(10.0, opr0); 1551 return true; 1552 1553 case AMDGPULibFunc::EI_LOG: 1554 Res0 = log(opr0); 1555 return true; 1556 1557 case AMDGPULibFunc::EI_LOG2: 1558 Res0 = log(opr0) / log(2.0); 1559 return true; 1560 1561 case AMDGPULibFunc::EI_LOG10: 1562 Res0 = log(opr0) / log(10.0); 1563 return true; 1564 1565 case AMDGPULibFunc::EI_RSQRT: 1566 Res0 = 1.0 / sqrt(opr0); 1567 return true; 1568 1569 case AMDGPULibFunc::EI_SIN: 1570 Res0 = sin(opr0); 1571 return true; 1572 1573 case AMDGPULibFunc::EI_SINH: 1574 Res0 = sinh(opr0); 1575 return true; 1576 1577 case AMDGPULibFunc::EI_SINPI: 1578 Res0 = sin(MATH_PI * opr0); 1579 return true; 1580 1581 case AMDGPULibFunc::EI_TAN: 1582 Res0 = tan(opr0); 1583 return true; 1584 1585 case AMDGPULibFunc::EI_TANH: 1586 Res0 = tanh(opr0); 1587 return true; 1588 1589 case AMDGPULibFunc::EI_TANPI: 1590 Res0 = tan(MATH_PI * opr0); 1591 return true; 1592 1593 // two-arg functions 1594 case AMDGPULibFunc::EI_POW: 1595 case AMDGPULibFunc::EI_POWR: 1596 Res0 = pow(opr0, opr1); 1597 return true; 1598 1599 case AMDGPULibFunc::EI_POWN: { 1600 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1601 double val = (double)iopr1->getSExtValue(); 1602 Res0 = pow(opr0, val); 1603 return true; 1604 } 1605 return false; 1606 } 1607 1608 case AMDGPULibFunc::EI_ROOTN: { 1609 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1610 double val = (double)iopr1->getSExtValue(); 1611 Res0 = pow(opr0, 1.0 / val); 1612 return true; 1613 } 1614 return false; 1615 } 1616 1617 // with ptr arg 1618 case AMDGPULibFunc::EI_SINCOS: 1619 Res0 = sin(opr0); 1620 Res1 = cos(opr0); 1621 return true; 1622 } 1623 1624 return false; 1625 } 1626 1627 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) { 1628 int numArgs = (int)aCI->arg_size(); 1629 if (numArgs > 3) 1630 return false; 1631 1632 Constant *copr0 = nullptr; 1633 Constant *copr1 = nullptr; 1634 if (numArgs > 0) { 1635 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr) 1636 return false; 1637 } 1638 1639 if (numArgs > 1) { 1640 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) { 1641 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS) 1642 return false; 1643 } 1644 } 1645 1646 // At this point, all arguments to aCI are constants. 1647 1648 // max vector size is 16, and sincos will generate two results. 1649 double DVal0[16], DVal1[16]; 1650 int FuncVecSize = getVecSize(FInfo); 1651 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS); 1652 if (FuncVecSize == 1) { 1653 if (!evaluateScalarMathFunc(FInfo, DVal0[0], DVal1[0], copr0, copr1)) { 1654 return false; 1655 } 1656 } else { 1657 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0); 1658 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1); 1659 for (int i = 0; i < FuncVecSize; ++i) { 1660 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr; 1661 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr; 1662 if (!evaluateScalarMathFunc(FInfo, DVal0[i], DVal1[i], celt0, celt1)) { 1663 return false; 1664 } 1665 } 1666 } 1667 1668 LLVMContext &context = aCI->getContext(); 1669 Constant *nval0, *nval1; 1670 if (FuncVecSize == 1) { 1671 nval0 = ConstantFP::get(aCI->getType(), DVal0[0]); 1672 if (hasTwoResults) 1673 nval1 = ConstantFP::get(aCI->getType(), DVal1[0]); 1674 } else { 1675 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1676 SmallVector <float, 0> FVal0, FVal1; 1677 for (int i = 0; i < FuncVecSize; ++i) 1678 FVal0.push_back((float)DVal0[i]); 1679 ArrayRef<float> tmp0(FVal0); 1680 nval0 = ConstantDataVector::get(context, tmp0); 1681 if (hasTwoResults) { 1682 for (int i = 0; i < FuncVecSize; ++i) 1683 FVal1.push_back((float)DVal1[i]); 1684 ArrayRef<float> tmp1(FVal1); 1685 nval1 = ConstantDataVector::get(context, tmp1); 1686 } 1687 } else { 1688 ArrayRef<double> tmp0(DVal0); 1689 nval0 = ConstantDataVector::get(context, tmp0); 1690 if (hasTwoResults) { 1691 ArrayRef<double> tmp1(DVal1); 1692 nval1 = ConstantDataVector::get(context, tmp1); 1693 } 1694 } 1695 } 1696 1697 if (hasTwoResults) { 1698 // sincos 1699 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS && 1700 "math function with ptr arg not supported yet"); 1701 new StoreInst(nval1, aCI->getArgOperand(1), aCI->getIterator()); 1702 } 1703 1704 replaceCall(aCI, nval0); 1705 return true; 1706 } 1707 1708 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F, 1709 FunctionAnalysisManager &AM) { 1710 AMDGPULibCalls Simplifier; 1711 Simplifier.initNativeFuncs(); 1712 Simplifier.initFunction(F, AM); 1713 1714 bool Changed = false; 1715 1716 LLVM_DEBUG(dbgs() << "AMDIC: process function "; 1717 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';); 1718 1719 for (auto &BB : F) { 1720 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) { 1721 // Ignore non-calls. 1722 CallInst *CI = dyn_cast<CallInst>(I); 1723 ++I; 1724 1725 if (CI) { 1726 if (Simplifier.fold(CI)) 1727 Changed = true; 1728 } 1729 } 1730 } 1731 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 1732 } 1733 1734 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F, 1735 FunctionAnalysisManager &AM) { 1736 if (UseNative.empty()) 1737 return PreservedAnalyses::all(); 1738 1739 AMDGPULibCalls Simplifier; 1740 Simplifier.initNativeFuncs(); 1741 Simplifier.initFunction(F, AM); 1742 1743 bool Changed = false; 1744 for (auto &BB : F) { 1745 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) { 1746 // Ignore non-calls. 1747 CallInst *CI = dyn_cast<CallInst>(I); 1748 ++I; 1749 if (CI && Simplifier.useNative(CI)) 1750 Changed = true; 1751 } 1752 } 1753 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 1754 } 1755