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/AliasAnalysis.h" 18 #include "llvm/Analysis/Loads.h" 19 #include "llvm/IR/IRBuilder.h" 20 #include "llvm/IR/IntrinsicInst.h" 21 #include "llvm/IR/IntrinsicsAMDGPU.h" 22 #include "llvm/InitializePasses.h" 23 #include "llvm/Target/TargetMachine.h" 24 25 #define DEBUG_TYPE "amdgpu-simplifylib" 26 27 using namespace llvm; 28 29 static cl::opt<bool> EnablePreLink("amdgpu-prelink", 30 cl::desc("Enable pre-link mode optimizations"), 31 cl::init(false), 32 cl::Hidden); 33 34 static cl::list<std::string> UseNative("amdgpu-use-native", 35 cl::desc("Comma separated list of functions to replace with native, or all"), 36 cl::CommaSeparated, cl::ValueOptional, 37 cl::Hidden); 38 39 #define MATH_PI numbers::pi 40 #define MATH_E numbers::e 41 #define MATH_SQRT2 numbers::sqrt2 42 #define MATH_SQRT1_2 numbers::inv_sqrt2 43 44 namespace llvm { 45 46 class AMDGPULibCalls { 47 private: 48 49 typedef llvm::AMDGPULibFunc FuncInfo; 50 51 const TargetMachine *TM; 52 53 // -fuse-native. 54 bool AllNative = false; 55 56 bool useNativeFunc(const StringRef F) const; 57 58 // Return a pointer (pointer expr) to the function if function definition with 59 // "FuncName" exists. It may create a new function prototype in pre-link mode. 60 FunctionCallee getFunction(Module *M, const FuncInfo &fInfo); 61 62 bool parseFunctionName(const StringRef &FMangledName, FuncInfo &FInfo); 63 64 bool TDOFold(CallInst *CI, const FuncInfo &FInfo); 65 66 /* Specialized optimizations */ 67 68 // recip (half or native) 69 bool fold_recip(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 70 71 // divide (half or native) 72 bool fold_divide(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 73 74 // pow/powr/pown 75 bool fold_pow(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 76 77 // rootn 78 bool fold_rootn(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 79 80 // fma/mad 81 bool fold_fma_mad(CallInst *CI, 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, 88 double& Res1, Constant *copr0, Constant *copr1, Constant *copr2); 89 bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo); 90 91 // sqrt 92 bool fold_sqrt(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 93 94 // sin/cos 95 bool fold_sincos(CallInst * CI, IRBuilder<> &B, AliasAnalysis * AA); 96 97 // __read_pipe/__write_pipe 98 bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 99 const FuncInfo &FInfo); 100 101 // llvm.amdgcn.wavefrontsize 102 bool fold_wavefrontsize(CallInst *CI, IRBuilder<> &B); 103 104 // Get insertion point at entry. 105 BasicBlock::iterator getEntryIns(CallInst * UI); 106 // Insert an Alloc instruction. 107 AllocaInst* insertAlloca(CallInst * UI, IRBuilder<> &B, const char *prefix); 108 // Get a scalar native builtin single argument FP function 109 FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo); 110 111 protected: 112 CallInst *CI; 113 114 bool isUnsafeMath(const CallInst *CI) const; 115 116 void replaceCall(Value *With) { 117 CI->replaceAllUsesWith(With); 118 CI->eraseFromParent(); 119 } 120 121 public: 122 AMDGPULibCalls(const TargetMachine *TM_ = nullptr) : TM(TM_) {} 123 124 bool fold(CallInst *CI, AliasAnalysis *AA = nullptr); 125 126 void initNativeFuncs(); 127 128 // Replace a normal math function call with that native version 129 bool useNative(CallInst *CI); 130 }; 131 132 } // end llvm namespace 133 134 namespace { 135 136 class AMDGPUSimplifyLibCalls : public FunctionPass { 137 138 AMDGPULibCalls Simplifier; 139 140 public: 141 static char ID; // Pass identification 142 143 AMDGPUSimplifyLibCalls(const TargetMachine *TM = nullptr) 144 : FunctionPass(ID), Simplifier(TM) { 145 initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry()); 146 } 147 148 void getAnalysisUsage(AnalysisUsage &AU) const override { 149 AU.addRequired<AAResultsWrapperPass>(); 150 } 151 152 bool runOnFunction(Function &M) override; 153 }; 154 155 class AMDGPUUseNativeCalls : public FunctionPass { 156 157 AMDGPULibCalls Simplifier; 158 159 public: 160 static char ID; // Pass identification 161 162 AMDGPUUseNativeCalls() : FunctionPass(ID) { 163 initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry()); 164 Simplifier.initNativeFuncs(); 165 } 166 167 bool runOnFunction(Function &F) override; 168 }; 169 170 } // end anonymous namespace. 171 172 char AMDGPUSimplifyLibCalls::ID = 0; 173 char AMDGPUUseNativeCalls::ID = 0; 174 175 INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib", 176 "Simplify well-known AMD library calls", false, false) 177 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 178 INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib", 179 "Simplify well-known AMD library calls", false, false) 180 181 INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative", 182 "Replace builtin math calls with that native versions.", 183 false, false) 184 185 template <typename IRB> 186 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg, 187 const Twine &Name = "") { 188 CallInst *R = B.CreateCall(Callee, Arg, Name); 189 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 190 R->setCallingConv(F->getCallingConv()); 191 return R; 192 } 193 194 template <typename IRB> 195 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1, 196 Value *Arg2, const Twine &Name = "") { 197 CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name); 198 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 199 R->setCallingConv(F->getCallingConv()); 200 return R; 201 } 202 203 // Data structures for table-driven optimizations. 204 // FuncTbl works for both f32 and f64 functions with 1 input argument 205 206 struct TableEntry { 207 double result; 208 double input; 209 }; 210 211 /* a list of {result, input} */ 212 static const TableEntry tbl_acos[] = { 213 {MATH_PI / 2.0, 0.0}, 214 {MATH_PI / 2.0, -0.0}, 215 {0.0, 1.0}, 216 {MATH_PI, -1.0} 217 }; 218 static const TableEntry tbl_acosh[] = { 219 {0.0, 1.0} 220 }; 221 static const TableEntry tbl_acospi[] = { 222 {0.5, 0.0}, 223 {0.5, -0.0}, 224 {0.0, 1.0}, 225 {1.0, -1.0} 226 }; 227 static const TableEntry tbl_asin[] = { 228 {0.0, 0.0}, 229 {-0.0, -0.0}, 230 {MATH_PI / 2.0, 1.0}, 231 {-MATH_PI / 2.0, -1.0} 232 }; 233 static const TableEntry tbl_asinh[] = { 234 {0.0, 0.0}, 235 {-0.0, -0.0} 236 }; 237 static const TableEntry tbl_asinpi[] = { 238 {0.0, 0.0}, 239 {-0.0, -0.0}, 240 {0.5, 1.0}, 241 {-0.5, -1.0} 242 }; 243 static const TableEntry tbl_atan[] = { 244 {0.0, 0.0}, 245 {-0.0, -0.0}, 246 {MATH_PI / 4.0, 1.0}, 247 {-MATH_PI / 4.0, -1.0} 248 }; 249 static const TableEntry tbl_atanh[] = { 250 {0.0, 0.0}, 251 {-0.0, -0.0} 252 }; 253 static const TableEntry tbl_atanpi[] = { 254 {0.0, 0.0}, 255 {-0.0, -0.0}, 256 {0.25, 1.0}, 257 {-0.25, -1.0} 258 }; 259 static const TableEntry tbl_cbrt[] = { 260 {0.0, 0.0}, 261 {-0.0, -0.0}, 262 {1.0, 1.0}, 263 {-1.0, -1.0}, 264 }; 265 static const TableEntry tbl_cos[] = { 266 {1.0, 0.0}, 267 {1.0, -0.0} 268 }; 269 static const TableEntry tbl_cosh[] = { 270 {1.0, 0.0}, 271 {1.0, -0.0} 272 }; 273 static const TableEntry tbl_cospi[] = { 274 {1.0, 0.0}, 275 {1.0, -0.0} 276 }; 277 static const TableEntry tbl_erfc[] = { 278 {1.0, 0.0}, 279 {1.0, -0.0} 280 }; 281 static const TableEntry tbl_erf[] = { 282 {0.0, 0.0}, 283 {-0.0, -0.0} 284 }; 285 static const TableEntry tbl_exp[] = { 286 {1.0, 0.0}, 287 {1.0, -0.0}, 288 {MATH_E, 1.0} 289 }; 290 static const TableEntry tbl_exp2[] = { 291 {1.0, 0.0}, 292 {1.0, -0.0}, 293 {2.0, 1.0} 294 }; 295 static const TableEntry tbl_exp10[] = { 296 {1.0, 0.0}, 297 {1.0, -0.0}, 298 {10.0, 1.0} 299 }; 300 static const TableEntry tbl_expm1[] = { 301 {0.0, 0.0}, 302 {-0.0, -0.0} 303 }; 304 static const TableEntry tbl_log[] = { 305 {0.0, 1.0}, 306 {1.0, MATH_E} 307 }; 308 static const TableEntry tbl_log2[] = { 309 {0.0, 1.0}, 310 {1.0, 2.0} 311 }; 312 static const TableEntry tbl_log10[] = { 313 {0.0, 1.0}, 314 {1.0, 10.0} 315 }; 316 static const TableEntry tbl_rsqrt[] = { 317 {1.0, 1.0}, 318 {MATH_SQRT1_2, 2.0} 319 }; 320 static const TableEntry tbl_sin[] = { 321 {0.0, 0.0}, 322 {-0.0, -0.0} 323 }; 324 static const TableEntry tbl_sinh[] = { 325 {0.0, 0.0}, 326 {-0.0, -0.0} 327 }; 328 static const TableEntry tbl_sinpi[] = { 329 {0.0, 0.0}, 330 {-0.0, -0.0} 331 }; 332 static const TableEntry tbl_sqrt[] = { 333 {0.0, 0.0}, 334 {1.0, 1.0}, 335 {MATH_SQRT2, 2.0} 336 }; 337 static const TableEntry tbl_tan[] = { 338 {0.0, 0.0}, 339 {-0.0, -0.0} 340 }; 341 static const TableEntry tbl_tanh[] = { 342 {0.0, 0.0}, 343 {-0.0, -0.0} 344 }; 345 static const TableEntry tbl_tanpi[] = { 346 {0.0, 0.0}, 347 {-0.0, -0.0} 348 }; 349 static const TableEntry tbl_tgamma[] = { 350 {1.0, 1.0}, 351 {1.0, 2.0}, 352 {2.0, 3.0}, 353 {6.0, 4.0} 354 }; 355 356 static bool HasNative(AMDGPULibFunc::EFuncId id) { 357 switch(id) { 358 case AMDGPULibFunc::EI_DIVIDE: 359 case AMDGPULibFunc::EI_COS: 360 case AMDGPULibFunc::EI_EXP: 361 case AMDGPULibFunc::EI_EXP2: 362 case AMDGPULibFunc::EI_EXP10: 363 case AMDGPULibFunc::EI_LOG: 364 case AMDGPULibFunc::EI_LOG2: 365 case AMDGPULibFunc::EI_LOG10: 366 case AMDGPULibFunc::EI_POWR: 367 case AMDGPULibFunc::EI_RECIP: 368 case AMDGPULibFunc::EI_RSQRT: 369 case AMDGPULibFunc::EI_SIN: 370 case AMDGPULibFunc::EI_SINCOS: 371 case AMDGPULibFunc::EI_SQRT: 372 case AMDGPULibFunc::EI_TAN: 373 return true; 374 default:; 375 } 376 return false; 377 } 378 379 using TableRef = ArrayRef<TableEntry>; 380 381 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) { 382 switch(id) { 383 case AMDGPULibFunc::EI_ACOS: return TableRef(tbl_acos); 384 case AMDGPULibFunc::EI_ACOSH: return TableRef(tbl_acosh); 385 case AMDGPULibFunc::EI_ACOSPI: return TableRef(tbl_acospi); 386 case AMDGPULibFunc::EI_ASIN: return TableRef(tbl_asin); 387 case AMDGPULibFunc::EI_ASINH: return TableRef(tbl_asinh); 388 case AMDGPULibFunc::EI_ASINPI: return TableRef(tbl_asinpi); 389 case AMDGPULibFunc::EI_ATAN: return TableRef(tbl_atan); 390 case AMDGPULibFunc::EI_ATANH: return TableRef(tbl_atanh); 391 case AMDGPULibFunc::EI_ATANPI: return TableRef(tbl_atanpi); 392 case AMDGPULibFunc::EI_CBRT: return TableRef(tbl_cbrt); 393 case AMDGPULibFunc::EI_NCOS: 394 case AMDGPULibFunc::EI_COS: return TableRef(tbl_cos); 395 case AMDGPULibFunc::EI_COSH: return TableRef(tbl_cosh); 396 case AMDGPULibFunc::EI_COSPI: return TableRef(tbl_cospi); 397 case AMDGPULibFunc::EI_ERFC: return TableRef(tbl_erfc); 398 case AMDGPULibFunc::EI_ERF: return TableRef(tbl_erf); 399 case AMDGPULibFunc::EI_EXP: return TableRef(tbl_exp); 400 case AMDGPULibFunc::EI_NEXP2: 401 case AMDGPULibFunc::EI_EXP2: return TableRef(tbl_exp2); 402 case AMDGPULibFunc::EI_EXP10: return TableRef(tbl_exp10); 403 case AMDGPULibFunc::EI_EXPM1: return TableRef(tbl_expm1); 404 case AMDGPULibFunc::EI_LOG: return TableRef(tbl_log); 405 case AMDGPULibFunc::EI_NLOG2: 406 case AMDGPULibFunc::EI_LOG2: return TableRef(tbl_log2); 407 case AMDGPULibFunc::EI_LOG10: return TableRef(tbl_log10); 408 case AMDGPULibFunc::EI_NRSQRT: 409 case AMDGPULibFunc::EI_RSQRT: return TableRef(tbl_rsqrt); 410 case AMDGPULibFunc::EI_NSIN: 411 case AMDGPULibFunc::EI_SIN: return TableRef(tbl_sin); 412 case AMDGPULibFunc::EI_SINH: return TableRef(tbl_sinh); 413 case AMDGPULibFunc::EI_SINPI: return TableRef(tbl_sinpi); 414 case AMDGPULibFunc::EI_NSQRT: 415 case AMDGPULibFunc::EI_SQRT: return TableRef(tbl_sqrt); 416 case AMDGPULibFunc::EI_TAN: return TableRef(tbl_tan); 417 case AMDGPULibFunc::EI_TANH: return TableRef(tbl_tanh); 418 case AMDGPULibFunc::EI_TANPI: return TableRef(tbl_tanpi); 419 case AMDGPULibFunc::EI_TGAMMA: return TableRef(tbl_tgamma); 420 default:; 421 } 422 return TableRef(); 423 } 424 425 static inline int getVecSize(const AMDGPULibFunc& FInfo) { 426 return FInfo.getLeads()[0].VectorSize; 427 } 428 429 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) { 430 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType; 431 } 432 433 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) { 434 // If we are doing PreLinkOpt, the function is external. So it is safe to 435 // use getOrInsertFunction() at this stage. 436 437 return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo) 438 : AMDGPULibFunc::getFunction(M, fInfo); 439 } 440 441 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName, 442 FuncInfo &FInfo) { 443 return AMDGPULibFunc::parse(FMangledName, FInfo); 444 } 445 446 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const { 447 if (auto Op = dyn_cast<FPMathOperator>(CI)) 448 if (Op->isFast()) 449 return true; 450 const Function *F = CI->getParent()->getParent(); 451 Attribute Attr = F->getFnAttribute("unsafe-fp-math"); 452 return Attr.getValueAsBool(); 453 } 454 455 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const { 456 return AllNative || llvm::is_contained(UseNative, F); 457 } 458 459 void AMDGPULibCalls::initNativeFuncs() { 460 AllNative = useNativeFunc("all") || 461 (UseNative.getNumOccurrences() && UseNative.size() == 1 && 462 UseNative.begin()->empty()); 463 } 464 465 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) { 466 bool native_sin = useNativeFunc("sin"); 467 bool native_cos = useNativeFunc("cos"); 468 469 if (native_sin && native_cos) { 470 Module *M = aCI->getModule(); 471 Value *opr0 = aCI->getArgOperand(0); 472 473 AMDGPULibFunc nf; 474 nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType; 475 nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize; 476 477 nf.setPrefix(AMDGPULibFunc::NATIVE); 478 nf.setId(AMDGPULibFunc::EI_SIN); 479 FunctionCallee sinExpr = getFunction(M, nf); 480 481 nf.setPrefix(AMDGPULibFunc::NATIVE); 482 nf.setId(AMDGPULibFunc::EI_COS); 483 FunctionCallee cosExpr = getFunction(M, nf); 484 if (sinExpr && cosExpr) { 485 Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI); 486 Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI); 487 new StoreInst(cosval, aCI->getArgOperand(1), aCI); 488 489 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 490 << " with native version of sin/cos"); 491 492 replaceCall(sinval); 493 return true; 494 } 495 } 496 return false; 497 } 498 499 bool AMDGPULibCalls::useNative(CallInst *aCI) { 500 CI = aCI; 501 Function *Callee = aCI->getCalledFunction(); 502 503 FuncInfo FInfo; 504 if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() || 505 FInfo.getPrefix() != AMDGPULibFunc::NOPFX || 506 getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) || 507 !(AllNative || useNativeFunc(FInfo.getName()))) { 508 return false; 509 } 510 511 if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS) 512 return sincosUseNative(aCI, FInfo); 513 514 FInfo.setPrefix(AMDGPULibFunc::NATIVE); 515 FunctionCallee F = getFunction(aCI->getModule(), FInfo); 516 if (!F) 517 return false; 518 519 aCI->setCalledFunction(F); 520 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 521 << " with native version"); 522 return true; 523 } 524 525 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe 526 // builtin, with appended type size and alignment arguments, where 2 or 4 527 // indicates the original number of arguments. The library has optimized version 528 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same 529 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N 530 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ..., 531 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4. 532 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 533 const FuncInfo &FInfo) { 534 auto *Callee = CI->getCalledFunction(); 535 if (!Callee->isDeclaration()) 536 return false; 537 538 assert(Callee->hasName() && "Invalid read_pipe/write_pipe function"); 539 auto *M = Callee->getParent(); 540 auto &Ctx = M->getContext(); 541 std::string Name = std::string(Callee->getName()); 542 auto NumArg = CI->arg_size(); 543 if (NumArg != 4 && NumArg != 6) 544 return false; 545 auto *PacketSize = CI->getArgOperand(NumArg - 2); 546 auto *PacketAlign = CI->getArgOperand(NumArg - 1); 547 if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign)) 548 return false; 549 unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue(); 550 Align Alignment = cast<ConstantInt>(PacketAlign)->getAlignValue(); 551 if (Alignment != Size) 552 return false; 553 554 Type *PtrElemTy; 555 if (Size <= 8) 556 PtrElemTy = Type::getIntNTy(Ctx, Size * 8); 557 else 558 PtrElemTy = FixedVectorType::get(Type::getInt64Ty(Ctx), Size / 8); 559 unsigned PtrArgLoc = CI->arg_size() - 3; 560 auto PtrArg = CI->getArgOperand(PtrArgLoc); 561 unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace(); 562 auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS); 563 564 SmallVector<llvm::Type *, 6> ArgTys; 565 for (unsigned I = 0; I != PtrArgLoc; ++I) 566 ArgTys.push_back(CI->getArgOperand(I)->getType()); 567 ArgTys.push_back(PtrTy); 568 569 Name = Name + "_" + std::to_string(Size); 570 auto *FTy = FunctionType::get(Callee->getReturnType(), 571 ArrayRef<Type *>(ArgTys), false); 572 AMDGPULibFunc NewLibFunc(Name, FTy); 573 FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc); 574 if (!F) 575 return false; 576 577 auto *BCast = B.CreatePointerCast(PtrArg, PtrTy); 578 SmallVector<Value *, 6> Args; 579 for (unsigned I = 0; I != PtrArgLoc; ++I) 580 Args.push_back(CI->getArgOperand(I)); 581 Args.push_back(BCast); 582 583 auto *NCI = B.CreateCall(F, Args); 584 NCI->setAttributes(CI->getAttributes()); 585 CI->replaceAllUsesWith(NCI); 586 CI->dropAllReferences(); 587 CI->eraseFromParent(); 588 589 return true; 590 } 591 592 // This function returns false if no change; return true otherwise. 593 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) { 594 this->CI = CI; 595 Function *Callee = CI->getCalledFunction(); 596 597 // Ignore indirect calls. 598 if (Callee == nullptr) 599 return false; 600 601 BasicBlock *BB = CI->getParent(); 602 LLVMContext &Context = CI->getParent()->getContext(); 603 IRBuilder<> B(Context); 604 605 // Set the builder to the instruction after the call. 606 B.SetInsertPoint(BB, CI->getIterator()); 607 608 // Copy fast flags from the original call. 609 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI)) 610 B.setFastMathFlags(FPOp->getFastMathFlags()); 611 612 switch (Callee->getIntrinsicID()) { 613 default: 614 break; 615 case Intrinsic::amdgcn_wavefrontsize: 616 return !EnablePreLink && fold_wavefrontsize(CI, B); 617 } 618 619 FuncInfo FInfo; 620 if (!parseFunctionName(Callee->getName(), FInfo)) 621 return false; 622 623 // Further check the number of arguments to see if they match. 624 if (CI->arg_size() != FInfo.getNumArgs()) 625 return false; 626 627 if (TDOFold(CI, FInfo)) 628 return true; 629 630 // Under unsafe-math, evaluate calls if possible. 631 // According to Brian Sumner, we can do this for all f32 function calls 632 // using host's double function calls. 633 if (isUnsafeMath(CI) && evaluateCall(CI, FInfo)) 634 return true; 635 636 // Specialized optimizations for each function call 637 switch (FInfo.getId()) { 638 case AMDGPULibFunc::EI_RECIP: 639 // skip vector function 640 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || 641 FInfo.getPrefix() == AMDGPULibFunc::HALF) && 642 "recip must be an either native or half function"); 643 return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo); 644 645 case AMDGPULibFunc::EI_DIVIDE: 646 // skip vector function 647 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || 648 FInfo.getPrefix() == AMDGPULibFunc::HALF) && 649 "divide must be an either native or half function"); 650 return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo); 651 652 case AMDGPULibFunc::EI_POW: 653 case AMDGPULibFunc::EI_POWR: 654 case AMDGPULibFunc::EI_POWN: 655 return fold_pow(CI, B, FInfo); 656 657 case AMDGPULibFunc::EI_ROOTN: 658 // skip vector function 659 return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo); 660 661 case AMDGPULibFunc::EI_FMA: 662 case AMDGPULibFunc::EI_MAD: 663 case AMDGPULibFunc::EI_NFMA: 664 // skip vector function 665 return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo); 666 667 case AMDGPULibFunc::EI_SQRT: 668 return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo); 669 case AMDGPULibFunc::EI_COS: 670 case AMDGPULibFunc::EI_SIN: 671 if ((getArgType(FInfo) == AMDGPULibFunc::F32 || 672 getArgType(FInfo) == AMDGPULibFunc::F64) 673 && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX)) 674 return fold_sincos(CI, B, AA); 675 676 break; 677 case AMDGPULibFunc::EI_READ_PIPE_2: 678 case AMDGPULibFunc::EI_READ_PIPE_4: 679 case AMDGPULibFunc::EI_WRITE_PIPE_2: 680 case AMDGPULibFunc::EI_WRITE_PIPE_4: 681 return fold_read_write_pipe(CI, B, FInfo); 682 683 default: 684 break; 685 } 686 687 return false; 688 } 689 690 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) { 691 // Table-Driven optimization 692 const TableRef tr = getOptTable(FInfo.getId()); 693 if (tr.empty()) 694 return false; 695 696 int const sz = (int)tr.size(); 697 const TableEntry * const ftbl = tr.data(); 698 Value *opr0 = CI->getArgOperand(0); 699 700 if (getVecSize(FInfo) > 1) { 701 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) { 702 SmallVector<double, 0> DVal; 703 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) { 704 ConstantFP *eltval = dyn_cast<ConstantFP>( 705 CV->getElementAsConstant((unsigned)eltNo)); 706 assert(eltval && "Non-FP arguments in math function!"); 707 bool found = false; 708 for (int i=0; i < sz; ++i) { 709 if (eltval->isExactlyValue(ftbl[i].input)) { 710 DVal.push_back(ftbl[i].result); 711 found = true; 712 break; 713 } 714 } 715 if (!found) { 716 // This vector constants not handled yet. 717 return false; 718 } 719 } 720 LLVMContext &context = CI->getParent()->getParent()->getContext(); 721 Constant *nval; 722 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 723 SmallVector<float, 0> FVal; 724 for (unsigned i = 0; i < DVal.size(); ++i) { 725 FVal.push_back((float)DVal[i]); 726 } 727 ArrayRef<float> tmp(FVal); 728 nval = ConstantDataVector::get(context, tmp); 729 } else { // F64 730 ArrayRef<double> tmp(DVal); 731 nval = ConstantDataVector::get(context, tmp); 732 } 733 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 734 replaceCall(nval); 735 return true; 736 } 737 } else { 738 // Scalar version 739 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) { 740 for (int i = 0; i < sz; ++i) { 741 if (CF->isExactlyValue(ftbl[i].input)) { 742 Value *nval = ConstantFP::get(CF->getType(), ftbl[i].result); 743 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 744 replaceCall(nval); 745 return true; 746 } 747 } 748 } 749 } 750 751 return false; 752 } 753 754 // [native_]half_recip(c) ==> 1.0/c 755 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B, 756 const FuncInfo &FInfo) { 757 Value *opr0 = CI->getArgOperand(0); 758 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) { 759 // Just create a normal div. Later, InstCombine will be able 760 // to compute the divide into a constant (avoid check float infinity 761 // or subnormal at this point). 762 Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0), 763 opr0, 764 "recip2div"); 765 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 766 replaceCall(nval); 767 return true; 768 } 769 return false; 770 } 771 772 // [native_]half_divide(x, c) ==> x/c 773 bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B, 774 const FuncInfo &FInfo) { 775 Value *opr0 = CI->getArgOperand(0); 776 Value *opr1 = CI->getArgOperand(1); 777 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0); 778 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1); 779 780 if ((CF0 && CF1) || // both are constants 781 (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32))) 782 // CF1 is constant && f32 divide 783 { 784 Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0), 785 opr1, "__div2recip"); 786 Value *nval = B.CreateFMul(opr0, nval1, "__div2mul"); 787 replaceCall(nval); 788 return true; 789 } 790 return false; 791 } 792 793 namespace llvm { 794 static double log2(double V) { 795 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L 796 return ::log2(V); 797 #else 798 return log(V) / numbers::ln2; 799 #endif 800 } 801 } 802 803 bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B, 804 const FuncInfo &FInfo) { 805 assert((FInfo.getId() == AMDGPULibFunc::EI_POW || 806 FInfo.getId() == AMDGPULibFunc::EI_POWR || 807 FInfo.getId() == AMDGPULibFunc::EI_POWN) && 808 "fold_pow: encounter a wrong function call"); 809 810 Value *opr0, *opr1; 811 ConstantFP *CF; 812 ConstantInt *CINT; 813 ConstantAggregateZero *CZero; 814 Type *eltType; 815 816 opr0 = CI->getArgOperand(0); 817 opr1 = CI->getArgOperand(1); 818 CZero = dyn_cast<ConstantAggregateZero>(opr1); 819 if (getVecSize(FInfo) == 1) { 820 eltType = opr0->getType(); 821 CF = dyn_cast<ConstantFP>(opr1); 822 CINT = dyn_cast<ConstantInt>(opr1); 823 } else { 824 VectorType *VTy = dyn_cast<VectorType>(opr0->getType()); 825 assert(VTy && "Oprand of vector function should be of vectortype"); 826 eltType = VTy->getElementType(); 827 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1); 828 829 // Now, only Handle vector const whose elements have the same value. 830 CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr; 831 CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr; 832 } 833 834 // No unsafe math , no constant argument, do nothing 835 if (!isUnsafeMath(CI) && !CF && !CINT && !CZero) 836 return false; 837 838 // 0x1111111 means that we don't do anything for this call. 839 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111); 840 841 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) { 842 // pow/powr/pown(x, 0) == 1 843 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n"); 844 Constant *cnval = ConstantFP::get(eltType, 1.0); 845 if (getVecSize(FInfo) > 1) { 846 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 847 } 848 replaceCall(cnval); 849 return true; 850 } 851 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) { 852 // pow/powr/pown(x, 1.0) = x 853 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n"); 854 replaceCall(opr0); 855 return true; 856 } 857 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) { 858 // pow/powr/pown(x, 2.0) = x*x 859 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0 860 << "\n"); 861 Value *nval = B.CreateFMul(opr0, opr0, "__pow2"); 862 replaceCall(nval); 863 return true; 864 } 865 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) { 866 // pow/powr/pown(x, -1.0) = 1.0/x 867 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n"); 868 Constant *cnval = ConstantFP::get(eltType, 1.0); 869 if (getVecSize(FInfo) > 1) { 870 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 871 } 872 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip"); 873 replaceCall(nval); 874 return true; 875 } 876 877 Module *M = CI->getModule(); 878 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) { 879 // pow[r](x, [-]0.5) = sqrt(x) 880 bool issqrt = CF->isExactlyValue(0.5); 881 if (FunctionCallee FPExpr = 882 getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT 883 : AMDGPULibFunc::EI_RSQRT, 884 FInfo))) { 885 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 886 << FInfo.getName().c_str() << "(" << *opr0 << ")\n"); 887 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt" 888 : "__pow2rsqrt"); 889 replaceCall(nval); 890 return true; 891 } 892 } 893 894 if (!isUnsafeMath(CI)) 895 return false; 896 897 // Unsafe Math optimization 898 899 // Remember that ci_opr1 is set if opr1 is integral 900 if (CF) { 901 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32) 902 ? (double)CF->getValueAPF().convertToFloat() 903 : CF->getValueAPF().convertToDouble(); 904 int ival = (int)dval; 905 if ((double)ival == dval) { 906 ci_opr1 = ival; 907 } else 908 ci_opr1 = 0x11111111; 909 } 910 911 // pow/powr/pown(x, c) = [1/](x*x*..x); where 912 // trunc(c) == c && the number of x == c && |c| <= 12 913 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1; 914 if (abs_opr1 <= 12) { 915 Constant *cnval; 916 Value *nval; 917 if (abs_opr1 == 0) { 918 cnval = ConstantFP::get(eltType, 1.0); 919 if (getVecSize(FInfo) > 1) { 920 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 921 } 922 nval = cnval; 923 } else { 924 Value *valx2 = nullptr; 925 nval = nullptr; 926 while (abs_opr1 > 0) { 927 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0; 928 if (abs_opr1 & 1) { 929 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2; 930 } 931 abs_opr1 >>= 1; 932 } 933 } 934 935 if (ci_opr1 < 0) { 936 cnval = ConstantFP::get(eltType, 1.0); 937 if (getVecSize(FInfo) > 1) { 938 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 939 } 940 nval = B.CreateFDiv(cnval, nval, "__1powprod"); 941 } 942 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 943 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0 944 << ")\n"); 945 replaceCall(nval); 946 return true; 947 } 948 949 // powr ---> exp2(y * log2(x)) 950 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31)) 951 FunctionCallee ExpExpr = 952 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo)); 953 if (!ExpExpr) 954 return false; 955 956 bool needlog = false; 957 bool needabs = false; 958 bool needcopysign = false; 959 Constant *cnval = nullptr; 960 if (getVecSize(FInfo) == 1) { 961 CF = dyn_cast<ConstantFP>(opr0); 962 963 if (CF) { 964 double V = (getArgType(FInfo) == AMDGPULibFunc::F32) 965 ? (double)CF->getValueAPF().convertToFloat() 966 : CF->getValueAPF().convertToDouble(); 967 968 V = log2(std::abs(V)); 969 cnval = ConstantFP::get(eltType, V); 970 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) && 971 CF->isNegative(); 972 } else { 973 needlog = true; 974 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR && 975 (!CF || CF->isNegative()); 976 } 977 } else { 978 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0); 979 980 if (!CDV) { 981 needlog = true; 982 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR; 983 } else { 984 assert ((int)CDV->getNumElements() == getVecSize(FInfo) && 985 "Wrong vector size detected"); 986 987 SmallVector<double, 0> DVal; 988 for (int i=0; i < getVecSize(FInfo); ++i) { 989 double V = (getArgType(FInfo) == AMDGPULibFunc::F32) 990 ? (double)CDV->getElementAsFloat(i) 991 : CDV->getElementAsDouble(i); 992 if (V < 0.0) needcopysign = true; 993 V = log2(std::abs(V)); 994 DVal.push_back(V); 995 } 996 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 997 SmallVector<float, 0> FVal; 998 for (unsigned i=0; i < DVal.size(); ++i) { 999 FVal.push_back((float)DVal[i]); 1000 } 1001 ArrayRef<float> tmp(FVal); 1002 cnval = ConstantDataVector::get(M->getContext(), tmp); 1003 } else { 1004 ArrayRef<double> tmp(DVal); 1005 cnval = ConstantDataVector::get(M->getContext(), tmp); 1006 } 1007 } 1008 } 1009 1010 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) { 1011 // We cannot handle corner cases for a general pow() function, give up 1012 // unless y is a constant integral value. Then proceed as if it were pown. 1013 if (getVecSize(FInfo) == 1) { 1014 if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) { 1015 double y = (getArgType(FInfo) == AMDGPULibFunc::F32) 1016 ? (double)CF->getValueAPF().convertToFloat() 1017 : CF->getValueAPF().convertToDouble(); 1018 if (y != (double)(int64_t)y) 1019 return false; 1020 } else 1021 return false; 1022 } else { 1023 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) { 1024 for (int i=0; i < getVecSize(FInfo); ++i) { 1025 double y = (getArgType(FInfo) == AMDGPULibFunc::F32) 1026 ? (double)CDV->getElementAsFloat(i) 1027 : CDV->getElementAsDouble(i); 1028 if (y != (double)(int64_t)y) 1029 return false; 1030 } 1031 } else 1032 return false; 1033 } 1034 } 1035 1036 Value *nval; 1037 if (needabs) { 1038 FunctionCallee AbsExpr = 1039 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS, FInfo)); 1040 if (!AbsExpr) 1041 return false; 1042 nval = CreateCallEx(B, AbsExpr, opr0, "__fabs"); 1043 } else { 1044 nval = cnval ? cnval : opr0; 1045 } 1046 if (needlog) { 1047 FunctionCallee LogExpr = 1048 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo)); 1049 if (!LogExpr) 1050 return false; 1051 nval = CreateCallEx(B,LogExpr, nval, "__log2"); 1052 } 1053 1054 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) { 1055 // convert int(32) to fp(f32 or f64) 1056 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F"); 1057 } 1058 nval = B.CreateFMul(opr1, nval, "__ylogx"); 1059 nval = CreateCallEx(B,ExpExpr, nval, "__exp2"); 1060 1061 if (needcopysign) { 1062 Value *opr_n; 1063 Type* rTy = opr0->getType(); 1064 Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty(); 1065 Type *nTy = nTyS; 1066 if (const auto *vTy = dyn_cast<FixedVectorType>(rTy)) 1067 nTy = FixedVectorType::get(nTyS, vTy); 1068 unsigned size = nTy->getScalarSizeInBits(); 1069 opr_n = CI->getArgOperand(1); 1070 if (opr_n->getType()->isIntegerTy()) 1071 opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou"); 1072 else 1073 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou"); 1074 1075 Value *sign = B.CreateShl(opr_n, size-1, "__yeven"); 1076 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign"); 1077 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign); 1078 nval = B.CreateBitCast(nval, opr0->getType()); 1079 } 1080 1081 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 1082 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n"); 1083 replaceCall(nval); 1084 1085 return true; 1086 } 1087 1088 bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B, 1089 const FuncInfo &FInfo) { 1090 Value *opr0 = CI->getArgOperand(0); 1091 Value *opr1 = CI->getArgOperand(1); 1092 1093 ConstantInt *CINT = dyn_cast<ConstantInt>(opr1); 1094 if (!CINT) { 1095 return false; 1096 } 1097 int ci_opr1 = (int)CINT->getSExtValue(); 1098 if (ci_opr1 == 1) { // rootn(x, 1) = x 1099 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n"); 1100 replaceCall(opr0); 1101 return true; 1102 } 1103 if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x) 1104 Module *M = CI->getModule(); 1105 if (FunctionCallee FPExpr = 1106 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) { 1107 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n"); 1108 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt"); 1109 replaceCall(nval); 1110 return true; 1111 } 1112 } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x) 1113 Module *M = CI->getModule(); 1114 if (FunctionCallee FPExpr = 1115 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) { 1116 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n"); 1117 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt"); 1118 replaceCall(nval); 1119 return true; 1120 } 1121 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x 1122 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n"); 1123 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0), 1124 opr0, 1125 "__rootn2div"); 1126 replaceCall(nval); 1127 return true; 1128 } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x) 1129 Module *M = CI->getModule(); 1130 if (FunctionCallee FPExpr = 1131 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) { 1132 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0 1133 << ")\n"); 1134 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt"); 1135 replaceCall(nval); 1136 return true; 1137 } 1138 } 1139 return false; 1140 } 1141 1142 bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B, 1143 const FuncInfo &FInfo) { 1144 Value *opr0 = CI->getArgOperand(0); 1145 Value *opr1 = CI->getArgOperand(1); 1146 Value *opr2 = CI->getArgOperand(2); 1147 1148 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0); 1149 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1); 1150 if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) { 1151 // fma/mad(a, b, c) = c if a=0 || b=0 1152 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n"); 1153 replaceCall(opr2); 1154 return true; 1155 } 1156 if (CF0 && CF0->isExactlyValue(1.0f)) { 1157 // fma/mad(a, b, c) = b+c if a=1 1158 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2 1159 << "\n"); 1160 Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd"); 1161 replaceCall(nval); 1162 return true; 1163 } 1164 if (CF1 && CF1->isExactlyValue(1.0f)) { 1165 // fma/mad(a, b, c) = a+c if b=1 1166 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2 1167 << "\n"); 1168 Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd"); 1169 replaceCall(nval); 1170 return true; 1171 } 1172 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) { 1173 if (CF->isZero()) { 1174 // fma/mad(a, b, c) = a*b if c=0 1175 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " 1176 << *opr1 << "\n"); 1177 Value *nval = B.CreateFMul(opr0, opr1, "fmamul"); 1178 replaceCall(nval); 1179 return true; 1180 } 1181 } 1182 1183 return false; 1184 } 1185 1186 // Get a scalar native builtin single argument FP function 1187 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M, 1188 const FuncInfo &FInfo) { 1189 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId())) 1190 return nullptr; 1191 FuncInfo nf = FInfo; 1192 nf.setPrefix(AMDGPULibFunc::NATIVE); 1193 return getFunction(M, nf); 1194 } 1195 1196 // fold sqrt -> native_sqrt (x) 1197 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B, 1198 const FuncInfo &FInfo) { 1199 if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) && 1200 (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) { 1201 if (FunctionCallee FPExpr = getNativeFunction( 1202 CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) { 1203 Value *opr0 = CI->getArgOperand(0); 1204 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 1205 << "sqrt(" << *opr0 << ")\n"); 1206 Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt"); 1207 replaceCall(nval); 1208 return true; 1209 } 1210 } 1211 return false; 1212 } 1213 1214 // fold sin, cos -> sincos. 1215 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B, 1216 AliasAnalysis *AA) { 1217 AMDGPULibFunc fInfo; 1218 if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo)) 1219 return false; 1220 1221 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN || 1222 fInfo.getId() == AMDGPULibFunc::EI_COS); 1223 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN; 1224 1225 Value *CArgVal = CI->getArgOperand(0); 1226 BasicBlock * const CBB = CI->getParent(); 1227 1228 int const MaxScan = 30; 1229 bool Changed = false; 1230 1231 { // fold in load value. 1232 LoadInst *LI = dyn_cast<LoadInst>(CArgVal); 1233 if (LI && LI->getParent() == CBB) { 1234 BasicBlock::iterator BBI = LI->getIterator(); 1235 Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA); 1236 if (AvailableVal) { 1237 Changed = true; 1238 CArgVal->replaceAllUsesWith(AvailableVal); 1239 if (CArgVal->getNumUses() == 0) 1240 LI->eraseFromParent(); 1241 CArgVal = CI->getArgOperand(0); 1242 } 1243 } 1244 } 1245 1246 Module *M = CI->getModule(); 1247 fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN); 1248 std::string const PairName = fInfo.mangle(); 1249 1250 CallInst *UI = nullptr; 1251 for (User* U : CArgVal->users()) { 1252 CallInst *XI = dyn_cast_or_null<CallInst>(U); 1253 if (!XI || XI == CI || XI->getParent() != CBB) 1254 continue; 1255 1256 Function *UCallee = XI->getCalledFunction(); 1257 if (!UCallee || !UCallee->getName().equals(PairName)) 1258 continue; 1259 1260 BasicBlock::iterator BBI = CI->getIterator(); 1261 if (BBI == CI->getParent()->begin()) 1262 break; 1263 --BBI; 1264 for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) { 1265 if (cast<Instruction>(BBI) == XI) { 1266 UI = XI; 1267 break; 1268 } 1269 } 1270 if (UI) break; 1271 } 1272 1273 if (!UI) 1274 return Changed; 1275 1276 // Merge the sin and cos. 1277 1278 // for OpenCL 2.0 we have only generic implementation of sincos 1279 // function. 1280 AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo); 1281 nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS); 1282 FunctionCallee Fsincos = getFunction(M, nf); 1283 if (!Fsincos) 1284 return Changed; 1285 1286 BasicBlock::iterator ItOld = B.GetInsertPoint(); 1287 AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_"); 1288 B.SetInsertPoint(UI); 1289 1290 Value *P = Alloc; 1291 Type *PTy = Fsincos.getFunctionType()->getParamType(1); 1292 // The allocaInst allocates the memory in private address space. This need 1293 // to be bitcasted to point to the address space of cos pointer type. 1294 // In OpenCL 2.0 this is generic, while in 1.2 that is private. 1295 if (PTy->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) 1296 P = B.CreateAddrSpaceCast(Alloc, PTy); 1297 CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P); 1298 1299 LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with " 1300 << *Call << "\n"); 1301 1302 if (!isSin) { // CI->cos, UI->sin 1303 B.SetInsertPoint(&*ItOld); 1304 UI->replaceAllUsesWith(&*Call); 1305 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc); 1306 CI->replaceAllUsesWith(Reload); 1307 UI->eraseFromParent(); 1308 CI->eraseFromParent(); 1309 } else { // CI->sin, UI->cos 1310 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc); 1311 UI->replaceAllUsesWith(Reload); 1312 CI->replaceAllUsesWith(Call); 1313 UI->eraseFromParent(); 1314 CI->eraseFromParent(); 1315 } 1316 return true; 1317 } 1318 1319 bool AMDGPULibCalls::fold_wavefrontsize(CallInst *CI, IRBuilder<> &B) { 1320 if (!TM) 1321 return false; 1322 1323 StringRef CPU = TM->getTargetCPU(); 1324 StringRef Features = TM->getTargetFeatureString(); 1325 if ((CPU.empty() || CPU.equals_insensitive("generic")) && 1326 (Features.empty() || !Features.contains_insensitive("wavefrontsize"))) 1327 return false; 1328 1329 Function *F = CI->getParent()->getParent(); 1330 const GCNSubtarget &ST = TM->getSubtarget<GCNSubtarget>(*F); 1331 unsigned N = ST.getWavefrontSize(); 1332 1333 LLVM_DEBUG(errs() << "AMDIC: fold_wavefrontsize (" << *CI << ") with " 1334 << N << "\n"); 1335 1336 CI->replaceAllUsesWith(ConstantInt::get(B.getInt32Ty(), N)); 1337 CI->eraseFromParent(); 1338 return true; 1339 } 1340 1341 // Get insertion point at entry. 1342 BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) { 1343 Function * Func = UI->getParent()->getParent(); 1344 BasicBlock * BB = &Func->getEntryBlock(); 1345 assert(BB && "Entry block not found!"); 1346 BasicBlock::iterator ItNew = BB->begin(); 1347 return ItNew; 1348 } 1349 1350 // Insert a AllocsInst at the beginning of function entry block. 1351 AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B, 1352 const char *prefix) { 1353 BasicBlock::iterator ItNew = getEntryIns(UI); 1354 Function *UCallee = UI->getCalledFunction(); 1355 Type *RetType = UCallee->getReturnType(); 1356 B.SetInsertPoint(&*ItNew); 1357 AllocaInst *Alloc = 1358 B.CreateAlloca(RetType, nullptr, std::string(prefix) + UI->getName()); 1359 Alloc->setAlignment( 1360 Align(UCallee->getParent()->getDataLayout().getTypeAllocSize(RetType))); 1361 return Alloc; 1362 } 1363 1364 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, 1365 double& Res0, double& Res1, 1366 Constant *copr0, Constant *copr1, 1367 Constant *copr2) { 1368 // By default, opr0/opr1/opr3 holds values of float/double type. 1369 // If they are not float/double, each function has to its 1370 // operand separately. 1371 double opr0=0.0, opr1=0.0, opr2=0.0; 1372 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0); 1373 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1); 1374 ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2); 1375 if (fpopr0) { 1376 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1377 ? fpopr0->getValueAPF().convertToDouble() 1378 : (double)fpopr0->getValueAPF().convertToFloat(); 1379 } 1380 1381 if (fpopr1) { 1382 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1383 ? fpopr1->getValueAPF().convertToDouble() 1384 : (double)fpopr1->getValueAPF().convertToFloat(); 1385 } 1386 1387 if (fpopr2) { 1388 opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1389 ? fpopr2->getValueAPF().convertToDouble() 1390 : (double)fpopr2->getValueAPF().convertToFloat(); 1391 } 1392 1393 switch (FInfo.getId()) { 1394 default : return false; 1395 1396 case AMDGPULibFunc::EI_ACOS: 1397 Res0 = acos(opr0); 1398 return true; 1399 1400 case AMDGPULibFunc::EI_ACOSH: 1401 // acosh(x) == log(x + sqrt(x*x - 1)) 1402 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0)); 1403 return true; 1404 1405 case AMDGPULibFunc::EI_ACOSPI: 1406 Res0 = acos(opr0) / MATH_PI; 1407 return true; 1408 1409 case AMDGPULibFunc::EI_ASIN: 1410 Res0 = asin(opr0); 1411 return true; 1412 1413 case AMDGPULibFunc::EI_ASINH: 1414 // asinh(x) == log(x + sqrt(x*x + 1)) 1415 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0)); 1416 return true; 1417 1418 case AMDGPULibFunc::EI_ASINPI: 1419 Res0 = asin(opr0) / MATH_PI; 1420 return true; 1421 1422 case AMDGPULibFunc::EI_ATAN: 1423 Res0 = atan(opr0); 1424 return true; 1425 1426 case AMDGPULibFunc::EI_ATANH: 1427 // atanh(x) == (log(x+1) - log(x-1))/2; 1428 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0; 1429 return true; 1430 1431 case AMDGPULibFunc::EI_ATANPI: 1432 Res0 = atan(opr0) / MATH_PI; 1433 return true; 1434 1435 case AMDGPULibFunc::EI_CBRT: 1436 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0); 1437 return true; 1438 1439 case AMDGPULibFunc::EI_COS: 1440 Res0 = cos(opr0); 1441 return true; 1442 1443 case AMDGPULibFunc::EI_COSH: 1444 Res0 = cosh(opr0); 1445 return true; 1446 1447 case AMDGPULibFunc::EI_COSPI: 1448 Res0 = cos(MATH_PI * opr0); 1449 return true; 1450 1451 case AMDGPULibFunc::EI_EXP: 1452 Res0 = exp(opr0); 1453 return true; 1454 1455 case AMDGPULibFunc::EI_EXP2: 1456 Res0 = pow(2.0, opr0); 1457 return true; 1458 1459 case AMDGPULibFunc::EI_EXP10: 1460 Res0 = pow(10.0, opr0); 1461 return true; 1462 1463 case AMDGPULibFunc::EI_EXPM1: 1464 Res0 = exp(opr0) - 1.0; 1465 return true; 1466 1467 case AMDGPULibFunc::EI_LOG: 1468 Res0 = log(opr0); 1469 return true; 1470 1471 case AMDGPULibFunc::EI_LOG2: 1472 Res0 = log(opr0) / log(2.0); 1473 return true; 1474 1475 case AMDGPULibFunc::EI_LOG10: 1476 Res0 = log(opr0) / log(10.0); 1477 return true; 1478 1479 case AMDGPULibFunc::EI_RSQRT: 1480 Res0 = 1.0 / sqrt(opr0); 1481 return true; 1482 1483 case AMDGPULibFunc::EI_SIN: 1484 Res0 = sin(opr0); 1485 return true; 1486 1487 case AMDGPULibFunc::EI_SINH: 1488 Res0 = sinh(opr0); 1489 return true; 1490 1491 case AMDGPULibFunc::EI_SINPI: 1492 Res0 = sin(MATH_PI * opr0); 1493 return true; 1494 1495 case AMDGPULibFunc::EI_SQRT: 1496 Res0 = sqrt(opr0); 1497 return true; 1498 1499 case AMDGPULibFunc::EI_TAN: 1500 Res0 = tan(opr0); 1501 return true; 1502 1503 case AMDGPULibFunc::EI_TANH: 1504 Res0 = tanh(opr0); 1505 return true; 1506 1507 case AMDGPULibFunc::EI_TANPI: 1508 Res0 = tan(MATH_PI * opr0); 1509 return true; 1510 1511 case AMDGPULibFunc::EI_RECIP: 1512 Res0 = 1.0 / opr0; 1513 return true; 1514 1515 // two-arg functions 1516 case AMDGPULibFunc::EI_DIVIDE: 1517 Res0 = opr0 / opr1; 1518 return true; 1519 1520 case AMDGPULibFunc::EI_POW: 1521 case AMDGPULibFunc::EI_POWR: 1522 Res0 = pow(opr0, opr1); 1523 return true; 1524 1525 case AMDGPULibFunc::EI_POWN: { 1526 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1527 double val = (double)iopr1->getSExtValue(); 1528 Res0 = pow(opr0, val); 1529 return true; 1530 } 1531 return false; 1532 } 1533 1534 case AMDGPULibFunc::EI_ROOTN: { 1535 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1536 double val = (double)iopr1->getSExtValue(); 1537 Res0 = pow(opr0, 1.0 / val); 1538 return true; 1539 } 1540 return false; 1541 } 1542 1543 // with ptr arg 1544 case AMDGPULibFunc::EI_SINCOS: 1545 Res0 = sin(opr0); 1546 Res1 = cos(opr0); 1547 return true; 1548 1549 // three-arg functions 1550 case AMDGPULibFunc::EI_FMA: 1551 case AMDGPULibFunc::EI_MAD: 1552 Res0 = opr0 * opr1 + opr2; 1553 return true; 1554 } 1555 1556 return false; 1557 } 1558 1559 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) { 1560 int numArgs = (int)aCI->arg_size(); 1561 if (numArgs > 3) 1562 return false; 1563 1564 Constant *copr0 = nullptr; 1565 Constant *copr1 = nullptr; 1566 Constant *copr2 = nullptr; 1567 if (numArgs > 0) { 1568 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr) 1569 return false; 1570 } 1571 1572 if (numArgs > 1) { 1573 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) { 1574 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS) 1575 return false; 1576 } 1577 } 1578 1579 if (numArgs > 2) { 1580 if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr) 1581 return false; 1582 } 1583 1584 // At this point, all arguments to aCI are constants. 1585 1586 // max vector size is 16, and sincos will generate two results. 1587 double DVal0[16], DVal1[16]; 1588 int FuncVecSize = getVecSize(FInfo); 1589 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS); 1590 if (FuncVecSize == 1) { 1591 if (!evaluateScalarMathFunc(FInfo, DVal0[0], 1592 DVal1[0], copr0, copr1, copr2)) { 1593 return false; 1594 } 1595 } else { 1596 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0); 1597 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1); 1598 ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2); 1599 for (int i = 0; i < FuncVecSize; ++i) { 1600 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr; 1601 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr; 1602 Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr; 1603 if (!evaluateScalarMathFunc(FInfo, DVal0[i], 1604 DVal1[i], celt0, celt1, celt2)) { 1605 return false; 1606 } 1607 } 1608 } 1609 1610 LLVMContext &context = CI->getParent()->getParent()->getContext(); 1611 Constant *nval0, *nval1; 1612 if (FuncVecSize == 1) { 1613 nval0 = ConstantFP::get(CI->getType(), DVal0[0]); 1614 if (hasTwoResults) 1615 nval1 = ConstantFP::get(CI->getType(), DVal1[0]); 1616 } else { 1617 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1618 SmallVector <float, 0> FVal0, FVal1; 1619 for (int i = 0; i < FuncVecSize; ++i) 1620 FVal0.push_back((float)DVal0[i]); 1621 ArrayRef<float> tmp0(FVal0); 1622 nval0 = ConstantDataVector::get(context, tmp0); 1623 if (hasTwoResults) { 1624 for (int i = 0; i < FuncVecSize; ++i) 1625 FVal1.push_back((float)DVal1[i]); 1626 ArrayRef<float> tmp1(FVal1); 1627 nval1 = ConstantDataVector::get(context, tmp1); 1628 } 1629 } else { 1630 ArrayRef<double> tmp0(DVal0); 1631 nval0 = ConstantDataVector::get(context, tmp0); 1632 if (hasTwoResults) { 1633 ArrayRef<double> tmp1(DVal1); 1634 nval1 = ConstantDataVector::get(context, tmp1); 1635 } 1636 } 1637 } 1638 1639 if (hasTwoResults) { 1640 // sincos 1641 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS && 1642 "math function with ptr arg not supported yet"); 1643 new StoreInst(nval1, aCI->getArgOperand(1), aCI); 1644 } 1645 1646 replaceCall(nval0); 1647 return true; 1648 } 1649 1650 // Public interface to the Simplify LibCalls pass. 1651 FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetMachine *TM) { 1652 return new AMDGPUSimplifyLibCalls(TM); 1653 } 1654 1655 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() { 1656 return new AMDGPUUseNativeCalls(); 1657 } 1658 1659 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) { 1660 if (skipFunction(F)) 1661 return false; 1662 1663 bool Changed = false; 1664 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 1665 1666 LLVM_DEBUG(dbgs() << "AMDIC: process function "; 1667 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';); 1668 1669 for (auto &BB : F) { 1670 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) { 1671 // Ignore non-calls. 1672 CallInst *CI = dyn_cast<CallInst>(I); 1673 ++I; 1674 // Ignore intrinsics that do not become real instructions. 1675 if (!CI || isa<DbgInfoIntrinsic>(CI) || CI->isLifetimeStartOrEnd()) 1676 continue; 1677 1678 // Ignore indirect calls. 1679 Function *Callee = CI->getCalledFunction(); 1680 if (Callee == nullptr) 1681 continue; 1682 1683 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n"; 1684 dbgs().flush()); 1685 if(Simplifier.fold(CI, AA)) 1686 Changed = true; 1687 } 1688 } 1689 return Changed; 1690 } 1691 1692 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F, 1693 FunctionAnalysisManager &AM) { 1694 AMDGPULibCalls Simplifier(&TM); 1695 Simplifier.initNativeFuncs(); 1696 1697 bool Changed = false; 1698 auto AA = &AM.getResult<AAManager>(F); 1699 1700 LLVM_DEBUG(dbgs() << "AMDIC: process function "; 1701 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';); 1702 1703 for (auto &BB : F) { 1704 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) { 1705 // Ignore non-calls. 1706 CallInst *CI = dyn_cast<CallInst>(I); 1707 ++I; 1708 // Ignore intrinsics that do not become real instructions. 1709 if (!CI || isa<DbgInfoIntrinsic>(CI) || CI->isLifetimeStartOrEnd()) 1710 continue; 1711 1712 // Ignore indirect calls. 1713 Function *Callee = CI->getCalledFunction(); 1714 if (Callee == nullptr) 1715 continue; 1716 1717 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n"; 1718 dbgs().flush()); 1719 if (Simplifier.fold(CI, AA)) 1720 Changed = true; 1721 } 1722 } 1723 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 1724 } 1725 1726 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) { 1727 if (skipFunction(F) || UseNative.empty()) 1728 return false; 1729 1730 bool Changed = false; 1731 for (auto &BB : F) { 1732 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) { 1733 // Ignore non-calls. 1734 CallInst *CI = dyn_cast<CallInst>(I); 1735 ++I; 1736 if (!CI) continue; 1737 1738 // Ignore indirect calls. 1739 Function *Callee = CI->getCalledFunction(); 1740 if (Callee == nullptr) 1741 continue; 1742 1743 if (Simplifier.useNative(CI)) 1744 Changed = true; 1745 } 1746 } 1747 return Changed; 1748 } 1749 1750 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F, 1751 FunctionAnalysisManager &AM) { 1752 if (UseNative.empty()) 1753 return PreservedAnalyses::all(); 1754 1755 AMDGPULibCalls Simplifier; 1756 Simplifier.initNativeFuncs(); 1757 1758 bool Changed = false; 1759 for (auto &BB : F) { 1760 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) { 1761 // Ignore non-calls. 1762 CallInst *CI = dyn_cast<CallInst>(I); 1763 ++I; 1764 if (!CI) 1765 continue; 1766 1767 // Ignore indirect calls. 1768 Function *Callee = CI->getCalledFunction(); 1769 if (Callee == nullptr) 1770 continue; 1771 1772 if (Simplifier.useNative(CI)) 1773 Changed = true; 1774 } 1775 } 1776 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 1777 } 1778