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