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 #define DEBUG_TYPE "amdgpu-simplifylib" 15 16 #include "AMDGPU.h" 17 #include "AMDGPULibFunc.h" 18 #include "AMDGPUSubtarget.h" 19 #include "llvm/Analysis/AliasAnalysis.h" 20 #include "llvm/Analysis/Loads.h" 21 #include "llvm/ADT/StringSet.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/IR/Constants.h" 24 #include "llvm/IR/DerivedTypes.h" 25 #include "llvm/IR/Instructions.h" 26 #include "llvm/IR/Intrinsics.h" 27 #include "llvm/IR/IRBuilder.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/LLVMContext.h" 30 #include "llvm/IR/Module.h" 31 #include "llvm/IR/ValueSymbolTable.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/raw_ostream.h" 34 #include "llvm/Target/TargetMachine.h" 35 #include "llvm/Target/TargetOptions.h" 36 #include <vector> 37 #include <cmath> 38 39 using namespace llvm; 40 41 static cl::opt<bool> EnablePreLink("amdgpu-prelink", 42 cl::desc("Enable pre-link mode optimizations"), 43 cl::init(false), 44 cl::Hidden); 45 46 static cl::list<std::string> UseNative("amdgpu-use-native", 47 cl::desc("Comma separated list of functions to replace with native, or all"), 48 cl::CommaSeparated, cl::ValueOptional, 49 cl::Hidden); 50 51 #define MATH_PI 3.14159265358979323846264338327950288419716939937511 52 #define MATH_E 2.71828182845904523536028747135266249775724709369996 53 #define MATH_SQRT2 1.41421356237309504880168872420969807856967187537695 54 55 #define MATH_LOG2E 1.4426950408889634073599246810018921374266459541529859 56 #define MATH_LOG10E 0.4342944819032518276511289189166050822943970058036665 57 // Value of log2(10) 58 #define MATH_LOG2_10 3.3219280948873623478703194294893901758648313930245806 59 // Value of 1 / log2(10) 60 #define MATH_RLOG2_10 0.3010299956639811952137388947244930267681898814621085 61 // Value of 1 / M_LOG2E_F = 1 / log2(e) 62 #define MATH_RLOG2_E 0.6931471805599453094172321214581765680755001343602552 63 64 namespace llvm { 65 66 class AMDGPULibCalls { 67 private: 68 69 typedef llvm::AMDGPULibFunc FuncInfo; 70 71 const TargetMachine *TM; 72 73 // -fuse-native. 74 bool AllNative = false; 75 76 bool useNativeFunc(const StringRef F) const; 77 78 // Return a pointer (pointer expr) to the function if function defintion with 79 // "FuncName" exists. It may create a new function prototype in pre-link mode. 80 FunctionCallee getFunction(Module *M, const FuncInfo &fInfo); 81 82 // Replace a normal function with its native version. 83 bool replaceWithNative(CallInst *CI, const FuncInfo &FInfo); 84 85 bool parseFunctionName(const StringRef& FMangledName, 86 FuncInfo *FInfo=nullptr /*out*/); 87 88 bool TDOFold(CallInst *CI, const FuncInfo &FInfo); 89 90 /* Specialized optimizations */ 91 92 // recip (half or native) 93 bool fold_recip(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 94 95 // divide (half or native) 96 bool fold_divide(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 97 98 // pow/powr/pown 99 bool fold_pow(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 100 101 // rootn 102 bool fold_rootn(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 103 104 // fma/mad 105 bool fold_fma_mad(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 106 107 // -fuse-native for sincos 108 bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo); 109 110 // evaluate calls if calls' arguments are constants. 111 bool evaluateScalarMathFunc(FuncInfo &FInfo, double& Res0, 112 double& Res1, Constant *copr0, Constant *copr1, Constant *copr2); 113 bool evaluateCall(CallInst *aCI, FuncInfo &FInfo); 114 115 // exp 116 bool fold_exp(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 117 118 // exp2 119 bool fold_exp2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 120 121 // exp10 122 bool fold_exp10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 123 124 // log 125 bool fold_log(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 126 127 // log2 128 bool fold_log2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 129 130 // log10 131 bool fold_log10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 132 133 // sqrt 134 bool fold_sqrt(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo); 135 136 // sin/cos 137 bool fold_sincos(CallInst * CI, IRBuilder<> &B, AliasAnalysis * AA); 138 139 // __read_pipe/__write_pipe 140 bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, FuncInfo &FInfo); 141 142 // llvm.amdgcn.wavefrontsize 143 bool fold_wavefrontsize(CallInst *CI, IRBuilder<> &B); 144 145 // Get insertion point at entry. 146 BasicBlock::iterator getEntryIns(CallInst * UI); 147 // Insert an Alloc instruction. 148 AllocaInst* insertAlloca(CallInst * UI, IRBuilder<> &B, const char *prefix); 149 // Get a scalar native builtin signle argument FP function 150 FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo); 151 152 protected: 153 CallInst *CI; 154 155 bool isUnsafeMath(const CallInst *CI) const; 156 157 void replaceCall(Value *With) { 158 CI->replaceAllUsesWith(With); 159 CI->eraseFromParent(); 160 } 161 162 public: 163 AMDGPULibCalls(const TargetMachine *TM_ = nullptr) : TM(TM_) {} 164 165 bool fold(CallInst *CI, AliasAnalysis *AA = nullptr); 166 167 void initNativeFuncs(); 168 169 // Replace a normal math function call with that native version 170 bool useNative(CallInst *CI); 171 }; 172 173 } // end llvm namespace 174 175 namespace { 176 177 class AMDGPUSimplifyLibCalls : public FunctionPass { 178 179 const TargetOptions Options; 180 181 AMDGPULibCalls Simplifier; 182 183 public: 184 static char ID; // Pass identification 185 186 AMDGPUSimplifyLibCalls(const TargetOptions &Opt = TargetOptions(), 187 const TargetMachine *TM = nullptr) 188 : FunctionPass(ID), Options(Opt), Simplifier(TM) { 189 initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry()); 190 } 191 192 void getAnalysisUsage(AnalysisUsage &AU) const override { 193 AU.addRequired<AAResultsWrapperPass>(); 194 } 195 196 bool runOnFunction(Function &M) override; 197 }; 198 199 class AMDGPUUseNativeCalls : public FunctionPass { 200 201 AMDGPULibCalls Simplifier; 202 203 public: 204 static char ID; // Pass identification 205 206 AMDGPUUseNativeCalls() : FunctionPass(ID) { 207 initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry()); 208 Simplifier.initNativeFuncs(); 209 } 210 211 bool runOnFunction(Function &F) override; 212 }; 213 214 } // end anonymous namespace. 215 216 char AMDGPUSimplifyLibCalls::ID = 0; 217 char AMDGPUUseNativeCalls::ID = 0; 218 219 INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib", 220 "Simplify well-known AMD library calls", false, false) 221 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 222 INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib", 223 "Simplify well-known AMD library calls", false, false) 224 225 INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative", 226 "Replace builtin math calls with that native versions.", 227 false, false) 228 229 template <typename IRB> 230 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg, 231 const Twine &Name = "") { 232 CallInst *R = B.CreateCall(Callee, Arg, Name); 233 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 234 R->setCallingConv(F->getCallingConv()); 235 return R; 236 } 237 238 template <typename IRB> 239 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1, 240 Value *Arg2, const Twine &Name = "") { 241 CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name); 242 if (Function *F = dyn_cast<Function>(Callee.getCallee())) 243 R->setCallingConv(F->getCallingConv()); 244 return R; 245 } 246 247 // Data structures for table-driven optimizations. 248 // FuncTbl works for both f32 and f64 functions with 1 input argument 249 250 struct TableEntry { 251 double result; 252 double input; 253 }; 254 255 /* a list of {result, input} */ 256 static const TableEntry tbl_acos[] = { 257 {MATH_PI/2.0, 0.0}, 258 {MATH_PI/2.0, -0.0}, 259 {0.0, 1.0}, 260 {MATH_PI, -1.0} 261 }; 262 static const TableEntry tbl_acosh[] = { 263 {0.0, 1.0} 264 }; 265 static const TableEntry tbl_acospi[] = { 266 {0.5, 0.0}, 267 {0.5, -0.0}, 268 {0.0, 1.0}, 269 {1.0, -1.0} 270 }; 271 static const TableEntry tbl_asin[] = { 272 {0.0, 0.0}, 273 {-0.0, -0.0}, 274 {MATH_PI/2.0, 1.0}, 275 {-MATH_PI/2.0, -1.0} 276 }; 277 static const TableEntry tbl_asinh[] = { 278 {0.0, 0.0}, 279 {-0.0, -0.0} 280 }; 281 static const TableEntry tbl_asinpi[] = { 282 {0.0, 0.0}, 283 {-0.0, -0.0}, 284 {0.5, 1.0}, 285 {-0.5, -1.0} 286 }; 287 static const TableEntry tbl_atan[] = { 288 {0.0, 0.0}, 289 {-0.0, -0.0}, 290 {MATH_PI/4.0, 1.0}, 291 {-MATH_PI/4.0, -1.0} 292 }; 293 static const TableEntry tbl_atanh[] = { 294 {0.0, 0.0}, 295 {-0.0, -0.0} 296 }; 297 static const TableEntry tbl_atanpi[] = { 298 {0.0, 0.0}, 299 {-0.0, -0.0}, 300 {0.25, 1.0}, 301 {-0.25, -1.0} 302 }; 303 static const TableEntry tbl_cbrt[] = { 304 {0.0, 0.0}, 305 {-0.0, -0.0}, 306 {1.0, 1.0}, 307 {-1.0, -1.0}, 308 }; 309 static const TableEntry tbl_cos[] = { 310 {1.0, 0.0}, 311 {1.0, -0.0} 312 }; 313 static const TableEntry tbl_cosh[] = { 314 {1.0, 0.0}, 315 {1.0, -0.0} 316 }; 317 static const TableEntry tbl_cospi[] = { 318 {1.0, 0.0}, 319 {1.0, -0.0} 320 }; 321 static const TableEntry tbl_erfc[] = { 322 {1.0, 0.0}, 323 {1.0, -0.0} 324 }; 325 static const TableEntry tbl_erf[] = { 326 {0.0, 0.0}, 327 {-0.0, -0.0} 328 }; 329 static const TableEntry tbl_exp[] = { 330 {1.0, 0.0}, 331 {1.0, -0.0}, 332 {MATH_E, 1.0} 333 }; 334 static const TableEntry tbl_exp2[] = { 335 {1.0, 0.0}, 336 {1.0, -0.0}, 337 {2.0, 1.0} 338 }; 339 static const TableEntry tbl_exp10[] = { 340 {1.0, 0.0}, 341 {1.0, -0.0}, 342 {10.0, 1.0} 343 }; 344 static const TableEntry tbl_expm1[] = { 345 {0.0, 0.0}, 346 {-0.0, -0.0} 347 }; 348 static const TableEntry tbl_log[] = { 349 {0.0, 1.0}, 350 {1.0, MATH_E} 351 }; 352 static const TableEntry tbl_log2[] = { 353 {0.0, 1.0}, 354 {1.0, 2.0} 355 }; 356 static const TableEntry tbl_log10[] = { 357 {0.0, 1.0}, 358 {1.0, 10.0} 359 }; 360 static const TableEntry tbl_rsqrt[] = { 361 {1.0, 1.0}, 362 {1.0/MATH_SQRT2, 2.0} 363 }; 364 static const TableEntry tbl_sin[] = { 365 {0.0, 0.0}, 366 {-0.0, -0.0} 367 }; 368 static const TableEntry tbl_sinh[] = { 369 {0.0, 0.0}, 370 {-0.0, -0.0} 371 }; 372 static const TableEntry tbl_sinpi[] = { 373 {0.0, 0.0}, 374 {-0.0, -0.0} 375 }; 376 static const TableEntry tbl_sqrt[] = { 377 {0.0, 0.0}, 378 {1.0, 1.0}, 379 {MATH_SQRT2, 2.0} 380 }; 381 static const TableEntry tbl_tan[] = { 382 {0.0, 0.0}, 383 {-0.0, -0.0} 384 }; 385 static const TableEntry tbl_tanh[] = { 386 {0.0, 0.0}, 387 {-0.0, -0.0} 388 }; 389 static const TableEntry tbl_tanpi[] = { 390 {0.0, 0.0}, 391 {-0.0, -0.0} 392 }; 393 static const TableEntry tbl_tgamma[] = { 394 {1.0, 1.0}, 395 {1.0, 2.0}, 396 {2.0, 3.0}, 397 {6.0, 4.0} 398 }; 399 400 static bool HasNative(AMDGPULibFunc::EFuncId id) { 401 switch(id) { 402 case AMDGPULibFunc::EI_DIVIDE: 403 case AMDGPULibFunc::EI_COS: 404 case AMDGPULibFunc::EI_EXP: 405 case AMDGPULibFunc::EI_EXP2: 406 case AMDGPULibFunc::EI_EXP10: 407 case AMDGPULibFunc::EI_LOG: 408 case AMDGPULibFunc::EI_LOG2: 409 case AMDGPULibFunc::EI_LOG10: 410 case AMDGPULibFunc::EI_POWR: 411 case AMDGPULibFunc::EI_RECIP: 412 case AMDGPULibFunc::EI_RSQRT: 413 case AMDGPULibFunc::EI_SIN: 414 case AMDGPULibFunc::EI_SINCOS: 415 case AMDGPULibFunc::EI_SQRT: 416 case AMDGPULibFunc::EI_TAN: 417 return true; 418 default:; 419 } 420 return false; 421 } 422 423 struct TableRef { 424 size_t size; 425 const TableEntry *table; // variable size: from 0 to (size - 1) 426 427 TableRef() : size(0), table(nullptr) {} 428 429 template <size_t N> 430 TableRef(const TableEntry (&tbl)[N]) : size(N), table(&tbl[0]) {} 431 }; 432 433 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) { 434 switch(id) { 435 case AMDGPULibFunc::EI_ACOS: return TableRef(tbl_acos); 436 case AMDGPULibFunc::EI_ACOSH: return TableRef(tbl_acosh); 437 case AMDGPULibFunc::EI_ACOSPI: return TableRef(tbl_acospi); 438 case AMDGPULibFunc::EI_ASIN: return TableRef(tbl_asin); 439 case AMDGPULibFunc::EI_ASINH: return TableRef(tbl_asinh); 440 case AMDGPULibFunc::EI_ASINPI: return TableRef(tbl_asinpi); 441 case AMDGPULibFunc::EI_ATAN: return TableRef(tbl_atan); 442 case AMDGPULibFunc::EI_ATANH: return TableRef(tbl_atanh); 443 case AMDGPULibFunc::EI_ATANPI: return TableRef(tbl_atanpi); 444 case AMDGPULibFunc::EI_CBRT: return TableRef(tbl_cbrt); 445 case AMDGPULibFunc::EI_NCOS: 446 case AMDGPULibFunc::EI_COS: return TableRef(tbl_cos); 447 case AMDGPULibFunc::EI_COSH: return TableRef(tbl_cosh); 448 case AMDGPULibFunc::EI_COSPI: return TableRef(tbl_cospi); 449 case AMDGPULibFunc::EI_ERFC: return TableRef(tbl_erfc); 450 case AMDGPULibFunc::EI_ERF: return TableRef(tbl_erf); 451 case AMDGPULibFunc::EI_EXP: return TableRef(tbl_exp); 452 case AMDGPULibFunc::EI_NEXP2: 453 case AMDGPULibFunc::EI_EXP2: return TableRef(tbl_exp2); 454 case AMDGPULibFunc::EI_EXP10: return TableRef(tbl_exp10); 455 case AMDGPULibFunc::EI_EXPM1: return TableRef(tbl_expm1); 456 case AMDGPULibFunc::EI_LOG: return TableRef(tbl_log); 457 case AMDGPULibFunc::EI_NLOG2: 458 case AMDGPULibFunc::EI_LOG2: return TableRef(tbl_log2); 459 case AMDGPULibFunc::EI_LOG10: return TableRef(tbl_log10); 460 case AMDGPULibFunc::EI_NRSQRT: 461 case AMDGPULibFunc::EI_RSQRT: return TableRef(tbl_rsqrt); 462 case AMDGPULibFunc::EI_NSIN: 463 case AMDGPULibFunc::EI_SIN: return TableRef(tbl_sin); 464 case AMDGPULibFunc::EI_SINH: return TableRef(tbl_sinh); 465 case AMDGPULibFunc::EI_SINPI: return TableRef(tbl_sinpi); 466 case AMDGPULibFunc::EI_NSQRT: 467 case AMDGPULibFunc::EI_SQRT: return TableRef(tbl_sqrt); 468 case AMDGPULibFunc::EI_TAN: return TableRef(tbl_tan); 469 case AMDGPULibFunc::EI_TANH: return TableRef(tbl_tanh); 470 case AMDGPULibFunc::EI_TANPI: return TableRef(tbl_tanpi); 471 case AMDGPULibFunc::EI_TGAMMA: return TableRef(tbl_tgamma); 472 default:; 473 } 474 return TableRef(); 475 } 476 477 static inline int getVecSize(const AMDGPULibFunc& FInfo) { 478 return FInfo.getLeads()[0].VectorSize; 479 } 480 481 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) { 482 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType; 483 } 484 485 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) { 486 // If we are doing PreLinkOpt, the function is external. So it is safe to 487 // use getOrInsertFunction() at this stage. 488 489 return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo) 490 : AMDGPULibFunc::getFunction(M, fInfo); 491 } 492 493 bool AMDGPULibCalls::parseFunctionName(const StringRef& FMangledName, 494 FuncInfo *FInfo) { 495 return AMDGPULibFunc::parse(FMangledName, *FInfo); 496 } 497 498 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const { 499 if (auto Op = dyn_cast<FPMathOperator>(CI)) 500 if (Op->isFast()) 501 return true; 502 const Function *F = CI->getParent()->getParent(); 503 Attribute Attr = F->getFnAttribute("unsafe-fp-math"); 504 return Attr.getValueAsString() == "true"; 505 } 506 507 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const { 508 return AllNative || 509 std::find(UseNative.begin(), UseNative.end(), F) != UseNative.end(); 510 } 511 512 void AMDGPULibCalls::initNativeFuncs() { 513 AllNative = useNativeFunc("all") || 514 (UseNative.getNumOccurrences() && UseNative.size() == 1 && 515 UseNative.begin()->empty()); 516 } 517 518 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) { 519 bool native_sin = useNativeFunc("sin"); 520 bool native_cos = useNativeFunc("cos"); 521 522 if (native_sin && native_cos) { 523 Module *M = aCI->getModule(); 524 Value *opr0 = aCI->getArgOperand(0); 525 526 AMDGPULibFunc nf; 527 nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType; 528 nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize; 529 530 nf.setPrefix(AMDGPULibFunc::NATIVE); 531 nf.setId(AMDGPULibFunc::EI_SIN); 532 FunctionCallee sinExpr = getFunction(M, nf); 533 534 nf.setPrefix(AMDGPULibFunc::NATIVE); 535 nf.setId(AMDGPULibFunc::EI_COS); 536 FunctionCallee cosExpr = getFunction(M, nf); 537 if (sinExpr && cosExpr) { 538 Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI); 539 Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI); 540 new StoreInst(cosval, aCI->getArgOperand(1), aCI); 541 542 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 543 << " with native version of sin/cos"); 544 545 replaceCall(sinval); 546 return true; 547 } 548 } 549 return false; 550 } 551 552 bool AMDGPULibCalls::useNative(CallInst *aCI) { 553 CI = aCI; 554 Function *Callee = aCI->getCalledFunction(); 555 556 FuncInfo FInfo; 557 if (!parseFunctionName(Callee->getName(), &FInfo) || !FInfo.isMangled() || 558 FInfo.getPrefix() != AMDGPULibFunc::NOPFX || 559 getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) || 560 !(AllNative || useNativeFunc(FInfo.getName()))) { 561 return false; 562 } 563 564 if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS) 565 return sincosUseNative(aCI, FInfo); 566 567 FInfo.setPrefix(AMDGPULibFunc::NATIVE); 568 FunctionCallee F = getFunction(aCI->getModule(), FInfo); 569 if (!F) 570 return false; 571 572 aCI->setCalledFunction(F); 573 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI 574 << " with native version"); 575 return true; 576 } 577 578 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe 579 // builtin, with appended type size and alignment arguments, where 2 or 4 580 // indicates the original number of arguments. The library has optimized version 581 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same 582 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N 583 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ..., 584 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4. 585 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, 586 FuncInfo &FInfo) { 587 auto *Callee = CI->getCalledFunction(); 588 if (!Callee->isDeclaration()) 589 return false; 590 591 assert(Callee->hasName() && "Invalid read_pipe/write_pipe function"); 592 auto *M = Callee->getParent(); 593 auto &Ctx = M->getContext(); 594 std::string Name = Callee->getName(); 595 auto NumArg = CI->getNumArgOperands(); 596 if (NumArg != 4 && NumArg != 6) 597 return false; 598 auto *PacketSize = CI->getArgOperand(NumArg - 2); 599 auto *PacketAlign = CI->getArgOperand(NumArg - 1); 600 if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign)) 601 return false; 602 unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue(); 603 unsigned Align = cast<ConstantInt>(PacketAlign)->getZExtValue(); 604 if (Size != Align || !isPowerOf2_32(Size)) 605 return false; 606 607 Type *PtrElemTy; 608 if (Size <= 8) 609 PtrElemTy = Type::getIntNTy(Ctx, Size * 8); 610 else 611 PtrElemTy = VectorType::get(Type::getInt64Ty(Ctx), Size / 8); 612 unsigned PtrArgLoc = CI->getNumArgOperands() - 3; 613 auto PtrArg = CI->getArgOperand(PtrArgLoc); 614 unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace(); 615 auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS); 616 617 SmallVector<llvm::Type *, 6> ArgTys; 618 for (unsigned I = 0; I != PtrArgLoc; ++I) 619 ArgTys.push_back(CI->getArgOperand(I)->getType()); 620 ArgTys.push_back(PtrTy); 621 622 Name = Name + "_" + std::to_string(Size); 623 auto *FTy = FunctionType::get(Callee->getReturnType(), 624 ArrayRef<Type *>(ArgTys), false); 625 AMDGPULibFunc NewLibFunc(Name, FTy); 626 FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc); 627 if (!F) 628 return false; 629 630 auto *BCast = B.CreatePointerCast(PtrArg, PtrTy); 631 SmallVector<Value *, 6> Args; 632 for (unsigned I = 0; I != PtrArgLoc; ++I) 633 Args.push_back(CI->getArgOperand(I)); 634 Args.push_back(BCast); 635 636 auto *NCI = B.CreateCall(F, Args); 637 NCI->setAttributes(CI->getAttributes()); 638 CI->replaceAllUsesWith(NCI); 639 CI->dropAllReferences(); 640 CI->eraseFromParent(); 641 642 return true; 643 } 644 645 // This function returns false if no change; return true otherwise. 646 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) { 647 this->CI = CI; 648 Function *Callee = CI->getCalledFunction(); 649 650 // Ignore indirect calls. 651 if (Callee == 0) return false; 652 653 BasicBlock *BB = CI->getParent(); 654 LLVMContext &Context = CI->getParent()->getContext(); 655 IRBuilder<> B(Context); 656 657 // Set the builder to the instruction after the call. 658 B.SetInsertPoint(BB, CI->getIterator()); 659 660 // Copy fast flags from the original call. 661 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI)) 662 B.setFastMathFlags(FPOp->getFastMathFlags()); 663 664 switch (Callee->getIntrinsicID()) { 665 default: 666 break; 667 case Intrinsic::amdgcn_wavefrontsize: 668 return !EnablePreLink && fold_wavefrontsize(CI, B); 669 } 670 671 FuncInfo FInfo; 672 if (!parseFunctionName(Callee->getName(), &FInfo)) 673 return false; 674 675 // Further check the number of arguments to see if they match. 676 if (CI->getNumArgOperands() != FInfo.getNumArgs()) 677 return false; 678 679 if (TDOFold(CI, FInfo)) 680 return true; 681 682 // Under unsafe-math, evaluate calls if possible. 683 // According to Brian Sumner, we can do this for all f32 function calls 684 // using host's double function calls. 685 if (isUnsafeMath(CI) && evaluateCall(CI, FInfo)) 686 return true; 687 688 // Specilized optimizations for each function call 689 switch (FInfo.getId()) { 690 case AMDGPULibFunc::EI_RECIP: 691 // skip vector function 692 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || 693 FInfo.getPrefix() == AMDGPULibFunc::HALF) && 694 "recip must be an either native or half function"); 695 return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo); 696 697 case AMDGPULibFunc::EI_DIVIDE: 698 // skip vector function 699 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || 700 FInfo.getPrefix() == AMDGPULibFunc::HALF) && 701 "divide must be an either native or half function"); 702 return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo); 703 704 case AMDGPULibFunc::EI_POW: 705 case AMDGPULibFunc::EI_POWR: 706 case AMDGPULibFunc::EI_POWN: 707 return fold_pow(CI, B, FInfo); 708 709 case AMDGPULibFunc::EI_ROOTN: 710 // skip vector function 711 return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo); 712 713 case AMDGPULibFunc::EI_FMA: 714 case AMDGPULibFunc::EI_MAD: 715 case AMDGPULibFunc::EI_NFMA: 716 // skip vector function 717 return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo); 718 719 case AMDGPULibFunc::EI_SQRT: 720 return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo); 721 case AMDGPULibFunc::EI_COS: 722 case AMDGPULibFunc::EI_SIN: 723 if ((getArgType(FInfo) == AMDGPULibFunc::F32 || 724 getArgType(FInfo) == AMDGPULibFunc::F64) 725 && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX)) 726 return fold_sincos(CI, B, AA); 727 728 break; 729 case AMDGPULibFunc::EI_READ_PIPE_2: 730 case AMDGPULibFunc::EI_READ_PIPE_4: 731 case AMDGPULibFunc::EI_WRITE_PIPE_2: 732 case AMDGPULibFunc::EI_WRITE_PIPE_4: 733 return fold_read_write_pipe(CI, B, FInfo); 734 735 default: 736 break; 737 } 738 739 return false; 740 } 741 742 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) { 743 // Table-Driven optimization 744 const TableRef tr = getOptTable(FInfo.getId()); 745 if (tr.size==0) 746 return false; 747 748 int const sz = (int)tr.size; 749 const TableEntry * const ftbl = tr.table; 750 Value *opr0 = CI->getArgOperand(0); 751 752 if (getVecSize(FInfo) > 1) { 753 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) { 754 SmallVector<double, 0> DVal; 755 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) { 756 ConstantFP *eltval = dyn_cast<ConstantFP>( 757 CV->getElementAsConstant((unsigned)eltNo)); 758 assert(eltval && "Non-FP arguments in math function!"); 759 bool found = false; 760 for (int i=0; i < sz; ++i) { 761 if (eltval->isExactlyValue(ftbl[i].input)) { 762 DVal.push_back(ftbl[i].result); 763 found = true; 764 break; 765 } 766 } 767 if (!found) { 768 // This vector constants not handled yet. 769 return false; 770 } 771 } 772 LLVMContext &context = CI->getParent()->getParent()->getContext(); 773 Constant *nval; 774 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 775 SmallVector<float, 0> FVal; 776 for (unsigned i = 0; i < DVal.size(); ++i) { 777 FVal.push_back((float)DVal[i]); 778 } 779 ArrayRef<float> tmp(FVal); 780 nval = ConstantDataVector::get(context, tmp); 781 } else { // F64 782 ArrayRef<double> tmp(DVal); 783 nval = ConstantDataVector::get(context, tmp); 784 } 785 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 786 replaceCall(nval); 787 return true; 788 } 789 } else { 790 // Scalar version 791 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) { 792 for (int i = 0; i < sz; ++i) { 793 if (CF->isExactlyValue(ftbl[i].input)) { 794 Value *nval = ConstantFP::get(CF->getType(), ftbl[i].result); 795 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 796 replaceCall(nval); 797 return true; 798 } 799 } 800 } 801 } 802 803 return false; 804 } 805 806 bool AMDGPULibCalls::replaceWithNative(CallInst *CI, const FuncInfo &FInfo) { 807 Module *M = CI->getModule(); 808 if (getArgType(FInfo) != AMDGPULibFunc::F32 || 809 FInfo.getPrefix() != AMDGPULibFunc::NOPFX || 810 !HasNative(FInfo.getId())) 811 return false; 812 813 AMDGPULibFunc nf = FInfo; 814 nf.setPrefix(AMDGPULibFunc::NATIVE); 815 if (FunctionCallee FPExpr = getFunction(M, nf)) { 816 LLVM_DEBUG(dbgs() << "AMDIC: " << *CI << " ---> "); 817 818 CI->setCalledFunction(FPExpr); 819 820 LLVM_DEBUG(dbgs() << *CI << '\n'); 821 822 return true; 823 } 824 return false; 825 } 826 827 // [native_]half_recip(c) ==> 1.0/c 828 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B, 829 const FuncInfo &FInfo) { 830 Value *opr0 = CI->getArgOperand(0); 831 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) { 832 // Just create a normal div. Later, InstCombine will be able 833 // to compute the divide into a constant (avoid check float infinity 834 // or subnormal at this point). 835 Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0), 836 opr0, 837 "recip2div"); 838 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n"); 839 replaceCall(nval); 840 return true; 841 } 842 return false; 843 } 844 845 // [native_]half_divide(x, c) ==> x/c 846 bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B, 847 const FuncInfo &FInfo) { 848 Value *opr0 = CI->getArgOperand(0); 849 Value *opr1 = CI->getArgOperand(1); 850 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0); 851 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1); 852 853 if ((CF0 && CF1) || // both are constants 854 (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32))) 855 // CF1 is constant && f32 divide 856 { 857 Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0), 858 opr1, "__div2recip"); 859 Value *nval = B.CreateFMul(opr0, nval1, "__div2mul"); 860 replaceCall(nval); 861 return true; 862 } 863 return false; 864 } 865 866 namespace llvm { 867 static double log2(double V) { 868 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L 869 return ::log2(V); 870 #else 871 return log(V) / 0.693147180559945309417; 872 #endif 873 } 874 } 875 876 bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B, 877 const FuncInfo &FInfo) { 878 assert((FInfo.getId() == AMDGPULibFunc::EI_POW || 879 FInfo.getId() == AMDGPULibFunc::EI_POWR || 880 FInfo.getId() == AMDGPULibFunc::EI_POWN) && 881 "fold_pow: encounter a wrong function call"); 882 883 Value *opr0, *opr1; 884 ConstantFP *CF; 885 ConstantInt *CINT; 886 ConstantAggregateZero *CZero; 887 Type *eltType; 888 889 opr0 = CI->getArgOperand(0); 890 opr1 = CI->getArgOperand(1); 891 CZero = dyn_cast<ConstantAggregateZero>(opr1); 892 if (getVecSize(FInfo) == 1) { 893 eltType = opr0->getType(); 894 CF = dyn_cast<ConstantFP>(opr1); 895 CINT = dyn_cast<ConstantInt>(opr1); 896 } else { 897 VectorType *VTy = dyn_cast<VectorType>(opr0->getType()); 898 assert(VTy && "Oprand of vector function should be of vectortype"); 899 eltType = VTy->getElementType(); 900 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1); 901 902 // Now, only Handle vector const whose elements have the same value. 903 CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr; 904 CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr; 905 } 906 907 // No unsafe math , no constant argument, do nothing 908 if (!isUnsafeMath(CI) && !CF && !CINT && !CZero) 909 return false; 910 911 // 0x1111111 means that we don't do anything for this call. 912 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111); 913 914 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) { 915 // pow/powr/pown(x, 0) == 1 916 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n"); 917 Constant *cnval = ConstantFP::get(eltType, 1.0); 918 if (getVecSize(FInfo) > 1) { 919 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 920 } 921 replaceCall(cnval); 922 return true; 923 } 924 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) { 925 // pow/powr/pown(x, 1.0) = x 926 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n"); 927 replaceCall(opr0); 928 return true; 929 } 930 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) { 931 // pow/powr/pown(x, 2.0) = x*x 932 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0 933 << "\n"); 934 Value *nval = B.CreateFMul(opr0, opr0, "__pow2"); 935 replaceCall(nval); 936 return true; 937 } 938 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) { 939 // pow/powr/pown(x, -1.0) = 1.0/x 940 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n"); 941 Constant *cnval = ConstantFP::get(eltType, 1.0); 942 if (getVecSize(FInfo) > 1) { 943 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 944 } 945 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip"); 946 replaceCall(nval); 947 return true; 948 } 949 950 Module *M = CI->getModule(); 951 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) { 952 // pow[r](x, [-]0.5) = sqrt(x) 953 bool issqrt = CF->isExactlyValue(0.5); 954 if (FunctionCallee FPExpr = 955 getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT 956 : AMDGPULibFunc::EI_RSQRT, 957 FInfo))) { 958 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 959 << FInfo.getName().c_str() << "(" << *opr0 << ")\n"); 960 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt" 961 : "__pow2rsqrt"); 962 replaceCall(nval); 963 return true; 964 } 965 } 966 967 if (!isUnsafeMath(CI)) 968 return false; 969 970 // Unsafe Math optimization 971 972 // Remember that ci_opr1 is set if opr1 is integral 973 if (CF) { 974 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32) 975 ? (double)CF->getValueAPF().convertToFloat() 976 : CF->getValueAPF().convertToDouble(); 977 int ival = (int)dval; 978 if ((double)ival == dval) { 979 ci_opr1 = ival; 980 } else 981 ci_opr1 = 0x11111111; 982 } 983 984 // pow/powr/pown(x, c) = [1/](x*x*..x); where 985 // trunc(c) == c && the number of x == c && |c| <= 12 986 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1; 987 if (abs_opr1 <= 12) { 988 Constant *cnval; 989 Value *nval; 990 if (abs_opr1 == 0) { 991 cnval = ConstantFP::get(eltType, 1.0); 992 if (getVecSize(FInfo) > 1) { 993 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 994 } 995 nval = cnval; 996 } else { 997 Value *valx2 = nullptr; 998 nval = nullptr; 999 while (abs_opr1 > 0) { 1000 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0; 1001 if (abs_opr1 & 1) { 1002 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2; 1003 } 1004 abs_opr1 >>= 1; 1005 } 1006 } 1007 1008 if (ci_opr1 < 0) { 1009 cnval = ConstantFP::get(eltType, 1.0); 1010 if (getVecSize(FInfo) > 1) { 1011 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval); 1012 } 1013 nval = B.CreateFDiv(cnval, nval, "__1powprod"); 1014 } 1015 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 1016 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0 1017 << ")\n"); 1018 replaceCall(nval); 1019 return true; 1020 } 1021 1022 // powr ---> exp2(y * log2(x)) 1023 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31)) 1024 FunctionCallee ExpExpr = 1025 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo)); 1026 if (!ExpExpr) 1027 return false; 1028 1029 bool needlog = false; 1030 bool needabs = false; 1031 bool needcopysign = false; 1032 Constant *cnval = nullptr; 1033 if (getVecSize(FInfo) == 1) { 1034 CF = dyn_cast<ConstantFP>(opr0); 1035 1036 if (CF) { 1037 double V = (getArgType(FInfo) == AMDGPULibFunc::F32) 1038 ? (double)CF->getValueAPF().convertToFloat() 1039 : CF->getValueAPF().convertToDouble(); 1040 1041 V = log2(std::abs(V)); 1042 cnval = ConstantFP::get(eltType, V); 1043 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) && 1044 CF->isNegative(); 1045 } else { 1046 needlog = true; 1047 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR && 1048 (!CF || CF->isNegative()); 1049 } 1050 } else { 1051 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0); 1052 1053 if (!CDV) { 1054 needlog = true; 1055 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR; 1056 } else { 1057 assert ((int)CDV->getNumElements() == getVecSize(FInfo) && 1058 "Wrong vector size detected"); 1059 1060 SmallVector<double, 0> DVal; 1061 for (int i=0; i < getVecSize(FInfo); ++i) { 1062 double V = (getArgType(FInfo) == AMDGPULibFunc::F32) 1063 ? (double)CDV->getElementAsFloat(i) 1064 : CDV->getElementAsDouble(i); 1065 if (V < 0.0) needcopysign = true; 1066 V = log2(std::abs(V)); 1067 DVal.push_back(V); 1068 } 1069 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1070 SmallVector<float, 0> FVal; 1071 for (unsigned i=0; i < DVal.size(); ++i) { 1072 FVal.push_back((float)DVal[i]); 1073 } 1074 ArrayRef<float> tmp(FVal); 1075 cnval = ConstantDataVector::get(M->getContext(), tmp); 1076 } else { 1077 ArrayRef<double> tmp(DVal); 1078 cnval = ConstantDataVector::get(M->getContext(), tmp); 1079 } 1080 } 1081 } 1082 1083 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) { 1084 // We cannot handle corner cases for a general pow() function, give up 1085 // unless y is a constant integral value. Then proceed as if it were pown. 1086 if (getVecSize(FInfo) == 1) { 1087 if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) { 1088 double y = (getArgType(FInfo) == AMDGPULibFunc::F32) 1089 ? (double)CF->getValueAPF().convertToFloat() 1090 : CF->getValueAPF().convertToDouble(); 1091 if (y != (double)(int64_t)y) 1092 return false; 1093 } else 1094 return false; 1095 } else { 1096 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) { 1097 for (int i=0; i < getVecSize(FInfo); ++i) { 1098 double y = (getArgType(FInfo) == AMDGPULibFunc::F32) 1099 ? (double)CDV->getElementAsFloat(i) 1100 : CDV->getElementAsDouble(i); 1101 if (y != (double)(int64_t)y) 1102 return false; 1103 } 1104 } else 1105 return false; 1106 } 1107 } 1108 1109 Value *nval; 1110 if (needabs) { 1111 FunctionCallee AbsExpr = 1112 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS, FInfo)); 1113 if (!AbsExpr) 1114 return false; 1115 nval = CreateCallEx(B, AbsExpr, opr0, "__fabs"); 1116 } else { 1117 nval = cnval ? cnval : opr0; 1118 } 1119 if (needlog) { 1120 FunctionCallee LogExpr = 1121 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo)); 1122 if (!LogExpr) 1123 return false; 1124 nval = CreateCallEx(B,LogExpr, nval, "__log2"); 1125 } 1126 1127 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) { 1128 // convert int(32) to fp(f32 or f64) 1129 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F"); 1130 } 1131 nval = B.CreateFMul(opr1, nval, "__ylogx"); 1132 nval = CreateCallEx(B,ExpExpr, nval, "__exp2"); 1133 1134 if (needcopysign) { 1135 Value *opr_n; 1136 Type* rTy = opr0->getType(); 1137 Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty(); 1138 Type *nTy = nTyS; 1139 if (const VectorType *vTy = dyn_cast<VectorType>(rTy)) 1140 nTy = VectorType::get(nTyS, vTy->getNumElements()); 1141 unsigned size = nTy->getScalarSizeInBits(); 1142 opr_n = CI->getArgOperand(1); 1143 if (opr_n->getType()->isIntegerTy()) 1144 opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou"); 1145 else 1146 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou"); 1147 1148 Value *sign = B.CreateShl(opr_n, size-1, "__yeven"); 1149 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign"); 1150 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign); 1151 nval = B.CreateBitCast(nval, opr0->getType()); 1152 } 1153 1154 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 1155 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n"); 1156 replaceCall(nval); 1157 1158 return true; 1159 } 1160 1161 bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B, 1162 const FuncInfo &FInfo) { 1163 Value *opr0 = CI->getArgOperand(0); 1164 Value *opr1 = CI->getArgOperand(1); 1165 1166 ConstantInt *CINT = dyn_cast<ConstantInt>(opr1); 1167 if (!CINT) { 1168 return false; 1169 } 1170 int ci_opr1 = (int)CINT->getSExtValue(); 1171 if (ci_opr1 == 1) { // rootn(x, 1) = x 1172 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n"); 1173 replaceCall(opr0); 1174 return true; 1175 } 1176 if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x) 1177 std::vector<const Type*> ParamsTys; 1178 ParamsTys.push_back(opr0->getType()); 1179 Module *M = CI->getModule(); 1180 if (FunctionCallee FPExpr = 1181 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) { 1182 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n"); 1183 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt"); 1184 replaceCall(nval); 1185 return true; 1186 } 1187 } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x) 1188 Module *M = CI->getModule(); 1189 if (FunctionCallee FPExpr = 1190 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) { 1191 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n"); 1192 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt"); 1193 replaceCall(nval); 1194 return true; 1195 } 1196 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x 1197 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n"); 1198 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0), 1199 opr0, 1200 "__rootn2div"); 1201 replaceCall(nval); 1202 return true; 1203 } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x) 1204 std::vector<const Type*> ParamsTys; 1205 ParamsTys.push_back(opr0->getType()); 1206 Module *M = CI->getModule(); 1207 if (FunctionCallee FPExpr = 1208 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) { 1209 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0 1210 << ")\n"); 1211 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt"); 1212 replaceCall(nval); 1213 return true; 1214 } 1215 } 1216 return false; 1217 } 1218 1219 bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B, 1220 const FuncInfo &FInfo) { 1221 Value *opr0 = CI->getArgOperand(0); 1222 Value *opr1 = CI->getArgOperand(1); 1223 Value *opr2 = CI->getArgOperand(2); 1224 1225 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0); 1226 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1); 1227 if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) { 1228 // fma/mad(a, b, c) = c if a=0 || b=0 1229 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n"); 1230 replaceCall(opr2); 1231 return true; 1232 } 1233 if (CF0 && CF0->isExactlyValue(1.0f)) { 1234 // fma/mad(a, b, c) = b+c if a=1 1235 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2 1236 << "\n"); 1237 Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd"); 1238 replaceCall(nval); 1239 return true; 1240 } 1241 if (CF1 && CF1->isExactlyValue(1.0f)) { 1242 // fma/mad(a, b, c) = a+c if b=1 1243 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2 1244 << "\n"); 1245 Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd"); 1246 replaceCall(nval); 1247 return true; 1248 } 1249 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) { 1250 if (CF->isZero()) { 1251 // fma/mad(a, b, c) = a*b if c=0 1252 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " 1253 << *opr1 << "\n"); 1254 Value *nval = B.CreateFMul(opr0, opr1, "fmamul"); 1255 replaceCall(nval); 1256 return true; 1257 } 1258 } 1259 1260 return false; 1261 } 1262 1263 // Get a scalar native builtin signle argument FP function 1264 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M, 1265 const FuncInfo &FInfo) { 1266 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId())) 1267 return nullptr; 1268 FuncInfo nf = FInfo; 1269 nf.setPrefix(AMDGPULibFunc::NATIVE); 1270 return getFunction(M, nf); 1271 } 1272 1273 // fold sqrt -> native_sqrt (x) 1274 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B, 1275 const FuncInfo &FInfo) { 1276 if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) && 1277 (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) { 1278 if (FunctionCallee FPExpr = getNativeFunction( 1279 CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) { 1280 Value *opr0 = CI->getArgOperand(0); 1281 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " 1282 << "sqrt(" << *opr0 << ")\n"); 1283 Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt"); 1284 replaceCall(nval); 1285 return true; 1286 } 1287 } 1288 return false; 1289 } 1290 1291 // fold sin, cos -> sincos. 1292 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B, 1293 AliasAnalysis *AA) { 1294 AMDGPULibFunc fInfo; 1295 if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo)) 1296 return false; 1297 1298 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN || 1299 fInfo.getId() == AMDGPULibFunc::EI_COS); 1300 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN; 1301 1302 Value *CArgVal = CI->getArgOperand(0); 1303 BasicBlock * const CBB = CI->getParent(); 1304 1305 int const MaxScan = 30; 1306 1307 { // fold in load value. 1308 LoadInst *LI = dyn_cast<LoadInst>(CArgVal); 1309 if (LI && LI->getParent() == CBB) { 1310 BasicBlock::iterator BBI = LI->getIterator(); 1311 Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA); 1312 if (AvailableVal) { 1313 CArgVal->replaceAllUsesWith(AvailableVal); 1314 if (CArgVal->getNumUses() == 0) 1315 LI->eraseFromParent(); 1316 CArgVal = CI->getArgOperand(0); 1317 } 1318 } 1319 } 1320 1321 Module *M = CI->getModule(); 1322 fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN); 1323 std::string const PairName = fInfo.mangle(); 1324 1325 CallInst *UI = nullptr; 1326 for (User* U : CArgVal->users()) { 1327 CallInst *XI = dyn_cast_or_null<CallInst>(U); 1328 if (!XI || XI == CI || XI->getParent() != CBB) 1329 continue; 1330 1331 Function *UCallee = XI->getCalledFunction(); 1332 if (!UCallee || !UCallee->getName().equals(PairName)) 1333 continue; 1334 1335 BasicBlock::iterator BBI = CI->getIterator(); 1336 if (BBI == CI->getParent()->begin()) 1337 break; 1338 --BBI; 1339 for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) { 1340 if (cast<Instruction>(BBI) == XI) { 1341 UI = XI; 1342 break; 1343 } 1344 } 1345 if (UI) break; 1346 } 1347 1348 if (!UI) return false; 1349 1350 // Merge the sin and cos. 1351 1352 // for OpenCL 2.0 we have only generic implementation of sincos 1353 // function. 1354 AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo); 1355 nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS); 1356 FunctionCallee Fsincos = getFunction(M, nf); 1357 if (!Fsincos) return false; 1358 1359 BasicBlock::iterator ItOld = B.GetInsertPoint(); 1360 AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_"); 1361 B.SetInsertPoint(UI); 1362 1363 Value *P = Alloc; 1364 Type *PTy = Fsincos.getFunctionType()->getParamType(1); 1365 // The allocaInst allocates the memory in private address space. This need 1366 // to be bitcasted to point to the address space of cos pointer type. 1367 // In OpenCL 2.0 this is generic, while in 1.2 that is private. 1368 if (PTy->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) 1369 P = B.CreateAddrSpaceCast(Alloc, PTy); 1370 CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P); 1371 1372 LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with " 1373 << *Call << "\n"); 1374 1375 if (!isSin) { // CI->cos, UI->sin 1376 B.SetInsertPoint(&*ItOld); 1377 UI->replaceAllUsesWith(&*Call); 1378 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc); 1379 CI->replaceAllUsesWith(Reload); 1380 UI->eraseFromParent(); 1381 CI->eraseFromParent(); 1382 } else { // CI->sin, UI->cos 1383 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc); 1384 UI->replaceAllUsesWith(Reload); 1385 CI->replaceAllUsesWith(Call); 1386 UI->eraseFromParent(); 1387 CI->eraseFromParent(); 1388 } 1389 return true; 1390 } 1391 1392 bool AMDGPULibCalls::fold_wavefrontsize(CallInst *CI, IRBuilder<> &B) { 1393 if (!TM) 1394 return false; 1395 1396 StringRef CPU = TM->getTargetCPU(); 1397 StringRef Features = TM->getTargetFeatureString(); 1398 if ((CPU.empty() || CPU.equals_lower("generic")) && 1399 (Features.empty() || 1400 Features.find_lower("wavefrontsize") == StringRef::npos)) 1401 return false; 1402 1403 Function *F = CI->getParent()->getParent(); 1404 const GCNSubtarget &ST = TM->getSubtarget<GCNSubtarget>(*F); 1405 unsigned N = ST.getWavefrontSize(); 1406 1407 LLVM_DEBUG(errs() << "AMDIC: fold_wavefrontsize (" << *CI << ") with " 1408 << N << "\n"); 1409 1410 CI->replaceAllUsesWith(ConstantInt::get(B.getInt32Ty(), N)); 1411 CI->eraseFromParent(); 1412 return true; 1413 } 1414 1415 // Get insertion point at entry. 1416 BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) { 1417 Function * Func = UI->getParent()->getParent(); 1418 BasicBlock * BB = &Func->getEntryBlock(); 1419 assert(BB && "Entry block not found!"); 1420 BasicBlock::iterator ItNew = BB->begin(); 1421 return ItNew; 1422 } 1423 1424 // Insert a AllocsInst at the beginning of function entry block. 1425 AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B, 1426 const char *prefix) { 1427 BasicBlock::iterator ItNew = getEntryIns(UI); 1428 Function *UCallee = UI->getCalledFunction(); 1429 Type *RetType = UCallee->getReturnType(); 1430 B.SetInsertPoint(&*ItNew); 1431 AllocaInst *Alloc = B.CreateAlloca(RetType, 0, 1432 std::string(prefix) + UI->getName()); 1433 Alloc->setAlignment(UCallee->getParent()->getDataLayout() 1434 .getTypeAllocSize(RetType)); 1435 return Alloc; 1436 } 1437 1438 bool AMDGPULibCalls::evaluateScalarMathFunc(FuncInfo &FInfo, 1439 double& Res0, double& Res1, 1440 Constant *copr0, Constant *copr1, 1441 Constant *copr2) { 1442 // By default, opr0/opr1/opr3 holds values of float/double type. 1443 // If they are not float/double, each function has to its 1444 // operand separately. 1445 double opr0=0.0, opr1=0.0, opr2=0.0; 1446 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0); 1447 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1); 1448 ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2); 1449 if (fpopr0) { 1450 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1451 ? fpopr0->getValueAPF().convertToDouble() 1452 : (double)fpopr0->getValueAPF().convertToFloat(); 1453 } 1454 1455 if (fpopr1) { 1456 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1457 ? fpopr1->getValueAPF().convertToDouble() 1458 : (double)fpopr1->getValueAPF().convertToFloat(); 1459 } 1460 1461 if (fpopr2) { 1462 opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64) 1463 ? fpopr2->getValueAPF().convertToDouble() 1464 : (double)fpopr2->getValueAPF().convertToFloat(); 1465 } 1466 1467 switch (FInfo.getId()) { 1468 default : return false; 1469 1470 case AMDGPULibFunc::EI_ACOS: 1471 Res0 = acos(opr0); 1472 return true; 1473 1474 case AMDGPULibFunc::EI_ACOSH: 1475 // acosh(x) == log(x + sqrt(x*x - 1)) 1476 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0)); 1477 return true; 1478 1479 case AMDGPULibFunc::EI_ACOSPI: 1480 Res0 = acos(opr0) / MATH_PI; 1481 return true; 1482 1483 case AMDGPULibFunc::EI_ASIN: 1484 Res0 = asin(opr0); 1485 return true; 1486 1487 case AMDGPULibFunc::EI_ASINH: 1488 // asinh(x) == log(x + sqrt(x*x + 1)) 1489 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0)); 1490 return true; 1491 1492 case AMDGPULibFunc::EI_ASINPI: 1493 Res0 = asin(opr0) / MATH_PI; 1494 return true; 1495 1496 case AMDGPULibFunc::EI_ATAN: 1497 Res0 = atan(opr0); 1498 return true; 1499 1500 case AMDGPULibFunc::EI_ATANH: 1501 // atanh(x) == (log(x+1) - log(x-1))/2; 1502 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0; 1503 return true; 1504 1505 case AMDGPULibFunc::EI_ATANPI: 1506 Res0 = atan(opr0) / MATH_PI; 1507 return true; 1508 1509 case AMDGPULibFunc::EI_CBRT: 1510 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0); 1511 return true; 1512 1513 case AMDGPULibFunc::EI_COS: 1514 Res0 = cos(opr0); 1515 return true; 1516 1517 case AMDGPULibFunc::EI_COSH: 1518 Res0 = cosh(opr0); 1519 return true; 1520 1521 case AMDGPULibFunc::EI_COSPI: 1522 Res0 = cos(MATH_PI * opr0); 1523 return true; 1524 1525 case AMDGPULibFunc::EI_EXP: 1526 Res0 = exp(opr0); 1527 return true; 1528 1529 case AMDGPULibFunc::EI_EXP2: 1530 Res0 = pow(2.0, opr0); 1531 return true; 1532 1533 case AMDGPULibFunc::EI_EXP10: 1534 Res0 = pow(10.0, opr0); 1535 return true; 1536 1537 case AMDGPULibFunc::EI_EXPM1: 1538 Res0 = exp(opr0) - 1.0; 1539 return true; 1540 1541 case AMDGPULibFunc::EI_LOG: 1542 Res0 = log(opr0); 1543 return true; 1544 1545 case AMDGPULibFunc::EI_LOG2: 1546 Res0 = log(opr0) / log(2.0); 1547 return true; 1548 1549 case AMDGPULibFunc::EI_LOG10: 1550 Res0 = log(opr0) / log(10.0); 1551 return true; 1552 1553 case AMDGPULibFunc::EI_RSQRT: 1554 Res0 = 1.0 / sqrt(opr0); 1555 return true; 1556 1557 case AMDGPULibFunc::EI_SIN: 1558 Res0 = sin(opr0); 1559 return true; 1560 1561 case AMDGPULibFunc::EI_SINH: 1562 Res0 = sinh(opr0); 1563 return true; 1564 1565 case AMDGPULibFunc::EI_SINPI: 1566 Res0 = sin(MATH_PI * opr0); 1567 return true; 1568 1569 case AMDGPULibFunc::EI_SQRT: 1570 Res0 = sqrt(opr0); 1571 return true; 1572 1573 case AMDGPULibFunc::EI_TAN: 1574 Res0 = tan(opr0); 1575 return true; 1576 1577 case AMDGPULibFunc::EI_TANH: 1578 Res0 = tanh(opr0); 1579 return true; 1580 1581 case AMDGPULibFunc::EI_TANPI: 1582 Res0 = tan(MATH_PI * opr0); 1583 return true; 1584 1585 case AMDGPULibFunc::EI_RECIP: 1586 Res0 = 1.0 / opr0; 1587 return true; 1588 1589 // two-arg functions 1590 case AMDGPULibFunc::EI_DIVIDE: 1591 Res0 = opr0 / opr1; 1592 return true; 1593 1594 case AMDGPULibFunc::EI_POW: 1595 case AMDGPULibFunc::EI_POWR: 1596 Res0 = pow(opr0, opr1); 1597 return true; 1598 1599 case AMDGPULibFunc::EI_POWN: { 1600 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1601 double val = (double)iopr1->getSExtValue(); 1602 Res0 = pow(opr0, val); 1603 return true; 1604 } 1605 return false; 1606 } 1607 1608 case AMDGPULibFunc::EI_ROOTN: { 1609 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) { 1610 double val = (double)iopr1->getSExtValue(); 1611 Res0 = pow(opr0, 1.0 / val); 1612 return true; 1613 } 1614 return false; 1615 } 1616 1617 // with ptr arg 1618 case AMDGPULibFunc::EI_SINCOS: 1619 Res0 = sin(opr0); 1620 Res1 = cos(opr0); 1621 return true; 1622 1623 // three-arg functions 1624 case AMDGPULibFunc::EI_FMA: 1625 case AMDGPULibFunc::EI_MAD: 1626 Res0 = opr0 * opr1 + opr2; 1627 return true; 1628 } 1629 1630 return false; 1631 } 1632 1633 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, FuncInfo &FInfo) { 1634 int numArgs = (int)aCI->getNumArgOperands(); 1635 if (numArgs > 3) 1636 return false; 1637 1638 Constant *copr0 = nullptr; 1639 Constant *copr1 = nullptr; 1640 Constant *copr2 = nullptr; 1641 if (numArgs > 0) { 1642 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr) 1643 return false; 1644 } 1645 1646 if (numArgs > 1) { 1647 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) { 1648 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS) 1649 return false; 1650 } 1651 } 1652 1653 if (numArgs > 2) { 1654 if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr) 1655 return false; 1656 } 1657 1658 // At this point, all arguments to aCI are constants. 1659 1660 // max vector size is 16, and sincos will generate two results. 1661 double DVal0[16], DVal1[16]; 1662 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS); 1663 if (getVecSize(FInfo) == 1) { 1664 if (!evaluateScalarMathFunc(FInfo, DVal0[0], 1665 DVal1[0], copr0, copr1, copr2)) { 1666 return false; 1667 } 1668 } else { 1669 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0); 1670 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1); 1671 ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2); 1672 for (int i=0; i < getVecSize(FInfo); ++i) { 1673 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr; 1674 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr; 1675 Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr; 1676 if (!evaluateScalarMathFunc(FInfo, DVal0[i], 1677 DVal1[i], celt0, celt1, celt2)) { 1678 return false; 1679 } 1680 } 1681 } 1682 1683 LLVMContext &context = CI->getParent()->getParent()->getContext(); 1684 Constant *nval0, *nval1; 1685 if (getVecSize(FInfo) == 1) { 1686 nval0 = ConstantFP::get(CI->getType(), DVal0[0]); 1687 if (hasTwoResults) 1688 nval1 = ConstantFP::get(CI->getType(), DVal1[0]); 1689 } else { 1690 if (getArgType(FInfo) == AMDGPULibFunc::F32) { 1691 SmallVector <float, 0> FVal0, FVal1; 1692 for (int i=0; i < getVecSize(FInfo); ++i) 1693 FVal0.push_back((float)DVal0[i]); 1694 ArrayRef<float> tmp0(FVal0); 1695 nval0 = ConstantDataVector::get(context, tmp0); 1696 if (hasTwoResults) { 1697 for (int i=0; i < getVecSize(FInfo); ++i) 1698 FVal1.push_back((float)DVal1[i]); 1699 ArrayRef<float> tmp1(FVal1); 1700 nval1 = ConstantDataVector::get(context, tmp1); 1701 } 1702 } else { 1703 ArrayRef<double> tmp0(DVal0); 1704 nval0 = ConstantDataVector::get(context, tmp0); 1705 if (hasTwoResults) { 1706 ArrayRef<double> tmp1(DVal1); 1707 nval1 = ConstantDataVector::get(context, tmp1); 1708 } 1709 } 1710 } 1711 1712 if (hasTwoResults) { 1713 // sincos 1714 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS && 1715 "math function with ptr arg not supported yet"); 1716 new StoreInst(nval1, aCI->getArgOperand(1), aCI); 1717 } 1718 1719 replaceCall(nval0); 1720 return true; 1721 } 1722 1723 // Public interface to the Simplify LibCalls pass. 1724 FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetOptions &Opt, 1725 const TargetMachine *TM) { 1726 return new AMDGPUSimplifyLibCalls(Opt, TM); 1727 } 1728 1729 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() { 1730 return new AMDGPUUseNativeCalls(); 1731 } 1732 1733 static bool setFastFlags(Function &F, const TargetOptions &Options) { 1734 AttrBuilder B; 1735 1736 if (Options.UnsafeFPMath || Options.NoInfsFPMath) 1737 B.addAttribute("no-infs-fp-math", "true"); 1738 if (Options.UnsafeFPMath || Options.NoNaNsFPMath) 1739 B.addAttribute("no-nans-fp-math", "true"); 1740 if (Options.UnsafeFPMath) { 1741 B.addAttribute("less-precise-fpmad", "true"); 1742 B.addAttribute("unsafe-fp-math", "true"); 1743 } 1744 1745 if (!B.hasAttributes()) 1746 return false; 1747 1748 F.addAttributes(AttributeList::FunctionIndex, B); 1749 1750 return true; 1751 } 1752 1753 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) { 1754 if (skipFunction(F)) 1755 return false; 1756 1757 bool Changed = false; 1758 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 1759 1760 LLVM_DEBUG(dbgs() << "AMDIC: process function "; 1761 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';); 1762 1763 if (!EnablePreLink) 1764 Changed |= setFastFlags(F, Options); 1765 1766 for (auto &BB : F) { 1767 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) { 1768 // Ignore non-calls. 1769 CallInst *CI = dyn_cast<CallInst>(I); 1770 ++I; 1771 if (!CI) continue; 1772 1773 // Ignore indirect calls. 1774 Function *Callee = CI->getCalledFunction(); 1775 if (Callee == 0) continue; 1776 1777 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n"; 1778 dbgs().flush()); 1779 if(Simplifier.fold(CI, AA)) 1780 Changed = true; 1781 } 1782 } 1783 return Changed; 1784 } 1785 1786 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) { 1787 if (skipFunction(F) || UseNative.empty()) 1788 return false; 1789 1790 bool Changed = false; 1791 for (auto &BB : F) { 1792 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) { 1793 // Ignore non-calls. 1794 CallInst *CI = dyn_cast<CallInst>(I); 1795 ++I; 1796 if (!CI) continue; 1797 1798 // Ignore indirect calls. 1799 Function *Callee = CI->getCalledFunction(); 1800 if (Callee == 0) continue; 1801 1802 if(Simplifier.useNative(CI)) 1803 Changed = true; 1804 } 1805 } 1806 return Changed; 1807 } 1808