1 //===- AMDGPULibCalls.cpp -------------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file does AMD library function optimizations.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "amdgpu-simplifylib"
18 #include "AMDGPULibFunc.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/IRBuilder.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/ValueSymbolTable.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Target/TargetOptions.h"
39 static cl::opt<bool> EnablePreLink("amdgpu-prelink",
40 cl::desc("Enable pre-link mode optimizations"),
44 static cl::list<std::string> UseNative("amdgpu-use-native",
45 cl::desc("Comma separated list of functions to replace with native, or all"),
46 cl::CommaSeparated, cl::ValueOptional,
49 #define MATH_PI 3.14159265358979323846264338327950288419716939937511
50 #define MATH_E 2.71828182845904523536028747135266249775724709369996
51 #define MATH_SQRT2 1.41421356237309504880168872420969807856967187537695
53 #define MATH_LOG2E 1.4426950408889634073599246810018921374266459541529859
54 #define MATH_LOG10E 0.4342944819032518276511289189166050822943970058036665
56 #define MATH_LOG2_10 3.3219280948873623478703194294893901758648313930245806
57 // Value of 1 / log2(10)
58 #define MATH_RLOG2_10 0.3010299956639811952137388947244930267681898814621085
59 // Value of 1 / M_LOG2E_F = 1 / log2(e)
60 #define MATH_RLOG2_E 0.6931471805599453094172321214581765680755001343602552
64 class AMDGPULibCalls {
67 typedef llvm::AMDGPULibFunc FuncInfo;
70 bool AllNative = false;
72 bool useNativeFunc(const StringRef F) const;
74 // Return a pointer (pointer expr) to the function if function defintion with
75 // "FuncName" exists. It may create a new function prototype in pre-link mode.
76 Constant *getFunction(Module *M, const FuncInfo& fInfo);
78 // Replace a normal function with its native version.
79 bool replaceWithNative(CallInst *CI, const FuncInfo &FInfo);
81 bool parseFunctionName(const StringRef& FMangledName,
82 FuncInfo *FInfo=nullptr /*out*/);
84 bool TDOFold(CallInst *CI, const FuncInfo &FInfo);
86 /* Specialized optimizations */
88 // recip (half or native)
89 bool fold_recip(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
91 // divide (half or native)
92 bool fold_divide(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
95 bool fold_pow(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
98 bool fold_rootn(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
101 bool fold_fma_mad(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
103 // -fuse-native for sincos
104 bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo);
106 // evaluate calls if calls' arguments are constants.
107 bool evaluateScalarMathFunc(FuncInfo &FInfo, double& Res0,
108 double& Res1, Constant *copr0, Constant *copr1, Constant *copr2);
109 bool evaluateCall(CallInst *aCI, FuncInfo &FInfo);
112 bool fold_exp(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
115 bool fold_exp2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
118 bool fold_exp10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
121 bool fold_log(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
124 bool fold_log2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
127 bool fold_log10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
130 bool fold_sqrt(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
133 bool fold_sincos(CallInst * CI, IRBuilder<> &B, AliasAnalysis * AA);
135 // __read_pipe/__write_pipe
136 bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, FuncInfo &FInfo);
138 // Get insertion point at entry.
139 BasicBlock::iterator getEntryIns(CallInst * UI);
140 // Insert an Alloc instruction.
141 AllocaInst* insertAlloca(CallInst * UI, IRBuilder<> &B, const char *prefix);
142 // Get a scalar native builtin signle argument FP function
143 Constant* getNativeFunction(Module* M, const FuncInfo &FInfo);
148 bool isUnsafeMath(const CallInst *CI) const;
150 void replaceCall(Value *With) {
151 CI->replaceAllUsesWith(With);
152 CI->eraseFromParent();
156 bool fold(CallInst *CI, AliasAnalysis *AA = nullptr);
158 void initNativeFuncs();
160 // Replace a normal math function call with that native version
161 bool useNative(CallInst *CI);
164 } // end llvm namespace
168 class AMDGPUSimplifyLibCalls : public FunctionPass {
170 AMDGPULibCalls Simplifier;
172 const TargetOptions Options;
175 static char ID; // Pass identification
177 AMDGPUSimplifyLibCalls(const TargetOptions &Opt = TargetOptions())
178 : FunctionPass(ID), Options(Opt) {
179 initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
182 void getAnalysisUsage(AnalysisUsage &AU) const override {
183 AU.addRequired<AAResultsWrapperPass>();
186 bool runOnFunction(Function &M) override;
189 class AMDGPUUseNativeCalls : public FunctionPass {
191 AMDGPULibCalls Simplifier;
194 static char ID; // Pass identification
196 AMDGPUUseNativeCalls() : FunctionPass(ID) {
197 initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry());
198 Simplifier.initNativeFuncs();
201 bool runOnFunction(Function &F) override;
204 } // end anonymous namespace.
206 char AMDGPUSimplifyLibCalls::ID = 0;
207 char AMDGPUUseNativeCalls::ID = 0;
209 INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
210 "Simplify well-known AMD library calls", false, false)
211 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
212 INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
213 "Simplify well-known AMD library calls", false, false)
215 INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative",
216 "Replace builtin math calls with that native versions.",
219 template <typename IRB>
220 static CallInst *CreateCallEx(IRB &B, Value *Callee, Value *Arg,
221 const Twine &Name = "") {
222 CallInst *R = B.CreateCall(Callee, Arg, Name);
223 if (Function* F = dyn_cast<Function>(Callee))
224 R->setCallingConv(F->getCallingConv());
228 template <typename IRB>
229 static CallInst *CreateCallEx2(IRB &B, Value *Callee, Value *Arg1, Value *Arg2,
230 const Twine &Name = "") {
231 CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
232 if (Function* F = dyn_cast<Function>(Callee))
233 R->setCallingConv(F->getCallingConv());
237 // Data structures for table-driven optimizations.
238 // FuncTbl works for both f32 and f64 functions with 1 input argument
245 /* a list of {result, input} */
246 static const TableEntry tbl_acos[] = {
252 static const TableEntry tbl_acosh[] = {
255 static const TableEntry tbl_acospi[] = {
261 static const TableEntry tbl_asin[] = {
267 static const TableEntry tbl_asinh[] = {
271 static const TableEntry tbl_asinpi[] = {
277 static const TableEntry tbl_atan[] = {
283 static const TableEntry tbl_atanh[] = {
287 static const TableEntry tbl_atanpi[] = {
293 static const TableEntry tbl_cbrt[] = {
299 static const TableEntry tbl_cos[] = {
303 static const TableEntry tbl_cosh[] = {
307 static const TableEntry tbl_cospi[] = {
311 static const TableEntry tbl_erfc[] = {
315 static const TableEntry tbl_erf[] = {
319 static const TableEntry tbl_exp[] = {
324 static const TableEntry tbl_exp2[] = {
329 static const TableEntry tbl_exp10[] = {
334 static const TableEntry tbl_expm1[] = {
338 static const TableEntry tbl_log[] = {
342 static const TableEntry tbl_log2[] = {
346 static const TableEntry tbl_log10[] = {
350 static const TableEntry tbl_rsqrt[] = {
352 {1.0/MATH_SQRT2, 2.0}
354 static const TableEntry tbl_sin[] = {
358 static const TableEntry tbl_sinh[] = {
362 static const TableEntry tbl_sinpi[] = {
366 static const TableEntry tbl_sqrt[] = {
371 static const TableEntry tbl_tan[] = {
375 static const TableEntry tbl_tanh[] = {
379 static const TableEntry tbl_tanpi[] = {
383 static const TableEntry tbl_tgamma[] = {
390 static bool HasNative(AMDGPULibFunc::EFuncId id) {
392 case AMDGPULibFunc::EI_DIVIDE:
393 case AMDGPULibFunc::EI_COS:
394 case AMDGPULibFunc::EI_EXP:
395 case AMDGPULibFunc::EI_EXP2:
396 case AMDGPULibFunc::EI_EXP10:
397 case AMDGPULibFunc::EI_LOG:
398 case AMDGPULibFunc::EI_LOG2:
399 case AMDGPULibFunc::EI_LOG10:
400 case AMDGPULibFunc::EI_POWR:
401 case AMDGPULibFunc::EI_RECIP:
402 case AMDGPULibFunc::EI_RSQRT:
403 case AMDGPULibFunc::EI_SIN:
404 case AMDGPULibFunc::EI_SINCOS:
405 case AMDGPULibFunc::EI_SQRT:
406 case AMDGPULibFunc::EI_TAN:
415 const TableEntry *table; // variable size: from 0 to (size - 1)
417 TableRef() : size(0), table(nullptr) {}
420 TableRef(const TableEntry (&tbl)[N]) : size(N), table(&tbl[0]) {}
423 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
425 case AMDGPULibFunc::EI_ACOS: return TableRef(tbl_acos);
426 case AMDGPULibFunc::EI_ACOSH: return TableRef(tbl_acosh);
427 case AMDGPULibFunc::EI_ACOSPI: return TableRef(tbl_acospi);
428 case AMDGPULibFunc::EI_ASIN: return TableRef(tbl_asin);
429 case AMDGPULibFunc::EI_ASINH: return TableRef(tbl_asinh);
430 case AMDGPULibFunc::EI_ASINPI: return TableRef(tbl_asinpi);
431 case AMDGPULibFunc::EI_ATAN: return TableRef(tbl_atan);
432 case AMDGPULibFunc::EI_ATANH: return TableRef(tbl_atanh);
433 case AMDGPULibFunc::EI_ATANPI: return TableRef(tbl_atanpi);
434 case AMDGPULibFunc::EI_CBRT: return TableRef(tbl_cbrt);
435 case AMDGPULibFunc::EI_NCOS:
436 case AMDGPULibFunc::EI_COS: return TableRef(tbl_cos);
437 case AMDGPULibFunc::EI_COSH: return TableRef(tbl_cosh);
438 case AMDGPULibFunc::EI_COSPI: return TableRef(tbl_cospi);
439 case AMDGPULibFunc::EI_ERFC: return TableRef(tbl_erfc);
440 case AMDGPULibFunc::EI_ERF: return TableRef(tbl_erf);
441 case AMDGPULibFunc::EI_EXP: return TableRef(tbl_exp);
442 case AMDGPULibFunc::EI_NEXP2:
443 case AMDGPULibFunc::EI_EXP2: return TableRef(tbl_exp2);
444 case AMDGPULibFunc::EI_EXP10: return TableRef(tbl_exp10);
445 case AMDGPULibFunc::EI_EXPM1: return TableRef(tbl_expm1);
446 case AMDGPULibFunc::EI_LOG: return TableRef(tbl_log);
447 case AMDGPULibFunc::EI_NLOG2:
448 case AMDGPULibFunc::EI_LOG2: return TableRef(tbl_log2);
449 case AMDGPULibFunc::EI_LOG10: return TableRef(tbl_log10);
450 case AMDGPULibFunc::EI_NRSQRT:
451 case AMDGPULibFunc::EI_RSQRT: return TableRef(tbl_rsqrt);
452 case AMDGPULibFunc::EI_NSIN:
453 case AMDGPULibFunc::EI_SIN: return TableRef(tbl_sin);
454 case AMDGPULibFunc::EI_SINH: return TableRef(tbl_sinh);
455 case AMDGPULibFunc::EI_SINPI: return TableRef(tbl_sinpi);
456 case AMDGPULibFunc::EI_NSQRT:
457 case AMDGPULibFunc::EI_SQRT: return TableRef(tbl_sqrt);
458 case AMDGPULibFunc::EI_TAN: return TableRef(tbl_tan);
459 case AMDGPULibFunc::EI_TANH: return TableRef(tbl_tanh);
460 case AMDGPULibFunc::EI_TANPI: return TableRef(tbl_tanpi);
461 case AMDGPULibFunc::EI_TGAMMA: return TableRef(tbl_tgamma);
467 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
468 return FInfo.getLeads()[0].VectorSize;
471 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
472 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
475 Constant *AMDGPULibCalls::getFunction(Module *M, const FuncInfo& fInfo) {
476 // If we are doing PreLinkOpt, the function is external. So it is safe to
477 // use getOrInsertFunction() at this stage.
479 return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
480 : AMDGPULibFunc::getFunction(M, fInfo);
483 bool AMDGPULibCalls::parseFunctionName(const StringRef& FMangledName,
485 return AMDGPULibFunc::parse(FMangledName, *FInfo);
488 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const {
489 if (auto Op = dyn_cast<FPMathOperator>(CI))
492 const Function *F = CI->getParent()->getParent();
493 Attribute Attr = F->getFnAttribute("unsafe-fp-math");
494 return Attr.getValueAsString() == "true";
497 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
499 std::find(UseNative.begin(), UseNative.end(), F) != UseNative.end();
502 void AMDGPULibCalls::initNativeFuncs() {
503 AllNative = useNativeFunc("all") ||
504 (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
505 UseNative.begin()->empty());
508 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
509 bool native_sin = useNativeFunc("sin");
510 bool native_cos = useNativeFunc("cos");
512 if (native_sin && native_cos) {
513 Module *M = aCI->getModule();
514 Value *opr0 = aCI->getArgOperand(0);
517 nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
518 nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
520 nf.setPrefix(AMDGPULibFunc::NATIVE);
521 nf.setId(AMDGPULibFunc::EI_SIN);
522 Constant *sinExpr = getFunction(M, nf);
524 nf.setPrefix(AMDGPULibFunc::NATIVE);
525 nf.setId(AMDGPULibFunc::EI_COS);
526 Constant *cosExpr = getFunction(M, nf);
527 if (sinExpr && cosExpr) {
528 Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
529 Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
530 new StoreInst(cosval, aCI->getArgOperand(1), aCI);
532 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
533 << " with native version of sin/cos");
542 bool AMDGPULibCalls::useNative(CallInst *aCI) {
544 Function *Callee = aCI->getCalledFunction();
547 if (!parseFunctionName(Callee->getName(), &FInfo) || !FInfo.isMangled() ||
548 FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
549 getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
550 !(AllNative || useNativeFunc(FInfo.getName()))) {
554 if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
555 return sincosUseNative(aCI, FInfo);
557 FInfo.setPrefix(AMDGPULibFunc::NATIVE);
558 Constant *F = getFunction(aCI->getModule(), FInfo);
562 aCI->setCalledFunction(F);
563 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
564 << " with native version");
568 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
569 // builtin, with appended type size and alignment arguments, where 2 or 4
570 // indicates the original number of arguments. The library has optimized version
571 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
572 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
573 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
574 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
575 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
577 auto *Callee = CI->getCalledFunction();
578 if (!Callee->isDeclaration())
581 assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
582 auto *M = Callee->getParent();
583 auto &Ctx = M->getContext();
584 std::string Name = Callee->getName();
585 auto NumArg = CI->getNumArgOperands();
586 if (NumArg != 4 && NumArg != 6)
588 auto *PacketSize = CI->getArgOperand(NumArg - 2);
589 auto *PacketAlign = CI->getArgOperand(NumArg - 1);
590 if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign))
592 unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue();
593 unsigned Align = cast<ConstantInt>(PacketAlign)->getZExtValue();
594 if (Size != Align || !isPowerOf2_32(Size))
599 PtrElemTy = Type::getIntNTy(Ctx, Size * 8);
601 PtrElemTy = VectorType::get(Type::getInt64Ty(Ctx), Size / 8);
602 unsigned PtrArgLoc = CI->getNumArgOperands() - 3;
603 auto PtrArg = CI->getArgOperand(PtrArgLoc);
604 unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace();
605 auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS);
607 SmallVector<llvm::Type *, 6> ArgTys;
608 for (unsigned I = 0; I != PtrArgLoc; ++I)
609 ArgTys.push_back(CI->getArgOperand(I)->getType());
610 ArgTys.push_back(PtrTy);
612 Name = Name + "_" + std::to_string(Size);
613 auto *FTy = FunctionType::get(Callee->getReturnType(),
614 ArrayRef<Type *>(ArgTys), false);
615 AMDGPULibFunc NewLibFunc(Name, FTy);
616 auto *F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
620 auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
621 SmallVector<Value *, 6> Args;
622 for (unsigned I = 0; I != PtrArgLoc; ++I)
623 Args.push_back(CI->getArgOperand(I));
624 Args.push_back(BCast);
626 auto *NCI = B.CreateCall(F, Args);
627 NCI->setAttributes(CI->getAttributes());
628 CI->replaceAllUsesWith(NCI);
629 CI->dropAllReferences();
630 CI->eraseFromParent();
635 // This function returns false if no change; return true otherwise.
636 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) {
638 Function *Callee = CI->getCalledFunction();
640 // Ignore indirect calls.
641 if (Callee == 0) return false;
644 if (!parseFunctionName(Callee->getName(), &FInfo))
647 // Further check the number of arguments to see if they match.
648 if (CI->getNumArgOperands() != FInfo.getNumArgs())
651 BasicBlock *BB = CI->getParent();
652 LLVMContext &Context = CI->getParent()->getContext();
653 IRBuilder<> B(Context);
655 // Set the builder to the instruction after the call.
656 B.SetInsertPoint(BB, CI->getIterator());
658 // Copy fast flags from the original call.
659 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI))
660 B.setFastMathFlags(FPOp->getFastMathFlags());
662 if (TDOFold(CI, FInfo))
665 // Under unsafe-math, evaluate calls if possible.
666 // According to Brian Sumner, we can do this for all f32 function calls
667 // using host's double function calls.
668 if (isUnsafeMath(CI) && evaluateCall(CI, FInfo))
671 // Specilized optimizations for each function call
672 switch (FInfo.getId()) {
673 case AMDGPULibFunc::EI_RECIP:
674 // skip vector function
675 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
676 FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
677 "recip must be an either native or half function");
678 return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo);
680 case AMDGPULibFunc::EI_DIVIDE:
681 // skip vector function
682 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
683 FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
684 "divide must be an either native or half function");
685 return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo);
687 case AMDGPULibFunc::EI_POW:
688 case AMDGPULibFunc::EI_POWR:
689 case AMDGPULibFunc::EI_POWN:
690 return fold_pow(CI, B, FInfo);
692 case AMDGPULibFunc::EI_ROOTN:
693 // skip vector function
694 return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo);
696 case AMDGPULibFunc::EI_FMA:
697 case AMDGPULibFunc::EI_MAD:
698 case AMDGPULibFunc::EI_NFMA:
699 // skip vector function
700 return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo);
702 case AMDGPULibFunc::EI_SQRT:
703 return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo);
704 case AMDGPULibFunc::EI_COS:
705 case AMDGPULibFunc::EI_SIN:
706 if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
707 getArgType(FInfo) == AMDGPULibFunc::F64)
708 && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX))
709 return fold_sincos(CI, B, AA);
712 case AMDGPULibFunc::EI_READ_PIPE_2:
713 case AMDGPULibFunc::EI_READ_PIPE_4:
714 case AMDGPULibFunc::EI_WRITE_PIPE_2:
715 case AMDGPULibFunc::EI_WRITE_PIPE_4:
716 return fold_read_write_pipe(CI, B, FInfo);
725 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
726 // Table-Driven optimization
727 const TableRef tr = getOptTable(FInfo.getId());
731 int const sz = (int)tr.size;
732 const TableEntry * const ftbl = tr.table;
733 Value *opr0 = CI->getArgOperand(0);
735 if (getVecSize(FInfo) > 1) {
736 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
737 SmallVector<double, 0> DVal;
738 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
739 ConstantFP *eltval = dyn_cast<ConstantFP>(
740 CV->getElementAsConstant((unsigned)eltNo));
741 assert(eltval && "Non-FP arguments in math function!");
743 for (int i=0; i < sz; ++i) {
744 if (eltval->isExactlyValue(ftbl[i].input)) {
745 DVal.push_back(ftbl[i].result);
751 // This vector constants not handled yet.
755 LLVMContext &context = CI->getParent()->getParent()->getContext();
757 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
758 SmallVector<float, 0> FVal;
759 for (unsigned i = 0; i < DVal.size(); ++i) {
760 FVal.push_back((float)DVal[i]);
762 ArrayRef<float> tmp(FVal);
763 nval = ConstantDataVector::get(context, tmp);
765 ArrayRef<double> tmp(DVal);
766 nval = ConstantDataVector::get(context, tmp);
768 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
774 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
775 for (int i = 0; i < sz; ++i) {
776 if (CF->isExactlyValue(ftbl[i].input)) {
777 Value *nval = ConstantFP::get(CF->getType(), ftbl[i].result);
778 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
789 bool AMDGPULibCalls::replaceWithNative(CallInst *CI, const FuncInfo &FInfo) {
790 Module *M = CI->getModule();
791 if (getArgType(FInfo) != AMDGPULibFunc::F32 ||
792 FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
793 !HasNative(FInfo.getId()))
796 AMDGPULibFunc nf = FInfo;
797 nf.setPrefix(AMDGPULibFunc::NATIVE);
798 if (Constant *FPExpr = getFunction(M, nf)) {
799 LLVM_DEBUG(dbgs() << "AMDIC: " << *CI << " ---> ");
801 CI->setCalledFunction(FPExpr);
803 LLVM_DEBUG(dbgs() << *CI << '\n');
810 // [native_]half_recip(c) ==> 1.0/c
811 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B,
812 const FuncInfo &FInfo) {
813 Value *opr0 = CI->getArgOperand(0);
814 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
815 // Just create a normal div. Later, InstCombine will be able
816 // to compute the divide into a constant (avoid check float infinity
817 // or subnormal at this point).
818 Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0),
821 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
828 // [native_]half_divide(x, c) ==> x/c
829 bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B,
830 const FuncInfo &FInfo) {
831 Value *opr0 = CI->getArgOperand(0);
832 Value *opr1 = CI->getArgOperand(1);
833 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
834 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
836 if ((CF0 && CF1) || // both are constants
837 (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32)))
838 // CF1 is constant && f32 divide
840 Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0),
841 opr1, "__div2recip");
842 Value *nval = B.CreateFMul(opr0, nval1, "__div2mul");
850 static double log2(double V) {
851 #if _XOPEN_SOURCE >= 600 || _ISOC99_SOURCE || _POSIX_C_SOURCE >= 200112L
854 return log(V) / 0.693147180559945309417;
859 bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B,
860 const FuncInfo &FInfo) {
861 assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
862 FInfo.getId() == AMDGPULibFunc::EI_POWR ||
863 FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
864 "fold_pow: encounter a wrong function call");
869 ConstantAggregateZero *CZero;
872 opr0 = CI->getArgOperand(0);
873 opr1 = CI->getArgOperand(1);
874 CZero = dyn_cast<ConstantAggregateZero>(opr1);
875 if (getVecSize(FInfo) == 1) {
876 eltType = opr0->getType();
877 CF = dyn_cast<ConstantFP>(opr1);
878 CINT = dyn_cast<ConstantInt>(opr1);
880 VectorType *VTy = dyn_cast<VectorType>(opr0->getType());
881 assert(VTy && "Oprand of vector function should be of vectortype");
882 eltType = VTy->getElementType();
883 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1);
885 // Now, only Handle vector const whose elements have the same value.
886 CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr;
887 CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr;
890 // No unsafe math , no constant argument, do nothing
891 if (!isUnsafeMath(CI) && !CF && !CINT && !CZero)
894 // 0x1111111 means that we don't do anything for this call.
895 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
897 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) {
898 // pow/powr/pown(x, 0) == 1
899 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n");
900 Constant *cnval = ConstantFP::get(eltType, 1.0);
901 if (getVecSize(FInfo) > 1) {
902 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
907 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
908 // pow/powr/pown(x, 1.0) = x
909 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
913 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
914 // pow/powr/pown(x, 2.0) = x*x
915 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0
917 Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
921 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
922 // pow/powr/pown(x, -1.0) = 1.0/x
923 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n");
924 Constant *cnval = ConstantFP::get(eltType, 1.0);
925 if (getVecSize(FInfo) > 1) {
926 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
928 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
933 Module *M = CI->getModule();
934 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
935 // pow[r](x, [-]0.5) = sqrt(x)
936 bool issqrt = CF->isExactlyValue(0.5);
937 if (Constant *FPExpr = getFunction(M,
938 AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
939 : AMDGPULibFunc::EI_RSQRT, FInfo))) {
940 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
941 << FInfo.getName().c_str() << "(" << *opr0 << ")\n");
942 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
949 if (!isUnsafeMath(CI))
952 // Unsafe Math optimization
954 // Remember that ci_opr1 is set if opr1 is integral
956 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
957 ? (double)CF->getValueAPF().convertToFloat()
958 : CF->getValueAPF().convertToDouble();
959 int ival = (int)dval;
960 if ((double)ival == dval) {
963 ci_opr1 = 0x11111111;
966 // pow/powr/pown(x, c) = [1/](x*x*..x); where
967 // trunc(c) == c && the number of x == c && |c| <= 12
968 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
969 if (abs_opr1 <= 12) {
973 cnval = ConstantFP::get(eltType, 1.0);
974 if (getVecSize(FInfo) > 1) {
975 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
979 Value *valx2 = nullptr;
981 while (abs_opr1 > 0) {
982 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
984 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
991 cnval = ConstantFP::get(eltType, 1.0);
992 if (getVecSize(FInfo) > 1) {
993 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
995 nval = B.CreateFDiv(cnval, nval, "__1powprod");
997 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
998 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
1004 // powr ---> exp2(y * log2(x))
1005 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
1006 Constant *ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2,
1011 bool needlog = false;
1012 bool needabs = false;
1013 bool needcopysign = false;
1014 Constant *cnval = nullptr;
1015 if (getVecSize(FInfo) == 1) {
1016 CF = dyn_cast<ConstantFP>(opr0);
1019 double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1020 ? (double)CF->getValueAPF().convertToFloat()
1021 : CF->getValueAPF().convertToDouble();
1023 V = log2(std::abs(V));
1024 cnval = ConstantFP::get(eltType, V);
1025 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
1029 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR &&
1030 (!CF || CF->isNegative());
1033 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
1037 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1039 assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
1040 "Wrong vector size detected");
1042 SmallVector<double, 0> DVal;
1043 for (int i=0; i < getVecSize(FInfo); ++i) {
1044 double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1045 ? (double)CDV->getElementAsFloat(i)
1046 : CDV->getElementAsDouble(i);
1047 if (V < 0.0) needcopysign = true;
1048 V = log2(std::abs(V));
1051 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1052 SmallVector<float, 0> FVal;
1053 for (unsigned i=0; i < DVal.size(); ++i) {
1054 FVal.push_back((float)DVal[i]);
1056 ArrayRef<float> tmp(FVal);
1057 cnval = ConstantDataVector::get(M->getContext(), tmp);
1059 ArrayRef<double> tmp(DVal);
1060 cnval = ConstantDataVector::get(M->getContext(), tmp);
1065 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1066 // We cannot handle corner cases for a general pow() function, give up
1067 // unless y is a constant integral value. Then proceed as if it were pown.
1068 if (getVecSize(FInfo) == 1) {
1069 if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) {
1070 double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1071 ? (double)CF->getValueAPF().convertToFloat()
1072 : CF->getValueAPF().convertToDouble();
1073 if (y != (double)(int64_t)y)
1078 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) {
1079 for (int i=0; i < getVecSize(FInfo); ++i) {
1080 double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1081 ? (double)CDV->getElementAsFloat(i)
1082 : CDV->getElementAsDouble(i);
1083 if (y != (double)(int64_t)y)
1093 Constant *AbsExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS,
1097 nval = CreateCallEx(B, AbsExpr, opr0, "__fabs");
1099 nval = cnval ? cnval : opr0;
1102 Constant *LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2,
1106 nval = CreateCallEx(B,LogExpr, nval, "__log2");
1109 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1110 // convert int(32) to fp(f32 or f64)
1111 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1113 nval = B.CreateFMul(opr1, nval, "__ylogx");
1114 nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1118 Type* rTy = opr0->getType();
1119 Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty();
1121 if (const VectorType *vTy = dyn_cast<VectorType>(rTy))
1122 nTy = VectorType::get(nTyS, vTy->getNumElements());
1123 unsigned size = nTy->getScalarSizeInBits();
1124 opr_n = CI->getArgOperand(1);
1125 if (opr_n->getType()->isIntegerTy())
1126 opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou");
1128 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1130 Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1131 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1132 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1133 nval = B.CreateBitCast(nval, opr0->getType());
1136 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1137 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1143 bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B,
1144 const FuncInfo &FInfo) {
1145 Value *opr0 = CI->getArgOperand(0);
1146 Value *opr1 = CI->getArgOperand(1);
1148 ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1152 int ci_opr1 = (int)CINT->getSExtValue();
1153 if (ci_opr1 == 1) { // rootn(x, 1) = x
1154 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
1158 if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x)
1159 std::vector<const Type*> ParamsTys;
1160 ParamsTys.push_back(opr0->getType());
1161 Module *M = CI->getModule();
1162 if (Constant *FPExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT,
1164 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n");
1165 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1169 } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1170 Module *M = CI->getModule();
1171 if (Constant *FPExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT,
1173 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n");
1174 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1178 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1179 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n");
1180 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1185 } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x)
1186 std::vector<const Type*> ParamsTys;
1187 ParamsTys.push_back(opr0->getType());
1188 Module *M = CI->getModule();
1189 if (Constant *FPExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT,
1191 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0
1193 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1201 bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B,
1202 const FuncInfo &FInfo) {
1203 Value *opr0 = CI->getArgOperand(0);
1204 Value *opr1 = CI->getArgOperand(1);
1205 Value *opr2 = CI->getArgOperand(2);
1207 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
1208 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
1209 if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) {
1210 // fma/mad(a, b, c) = c if a=0 || b=0
1211 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n");
1215 if (CF0 && CF0->isExactlyValue(1.0f)) {
1216 // fma/mad(a, b, c) = b+c if a=1
1217 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2
1219 Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd");
1223 if (CF1 && CF1->isExactlyValue(1.0f)) {
1224 // fma/mad(a, b, c) = a+c if b=1
1225 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2
1227 Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd");
1231 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) {
1233 // fma/mad(a, b, c) = a*b if c=0
1234 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * "
1236 Value *nval = B.CreateFMul(opr0, opr1, "fmamul");
1245 // Get a scalar native builtin signle argument FP function
1246 Constant* AMDGPULibCalls::getNativeFunction(Module* M, const FuncInfo& FInfo) {
1247 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1249 FuncInfo nf = FInfo;
1250 nf.setPrefix(AMDGPULibFunc::NATIVE);
1251 return getFunction(M, nf);
1254 // fold sqrt -> native_sqrt (x)
1255 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
1256 const FuncInfo &FInfo) {
1257 if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1258 (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1259 if (Constant *FPExpr = getNativeFunction(
1260 CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1261 Value *opr0 = CI->getArgOperand(0);
1262 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1263 << "sqrt(" << *opr0 << ")\n");
1264 Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1272 // fold sin, cos -> sincos.
1273 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
1274 AliasAnalysis *AA) {
1275 AMDGPULibFunc fInfo;
1276 if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo))
1279 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1280 fInfo.getId() == AMDGPULibFunc::EI_COS);
1281 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1283 Value *CArgVal = CI->getArgOperand(0);
1284 BasicBlock * const CBB = CI->getParent();
1286 int const MaxScan = 30;
1288 { // fold in load value.
1289 LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
1290 if (LI && LI->getParent() == CBB) {
1291 BasicBlock::iterator BBI = LI->getIterator();
1292 Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
1294 CArgVal->replaceAllUsesWith(AvailableVal);
1295 if (CArgVal->getNumUses() == 0)
1296 LI->eraseFromParent();
1297 CArgVal = CI->getArgOperand(0);
1302 Module *M = CI->getModule();
1303 fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN);
1304 std::string const PairName = fInfo.mangle();
1306 CallInst *UI = nullptr;
1307 for (User* U : CArgVal->users()) {
1308 CallInst *XI = dyn_cast_or_null<CallInst>(U);
1309 if (!XI || XI == CI || XI->getParent() != CBB)
1312 Function *UCallee = XI->getCalledFunction();
1313 if (!UCallee || !UCallee->getName().equals(PairName))
1316 BasicBlock::iterator BBI = CI->getIterator();
1317 if (BBI == CI->getParent()->begin())
1320 for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
1321 if (cast<Instruction>(BBI) == XI) {
1329 if (!UI) return false;
1331 // Merge the sin and cos.
1333 // for OpenCL 2.0 we have only generic implementation of sincos
1335 AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
1336 const AMDGPUAS AS = AMDGPU::getAMDGPUAS(*M);
1337 nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AS.FLAT_ADDRESS);
1338 Function *Fsincos = dyn_cast_or_null<Function>(getFunction(M, nf));
1339 if (!Fsincos) return false;
1341 BasicBlock::iterator ItOld = B.GetInsertPoint();
1342 AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
1343 B.SetInsertPoint(UI);
1346 Type *PTy = Fsincos->getFunctionType()->getParamType(1);
1347 // The allocaInst allocates the memory in private address space. This need
1348 // to be bitcasted to point to the address space of cos pointer type.
1349 // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1350 if (PTy->getPointerAddressSpace() != AS.PRIVATE_ADDRESS)
1351 P = B.CreateAddrSpaceCast(Alloc, PTy);
1352 CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P);
1354 LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with "
1357 if (!isSin) { // CI->cos, UI->sin
1358 B.SetInsertPoint(&*ItOld);
1359 UI->replaceAllUsesWith(&*Call);
1360 Instruction *Reload = B.CreateLoad(Alloc);
1361 CI->replaceAllUsesWith(Reload);
1362 UI->eraseFromParent();
1363 CI->eraseFromParent();
1364 } else { // CI->sin, UI->cos
1365 Instruction *Reload = B.CreateLoad(Alloc);
1366 UI->replaceAllUsesWith(Reload);
1367 CI->replaceAllUsesWith(Call);
1368 UI->eraseFromParent();
1369 CI->eraseFromParent();
1374 // Get insertion point at entry.
1375 BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) {
1376 Function * Func = UI->getParent()->getParent();
1377 BasicBlock * BB = &Func->getEntryBlock();
1378 assert(BB && "Entry block not found!");
1379 BasicBlock::iterator ItNew = BB->begin();
1383 // Insert a AllocsInst at the beginning of function entry block.
1384 AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B,
1385 const char *prefix) {
1386 BasicBlock::iterator ItNew = getEntryIns(UI);
1387 Function *UCallee = UI->getCalledFunction();
1388 Type *RetType = UCallee->getReturnType();
1389 B.SetInsertPoint(&*ItNew);
1390 AllocaInst *Alloc = B.CreateAlloca(RetType, 0,
1391 std::string(prefix) + UI->getName());
1392 Alloc->setAlignment(UCallee->getParent()->getDataLayout()
1393 .getTypeAllocSize(RetType));
1397 bool AMDGPULibCalls::evaluateScalarMathFunc(FuncInfo &FInfo,
1398 double& Res0, double& Res1,
1399 Constant *copr0, Constant *copr1,
1401 // By default, opr0/opr1/opr3 holds values of float/double type.
1402 // If they are not float/double, each function has to its
1403 // operand separately.
1404 double opr0=0.0, opr1=0.0, opr2=0.0;
1405 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1406 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1407 ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2);
1409 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1410 ? fpopr0->getValueAPF().convertToDouble()
1411 : (double)fpopr0->getValueAPF().convertToFloat();
1415 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1416 ? fpopr1->getValueAPF().convertToDouble()
1417 : (double)fpopr1->getValueAPF().convertToFloat();
1421 opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1422 ? fpopr2->getValueAPF().convertToDouble()
1423 : (double)fpopr2->getValueAPF().convertToFloat();
1426 switch (FInfo.getId()) {
1427 default : return false;
1429 case AMDGPULibFunc::EI_ACOS:
1433 case AMDGPULibFunc::EI_ACOSH:
1434 // acosh(x) == log(x + sqrt(x*x - 1))
1435 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1438 case AMDGPULibFunc::EI_ACOSPI:
1439 Res0 = acos(opr0) / MATH_PI;
1442 case AMDGPULibFunc::EI_ASIN:
1446 case AMDGPULibFunc::EI_ASINH:
1447 // asinh(x) == log(x + sqrt(x*x + 1))
1448 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1451 case AMDGPULibFunc::EI_ASINPI:
1452 Res0 = asin(opr0) / MATH_PI;
1455 case AMDGPULibFunc::EI_ATAN:
1459 case AMDGPULibFunc::EI_ATANH:
1460 // atanh(x) == (log(x+1) - log(x-1))/2;
1461 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1464 case AMDGPULibFunc::EI_ATANPI:
1465 Res0 = atan(opr0) / MATH_PI;
1468 case AMDGPULibFunc::EI_CBRT:
1469 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1472 case AMDGPULibFunc::EI_COS:
1476 case AMDGPULibFunc::EI_COSH:
1480 case AMDGPULibFunc::EI_COSPI:
1481 Res0 = cos(MATH_PI * opr0);
1484 case AMDGPULibFunc::EI_EXP:
1488 case AMDGPULibFunc::EI_EXP2:
1489 Res0 = pow(2.0, opr0);
1492 case AMDGPULibFunc::EI_EXP10:
1493 Res0 = pow(10.0, opr0);
1496 case AMDGPULibFunc::EI_EXPM1:
1497 Res0 = exp(opr0) - 1.0;
1500 case AMDGPULibFunc::EI_LOG:
1504 case AMDGPULibFunc::EI_LOG2:
1505 Res0 = log(opr0) / log(2.0);
1508 case AMDGPULibFunc::EI_LOG10:
1509 Res0 = log(opr0) / log(10.0);
1512 case AMDGPULibFunc::EI_RSQRT:
1513 Res0 = 1.0 / sqrt(opr0);
1516 case AMDGPULibFunc::EI_SIN:
1520 case AMDGPULibFunc::EI_SINH:
1524 case AMDGPULibFunc::EI_SINPI:
1525 Res0 = sin(MATH_PI * opr0);
1528 case AMDGPULibFunc::EI_SQRT:
1532 case AMDGPULibFunc::EI_TAN:
1536 case AMDGPULibFunc::EI_TANH:
1540 case AMDGPULibFunc::EI_TANPI:
1541 Res0 = tan(MATH_PI * opr0);
1544 case AMDGPULibFunc::EI_RECIP:
1548 // two-arg functions
1549 case AMDGPULibFunc::EI_DIVIDE:
1553 case AMDGPULibFunc::EI_POW:
1554 case AMDGPULibFunc::EI_POWR:
1555 Res0 = pow(opr0, opr1);
1558 case AMDGPULibFunc::EI_POWN: {
1559 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1560 double val = (double)iopr1->getSExtValue();
1561 Res0 = pow(opr0, val);
1567 case AMDGPULibFunc::EI_ROOTN: {
1568 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1569 double val = (double)iopr1->getSExtValue();
1570 Res0 = pow(opr0, 1.0 / val);
1577 case AMDGPULibFunc::EI_SINCOS:
1582 // three-arg functions
1583 case AMDGPULibFunc::EI_FMA:
1584 case AMDGPULibFunc::EI_MAD:
1585 Res0 = opr0 * opr1 + opr2;
1592 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, FuncInfo &FInfo) {
1593 int numArgs = (int)aCI->getNumArgOperands();
1597 Constant *copr0 = nullptr;
1598 Constant *copr1 = nullptr;
1599 Constant *copr2 = nullptr;
1601 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1606 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1607 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1613 if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr)
1617 // At this point, all arguments to aCI are constants.
1619 // max vector size is 16, and sincos will generate two results.
1620 double DVal0[16], DVal1[16];
1621 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1622 if (getVecSize(FInfo) == 1) {
1623 if (!evaluateScalarMathFunc(FInfo, DVal0[0],
1624 DVal1[0], copr0, copr1, copr2)) {
1628 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1629 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1630 ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2);
1631 for (int i=0; i < getVecSize(FInfo); ++i) {
1632 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1633 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1634 Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr;
1635 if (!evaluateScalarMathFunc(FInfo, DVal0[i],
1636 DVal1[i], celt0, celt1, celt2)) {
1642 LLVMContext &context = CI->getParent()->getParent()->getContext();
1643 Constant *nval0, *nval1;
1644 if (getVecSize(FInfo) == 1) {
1645 nval0 = ConstantFP::get(CI->getType(), DVal0[0]);
1647 nval1 = ConstantFP::get(CI->getType(), DVal1[0]);
1649 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1650 SmallVector <float, 0> FVal0, FVal1;
1651 for (int i=0; i < getVecSize(FInfo); ++i)
1652 FVal0.push_back((float)DVal0[i]);
1653 ArrayRef<float> tmp0(FVal0);
1654 nval0 = ConstantDataVector::get(context, tmp0);
1655 if (hasTwoResults) {
1656 for (int i=0; i < getVecSize(FInfo); ++i)
1657 FVal1.push_back((float)DVal1[i]);
1658 ArrayRef<float> tmp1(FVal1);
1659 nval1 = ConstantDataVector::get(context, tmp1);
1662 ArrayRef<double> tmp0(DVal0);
1663 nval0 = ConstantDataVector::get(context, tmp0);
1664 if (hasTwoResults) {
1665 ArrayRef<double> tmp1(DVal1);
1666 nval1 = ConstantDataVector::get(context, tmp1);
1671 if (hasTwoResults) {
1673 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1674 "math function with ptr arg not supported yet");
1675 new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1682 // Public interface to the Simplify LibCalls pass.
1683 FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetOptions &Opt) {
1684 return new AMDGPUSimplifyLibCalls(Opt);
1687 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
1688 return new AMDGPUUseNativeCalls();
1691 static bool setFastFlags(Function &F, const TargetOptions &Options) {
1694 if (Options.UnsafeFPMath || Options.NoInfsFPMath)
1695 B.addAttribute("no-infs-fp-math", "true");
1696 if (Options.UnsafeFPMath || Options.NoNaNsFPMath)
1697 B.addAttribute("no-nans-fp-math", "true");
1698 if (Options.UnsafeFPMath) {
1699 B.addAttribute("less-precise-fpmad", "true");
1700 B.addAttribute("unsafe-fp-math", "true");
1703 if (!B.hasAttributes())
1706 F.addAttributes(AttributeList::FunctionIndex, B);
1711 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
1712 if (skipFunction(F))
1715 bool Changed = false;
1716 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1718 LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1719 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1722 Changed |= setFastFlags(F, Options);
1724 for (auto &BB : F) {
1725 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1726 // Ignore non-calls.
1727 CallInst *CI = dyn_cast<CallInst>(I);
1731 // Ignore indirect calls.
1732 Function *Callee = CI->getCalledFunction();
1733 if (Callee == 0) continue;
1735 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1737 if(Simplifier.fold(CI, AA))
1744 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) {
1745 if (skipFunction(F) || UseNative.empty())
1748 bool Changed = false;
1749 for (auto &BB : F) {
1750 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1751 // Ignore non-calls.
1752 CallInst *CI = dyn_cast<CallInst>(I);
1756 // Ignore indirect calls.
1757 Function *Callee = CI->getCalledFunction();
1758 if (Callee == 0) continue;
1760 if(Simplifier.useNative(CI))