xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp (revision 930996e9e442d69a321cd5c945543d37c97f4c0e)
1 //===- AMDGPULibCalls.cpp -------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file does AMD library function optimizations.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "AMDGPU.h"
15 #include "AMDGPULibFunc.h"
16 #include "GCNSubtarget.h"
17 #include "llvm/Analysis/AssumptionCache.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/AttributeMask.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/IntrinsicsAMDGPU.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/InitializePasses.h"
27 #include <cmath>
28 
29 #define DEBUG_TYPE "amdgpu-simplifylib"
30 
31 using namespace llvm;
32 using namespace llvm::PatternMatch;
33 
34 static cl::opt<bool> EnablePreLink("amdgpu-prelink",
35   cl::desc("Enable pre-link mode optimizations"),
36   cl::init(false),
37   cl::Hidden);
38 
39 static cl::list<std::string> UseNative("amdgpu-use-native",
40   cl::desc("Comma separated list of functions to replace with native, or all"),
41   cl::CommaSeparated, cl::ValueOptional,
42   cl::Hidden);
43 
44 #define MATH_PI      numbers::pi
45 #define MATH_E       numbers::e
46 #define MATH_SQRT2   numbers::sqrt2
47 #define MATH_SQRT1_2 numbers::inv_sqrt2
48 
49 namespace llvm {
50 
51 class AMDGPULibCalls {
52 private:
53   const TargetLibraryInfo *TLInfo = nullptr;
54   AssumptionCache *AC = nullptr;
55   DominatorTree *DT = nullptr;
56 
57   typedef llvm::AMDGPULibFunc FuncInfo;
58 
59   bool UnsafeFPMath = false;
60 
61   // -fuse-native.
62   bool AllNative = false;
63 
64   bool useNativeFunc(const StringRef F) const;
65 
66   // Return a pointer (pointer expr) to the function if function definition with
67   // "FuncName" exists. It may create a new function prototype in pre-link mode.
68   FunctionCallee getFunction(Module *M, const FuncInfo &fInfo);
69 
70   bool parseFunctionName(const StringRef &FMangledName, FuncInfo &FInfo);
71 
72   bool TDOFold(CallInst *CI, const FuncInfo &FInfo);
73 
74   /* Specialized optimizations */
75 
76   // pow/powr/pown
77   bool fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
78 
79   // rootn
80   bool fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
81 
82   // -fuse-native for sincos
83   bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo);
84 
85   // evaluate calls if calls' arguments are constants.
86   bool evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0, double &Res1,
87                               Constant *copr0, Constant *copr1);
88   bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo);
89 
90   /// Insert a value to sincos function \p Fsincos. Returns (value of sin, value
91   /// of cos, sincos call).
92   std::tuple<Value *, Value *, Value *> insertSinCos(Value *Arg,
93                                                      FastMathFlags FMF,
94                                                      IRBuilder<> &B,
95                                                      FunctionCallee Fsincos);
96 
97   // sin/cos
98   bool fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
99 
100   // __read_pipe/__write_pipe
101   bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
102                             const FuncInfo &FInfo);
103 
104   // Get a scalar native builtin single argument FP function
105   FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo);
106 
107   /// Substitute a call to a known libcall with an intrinsic call. If \p
108   /// AllowMinSize is true, allow the replacement in a minsize function.
109   bool shouldReplaceLibcallWithIntrinsic(const CallInst *CI,
110                                          bool AllowMinSizeF32 = false,
111                                          bool AllowF64 = false,
112                                          bool AllowStrictFP = false);
113   void replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI,
114                                          Intrinsic::ID IntrID);
115 
116   bool tryReplaceLibcallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI,
117                                             Intrinsic::ID IntrID,
118                                             bool AllowMinSizeF32 = false,
119                                             bool AllowF64 = false,
120                                             bool AllowStrictFP = false);
121 
122 protected:
123   bool isUnsafeMath(const FPMathOperator *FPOp) const;
124   bool isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const;
125 
126   bool canIncreasePrecisionOfConstantFold(const FPMathOperator *FPOp) const;
127 
128   static void replaceCall(Instruction *I, Value *With) {
129     I->replaceAllUsesWith(With);
130     I->eraseFromParent();
131   }
132 
133   static void replaceCall(FPMathOperator *I, Value *With) {
134     replaceCall(cast<Instruction>(I), With);
135   }
136 
137 public:
138   AMDGPULibCalls() {}
139 
140   bool fold(CallInst *CI);
141 
142   void initFunction(Function &F, FunctionAnalysisManager &FAM);
143   void initNativeFuncs();
144 
145   // Replace a normal math function call with that native version
146   bool useNative(CallInst *CI);
147 };
148 
149 } // end llvm namespace
150 
151 template <typename IRB>
152 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg,
153                               const Twine &Name = "") {
154   CallInst *R = B.CreateCall(Callee, Arg, Name);
155   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
156     R->setCallingConv(F->getCallingConv());
157   return R;
158 }
159 
160 template <typename IRB>
161 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1,
162                                Value *Arg2, const Twine &Name = "") {
163   CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
164   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
165     R->setCallingConv(F->getCallingConv());
166   return R;
167 }
168 
169 static FunctionType *getPownType(FunctionType *FT) {
170   Type *PowNExpTy = Type::getInt32Ty(FT->getContext());
171   if (VectorType *VecTy = dyn_cast<VectorType>(FT->getReturnType()))
172     PowNExpTy = VectorType::get(PowNExpTy, VecTy->getElementCount());
173 
174   return FunctionType::get(FT->getReturnType(),
175                            {FT->getParamType(0), PowNExpTy}, false);
176 }
177 
178 //  Data structures for table-driven optimizations.
179 //  FuncTbl works for both f32 and f64 functions with 1 input argument
180 
181 struct TableEntry {
182   double   result;
183   double   input;
184 };
185 
186 /* a list of {result, input} */
187 static const TableEntry tbl_acos[] = {
188   {MATH_PI / 2.0, 0.0},
189   {MATH_PI / 2.0, -0.0},
190   {0.0, 1.0},
191   {MATH_PI, -1.0}
192 };
193 static const TableEntry tbl_acosh[] = {
194   {0.0, 1.0}
195 };
196 static const TableEntry tbl_acospi[] = {
197   {0.5, 0.0},
198   {0.5, -0.0},
199   {0.0, 1.0},
200   {1.0, -1.0}
201 };
202 static const TableEntry tbl_asin[] = {
203   {0.0, 0.0},
204   {-0.0, -0.0},
205   {MATH_PI / 2.0, 1.0},
206   {-MATH_PI / 2.0, -1.0}
207 };
208 static const TableEntry tbl_asinh[] = {
209   {0.0, 0.0},
210   {-0.0, -0.0}
211 };
212 static const TableEntry tbl_asinpi[] = {
213   {0.0, 0.0},
214   {-0.0, -0.0},
215   {0.5, 1.0},
216   {-0.5, -1.0}
217 };
218 static const TableEntry tbl_atan[] = {
219   {0.0, 0.0},
220   {-0.0, -0.0},
221   {MATH_PI / 4.0, 1.0},
222   {-MATH_PI / 4.0, -1.0}
223 };
224 static const TableEntry tbl_atanh[] = {
225   {0.0, 0.0},
226   {-0.0, -0.0}
227 };
228 static const TableEntry tbl_atanpi[] = {
229   {0.0, 0.0},
230   {-0.0, -0.0},
231   {0.25, 1.0},
232   {-0.25, -1.0}
233 };
234 static const TableEntry tbl_cbrt[] = {
235   {0.0, 0.0},
236   {-0.0, -0.0},
237   {1.0, 1.0},
238   {-1.0, -1.0},
239 };
240 static const TableEntry tbl_cos[] = {
241   {1.0, 0.0},
242   {1.0, -0.0}
243 };
244 static const TableEntry tbl_cosh[] = {
245   {1.0, 0.0},
246   {1.0, -0.0}
247 };
248 static const TableEntry tbl_cospi[] = {
249   {1.0, 0.0},
250   {1.0, -0.0}
251 };
252 static const TableEntry tbl_erfc[] = {
253   {1.0, 0.0},
254   {1.0, -0.0}
255 };
256 static const TableEntry tbl_erf[] = {
257   {0.0, 0.0},
258   {-0.0, -0.0}
259 };
260 static const TableEntry tbl_exp[] = {
261   {1.0, 0.0},
262   {1.0, -0.0},
263   {MATH_E, 1.0}
264 };
265 static const TableEntry tbl_exp2[] = {
266   {1.0, 0.0},
267   {1.0, -0.0},
268   {2.0, 1.0}
269 };
270 static const TableEntry tbl_exp10[] = {
271   {1.0, 0.0},
272   {1.0, -0.0},
273   {10.0, 1.0}
274 };
275 static const TableEntry tbl_expm1[] = {
276   {0.0, 0.0},
277   {-0.0, -0.0}
278 };
279 static const TableEntry tbl_log[] = {
280   {0.0, 1.0},
281   {1.0, MATH_E}
282 };
283 static const TableEntry tbl_log2[] = {
284   {0.0, 1.0},
285   {1.0, 2.0}
286 };
287 static const TableEntry tbl_log10[] = {
288   {0.0, 1.0},
289   {1.0, 10.0}
290 };
291 static const TableEntry tbl_rsqrt[] = {
292   {1.0, 1.0},
293   {MATH_SQRT1_2, 2.0}
294 };
295 static const TableEntry tbl_sin[] = {
296   {0.0, 0.0},
297   {-0.0, -0.0}
298 };
299 static const TableEntry tbl_sinh[] = {
300   {0.0, 0.0},
301   {-0.0, -0.0}
302 };
303 static const TableEntry tbl_sinpi[] = {
304   {0.0, 0.0},
305   {-0.0, -0.0}
306 };
307 static const TableEntry tbl_sqrt[] = {
308   {0.0, 0.0},
309   {1.0, 1.0},
310   {MATH_SQRT2, 2.0}
311 };
312 static const TableEntry tbl_tan[] = {
313   {0.0, 0.0},
314   {-0.0, -0.0}
315 };
316 static const TableEntry tbl_tanh[] = {
317   {0.0, 0.0},
318   {-0.0, -0.0}
319 };
320 static const TableEntry tbl_tanpi[] = {
321   {0.0, 0.0},
322   {-0.0, -0.0}
323 };
324 static const TableEntry tbl_tgamma[] = {
325   {1.0, 1.0},
326   {1.0, 2.0},
327   {2.0, 3.0},
328   {6.0, 4.0}
329 };
330 
331 static bool HasNative(AMDGPULibFunc::EFuncId id) {
332   switch(id) {
333   case AMDGPULibFunc::EI_DIVIDE:
334   case AMDGPULibFunc::EI_COS:
335   case AMDGPULibFunc::EI_EXP:
336   case AMDGPULibFunc::EI_EXP2:
337   case AMDGPULibFunc::EI_EXP10:
338   case AMDGPULibFunc::EI_LOG:
339   case AMDGPULibFunc::EI_LOG2:
340   case AMDGPULibFunc::EI_LOG10:
341   case AMDGPULibFunc::EI_POWR:
342   case AMDGPULibFunc::EI_RECIP:
343   case AMDGPULibFunc::EI_RSQRT:
344   case AMDGPULibFunc::EI_SIN:
345   case AMDGPULibFunc::EI_SINCOS:
346   case AMDGPULibFunc::EI_SQRT:
347   case AMDGPULibFunc::EI_TAN:
348     return true;
349   default:;
350   }
351   return false;
352 }
353 
354 using TableRef = ArrayRef<TableEntry>;
355 
356 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
357   switch(id) {
358   case AMDGPULibFunc::EI_ACOS:    return TableRef(tbl_acos);
359   case AMDGPULibFunc::EI_ACOSH:   return TableRef(tbl_acosh);
360   case AMDGPULibFunc::EI_ACOSPI:  return TableRef(tbl_acospi);
361   case AMDGPULibFunc::EI_ASIN:    return TableRef(tbl_asin);
362   case AMDGPULibFunc::EI_ASINH:   return TableRef(tbl_asinh);
363   case AMDGPULibFunc::EI_ASINPI:  return TableRef(tbl_asinpi);
364   case AMDGPULibFunc::EI_ATAN:    return TableRef(tbl_atan);
365   case AMDGPULibFunc::EI_ATANH:   return TableRef(tbl_atanh);
366   case AMDGPULibFunc::EI_ATANPI:  return TableRef(tbl_atanpi);
367   case AMDGPULibFunc::EI_CBRT:    return TableRef(tbl_cbrt);
368   case AMDGPULibFunc::EI_NCOS:
369   case AMDGPULibFunc::EI_COS:     return TableRef(tbl_cos);
370   case AMDGPULibFunc::EI_COSH:    return TableRef(tbl_cosh);
371   case AMDGPULibFunc::EI_COSPI:   return TableRef(tbl_cospi);
372   case AMDGPULibFunc::EI_ERFC:    return TableRef(tbl_erfc);
373   case AMDGPULibFunc::EI_ERF:     return TableRef(tbl_erf);
374   case AMDGPULibFunc::EI_EXP:     return TableRef(tbl_exp);
375   case AMDGPULibFunc::EI_NEXP2:
376   case AMDGPULibFunc::EI_EXP2:    return TableRef(tbl_exp2);
377   case AMDGPULibFunc::EI_EXP10:   return TableRef(tbl_exp10);
378   case AMDGPULibFunc::EI_EXPM1:   return TableRef(tbl_expm1);
379   case AMDGPULibFunc::EI_LOG:     return TableRef(tbl_log);
380   case AMDGPULibFunc::EI_NLOG2:
381   case AMDGPULibFunc::EI_LOG2:    return TableRef(tbl_log2);
382   case AMDGPULibFunc::EI_LOG10:   return TableRef(tbl_log10);
383   case AMDGPULibFunc::EI_NRSQRT:
384   case AMDGPULibFunc::EI_RSQRT:   return TableRef(tbl_rsqrt);
385   case AMDGPULibFunc::EI_NSIN:
386   case AMDGPULibFunc::EI_SIN:     return TableRef(tbl_sin);
387   case AMDGPULibFunc::EI_SINH:    return TableRef(tbl_sinh);
388   case AMDGPULibFunc::EI_SINPI:   return TableRef(tbl_sinpi);
389   case AMDGPULibFunc::EI_NSQRT:
390   case AMDGPULibFunc::EI_SQRT:    return TableRef(tbl_sqrt);
391   case AMDGPULibFunc::EI_TAN:     return TableRef(tbl_tan);
392   case AMDGPULibFunc::EI_TANH:    return TableRef(tbl_tanh);
393   case AMDGPULibFunc::EI_TANPI:   return TableRef(tbl_tanpi);
394   case AMDGPULibFunc::EI_TGAMMA:  return TableRef(tbl_tgamma);
395   default:;
396   }
397   return TableRef();
398 }
399 
400 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
401   return FInfo.getLeads()[0].VectorSize;
402 }
403 
404 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
405   return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
406 }
407 
408 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) {
409   // If we are doing PreLinkOpt, the function is external. So it is safe to
410   // use getOrInsertFunction() at this stage.
411 
412   return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
413                        : AMDGPULibFunc::getFunction(M, fInfo);
414 }
415 
416 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName,
417                                        FuncInfo &FInfo) {
418   return AMDGPULibFunc::parse(FMangledName, FInfo);
419 }
420 
421 bool AMDGPULibCalls::isUnsafeMath(const FPMathOperator *FPOp) const {
422   return UnsafeFPMath || FPOp->isFast();
423 }
424 
425 bool AMDGPULibCalls::isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const {
426   return UnsafeFPMath ||
427          (FPOp->hasApproxFunc() && FPOp->hasNoNaNs() && FPOp->hasNoInfs());
428 }
429 
430 bool AMDGPULibCalls::canIncreasePrecisionOfConstantFold(
431     const FPMathOperator *FPOp) const {
432   // TODO: Refine to approxFunc or contract
433   return isUnsafeMath(FPOp);
434 }
435 
436 void AMDGPULibCalls::initFunction(Function &F, FunctionAnalysisManager &FAM) {
437   UnsafeFPMath = F.getFnAttribute("unsafe-fp-math").getValueAsBool();
438   AC = &FAM.getResult<AssumptionAnalysis>(F);
439   TLInfo = &FAM.getResult<TargetLibraryAnalysis>(F);
440   DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
441 }
442 
443 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
444   return AllNative || llvm::is_contained(UseNative, F);
445 }
446 
447 void AMDGPULibCalls::initNativeFuncs() {
448   AllNative = useNativeFunc("all") ||
449               (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
450                UseNative.begin()->empty());
451 }
452 
453 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
454   bool native_sin = useNativeFunc("sin");
455   bool native_cos = useNativeFunc("cos");
456 
457   if (native_sin && native_cos) {
458     Module *M = aCI->getModule();
459     Value *opr0 = aCI->getArgOperand(0);
460 
461     AMDGPULibFunc nf;
462     nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
463     nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
464 
465     nf.setPrefix(AMDGPULibFunc::NATIVE);
466     nf.setId(AMDGPULibFunc::EI_SIN);
467     FunctionCallee sinExpr = getFunction(M, nf);
468 
469     nf.setPrefix(AMDGPULibFunc::NATIVE);
470     nf.setId(AMDGPULibFunc::EI_COS);
471     FunctionCallee cosExpr = getFunction(M, nf);
472     if (sinExpr && cosExpr) {
473       Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
474       Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
475       new StoreInst(cosval, aCI->getArgOperand(1), aCI);
476 
477       DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
478                                           << " with native version of sin/cos");
479 
480       replaceCall(aCI, sinval);
481       return true;
482     }
483   }
484   return false;
485 }
486 
487 bool AMDGPULibCalls::useNative(CallInst *aCI) {
488   Function *Callee = aCI->getCalledFunction();
489   if (!Callee || aCI->isNoBuiltin())
490     return false;
491 
492   FuncInfo FInfo;
493   if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() ||
494       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
495       getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
496       !(AllNative || useNativeFunc(FInfo.getName()))) {
497     return false;
498   }
499 
500   if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
501     return sincosUseNative(aCI, FInfo);
502 
503   FInfo.setPrefix(AMDGPULibFunc::NATIVE);
504   FunctionCallee F = getFunction(aCI->getModule(), FInfo);
505   if (!F)
506     return false;
507 
508   aCI->setCalledFunction(F);
509   DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
510                                       << " with native version");
511   return true;
512 }
513 
514 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
515 // builtin, with appended type size and alignment arguments, where 2 or 4
516 // indicates the original number of arguments. The library has optimized version
517 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
518 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
519 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
520 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
521 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
522                                           const FuncInfo &FInfo) {
523   auto *Callee = CI->getCalledFunction();
524   if (!Callee->isDeclaration())
525     return false;
526 
527   assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
528   auto *M = Callee->getParent();
529   std::string Name = std::string(Callee->getName());
530   auto NumArg = CI->arg_size();
531   if (NumArg != 4 && NumArg != 6)
532     return false;
533   ConstantInt *PacketSize =
534       dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 2));
535   ConstantInt *PacketAlign =
536       dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 1));
537   if (!PacketSize || !PacketAlign)
538     return false;
539 
540   unsigned Size = PacketSize->getZExtValue();
541   Align Alignment = PacketAlign->getAlignValue();
542   if (Alignment != Size)
543     return false;
544 
545   unsigned PtrArgLoc = CI->arg_size() - 3;
546   Value *PtrArg = CI->getArgOperand(PtrArgLoc);
547   Type *PtrTy = PtrArg->getType();
548 
549   SmallVector<llvm::Type *, 6> ArgTys;
550   for (unsigned I = 0; I != PtrArgLoc; ++I)
551     ArgTys.push_back(CI->getArgOperand(I)->getType());
552   ArgTys.push_back(PtrTy);
553 
554   Name = Name + "_" + std::to_string(Size);
555   auto *FTy = FunctionType::get(Callee->getReturnType(),
556                                 ArrayRef<Type *>(ArgTys), false);
557   AMDGPULibFunc NewLibFunc(Name, FTy);
558   FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
559   if (!F)
560     return false;
561 
562   SmallVector<Value *, 6> Args;
563   for (unsigned I = 0; I != PtrArgLoc; ++I)
564     Args.push_back(CI->getArgOperand(I));
565   Args.push_back(PtrArg);
566 
567   auto *NCI = B.CreateCall(F, Args);
568   NCI->setAttributes(CI->getAttributes());
569   CI->replaceAllUsesWith(NCI);
570   CI->dropAllReferences();
571   CI->eraseFromParent();
572 
573   return true;
574 }
575 
576 static bool isKnownIntegral(const Value *V, const DataLayout &DL,
577                             FastMathFlags FMF) {
578   if (isa<UndefValue>(V))
579     return true;
580 
581   if (const ConstantFP *CF = dyn_cast<ConstantFP>(V))
582     return CF->getValueAPF().isInteger();
583 
584   if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) {
585     for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) {
586       Constant *ConstElt = CDV->getElementAsConstant(i);
587       if (isa<UndefValue>(ConstElt))
588         continue;
589       const ConstantFP *CFP = dyn_cast<ConstantFP>(ConstElt);
590       if (!CFP || !CFP->getValue().isInteger())
591         return false;
592     }
593 
594     return true;
595   }
596 
597   const Instruction *I = dyn_cast<Instruction>(V);
598   if (!I)
599     return false;
600 
601   switch (I->getOpcode()) {
602   case Instruction::SIToFP:
603   case Instruction::UIToFP:
604     // TODO: Could check nofpclass(inf) on incoming argument
605     if (FMF.noInfs())
606       return true;
607 
608     // Need to check int size cannot produce infinity, which computeKnownFPClass
609     // knows how to do already.
610     return isKnownNeverInfinity(I, /*Depth=*/0, SimplifyQuery(DL));
611   case Instruction::Call: {
612     const CallInst *CI = cast<CallInst>(I);
613     switch (CI->getIntrinsicID()) {
614     case Intrinsic::trunc:
615     case Intrinsic::floor:
616     case Intrinsic::ceil:
617     case Intrinsic::rint:
618     case Intrinsic::nearbyint:
619     case Intrinsic::round:
620     case Intrinsic::roundeven:
621       return (FMF.noInfs() && FMF.noNaNs()) ||
622              isKnownNeverInfOrNaN(I, /*Depth=*/0, SimplifyQuery(DL));
623     default:
624       break;
625     }
626 
627     break;
628   }
629   default:
630     break;
631   }
632 
633   return false;
634 }
635 
636 // This function returns false if no change; return true otherwise.
637 bool AMDGPULibCalls::fold(CallInst *CI) {
638   Function *Callee = CI->getCalledFunction();
639   // Ignore indirect calls.
640   if (!Callee || Callee->isIntrinsic() || CI->isNoBuiltin())
641     return false;
642 
643   FuncInfo FInfo;
644   if (!parseFunctionName(Callee->getName(), FInfo))
645     return false;
646 
647   // Further check the number of arguments to see if they match.
648   // TODO: Check calling convention matches too
649   if (!FInfo.isCompatibleSignature(CI->getFunctionType()))
650     return false;
651 
652   LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << '\n');
653 
654   if (TDOFold(CI, FInfo))
655     return true;
656 
657   IRBuilder<> B(CI);
658 
659   if (FPMathOperator *FPOp = dyn_cast<FPMathOperator>(CI)) {
660     // Under unsafe-math, evaluate calls if possible.
661     // According to Brian Sumner, we can do this for all f32 function calls
662     // using host's double function calls.
663     if (canIncreasePrecisionOfConstantFold(FPOp) && evaluateCall(CI, FInfo))
664       return true;
665 
666     // Copy fast flags from the original call.
667     FastMathFlags FMF = FPOp->getFastMathFlags();
668     B.setFastMathFlags(FMF);
669 
670     // Specialized optimizations for each function call.
671     //
672     // TODO: Handle native functions
673     switch (FInfo.getId()) {
674     case AMDGPULibFunc::EI_EXP:
675       if (FMF.none())
676         return false;
677       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp,
678                                                   FMF.approxFunc());
679     case AMDGPULibFunc::EI_EXP2:
680       if (FMF.none())
681         return false;
682       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp2,
683                                                   FMF.approxFunc());
684     case AMDGPULibFunc::EI_LOG:
685       if (FMF.none())
686         return false;
687       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log,
688                                                   FMF.approxFunc());
689     case AMDGPULibFunc::EI_LOG2:
690       if (FMF.none())
691         return false;
692       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log2,
693                                                   FMF.approxFunc());
694     case AMDGPULibFunc::EI_LOG10:
695       if (FMF.none())
696         return false;
697       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log10,
698                                                   FMF.approxFunc());
699     case AMDGPULibFunc::EI_FMIN:
700       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::minnum,
701                                                   true, true);
702     case AMDGPULibFunc::EI_FMAX:
703       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::maxnum,
704                                                   true, true);
705     case AMDGPULibFunc::EI_FMA:
706       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fma, true,
707                                                   true);
708     case AMDGPULibFunc::EI_MAD:
709       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fmuladd,
710                                                   true, true);
711     case AMDGPULibFunc::EI_FABS:
712       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fabs, true,
713                                                   true, true);
714     case AMDGPULibFunc::EI_COPYSIGN:
715       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::copysign,
716                                                   true, true, true);
717     case AMDGPULibFunc::EI_FLOOR:
718       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::floor, true,
719                                                   true);
720     case AMDGPULibFunc::EI_CEIL:
721       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::ceil, true,
722                                                   true);
723     case AMDGPULibFunc::EI_TRUNC:
724       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::trunc, true,
725                                                   true);
726     case AMDGPULibFunc::EI_RINT:
727       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::rint, true,
728                                                   true);
729     case AMDGPULibFunc::EI_ROUND:
730       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::round, true,
731                                                   true);
732     case AMDGPULibFunc::EI_LDEXP: {
733       if (!shouldReplaceLibcallWithIntrinsic(CI, true, true))
734         return false;
735 
736       Value *Arg1 = CI->getArgOperand(1);
737       if (VectorType *VecTy = dyn_cast<VectorType>(CI->getType());
738           VecTy && !isa<VectorType>(Arg1->getType())) {
739         Value *SplatArg1 = B.CreateVectorSplat(VecTy->getElementCount(), Arg1);
740         CI->setArgOperand(1, SplatArg1);
741       }
742 
743       CI->setCalledFunction(Intrinsic::getDeclaration(
744           CI->getModule(), Intrinsic::ldexp,
745           {CI->getType(), CI->getArgOperand(1)->getType()}));
746       return true;
747     }
748     case AMDGPULibFunc::EI_POW: {
749       Module *M = Callee->getParent();
750       AMDGPULibFunc PowrInfo(AMDGPULibFunc::EI_POWR, FInfo);
751       FunctionCallee PowrFunc = getFunction(M, PowrInfo);
752       CallInst *Call = cast<CallInst>(FPOp);
753 
754       // pow(x, y) -> powr(x, y) for x >= -0.0
755       // TODO: Account for flags on current call
756       if (PowrFunc &&
757           cannotBeOrderedLessThanZero(
758               FPOp->getOperand(0), /*Depth=*/0,
759               SimplifyQuery(M->getDataLayout(), TLInfo, DT, AC, Call))) {
760         Call->setCalledFunction(PowrFunc);
761         return fold_pow(FPOp, B, PowrInfo) || true;
762       }
763 
764       // pow(x, y) -> pown(x, y) for known integral y
765       if (isKnownIntegral(FPOp->getOperand(1), M->getDataLayout(),
766                           FPOp->getFastMathFlags())) {
767         FunctionType *PownType = getPownType(CI->getFunctionType());
768         AMDGPULibFunc PownInfo(AMDGPULibFunc::EI_POWN, PownType, true);
769         FunctionCallee PownFunc = getFunction(M, PownInfo);
770         if (PownFunc) {
771           // TODO: If the incoming integral value is an sitofp/uitofp, it won't
772           // fold out without a known range. We can probably take the source
773           // value directly.
774           Value *CastedArg =
775               B.CreateFPToSI(FPOp->getOperand(1), PownType->getParamType(1));
776           // Have to drop any nofpclass attributes on the original call site.
777           Call->removeParamAttrs(
778               1, AttributeFuncs::typeIncompatible(CastedArg->getType()));
779           Call->setCalledFunction(PownFunc);
780           Call->setArgOperand(1, CastedArg);
781           return fold_pow(FPOp, B, PownInfo) || true;
782         }
783       }
784 
785       return fold_pow(FPOp, B, FInfo);
786     }
787     case AMDGPULibFunc::EI_POWR:
788     case AMDGPULibFunc::EI_POWN:
789       return fold_pow(FPOp, B, FInfo);
790     case AMDGPULibFunc::EI_ROOTN:
791       return fold_rootn(FPOp, B, FInfo);
792     case AMDGPULibFunc::EI_SQRT:
793       // TODO: Allow with strictfp + constrained intrinsic
794       return tryReplaceLibcallWithSimpleIntrinsic(
795           B, CI, Intrinsic::sqrt, true, true, /*AllowStrictFP=*/false);
796     case AMDGPULibFunc::EI_COS:
797     case AMDGPULibFunc::EI_SIN:
798       return fold_sincos(FPOp, B, FInfo);
799     default:
800       break;
801     }
802   } else {
803     // Specialized optimizations for each function call
804     switch (FInfo.getId()) {
805     case AMDGPULibFunc::EI_READ_PIPE_2:
806     case AMDGPULibFunc::EI_READ_PIPE_4:
807     case AMDGPULibFunc::EI_WRITE_PIPE_2:
808     case AMDGPULibFunc::EI_WRITE_PIPE_4:
809       return fold_read_write_pipe(CI, B, FInfo);
810     default:
811       break;
812     }
813   }
814 
815   return false;
816 }
817 
818 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
819   // Table-Driven optimization
820   const TableRef tr = getOptTable(FInfo.getId());
821   if (tr.empty())
822     return false;
823 
824   int const sz = (int)tr.size();
825   Value *opr0 = CI->getArgOperand(0);
826 
827   if (getVecSize(FInfo) > 1) {
828     if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
829       SmallVector<double, 0> DVal;
830       for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
831         ConstantFP *eltval = dyn_cast<ConstantFP>(
832                                CV->getElementAsConstant((unsigned)eltNo));
833         assert(eltval && "Non-FP arguments in math function!");
834         bool found = false;
835         for (int i=0; i < sz; ++i) {
836           if (eltval->isExactlyValue(tr[i].input)) {
837             DVal.push_back(tr[i].result);
838             found = true;
839             break;
840           }
841         }
842         if (!found) {
843           // This vector constants not handled yet.
844           return false;
845         }
846       }
847       LLVMContext &context = CI->getParent()->getParent()->getContext();
848       Constant *nval;
849       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
850         SmallVector<float, 0> FVal;
851         for (unsigned i = 0; i < DVal.size(); ++i) {
852           FVal.push_back((float)DVal[i]);
853         }
854         ArrayRef<float> tmp(FVal);
855         nval = ConstantDataVector::get(context, tmp);
856       } else { // F64
857         ArrayRef<double> tmp(DVal);
858         nval = ConstantDataVector::get(context, tmp);
859       }
860       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
861       replaceCall(CI, nval);
862       return true;
863     }
864   } else {
865     // Scalar version
866     if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
867       for (int i = 0; i < sz; ++i) {
868         if (CF->isExactlyValue(tr[i].input)) {
869           Value *nval = ConstantFP::get(CF->getType(), tr[i].result);
870           LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
871           replaceCall(CI, nval);
872           return true;
873         }
874       }
875     }
876   }
877 
878   return false;
879 }
880 
881 namespace llvm {
882 static double log2(double V) {
883 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L
884   return ::log2(V);
885 #else
886   return log(V) / numbers::ln2;
887 #endif
888 }
889 }
890 
891 bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B,
892                               const FuncInfo &FInfo) {
893   assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
894           FInfo.getId() == AMDGPULibFunc::EI_POWR ||
895           FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
896          "fold_pow: encounter a wrong function call");
897 
898   Module *M = B.GetInsertBlock()->getModule();
899   Type *eltType = FPOp->getType()->getScalarType();
900   Value *opr0 = FPOp->getOperand(0);
901   Value *opr1 = FPOp->getOperand(1);
902 
903   const APFloat *CF = nullptr;
904   const APInt *CINT = nullptr;
905   if (!match(opr1, m_APFloatAllowUndef(CF)))
906     match(opr1, m_APIntAllowUndef(CINT));
907 
908   // 0x1111111 means that we don't do anything for this call.
909   int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
910 
911   if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0)) {
912     //  pow/powr/pown(x, 0) == 1
913     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1\n");
914     Constant *cnval = ConstantFP::get(eltType, 1.0);
915     if (getVecSize(FInfo) > 1) {
916       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
917     }
918     replaceCall(FPOp, cnval);
919     return true;
920   }
921   if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
922     // pow/powr/pown(x, 1.0) = x
923     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n");
924     replaceCall(FPOp, opr0);
925     return true;
926   }
927   if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
928     // pow/powr/pown(x, 2.0) = x*x
929     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << " * "
930                       << *opr0 << "\n");
931     Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
932     replaceCall(FPOp, nval);
933     return true;
934   }
935   if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
936     // pow/powr/pown(x, -1.0) = 1.0/x
937     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1 / " << *opr0 << "\n");
938     Constant *cnval = ConstantFP::get(eltType, 1.0);
939     if (getVecSize(FInfo) > 1) {
940       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
941     }
942     Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
943     replaceCall(FPOp, nval);
944     return true;
945   }
946 
947   if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
948     // pow[r](x, [-]0.5) = sqrt(x)
949     bool issqrt = CF->isExactlyValue(0.5);
950     if (FunctionCallee FPExpr =
951             getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
952                                                 : AMDGPULibFunc::EI_RSQRT,
953                                          FInfo))) {
954       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << FInfo.getName()
955                         << '(' << *opr0 << ")\n");
956       Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
957                                                         : "__pow2rsqrt");
958       replaceCall(FPOp, nval);
959       return true;
960     }
961   }
962 
963   if (!isUnsafeFiniteOnlyMath(FPOp))
964     return false;
965 
966   // Unsafe Math optimization
967 
968   // Remember that ci_opr1 is set if opr1 is integral
969   if (CF) {
970     double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
971                       ? (double)CF->convertToFloat()
972                       : CF->convertToDouble();
973     int ival = (int)dval;
974     if ((double)ival == dval) {
975       ci_opr1 = ival;
976     } else
977       ci_opr1 = 0x11111111;
978   }
979 
980   // pow/powr/pown(x, c) = [1/](x*x*..x); where
981   //   trunc(c) == c && the number of x == c && |c| <= 12
982   unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
983   if (abs_opr1 <= 12) {
984     Constant *cnval;
985     Value *nval;
986     if (abs_opr1 == 0) {
987       cnval = ConstantFP::get(eltType, 1.0);
988       if (getVecSize(FInfo) > 1) {
989         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
990       }
991       nval = cnval;
992     } else {
993       Value *valx2 = nullptr;
994       nval = nullptr;
995       while (abs_opr1 > 0) {
996         valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
997         if (abs_opr1 & 1) {
998           nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
999         }
1000         abs_opr1 >>= 1;
1001       }
1002     }
1003 
1004     if (ci_opr1 < 0) {
1005       cnval = ConstantFP::get(eltType, 1.0);
1006       if (getVecSize(FInfo) > 1) {
1007         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
1008       }
1009       nval = B.CreateFDiv(cnval, nval, "__1powprod");
1010     }
1011     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1012                       << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
1013                       << ")\n");
1014     replaceCall(FPOp, nval);
1015     return true;
1016   }
1017 
1018   // If we should use the generic intrinsic instead of emitting a libcall
1019   const bool ShouldUseIntrinsic = eltType->isFloatTy() || eltType->isHalfTy();
1020 
1021   // powr ---> exp2(y * log2(x))
1022   // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
1023   FunctionCallee ExpExpr;
1024   if (ShouldUseIntrinsic)
1025     ExpExpr = Intrinsic::getDeclaration(M, Intrinsic::exp2, {FPOp->getType()});
1026   else {
1027     ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo));
1028     if (!ExpExpr)
1029       return false;
1030   }
1031 
1032   bool needlog = false;
1033   bool needabs = false;
1034   bool needcopysign = false;
1035   Constant *cnval = nullptr;
1036   if (getVecSize(FInfo) == 1) {
1037     CF = nullptr;
1038     match(opr0, m_APFloatAllowUndef(CF));
1039 
1040     if (CF) {
1041       double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1042                      ? (double)CF->convertToFloat()
1043                      : CF->convertToDouble();
1044 
1045       V = log2(std::abs(V));
1046       cnval = ConstantFP::get(eltType, V);
1047       needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
1048                      CF->isNegative();
1049     } else {
1050       needlog = true;
1051       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1052     }
1053   } else {
1054     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
1055 
1056     if (!CDV) {
1057       needlog = true;
1058       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1059     } else {
1060       assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
1061               "Wrong vector size detected");
1062 
1063       SmallVector<double, 0> DVal;
1064       for (int i=0; i < getVecSize(FInfo); ++i) {
1065         double V = CDV->getElementAsAPFloat(i).convertToDouble();
1066         if (V < 0.0) needcopysign = true;
1067         V = log2(std::abs(V));
1068         DVal.push_back(V);
1069       }
1070       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1071         SmallVector<float, 0> FVal;
1072         for (unsigned i=0; i < DVal.size(); ++i) {
1073           FVal.push_back((float)DVal[i]);
1074         }
1075         ArrayRef<float> tmp(FVal);
1076         cnval = ConstantDataVector::get(M->getContext(), tmp);
1077       } else {
1078         ArrayRef<double> tmp(DVal);
1079         cnval = ConstantDataVector::get(M->getContext(), tmp);
1080       }
1081     }
1082   }
1083 
1084   if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1085     // We cannot handle corner cases for a general pow() function, give up
1086     // unless y is a constant integral value. Then proceed as if it were pown.
1087     if (!isKnownIntegral(opr1, M->getDataLayout(), FPOp->getFastMathFlags()))
1088       return false;
1089   }
1090 
1091   Value *nval;
1092   if (needabs) {
1093     nval = B.CreateUnaryIntrinsic(Intrinsic::fabs, opr0, nullptr, "__fabs");
1094   } else {
1095     nval = cnval ? cnval : opr0;
1096   }
1097   if (needlog) {
1098     FunctionCallee LogExpr;
1099     if (ShouldUseIntrinsic) {
1100       LogExpr =
1101           Intrinsic::getDeclaration(M, Intrinsic::log2, {FPOp->getType()});
1102     } else {
1103       LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo));
1104       if (!LogExpr)
1105         return false;
1106     }
1107 
1108     nval = CreateCallEx(B,LogExpr, nval, "__log2");
1109   }
1110 
1111   if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1112     // convert int(32) to fp(f32 or f64)
1113     opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1114   }
1115   nval = B.CreateFMul(opr1, nval, "__ylogx");
1116   nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1117 
1118   if (needcopysign) {
1119     Value *opr_n;
1120     Type* rTy = opr0->getType();
1121     Type* nTyS = B.getIntNTy(eltType->getPrimitiveSizeInBits());
1122     Type *nTy = nTyS;
1123     if (const auto *vTy = dyn_cast<FixedVectorType>(rTy))
1124       nTy = FixedVectorType::get(nTyS, vTy);
1125     unsigned size = nTy->getScalarSizeInBits();
1126     opr_n = FPOp->getOperand(1);
1127     if (opr_n->getType()->isIntegerTy())
1128       opr_n = B.CreateZExtOrTrunc(opr_n, nTy, "__ytou");
1129     else
1130       opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1131 
1132     Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1133     sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1134     nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1135     nval = B.CreateBitCast(nval, opr0->getType());
1136   }
1137 
1138   LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1139                     << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1140   replaceCall(FPOp, nval);
1141 
1142   return true;
1143 }
1144 
1145 bool AMDGPULibCalls::fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B,
1146                                 const FuncInfo &FInfo) {
1147   // skip vector function
1148   if (getVecSize(FInfo) != 1)
1149     return false;
1150 
1151   Value *opr0 = FPOp->getOperand(0);
1152   Value *opr1 = FPOp->getOperand(1);
1153 
1154   ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1155   if (!CINT) {
1156     return false;
1157   }
1158   int ci_opr1 = (int)CINT->getSExtValue();
1159   if (ci_opr1 == 1) {  // rootn(x, 1) = x
1160     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n");
1161     replaceCall(FPOp, opr0);
1162     return true;
1163   }
1164 
1165   Module *M = B.GetInsertBlock()->getModule();
1166   if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x)
1167     if (FunctionCallee FPExpr =
1168             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1169       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> sqrt(" << *opr0
1170                         << ")\n");
1171       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1172       replaceCall(FPOp, nval);
1173       return true;
1174     }
1175   } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1176     if (FunctionCallee FPExpr =
1177             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1178       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> cbrt(" << *opr0
1179                         << ")\n");
1180       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1181       replaceCall(FPOp, nval);
1182       return true;
1183     }
1184   } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1185     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1.0 / " << *opr0 << "\n");
1186     Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1187                                opr0,
1188                                "__rootn2div");
1189     replaceCall(FPOp, nval);
1190     return true;
1191   } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x)
1192     if (FunctionCallee FPExpr =
1193             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) {
1194       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> rsqrt(" << *opr0
1195                         << ")\n");
1196       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1197       replaceCall(FPOp, nval);
1198       return true;
1199     }
1200   }
1201   return false;
1202 }
1203 
1204 // Get a scalar native builtin single argument FP function
1205 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1206                                                  const FuncInfo &FInfo) {
1207   if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1208     return nullptr;
1209   FuncInfo nf = FInfo;
1210   nf.setPrefix(AMDGPULibFunc::NATIVE);
1211   return getFunction(M, nf);
1212 }
1213 
1214 // Some library calls are just wrappers around llvm intrinsics, but compiled
1215 // conservatively. Preserve the flags from the original call site by
1216 // substituting them with direct calls with all the flags.
1217 bool AMDGPULibCalls::shouldReplaceLibcallWithIntrinsic(const CallInst *CI,
1218                                                        bool AllowMinSizeF32,
1219                                                        bool AllowF64,
1220                                                        bool AllowStrictFP) {
1221   Type *FltTy = CI->getType()->getScalarType();
1222   const bool IsF32 = FltTy->isFloatTy();
1223 
1224   // f64 intrinsics aren't implemented for most operations.
1225   if (!IsF32 && !FltTy->isHalfTy() && (!AllowF64 || !FltTy->isDoubleTy()))
1226     return false;
1227 
1228   // We're implicitly inlining by replacing the libcall with the intrinsic, so
1229   // don't do it for noinline call sites.
1230   if (CI->isNoInline())
1231     return false;
1232 
1233   const Function *ParentF = CI->getFunction();
1234   // TODO: Handle strictfp
1235   if (!AllowStrictFP && ParentF->hasFnAttribute(Attribute::StrictFP))
1236     return false;
1237 
1238   if (IsF32 && !AllowMinSizeF32 && ParentF->hasMinSize())
1239     return false;
1240   return true;
1241 }
1242 
1243 void AMDGPULibCalls::replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B,
1244                                                        CallInst *CI,
1245                                                        Intrinsic::ID IntrID) {
1246   if (CI->arg_size() == 2) {
1247     Value *Arg0 = CI->getArgOperand(0);
1248     Value *Arg1 = CI->getArgOperand(1);
1249     VectorType *Arg0VecTy = dyn_cast<VectorType>(Arg0->getType());
1250     VectorType *Arg1VecTy = dyn_cast<VectorType>(Arg1->getType());
1251     if (Arg0VecTy && !Arg1VecTy) {
1252       Value *SplatRHS = B.CreateVectorSplat(Arg0VecTy->getElementCount(), Arg1);
1253       CI->setArgOperand(1, SplatRHS);
1254     } else if (!Arg0VecTy && Arg1VecTy) {
1255       Value *SplatLHS = B.CreateVectorSplat(Arg1VecTy->getElementCount(), Arg0);
1256       CI->setArgOperand(0, SplatLHS);
1257     }
1258   }
1259 
1260   CI->setCalledFunction(
1261       Intrinsic::getDeclaration(CI->getModule(), IntrID, {CI->getType()}));
1262 }
1263 
1264 bool AMDGPULibCalls::tryReplaceLibcallWithSimpleIntrinsic(
1265     IRBuilder<> &B, CallInst *CI, Intrinsic::ID IntrID, bool AllowMinSizeF32,
1266     bool AllowF64, bool AllowStrictFP) {
1267   if (!shouldReplaceLibcallWithIntrinsic(CI, AllowMinSizeF32, AllowF64,
1268                                          AllowStrictFP))
1269     return false;
1270   replaceLibCallWithSimpleIntrinsic(B, CI, IntrID);
1271   return true;
1272 }
1273 
1274 std::tuple<Value *, Value *, Value *>
1275 AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B,
1276                              FunctionCallee Fsincos) {
1277   DebugLoc DL = B.getCurrentDebugLocation();
1278   Function *F = B.GetInsertBlock()->getParent();
1279   B.SetInsertPointPastAllocas(F);
1280 
1281   AllocaInst *Alloc = B.CreateAlloca(Arg->getType(), nullptr, "__sincos_");
1282 
1283   if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1284     // If the argument is an instruction, it must dominate all uses so put our
1285     // sincos call there. Otherwise, right after the allocas works well enough
1286     // if it's an argument or constant.
1287 
1288     B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1289 
1290     // SetInsertPoint unwelcomely always tries to set the debug loc.
1291     B.SetCurrentDebugLocation(DL);
1292   }
1293 
1294   Type *CosPtrTy = Fsincos.getFunctionType()->getParamType(1);
1295 
1296   // The allocaInst allocates the memory in private address space. This need
1297   // to be addrspacecasted to point to the address space of cos pointer type.
1298   // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1299   Value *CastAlloc = B.CreateAddrSpaceCast(Alloc, CosPtrTy);
1300 
1301   CallInst *SinCos = CreateCallEx2(B, Fsincos, Arg, CastAlloc);
1302 
1303   // TODO: Is it worth trying to preserve the location for the cos calls for the
1304   // load?
1305 
1306   LoadInst *LoadCos = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1307   return {SinCos, LoadCos, SinCos};
1308 }
1309 
1310 // fold sin, cos -> sincos.
1311 bool AMDGPULibCalls::fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B,
1312                                  const FuncInfo &fInfo) {
1313   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1314          fInfo.getId() == AMDGPULibFunc::EI_COS);
1315 
1316   if ((getArgType(fInfo) != AMDGPULibFunc::F32 &&
1317        getArgType(fInfo) != AMDGPULibFunc::F64) ||
1318       fInfo.getPrefix() != AMDGPULibFunc::NOPFX)
1319     return false;
1320 
1321   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1322 
1323   Value *CArgVal = FPOp->getOperand(0);
1324   CallInst *CI = cast<CallInst>(FPOp);
1325 
1326   Function *F = B.GetInsertBlock()->getParent();
1327   Module *M = F->getParent();
1328 
1329   // Merge the sin and cos. For OpenCL 2.0, there may only be a generic pointer
1330   // implementation. Prefer the private form if available.
1331   AMDGPULibFunc SinCosLibFuncPrivate(AMDGPULibFunc::EI_SINCOS, fInfo);
1332   SinCosLibFuncPrivate.getLeads()[0].PtrKind =
1333       AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::PRIVATE_ADDRESS);
1334 
1335   AMDGPULibFunc SinCosLibFuncGeneric(AMDGPULibFunc::EI_SINCOS, fInfo);
1336   SinCosLibFuncGeneric.getLeads()[0].PtrKind =
1337       AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1338 
1339   FunctionCallee FSinCosPrivate = getFunction(M, SinCosLibFuncPrivate);
1340   FunctionCallee FSinCosGeneric = getFunction(M, SinCosLibFuncGeneric);
1341   FunctionCallee FSinCos = FSinCosPrivate ? FSinCosPrivate : FSinCosGeneric;
1342   if (!FSinCos)
1343     return false;
1344 
1345   SmallVector<CallInst *> SinCalls;
1346   SmallVector<CallInst *> CosCalls;
1347   SmallVector<CallInst *> SinCosCalls;
1348   FuncInfo PartnerInfo(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN,
1349                        fInfo);
1350   const std::string PairName = PartnerInfo.mangle();
1351 
1352   StringRef SinName = isSin ? CI->getCalledFunction()->getName() : PairName;
1353   StringRef CosName = isSin ? PairName : CI->getCalledFunction()->getName();
1354   const std::string SinCosPrivateName = SinCosLibFuncPrivate.mangle();
1355   const std::string SinCosGenericName = SinCosLibFuncGeneric.mangle();
1356 
1357   // Intersect the two sets of flags.
1358   FastMathFlags FMF = FPOp->getFastMathFlags();
1359   MDNode *FPMath = CI->getMetadata(LLVMContext::MD_fpmath);
1360 
1361   SmallVector<DILocation *> MergeDbgLocs = {CI->getDebugLoc()};
1362 
1363   for (User* U : CArgVal->users()) {
1364     CallInst *XI = dyn_cast<CallInst>(U);
1365     if (!XI || XI->getFunction() != F || XI->isNoBuiltin())
1366       continue;
1367 
1368     Function *UCallee = XI->getCalledFunction();
1369     if (!UCallee)
1370       continue;
1371 
1372     bool Handled = true;
1373 
1374     if (UCallee->getName() == SinName)
1375       SinCalls.push_back(XI);
1376     else if (UCallee->getName() == CosName)
1377       CosCalls.push_back(XI);
1378     else if (UCallee->getName() == SinCosPrivateName ||
1379              UCallee->getName() == SinCosGenericName)
1380       SinCosCalls.push_back(XI);
1381     else
1382       Handled = false;
1383 
1384     if (Handled) {
1385       MergeDbgLocs.push_back(XI->getDebugLoc());
1386       auto *OtherOp = cast<FPMathOperator>(XI);
1387       FMF &= OtherOp->getFastMathFlags();
1388       FPMath = MDNode::getMostGenericFPMath(
1389           FPMath, XI->getMetadata(LLVMContext::MD_fpmath));
1390     }
1391   }
1392 
1393   if (SinCalls.empty() || CosCalls.empty())
1394     return false;
1395 
1396   B.setFastMathFlags(FMF);
1397   B.setDefaultFPMathTag(FPMath);
1398   DILocation *DbgLoc = DILocation::getMergedLocations(MergeDbgLocs);
1399   B.SetCurrentDebugLocation(DbgLoc);
1400 
1401   auto [Sin, Cos, SinCos] = insertSinCos(CArgVal, FMF, B, FSinCos);
1402 
1403   auto replaceTrigInsts = [](ArrayRef<CallInst *> Calls, Value *Res) {
1404     for (CallInst *C : Calls)
1405       C->replaceAllUsesWith(Res);
1406 
1407     // Leave the other dead instructions to avoid clobbering iterators.
1408   };
1409 
1410   replaceTrigInsts(SinCalls, Sin);
1411   replaceTrigInsts(CosCalls, Cos);
1412   replaceTrigInsts(SinCosCalls, SinCos);
1413 
1414   // It's safe to delete the original now.
1415   CI->eraseFromParent();
1416   return true;
1417 }
1418 
1419 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0,
1420                                             double &Res1, Constant *copr0,
1421                                             Constant *copr1) {
1422   // By default, opr0/opr1/opr3 holds values of float/double type.
1423   // If they are not float/double, each function has to its
1424   // operand separately.
1425   double opr0 = 0.0, opr1 = 0.0;
1426   ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1427   ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1428   if (fpopr0) {
1429     opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1430              ? fpopr0->getValueAPF().convertToDouble()
1431              : (double)fpopr0->getValueAPF().convertToFloat();
1432   }
1433 
1434   if (fpopr1) {
1435     opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1436              ? fpopr1->getValueAPF().convertToDouble()
1437              : (double)fpopr1->getValueAPF().convertToFloat();
1438   }
1439 
1440   switch (FInfo.getId()) {
1441   default : return false;
1442 
1443   case AMDGPULibFunc::EI_ACOS:
1444     Res0 = acos(opr0);
1445     return true;
1446 
1447   case AMDGPULibFunc::EI_ACOSH:
1448     // acosh(x) == log(x + sqrt(x*x - 1))
1449     Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1450     return true;
1451 
1452   case AMDGPULibFunc::EI_ACOSPI:
1453     Res0 = acos(opr0) / MATH_PI;
1454     return true;
1455 
1456   case AMDGPULibFunc::EI_ASIN:
1457     Res0 = asin(opr0);
1458     return true;
1459 
1460   case AMDGPULibFunc::EI_ASINH:
1461     // asinh(x) == log(x + sqrt(x*x + 1))
1462     Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1463     return true;
1464 
1465   case AMDGPULibFunc::EI_ASINPI:
1466     Res0 = asin(opr0) / MATH_PI;
1467     return true;
1468 
1469   case AMDGPULibFunc::EI_ATAN:
1470     Res0 = atan(opr0);
1471     return true;
1472 
1473   case AMDGPULibFunc::EI_ATANH:
1474     // atanh(x) == (log(x+1) - log(x-1))/2;
1475     Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1476     return true;
1477 
1478   case AMDGPULibFunc::EI_ATANPI:
1479     Res0 = atan(opr0) / MATH_PI;
1480     return true;
1481 
1482   case AMDGPULibFunc::EI_CBRT:
1483     Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1484     return true;
1485 
1486   case AMDGPULibFunc::EI_COS:
1487     Res0 = cos(opr0);
1488     return true;
1489 
1490   case AMDGPULibFunc::EI_COSH:
1491     Res0 = cosh(opr0);
1492     return true;
1493 
1494   case AMDGPULibFunc::EI_COSPI:
1495     Res0 = cos(MATH_PI * opr0);
1496     return true;
1497 
1498   case AMDGPULibFunc::EI_EXP:
1499     Res0 = exp(opr0);
1500     return true;
1501 
1502   case AMDGPULibFunc::EI_EXP2:
1503     Res0 = pow(2.0, opr0);
1504     return true;
1505 
1506   case AMDGPULibFunc::EI_EXP10:
1507     Res0 = pow(10.0, opr0);
1508     return true;
1509 
1510   case AMDGPULibFunc::EI_LOG:
1511     Res0 = log(opr0);
1512     return true;
1513 
1514   case AMDGPULibFunc::EI_LOG2:
1515     Res0 = log(opr0) / log(2.0);
1516     return true;
1517 
1518   case AMDGPULibFunc::EI_LOG10:
1519     Res0 = log(opr0) / log(10.0);
1520     return true;
1521 
1522   case AMDGPULibFunc::EI_RSQRT:
1523     Res0 = 1.0 / sqrt(opr0);
1524     return true;
1525 
1526   case AMDGPULibFunc::EI_SIN:
1527     Res0 = sin(opr0);
1528     return true;
1529 
1530   case AMDGPULibFunc::EI_SINH:
1531     Res0 = sinh(opr0);
1532     return true;
1533 
1534   case AMDGPULibFunc::EI_SINPI:
1535     Res0 = sin(MATH_PI * opr0);
1536     return true;
1537 
1538   case AMDGPULibFunc::EI_TAN:
1539     Res0 = tan(opr0);
1540     return true;
1541 
1542   case AMDGPULibFunc::EI_TANH:
1543     Res0 = tanh(opr0);
1544     return true;
1545 
1546   case AMDGPULibFunc::EI_TANPI:
1547     Res0 = tan(MATH_PI * opr0);
1548     return true;
1549 
1550   // two-arg functions
1551   case AMDGPULibFunc::EI_POW:
1552   case AMDGPULibFunc::EI_POWR:
1553     Res0 = pow(opr0, opr1);
1554     return true;
1555 
1556   case AMDGPULibFunc::EI_POWN: {
1557     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1558       double val = (double)iopr1->getSExtValue();
1559       Res0 = pow(opr0, val);
1560       return true;
1561     }
1562     return false;
1563   }
1564 
1565   case AMDGPULibFunc::EI_ROOTN: {
1566     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1567       double val = (double)iopr1->getSExtValue();
1568       Res0 = pow(opr0, 1.0 / val);
1569       return true;
1570     }
1571     return false;
1572   }
1573 
1574   // with ptr arg
1575   case AMDGPULibFunc::EI_SINCOS:
1576     Res0 = sin(opr0);
1577     Res1 = cos(opr0);
1578     return true;
1579   }
1580 
1581   return false;
1582 }
1583 
1584 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) {
1585   int numArgs = (int)aCI->arg_size();
1586   if (numArgs > 3)
1587     return false;
1588 
1589   Constant *copr0 = nullptr;
1590   Constant *copr1 = nullptr;
1591   if (numArgs > 0) {
1592     if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1593       return false;
1594   }
1595 
1596   if (numArgs > 1) {
1597     if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1598       if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1599         return false;
1600     }
1601   }
1602 
1603   // At this point, all arguments to aCI are constants.
1604 
1605   // max vector size is 16, and sincos will generate two results.
1606   double DVal0[16], DVal1[16];
1607   int FuncVecSize = getVecSize(FInfo);
1608   bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1609   if (FuncVecSize == 1) {
1610     if (!evaluateScalarMathFunc(FInfo, DVal0[0], DVal1[0], copr0, copr1)) {
1611       return false;
1612     }
1613   } else {
1614     ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1615     ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1616     for (int i = 0; i < FuncVecSize; ++i) {
1617       Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1618       Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1619       if (!evaluateScalarMathFunc(FInfo, DVal0[i], DVal1[i], celt0, celt1)) {
1620         return false;
1621       }
1622     }
1623   }
1624 
1625   LLVMContext &context = aCI->getContext();
1626   Constant *nval0, *nval1;
1627   if (FuncVecSize == 1) {
1628     nval0 = ConstantFP::get(aCI->getType(), DVal0[0]);
1629     if (hasTwoResults)
1630       nval1 = ConstantFP::get(aCI->getType(), DVal1[0]);
1631   } else {
1632     if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1633       SmallVector <float, 0> FVal0, FVal1;
1634       for (int i = 0; i < FuncVecSize; ++i)
1635         FVal0.push_back((float)DVal0[i]);
1636       ArrayRef<float> tmp0(FVal0);
1637       nval0 = ConstantDataVector::get(context, tmp0);
1638       if (hasTwoResults) {
1639         for (int i = 0; i < FuncVecSize; ++i)
1640           FVal1.push_back((float)DVal1[i]);
1641         ArrayRef<float> tmp1(FVal1);
1642         nval1 = ConstantDataVector::get(context, tmp1);
1643       }
1644     } else {
1645       ArrayRef<double> tmp0(DVal0);
1646       nval0 = ConstantDataVector::get(context, tmp0);
1647       if (hasTwoResults) {
1648         ArrayRef<double> tmp1(DVal1);
1649         nval1 = ConstantDataVector::get(context, tmp1);
1650       }
1651     }
1652   }
1653 
1654   if (hasTwoResults) {
1655     // sincos
1656     assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1657            "math function with ptr arg not supported yet");
1658     new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1659   }
1660 
1661   replaceCall(aCI, nval0);
1662   return true;
1663 }
1664 
1665 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F,
1666                                                   FunctionAnalysisManager &AM) {
1667   AMDGPULibCalls Simplifier;
1668   Simplifier.initNativeFuncs();
1669   Simplifier.initFunction(F, AM);
1670 
1671   bool Changed = false;
1672 
1673   LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1674              F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1675 
1676   for (auto &BB : F) {
1677     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1678       // Ignore non-calls.
1679       CallInst *CI = dyn_cast<CallInst>(I);
1680       ++I;
1681 
1682       if (CI) {
1683         if (Simplifier.fold(CI))
1684           Changed = true;
1685       }
1686     }
1687   }
1688   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1689 }
1690 
1691 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F,
1692                                                 FunctionAnalysisManager &AM) {
1693   if (UseNative.empty())
1694     return PreservedAnalyses::all();
1695 
1696   AMDGPULibCalls Simplifier;
1697   Simplifier.initNativeFuncs();
1698   Simplifier.initFunction(F, AM);
1699 
1700   bool Changed = false;
1701   for (auto &BB : F) {
1702     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1703       // Ignore non-calls.
1704       CallInst *CI = dyn_cast<CallInst>(I);
1705       ++I;
1706       if (CI && Simplifier.useNative(CI))
1707         Changed = true;
1708     }
1709   }
1710   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1711 }
1712