xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp (revision 770cf0a5f02dc8983a89c6568d741fbc25baa999)
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 "llvm/Analysis/AssumptionCache.h"
17 #include "llvm/Analysis/TargetLibraryInfo.h"
18 #include "llvm/Analysis/ValueTracking.h"
19 #include "llvm/IR/AttributeMask.h"
20 #include "llvm/IR/Dominators.h"
21 #include "llvm/IR/IRBuilder.h"
22 #include "llvm/IR/MDBuilder.h"
23 #include "llvm/IR/PatternMatch.h"
24 #include <cmath>
25 
26 #define DEBUG_TYPE "amdgpu-simplifylib"
27 
28 using namespace llvm;
29 using namespace llvm::PatternMatch;
30 
31 static cl::opt<bool> EnablePreLink("amdgpu-prelink",
32   cl::desc("Enable pre-link mode optimizations"),
33   cl::init(false),
34   cl::Hidden);
35 
36 static cl::list<std::string> UseNative("amdgpu-use-native",
37   cl::desc("Comma separated list of functions to replace with native, or all"),
38   cl::CommaSeparated, cl::ValueOptional,
39   cl::Hidden);
40 
41 #define MATH_PI      numbers::pi
42 #define MATH_E       numbers::e
43 #define MATH_SQRT2   numbers::sqrt2
44 #define MATH_SQRT1_2 numbers::inv_sqrt2
45 
46 namespace llvm {
47 
48 class AMDGPULibCalls {
49 private:
50   const TargetLibraryInfo *TLInfo = nullptr;
51   AssumptionCache *AC = nullptr;
52   DominatorTree *DT = nullptr;
53 
54   using FuncInfo = llvm::AMDGPULibFunc;
55 
56   bool UnsafeFPMath = false;
57 
58   // -fuse-native.
59   bool AllNative = false;
60 
61   bool useNativeFunc(const StringRef F) const;
62 
63   // Return a pointer (pointer expr) to the function if function definition with
64   // "FuncName" exists. It may create a new function prototype in pre-link mode.
65   FunctionCallee getFunction(Module *M, const FuncInfo &fInfo);
66 
67   bool parseFunctionName(const StringRef &FMangledName, FuncInfo &FInfo);
68 
69   bool TDOFold(CallInst *CI, const FuncInfo &FInfo);
70 
71   /* Specialized optimizations */
72 
73   // pow/powr/pown
74   bool fold_pow(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
75 
76   // rootn
77   bool fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
78 
79   // -fuse-native for sincos
80   bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo);
81 
82   // evaluate calls if calls' arguments are constants.
83   bool evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0, double &Res1,
84                               Constant *copr0, Constant *copr1);
85   bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo);
86 
87   /// Insert a value to sincos function \p Fsincos. Returns (value of sin, value
88   /// of cos, sincos call).
89   std::tuple<Value *, Value *, Value *> insertSinCos(Value *Arg,
90                                                      FastMathFlags FMF,
91                                                      IRBuilder<> &B,
92                                                      FunctionCallee Fsincos);
93 
94   // sin/cos
95   bool fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
96 
97   // __read_pipe/__write_pipe
98   bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
99                             const FuncInfo &FInfo);
100 
101   // Get a scalar native builtin single argument FP function
102   FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo);
103 
104   /// Substitute a call to a known libcall with an intrinsic call. If \p
105   /// AllowMinSize is true, allow the replacement in a minsize function.
106   bool shouldReplaceLibcallWithIntrinsic(const CallInst *CI,
107                                          bool AllowMinSizeF32 = false,
108                                          bool AllowF64 = false,
109                                          bool AllowStrictFP = false);
110   void replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI,
111                                          Intrinsic::ID IntrID);
112 
113   bool tryReplaceLibcallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI,
114                                             Intrinsic::ID IntrID,
115                                             bool AllowMinSizeF32 = false,
116                                             bool AllowF64 = false,
117                                             bool AllowStrictFP = false);
118 
119 protected:
120   bool isUnsafeMath(const FPMathOperator *FPOp) const;
121   bool isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const;
122 
123   bool canIncreasePrecisionOfConstantFold(const FPMathOperator *FPOp) const;
124 
125   static void replaceCall(Instruction *I, Value *With) {
126     I->replaceAllUsesWith(With);
127     I->eraseFromParent();
128   }
129 
130   static void replaceCall(FPMathOperator *I, Value *With) {
131     replaceCall(cast<Instruction>(I), With);
132   }
133 
134 public:
135   AMDGPULibCalls() = default;
136 
137   bool fold(CallInst *CI);
138 
139   void initFunction(Function &F, FunctionAnalysisManager &FAM);
140   void initNativeFuncs();
141 
142   // Replace a normal math function call with that native version
143   bool useNative(CallInst *CI);
144 };
145 
146 } // end namespace llvm
147 
148 template <typename IRB>
149 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg,
150                               const Twine &Name = "") {
151   CallInst *R = B.CreateCall(Callee, Arg, Name);
152   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
153     R->setCallingConv(F->getCallingConv());
154   return R;
155 }
156 
157 template <typename IRB>
158 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1,
159                                Value *Arg2, const Twine &Name = "") {
160   CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
161   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
162     R->setCallingConv(F->getCallingConv());
163   return R;
164 }
165 
166 static FunctionType *getPownType(FunctionType *FT) {
167   Type *PowNExpTy = Type::getInt32Ty(FT->getContext());
168   if (VectorType *VecTy = dyn_cast<VectorType>(FT->getReturnType()))
169     PowNExpTy = VectorType::get(PowNExpTy, VecTy->getElementCount());
170 
171   return FunctionType::get(FT->getReturnType(),
172                            {FT->getParamType(0), PowNExpTy}, false);
173 }
174 
175 //  Data structures for table-driven optimizations.
176 //  FuncTbl works for both f32 and f64 functions with 1 input argument
177 
178 struct TableEntry {
179   double   result;
180   double   input;
181 };
182 
183 /* a list of {result, input} */
184 static const TableEntry tbl_acos[] = {
185   {MATH_PI / 2.0, 0.0},
186   {MATH_PI / 2.0, -0.0},
187   {0.0, 1.0},
188   {MATH_PI, -1.0}
189 };
190 static const TableEntry tbl_acosh[] = {
191   {0.0, 1.0}
192 };
193 static const TableEntry tbl_acospi[] = {
194   {0.5, 0.0},
195   {0.5, -0.0},
196   {0.0, 1.0},
197   {1.0, -1.0}
198 };
199 static const TableEntry tbl_asin[] = {
200   {0.0, 0.0},
201   {-0.0, -0.0},
202   {MATH_PI / 2.0, 1.0},
203   {-MATH_PI / 2.0, -1.0}
204 };
205 static const TableEntry tbl_asinh[] = {
206   {0.0, 0.0},
207   {-0.0, -0.0}
208 };
209 static const TableEntry tbl_asinpi[] = {
210   {0.0, 0.0},
211   {-0.0, -0.0},
212   {0.5, 1.0},
213   {-0.5, -1.0}
214 };
215 static const TableEntry tbl_atan[] = {
216   {0.0, 0.0},
217   {-0.0, -0.0},
218   {MATH_PI / 4.0, 1.0},
219   {-MATH_PI / 4.0, -1.0}
220 };
221 static const TableEntry tbl_atanh[] = {
222   {0.0, 0.0},
223   {-0.0, -0.0}
224 };
225 static const TableEntry tbl_atanpi[] = {
226   {0.0, 0.0},
227   {-0.0, -0.0},
228   {0.25, 1.0},
229   {-0.25, -1.0}
230 };
231 static const TableEntry tbl_cbrt[] = {
232   {0.0, 0.0},
233   {-0.0, -0.0},
234   {1.0, 1.0},
235   {-1.0, -1.0},
236 };
237 static const TableEntry tbl_cos[] = {
238   {1.0, 0.0},
239   {1.0, -0.0}
240 };
241 static const TableEntry tbl_cosh[] = {
242   {1.0, 0.0},
243   {1.0, -0.0}
244 };
245 static const TableEntry tbl_cospi[] = {
246   {1.0, 0.0},
247   {1.0, -0.0}
248 };
249 static const TableEntry tbl_erfc[] = {
250   {1.0, 0.0},
251   {1.0, -0.0}
252 };
253 static const TableEntry tbl_erf[] = {
254   {0.0, 0.0},
255   {-0.0, -0.0}
256 };
257 static const TableEntry tbl_exp[] = {
258   {1.0, 0.0},
259   {1.0, -0.0},
260   {MATH_E, 1.0}
261 };
262 static const TableEntry tbl_exp2[] = {
263   {1.0, 0.0},
264   {1.0, -0.0},
265   {2.0, 1.0}
266 };
267 static const TableEntry tbl_exp10[] = {
268   {1.0, 0.0},
269   {1.0, -0.0},
270   {10.0, 1.0}
271 };
272 static const TableEntry tbl_expm1[] = {
273   {0.0, 0.0},
274   {-0.0, -0.0}
275 };
276 static const TableEntry tbl_log[] = {
277   {0.0, 1.0},
278   {1.0, MATH_E}
279 };
280 static const TableEntry tbl_log2[] = {
281   {0.0, 1.0},
282   {1.0, 2.0}
283 };
284 static const TableEntry tbl_log10[] = {
285   {0.0, 1.0},
286   {1.0, 10.0}
287 };
288 static const TableEntry tbl_rsqrt[] = {
289   {1.0, 1.0},
290   {MATH_SQRT1_2, 2.0}
291 };
292 static const TableEntry tbl_sin[] = {
293   {0.0, 0.0},
294   {-0.0, -0.0}
295 };
296 static const TableEntry tbl_sinh[] = {
297   {0.0, 0.0},
298   {-0.0, -0.0}
299 };
300 static const TableEntry tbl_sinpi[] = {
301   {0.0, 0.0},
302   {-0.0, -0.0}
303 };
304 static const TableEntry tbl_sqrt[] = {
305   {0.0, 0.0},
306   {1.0, 1.0},
307   {MATH_SQRT2, 2.0}
308 };
309 static const TableEntry tbl_tan[] = {
310   {0.0, 0.0},
311   {-0.0, -0.0}
312 };
313 static const TableEntry tbl_tanh[] = {
314   {0.0, 0.0},
315   {-0.0, -0.0}
316 };
317 static const TableEntry tbl_tanpi[] = {
318   {0.0, 0.0},
319   {-0.0, -0.0}
320 };
321 static const TableEntry tbl_tgamma[] = {
322   {1.0, 1.0},
323   {1.0, 2.0},
324   {2.0, 3.0},
325   {6.0, 4.0}
326 };
327 
328 static bool HasNative(AMDGPULibFunc::EFuncId id) {
329   switch(id) {
330   case AMDGPULibFunc::EI_DIVIDE:
331   case AMDGPULibFunc::EI_COS:
332   case AMDGPULibFunc::EI_EXP:
333   case AMDGPULibFunc::EI_EXP2:
334   case AMDGPULibFunc::EI_EXP10:
335   case AMDGPULibFunc::EI_LOG:
336   case AMDGPULibFunc::EI_LOG2:
337   case AMDGPULibFunc::EI_LOG10:
338   case AMDGPULibFunc::EI_POWR:
339   case AMDGPULibFunc::EI_RECIP:
340   case AMDGPULibFunc::EI_RSQRT:
341   case AMDGPULibFunc::EI_SIN:
342   case AMDGPULibFunc::EI_SINCOS:
343   case AMDGPULibFunc::EI_SQRT:
344   case AMDGPULibFunc::EI_TAN:
345     return true;
346   default:;
347   }
348   return false;
349 }
350 
351 using TableRef = ArrayRef<TableEntry>;
352 
353 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
354   switch(id) {
355   case AMDGPULibFunc::EI_ACOS:    return TableRef(tbl_acos);
356   case AMDGPULibFunc::EI_ACOSH:   return TableRef(tbl_acosh);
357   case AMDGPULibFunc::EI_ACOSPI:  return TableRef(tbl_acospi);
358   case AMDGPULibFunc::EI_ASIN:    return TableRef(tbl_asin);
359   case AMDGPULibFunc::EI_ASINH:   return TableRef(tbl_asinh);
360   case AMDGPULibFunc::EI_ASINPI:  return TableRef(tbl_asinpi);
361   case AMDGPULibFunc::EI_ATAN:    return TableRef(tbl_atan);
362   case AMDGPULibFunc::EI_ATANH:   return TableRef(tbl_atanh);
363   case AMDGPULibFunc::EI_ATANPI:  return TableRef(tbl_atanpi);
364   case AMDGPULibFunc::EI_CBRT:    return TableRef(tbl_cbrt);
365   case AMDGPULibFunc::EI_NCOS:
366   case AMDGPULibFunc::EI_COS:     return TableRef(tbl_cos);
367   case AMDGPULibFunc::EI_COSH:    return TableRef(tbl_cosh);
368   case AMDGPULibFunc::EI_COSPI:   return TableRef(tbl_cospi);
369   case AMDGPULibFunc::EI_ERFC:    return TableRef(tbl_erfc);
370   case AMDGPULibFunc::EI_ERF:     return TableRef(tbl_erf);
371   case AMDGPULibFunc::EI_EXP:     return TableRef(tbl_exp);
372   case AMDGPULibFunc::EI_NEXP2:
373   case AMDGPULibFunc::EI_EXP2:    return TableRef(tbl_exp2);
374   case AMDGPULibFunc::EI_EXP10:   return TableRef(tbl_exp10);
375   case AMDGPULibFunc::EI_EXPM1:   return TableRef(tbl_expm1);
376   case AMDGPULibFunc::EI_LOG:     return TableRef(tbl_log);
377   case AMDGPULibFunc::EI_NLOG2:
378   case AMDGPULibFunc::EI_LOG2:    return TableRef(tbl_log2);
379   case AMDGPULibFunc::EI_LOG10:   return TableRef(tbl_log10);
380   case AMDGPULibFunc::EI_NRSQRT:
381   case AMDGPULibFunc::EI_RSQRT:   return TableRef(tbl_rsqrt);
382   case AMDGPULibFunc::EI_NSIN:
383   case AMDGPULibFunc::EI_SIN:     return TableRef(tbl_sin);
384   case AMDGPULibFunc::EI_SINH:    return TableRef(tbl_sinh);
385   case AMDGPULibFunc::EI_SINPI:   return TableRef(tbl_sinpi);
386   case AMDGPULibFunc::EI_NSQRT:
387   case AMDGPULibFunc::EI_SQRT:    return TableRef(tbl_sqrt);
388   case AMDGPULibFunc::EI_TAN:     return TableRef(tbl_tan);
389   case AMDGPULibFunc::EI_TANH:    return TableRef(tbl_tanh);
390   case AMDGPULibFunc::EI_TANPI:   return TableRef(tbl_tanpi);
391   case AMDGPULibFunc::EI_TGAMMA:  return TableRef(tbl_tgamma);
392   default:;
393   }
394   return TableRef();
395 }
396 
397 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
398   return FInfo.getLeads()[0].VectorSize;
399 }
400 
401 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
402   return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
403 }
404 
405 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) {
406   // If we are doing PreLinkOpt, the function is external. So it is safe to
407   // use getOrInsertFunction() at this stage.
408 
409   return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
410                        : AMDGPULibFunc::getFunction(M, fInfo);
411 }
412 
413 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName,
414                                        FuncInfo &FInfo) {
415   return AMDGPULibFunc::parse(FMangledName, FInfo);
416 }
417 
418 bool AMDGPULibCalls::isUnsafeMath(const FPMathOperator *FPOp) const {
419   return UnsafeFPMath || FPOp->isFast();
420 }
421 
422 bool AMDGPULibCalls::isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const {
423   return UnsafeFPMath ||
424          (FPOp->hasApproxFunc() && FPOp->hasNoNaNs() && FPOp->hasNoInfs());
425 }
426 
427 bool AMDGPULibCalls::canIncreasePrecisionOfConstantFold(
428     const FPMathOperator *FPOp) const {
429   // TODO: Refine to approxFunc or contract
430   return isUnsafeMath(FPOp);
431 }
432 
433 void AMDGPULibCalls::initFunction(Function &F, FunctionAnalysisManager &FAM) {
434   UnsafeFPMath = F.getFnAttribute("unsafe-fp-math").getValueAsBool();
435   AC = &FAM.getResult<AssumptionAnalysis>(F);
436   TLInfo = &FAM.getResult<TargetLibraryAnalysis>(F);
437   DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
438 }
439 
440 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
441   return AllNative || llvm::is_contained(UseNative, F);
442 }
443 
444 void AMDGPULibCalls::initNativeFuncs() {
445   AllNative = useNativeFunc("all") ||
446               (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
447                UseNative.begin()->empty());
448 }
449 
450 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
451   bool native_sin = useNativeFunc("sin");
452   bool native_cos = useNativeFunc("cos");
453 
454   if (native_sin && native_cos) {
455     Module *M = aCI->getModule();
456     Value *opr0 = aCI->getArgOperand(0);
457 
458     AMDGPULibFunc nf;
459     nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
460     nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
461 
462     nf.setPrefix(AMDGPULibFunc::NATIVE);
463     nf.setId(AMDGPULibFunc::EI_SIN);
464     FunctionCallee sinExpr = getFunction(M, nf);
465 
466     nf.setPrefix(AMDGPULibFunc::NATIVE);
467     nf.setId(AMDGPULibFunc::EI_COS);
468     FunctionCallee cosExpr = getFunction(M, nf);
469     if (sinExpr && cosExpr) {
470       Value *sinval =
471           CallInst::Create(sinExpr, opr0, "splitsin", aCI->getIterator());
472       Value *cosval =
473           CallInst::Create(cosExpr, opr0, "splitcos", aCI->getIterator());
474       new StoreInst(cosval, aCI->getArgOperand(1), aCI->getIterator());
475 
476       DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
477                                           << " with native version of sin/cos");
478 
479       replaceCall(aCI, sinval);
480       return true;
481     }
482   }
483   return false;
484 }
485 
486 bool AMDGPULibCalls::useNative(CallInst *aCI) {
487   Function *Callee = aCI->getCalledFunction();
488   if (!Callee || aCI->isNoBuiltin())
489     return false;
490 
491   FuncInfo FInfo;
492   if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() ||
493       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
494       getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
495       !(AllNative || useNativeFunc(FInfo.getName()))) {
496     return false;
497   }
498 
499   if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
500     return sincosUseNative(aCI, FInfo);
501 
502   FInfo.setPrefix(AMDGPULibFunc::NATIVE);
503   FunctionCallee F = getFunction(aCI->getModule(), FInfo);
504   if (!F)
505     return false;
506 
507   aCI->setCalledFunction(F);
508   DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
509                                       << " with native version");
510   return true;
511 }
512 
513 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
514 // builtin, with appended type size and alignment arguments, where 2 or 4
515 // indicates the original number of arguments. The library has optimized version
516 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
517 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
518 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
519 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
520 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
521                                           const FuncInfo &FInfo) {
522   auto *Callee = CI->getCalledFunction();
523   if (!Callee->isDeclaration())
524     return false;
525 
526   assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
527   auto *M = Callee->getParent();
528   std::string Name = std::string(Callee->getName());
529   auto NumArg = CI->arg_size();
530   if (NumArg != 4 && NumArg != 6)
531     return false;
532   ConstantInt *PacketSize =
533       dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 2));
534   ConstantInt *PacketAlign =
535       dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 1));
536   if (!PacketSize || !PacketAlign)
537     return false;
538 
539   unsigned Size = PacketSize->getZExtValue();
540   Align Alignment = PacketAlign->getAlignValue();
541   if (Alignment != Size)
542     return false;
543 
544   unsigned PtrArgLoc = CI->arg_size() - 3;
545   Value *PtrArg = CI->getArgOperand(PtrArgLoc);
546   Type *PtrTy = PtrArg->getType();
547 
548   SmallVector<llvm::Type *, 6> ArgTys;
549   for (unsigned I = 0; I != PtrArgLoc; ++I)
550     ArgTys.push_back(CI->getArgOperand(I)->getType());
551   ArgTys.push_back(PtrTy);
552 
553   Name = Name + "_" + std::to_string(Size);
554   auto *FTy = FunctionType::get(Callee->getReturnType(),
555                                 ArrayRef<Type *>(ArgTys), false);
556   AMDGPULibFunc NewLibFunc(Name, FTy);
557   FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
558   if (!F)
559     return false;
560 
561   SmallVector<Value *, 6> Args;
562   for (unsigned I = 0; I != PtrArgLoc; ++I)
563     Args.push_back(CI->getArgOperand(I));
564   Args.push_back(PtrArg);
565 
566   auto *NCI = B.CreateCall(F, Args);
567   NCI->setAttributes(CI->getAttributes());
568   CI->replaceAllUsesWith(NCI);
569   CI->dropAllReferences();
570   CI->eraseFromParent();
571 
572   return true;
573 }
574 
575 static bool isKnownIntegral(const Value *V, const DataLayout &DL,
576                             FastMathFlags FMF) {
577   if (isa<PoisonValue>(V))
578     return true;
579   if (isa<UndefValue>(V))
580     return false;
581 
582   if (const ConstantFP *CF = dyn_cast<ConstantFP>(V))
583     return CF->getValueAPF().isInteger();
584 
585   auto *VFVTy = dyn_cast<FixedVectorType>(V->getType());
586   const Constant *CV = dyn_cast<Constant>(V);
587   if (VFVTy && CV) {
588     unsigned NumElts = VFVTy->getNumElements();
589     for (unsigned i = 0; i != NumElts; ++i) {
590       Constant *Elt = CV->getAggregateElement(i);
591       if (!Elt)
592         return false;
593       if (isa<PoisonValue>(Elt))
594         continue;
595 
596       const ConstantFP *CFP = dyn_cast<ConstantFP>(Elt);
597       if (!CFP || !CFP->getValue().isInteger())
598         return false;
599     }
600 
601     return true;
602   }
603 
604   const Instruction *I = dyn_cast<Instruction>(V);
605   if (!I)
606     return false;
607 
608   switch (I->getOpcode()) {
609   case Instruction::SIToFP:
610   case Instruction::UIToFP:
611     // TODO: Could check nofpclass(inf) on incoming argument
612     if (FMF.noInfs())
613       return true;
614 
615     // Need to check int size cannot produce infinity, which computeKnownFPClass
616     // knows how to do already.
617     return isKnownNeverInfinity(I, SimplifyQuery(DL));
618   case Instruction::Call: {
619     const CallInst *CI = cast<CallInst>(I);
620     switch (CI->getIntrinsicID()) {
621     case Intrinsic::trunc:
622     case Intrinsic::floor:
623     case Intrinsic::ceil:
624     case Intrinsic::rint:
625     case Intrinsic::nearbyint:
626     case Intrinsic::round:
627     case Intrinsic::roundeven:
628       return (FMF.noInfs() && FMF.noNaNs()) ||
629              isKnownNeverInfOrNaN(I, SimplifyQuery(DL));
630     default:
631       break;
632     }
633 
634     break;
635   }
636   default:
637     break;
638   }
639 
640   return false;
641 }
642 
643 // This function returns false if no change; return true otherwise.
644 bool AMDGPULibCalls::fold(CallInst *CI) {
645   Function *Callee = CI->getCalledFunction();
646   // Ignore indirect calls.
647   if (!Callee || Callee->isIntrinsic() || CI->isNoBuiltin())
648     return false;
649 
650   FuncInfo FInfo;
651   if (!parseFunctionName(Callee->getName(), FInfo))
652     return false;
653 
654   // Further check the number of arguments to see if they match.
655   // TODO: Check calling convention matches too
656   if (!FInfo.isCompatibleSignature(*Callee->getParent(), CI->getFunctionType()))
657     return false;
658 
659   LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << '\n');
660 
661   if (TDOFold(CI, FInfo))
662     return true;
663 
664   IRBuilder<> B(CI);
665   if (CI->isStrictFP())
666     B.setIsFPConstrained(true);
667 
668   if (FPMathOperator *FPOp = dyn_cast<FPMathOperator>(CI)) {
669     // Under unsafe-math, evaluate calls if possible.
670     // According to Brian Sumner, we can do this for all f32 function calls
671     // using host's double function calls.
672     if (canIncreasePrecisionOfConstantFold(FPOp) && evaluateCall(CI, FInfo))
673       return true;
674 
675     // Copy fast flags from the original call.
676     FastMathFlags FMF = FPOp->getFastMathFlags();
677     B.setFastMathFlags(FMF);
678 
679     // Specialized optimizations for each function call.
680     //
681     // TODO: Handle native functions
682     switch (FInfo.getId()) {
683     case AMDGPULibFunc::EI_EXP:
684       if (FMF.none())
685         return false;
686       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp,
687                                                   FMF.approxFunc());
688     case AMDGPULibFunc::EI_EXP2:
689       if (FMF.none())
690         return false;
691       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp2,
692                                                   FMF.approxFunc());
693     case AMDGPULibFunc::EI_LOG:
694       if (FMF.none())
695         return false;
696       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log,
697                                                   FMF.approxFunc());
698     case AMDGPULibFunc::EI_LOG2:
699       if (FMF.none())
700         return false;
701       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log2,
702                                                   FMF.approxFunc());
703     case AMDGPULibFunc::EI_LOG10:
704       if (FMF.none())
705         return false;
706       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log10,
707                                                   FMF.approxFunc());
708     case AMDGPULibFunc::EI_FMIN:
709       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::minnum,
710                                                   true, true);
711     case AMDGPULibFunc::EI_FMAX:
712       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::maxnum,
713                                                   true, true);
714     case AMDGPULibFunc::EI_FMA:
715       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fma, true,
716                                                   true);
717     case AMDGPULibFunc::EI_MAD:
718       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fmuladd,
719                                                   true, true);
720     case AMDGPULibFunc::EI_FABS:
721       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fabs, true,
722                                                   true, true);
723     case AMDGPULibFunc::EI_COPYSIGN:
724       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::copysign,
725                                                   true, true, true);
726     case AMDGPULibFunc::EI_FLOOR:
727       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::floor, true,
728                                                   true);
729     case AMDGPULibFunc::EI_CEIL:
730       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::ceil, true,
731                                                   true);
732     case AMDGPULibFunc::EI_TRUNC:
733       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::trunc, true,
734                                                   true);
735     case AMDGPULibFunc::EI_RINT:
736       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::rint, true,
737                                                   true);
738     case AMDGPULibFunc::EI_ROUND:
739       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::round, true,
740                                                   true);
741     case AMDGPULibFunc::EI_LDEXP: {
742       if (!shouldReplaceLibcallWithIntrinsic(CI, true, true))
743         return false;
744 
745       Value *Arg1 = CI->getArgOperand(1);
746       if (VectorType *VecTy = dyn_cast<VectorType>(CI->getType());
747           VecTy && !isa<VectorType>(Arg1->getType())) {
748         Value *SplatArg1 = B.CreateVectorSplat(VecTy->getElementCount(), Arg1);
749         CI->setArgOperand(1, SplatArg1);
750       }
751 
752       CI->setCalledFunction(Intrinsic::getOrInsertDeclaration(
753           CI->getModule(), Intrinsic::ldexp,
754           {CI->getType(), CI->getArgOperand(1)->getType()}));
755       return true;
756     }
757     case AMDGPULibFunc::EI_POW: {
758       Module *M = Callee->getParent();
759       AMDGPULibFunc PowrInfo(AMDGPULibFunc::EI_POWR, FInfo);
760       FunctionCallee PowrFunc = getFunction(M, PowrInfo);
761       CallInst *Call = cast<CallInst>(FPOp);
762 
763       // pow(x, y) -> powr(x, y) for x >= -0.0
764       // TODO: Account for flags on current call
765       if (PowrFunc &&
766           cannotBeOrderedLessThanZero(
767               FPOp->getOperand(0),
768               SimplifyQuery(M->getDataLayout(), TLInfo, DT, AC, Call))) {
769         Call->setCalledFunction(PowrFunc);
770         return fold_pow(FPOp, B, PowrInfo) || true;
771       }
772 
773       // pow(x, y) -> pown(x, y) for known integral y
774       if (isKnownIntegral(FPOp->getOperand(1), M->getDataLayout(),
775                           FPOp->getFastMathFlags())) {
776         FunctionType *PownType = getPownType(CI->getFunctionType());
777         AMDGPULibFunc PownInfo(AMDGPULibFunc::EI_POWN, PownType, true);
778         FunctionCallee PownFunc = getFunction(M, PownInfo);
779         if (PownFunc) {
780           // TODO: If the incoming integral value is an sitofp/uitofp, it won't
781           // fold out without a known range. We can probably take the source
782           // value directly.
783           Value *CastedArg =
784               B.CreateFPToSI(FPOp->getOperand(1), PownType->getParamType(1));
785           // Have to drop any nofpclass attributes on the original call site.
786           Call->removeParamAttrs(
787               1, AttributeFuncs::typeIncompatible(CastedArg->getType(),
788                                                   Call->getParamAttributes(1)));
789           Call->setCalledFunction(PownFunc);
790           Call->setArgOperand(1, CastedArg);
791           return fold_pow(FPOp, B, PownInfo) || true;
792         }
793       }
794 
795       return fold_pow(FPOp, B, FInfo);
796     }
797     case AMDGPULibFunc::EI_POWR:
798     case AMDGPULibFunc::EI_POWN:
799       return fold_pow(FPOp, B, FInfo);
800     case AMDGPULibFunc::EI_ROOTN:
801       return fold_rootn(FPOp, B, FInfo);
802     case AMDGPULibFunc::EI_SQRT:
803       // TODO: Allow with strictfp + constrained intrinsic
804       return tryReplaceLibcallWithSimpleIntrinsic(
805           B, CI, Intrinsic::sqrt, true, true, /*AllowStrictFP=*/false);
806     case AMDGPULibFunc::EI_COS:
807     case AMDGPULibFunc::EI_SIN:
808       return fold_sincos(FPOp, B, FInfo);
809     default:
810       break;
811     }
812   } else {
813     // Specialized optimizations for each function call
814     switch (FInfo.getId()) {
815     case AMDGPULibFunc::EI_READ_PIPE_2:
816     case AMDGPULibFunc::EI_READ_PIPE_4:
817     case AMDGPULibFunc::EI_WRITE_PIPE_2:
818     case AMDGPULibFunc::EI_WRITE_PIPE_4:
819       return fold_read_write_pipe(CI, B, FInfo);
820     default:
821       break;
822     }
823   }
824 
825   return false;
826 }
827 
828 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
829   // Table-Driven optimization
830   const TableRef tr = getOptTable(FInfo.getId());
831   if (tr.empty())
832     return false;
833 
834   int const sz = (int)tr.size();
835   Value *opr0 = CI->getArgOperand(0);
836 
837   if (getVecSize(FInfo) > 1) {
838     if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
839       SmallVector<double, 0> DVal;
840       for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
841         ConstantFP *eltval = dyn_cast<ConstantFP>(
842                                CV->getElementAsConstant((unsigned)eltNo));
843         assert(eltval && "Non-FP arguments in math function!");
844         bool found = false;
845         for (int i=0; i < sz; ++i) {
846           if (eltval->isExactlyValue(tr[i].input)) {
847             DVal.push_back(tr[i].result);
848             found = true;
849             break;
850           }
851         }
852         if (!found) {
853           // This vector constants not handled yet.
854           return false;
855         }
856       }
857       LLVMContext &context = CI->getParent()->getParent()->getContext();
858       Constant *nval;
859       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
860         SmallVector<float, 0> FVal;
861         for (double D : DVal)
862           FVal.push_back((float)D);
863         ArrayRef<float> tmp(FVal);
864         nval = ConstantDataVector::get(context, tmp);
865       } else { // F64
866         ArrayRef<double> tmp(DVal);
867         nval = ConstantDataVector::get(context, tmp);
868       }
869       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
870       replaceCall(CI, nval);
871       return true;
872     }
873   } else {
874     // Scalar version
875     if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
876       for (int i = 0; i < sz; ++i) {
877         if (CF->isExactlyValue(tr[i].input)) {
878           Value *nval = ConstantFP::get(CF->getType(), tr[i].result);
879           LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
880           replaceCall(CI, nval);
881           return true;
882         }
883       }
884     }
885   }
886 
887   return false;
888 }
889 
890 namespace llvm {
891 static double log2(double V) {
892 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L
893   return ::log2(V);
894 #else
895   return log(V) / numbers::ln2;
896 #endif
897 }
898 } // namespace llvm
899 
900 bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B,
901                               const FuncInfo &FInfo) {
902   assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
903           FInfo.getId() == AMDGPULibFunc::EI_POWR ||
904           FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
905          "fold_pow: encounter a wrong function call");
906 
907   Module *M = B.GetInsertBlock()->getModule();
908   Type *eltType = FPOp->getType()->getScalarType();
909   Value *opr0 = FPOp->getOperand(0);
910   Value *opr1 = FPOp->getOperand(1);
911 
912   const APFloat *CF = nullptr;
913   const APInt *CINT = nullptr;
914   if (!match(opr1, m_APFloatAllowPoison(CF)))
915     match(opr1, m_APIntAllowPoison(CINT));
916 
917   // 0x1111111 means that we don't do anything for this call.
918   int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
919 
920   if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0)) {
921     //  pow/powr/pown(x, 0) == 1
922     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1\n");
923     Constant *cnval = ConstantFP::get(eltType, 1.0);
924     if (getVecSize(FInfo) > 1) {
925       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
926     }
927     replaceCall(FPOp, cnval);
928     return true;
929   }
930   if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
931     // pow/powr/pown(x, 1.0) = x
932     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n");
933     replaceCall(FPOp, opr0);
934     return true;
935   }
936   if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
937     // pow/powr/pown(x, 2.0) = x*x
938     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << " * "
939                       << *opr0 << "\n");
940     Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
941     replaceCall(FPOp, nval);
942     return true;
943   }
944   if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
945     // pow/powr/pown(x, -1.0) = 1.0/x
946     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1 / " << *opr0 << "\n");
947     Constant *cnval = ConstantFP::get(eltType, 1.0);
948     if (getVecSize(FInfo) > 1) {
949       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
950     }
951     Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
952     replaceCall(FPOp, nval);
953     return true;
954   }
955 
956   if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
957     // pow[r](x, [-]0.5) = sqrt(x)
958     bool issqrt = CF->isExactlyValue(0.5);
959     if (FunctionCallee FPExpr =
960             getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
961                                                 : AMDGPULibFunc::EI_RSQRT,
962                                          FInfo))) {
963       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << FInfo.getName()
964                         << '(' << *opr0 << ")\n");
965       Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
966                                                         : "__pow2rsqrt");
967       replaceCall(FPOp, nval);
968       return true;
969     }
970   }
971 
972   if (!isUnsafeFiniteOnlyMath(FPOp))
973     return false;
974 
975   // Unsafe Math optimization
976 
977   // Remember that ci_opr1 is set if opr1 is integral
978   if (CF) {
979     double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
980                       ? (double)CF->convertToFloat()
981                       : CF->convertToDouble();
982     int ival = (int)dval;
983     if ((double)ival == dval) {
984       ci_opr1 = ival;
985     } else
986       ci_opr1 = 0x11111111;
987   }
988 
989   // pow/powr/pown(x, c) = [1/](x*x*..x); where
990   //   trunc(c) == c && the number of x == c && |c| <= 12
991   unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
992   if (abs_opr1 <= 12) {
993     Constant *cnval;
994     Value *nval;
995     if (abs_opr1 == 0) {
996       cnval = ConstantFP::get(eltType, 1.0);
997       if (getVecSize(FInfo) > 1) {
998         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
999       }
1000       nval = cnval;
1001     } else {
1002       Value *valx2 = nullptr;
1003       nval = nullptr;
1004       while (abs_opr1 > 0) {
1005         valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
1006         if (abs_opr1 & 1) {
1007           nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
1008         }
1009         abs_opr1 >>= 1;
1010       }
1011     }
1012 
1013     if (ci_opr1 < 0) {
1014       cnval = ConstantFP::get(eltType, 1.0);
1015       if (getVecSize(FInfo) > 1) {
1016         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
1017       }
1018       nval = B.CreateFDiv(cnval, nval, "__1powprod");
1019     }
1020     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1021                       << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
1022                       << ")\n");
1023     replaceCall(FPOp, nval);
1024     return true;
1025   }
1026 
1027   // If we should use the generic intrinsic instead of emitting a libcall
1028   const bool ShouldUseIntrinsic = eltType->isFloatTy() || eltType->isHalfTy();
1029 
1030   // powr ---> exp2(y * log2(x))
1031   // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
1032   FunctionCallee ExpExpr;
1033   if (ShouldUseIntrinsic)
1034     ExpExpr = Intrinsic::getOrInsertDeclaration(M, Intrinsic::exp2,
1035                                                 {FPOp->getType()});
1036   else {
1037     ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo));
1038     if (!ExpExpr)
1039       return false;
1040   }
1041 
1042   bool needlog = false;
1043   bool needabs = false;
1044   bool needcopysign = false;
1045   Constant *cnval = nullptr;
1046   if (getVecSize(FInfo) == 1) {
1047     CF = nullptr;
1048     match(opr0, m_APFloatAllowPoison(CF));
1049 
1050     if (CF) {
1051       double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1052                      ? (double)CF->convertToFloat()
1053                      : CF->convertToDouble();
1054 
1055       V = log2(std::abs(V));
1056       cnval = ConstantFP::get(eltType, V);
1057       needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
1058                      CF->isNegative();
1059     } else {
1060       needlog = true;
1061       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1062     }
1063   } else {
1064     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
1065 
1066     if (!CDV) {
1067       needlog = true;
1068       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1069     } else {
1070       assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
1071               "Wrong vector size detected");
1072 
1073       SmallVector<double, 0> DVal;
1074       for (int i=0; i < getVecSize(FInfo); ++i) {
1075         double V = CDV->getElementAsAPFloat(i).convertToDouble();
1076         if (V < 0.0) needcopysign = true;
1077         V = log2(std::abs(V));
1078         DVal.push_back(V);
1079       }
1080       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1081         SmallVector<float, 0> FVal;
1082         for (double D : DVal)
1083           FVal.push_back((float)D);
1084         ArrayRef<float> tmp(FVal);
1085         cnval = ConstantDataVector::get(M->getContext(), tmp);
1086       } else {
1087         ArrayRef<double> tmp(DVal);
1088         cnval = ConstantDataVector::get(M->getContext(), tmp);
1089       }
1090     }
1091   }
1092 
1093   if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1094     // We cannot handle corner cases for a general pow() function, give up
1095     // unless y is a constant integral value. Then proceed as if it were pown.
1096     if (!isKnownIntegral(opr1, M->getDataLayout(), FPOp->getFastMathFlags()))
1097       return false;
1098   }
1099 
1100   Value *nval;
1101   if (needabs) {
1102     nval = B.CreateUnaryIntrinsic(Intrinsic::fabs, opr0, nullptr, "__fabs");
1103   } else {
1104     nval = cnval ? cnval : opr0;
1105   }
1106   if (needlog) {
1107     FunctionCallee LogExpr;
1108     if (ShouldUseIntrinsic) {
1109       LogExpr = Intrinsic::getOrInsertDeclaration(M, Intrinsic::log2,
1110                                                   {FPOp->getType()});
1111     } else {
1112       LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo));
1113       if (!LogExpr)
1114         return false;
1115     }
1116 
1117     nval = CreateCallEx(B,LogExpr, nval, "__log2");
1118   }
1119 
1120   if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1121     // convert int(32) to fp(f32 or f64)
1122     opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1123   }
1124   nval = B.CreateFMul(opr1, nval, "__ylogx");
1125   nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1126 
1127   if (needcopysign) {
1128     Type* nTyS = B.getIntNTy(eltType->getPrimitiveSizeInBits());
1129     Type *nTy = FPOp->getType()->getWithNewType(nTyS);
1130     unsigned size = nTy->getScalarSizeInBits();
1131     Value *opr_n = FPOp->getOperand(1);
1132     if (opr_n->getType()->getScalarType()->isIntegerTy())
1133       opr_n = B.CreateZExtOrTrunc(opr_n, nTy, "__ytou");
1134     else
1135       opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1136 
1137     Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1138     sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1139     nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1140     nval = B.CreateBitCast(nval, opr0->getType());
1141   }
1142 
1143   LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1144                     << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1145   replaceCall(FPOp, nval);
1146 
1147   return true;
1148 }
1149 
1150 bool AMDGPULibCalls::fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B,
1151                                 const FuncInfo &FInfo) {
1152   Value *opr0 = FPOp->getOperand(0);
1153   Value *opr1 = FPOp->getOperand(1);
1154 
1155   const APInt *CINT = nullptr;
1156   if (!match(opr1, m_APIntAllowPoison(CINT)))
1157     return false;
1158 
1159   Function *Parent = B.GetInsertBlock()->getParent();
1160 
1161   int ci_opr1 = (int)CINT->getSExtValue();
1162   if (ci_opr1 == 1 && !Parent->hasFnAttribute(Attribute::StrictFP)) {
1163     // rootn(x, 1) = x
1164     //
1165     // TODO: Insert constrained canonicalize for strictfp case.
1166     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << '\n');
1167     replaceCall(FPOp, opr0);
1168     return true;
1169   }
1170 
1171   Module *M = B.GetInsertBlock()->getModule();
1172 
1173   CallInst *CI = cast<CallInst>(FPOp);
1174   if (ci_opr1 == 2 &&
1175       shouldReplaceLibcallWithIntrinsic(CI,
1176                                         /*AllowMinSizeF32=*/true,
1177                                         /*AllowF64=*/true)) {
1178     // rootn(x, 2) = sqrt(x)
1179     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> sqrt(" << *opr0 << ")\n");
1180 
1181     CallInst *NewCall = B.CreateUnaryIntrinsic(Intrinsic::sqrt, opr0, CI);
1182     NewCall->takeName(CI);
1183 
1184     // OpenCL rootn has a looser ulp of 2 requirement than sqrt, so add some
1185     // metadata.
1186     MDBuilder MDHelper(M->getContext());
1187     MDNode *FPMD = MDHelper.createFPMath(std::max(FPOp->getFPAccuracy(), 2.0f));
1188     NewCall->setMetadata(LLVMContext::MD_fpmath, FPMD);
1189 
1190     replaceCall(CI, NewCall);
1191     return true;
1192   }
1193 
1194   if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1195     if (FunctionCallee FPExpr =
1196             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1197       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> cbrt(" << *opr0
1198                         << ")\n");
1199       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1200       replaceCall(FPOp, nval);
1201       return true;
1202     }
1203   } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1204     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1.0 / " << *opr0 << "\n");
1205     Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1206                                opr0,
1207                                "__rootn2div");
1208     replaceCall(FPOp, nval);
1209     return true;
1210   }
1211 
1212   if (ci_opr1 == -2 &&
1213       shouldReplaceLibcallWithIntrinsic(CI,
1214                                         /*AllowMinSizeF32=*/true,
1215                                         /*AllowF64=*/true)) {
1216     // rootn(x, -2) = rsqrt(x)
1217 
1218     // The original rootn had looser ulp requirements than the resultant sqrt
1219     // and fdiv.
1220     MDBuilder MDHelper(M->getContext());
1221     MDNode *FPMD = MDHelper.createFPMath(std::max(FPOp->getFPAccuracy(), 2.0f));
1222 
1223     // TODO: Could handle strictfp but need to fix strict sqrt emission
1224     FastMathFlags FMF = FPOp->getFastMathFlags();
1225     FMF.setAllowContract(true);
1226 
1227     CallInst *Sqrt = B.CreateUnaryIntrinsic(Intrinsic::sqrt, opr0, CI);
1228     Instruction *RSqrt = cast<Instruction>(
1229         B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0), Sqrt));
1230     Sqrt->setFastMathFlags(FMF);
1231     RSqrt->setFastMathFlags(FMF);
1232     RSqrt->setMetadata(LLVMContext::MD_fpmath, FPMD);
1233 
1234     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> rsqrt(" << *opr0
1235                       << ")\n");
1236     replaceCall(CI, RSqrt);
1237     return true;
1238   }
1239 
1240   return false;
1241 }
1242 
1243 // Get a scalar native builtin single argument FP function
1244 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1245                                                  const FuncInfo &FInfo) {
1246   if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1247     return nullptr;
1248   FuncInfo nf = FInfo;
1249   nf.setPrefix(AMDGPULibFunc::NATIVE);
1250   return getFunction(M, nf);
1251 }
1252 
1253 // Some library calls are just wrappers around llvm intrinsics, but compiled
1254 // conservatively. Preserve the flags from the original call site by
1255 // substituting them with direct calls with all the flags.
1256 bool AMDGPULibCalls::shouldReplaceLibcallWithIntrinsic(const CallInst *CI,
1257                                                        bool AllowMinSizeF32,
1258                                                        bool AllowF64,
1259                                                        bool AllowStrictFP) {
1260   Type *FltTy = CI->getType()->getScalarType();
1261   const bool IsF32 = FltTy->isFloatTy();
1262 
1263   // f64 intrinsics aren't implemented for most operations.
1264   if (!IsF32 && !FltTy->isHalfTy() && (!AllowF64 || !FltTy->isDoubleTy()))
1265     return false;
1266 
1267   // We're implicitly inlining by replacing the libcall with the intrinsic, so
1268   // don't do it for noinline call sites.
1269   if (CI->isNoInline())
1270     return false;
1271 
1272   const Function *ParentF = CI->getFunction();
1273   // TODO: Handle strictfp
1274   if (!AllowStrictFP && ParentF->hasFnAttribute(Attribute::StrictFP))
1275     return false;
1276 
1277   if (IsF32 && !AllowMinSizeF32 && ParentF->hasMinSize())
1278     return false;
1279   return true;
1280 }
1281 
1282 void AMDGPULibCalls::replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B,
1283                                                        CallInst *CI,
1284                                                        Intrinsic::ID IntrID) {
1285   if (CI->arg_size() == 2) {
1286     Value *Arg0 = CI->getArgOperand(0);
1287     Value *Arg1 = CI->getArgOperand(1);
1288     VectorType *Arg0VecTy = dyn_cast<VectorType>(Arg0->getType());
1289     VectorType *Arg1VecTy = dyn_cast<VectorType>(Arg1->getType());
1290     if (Arg0VecTy && !Arg1VecTy) {
1291       Value *SplatRHS = B.CreateVectorSplat(Arg0VecTy->getElementCount(), Arg1);
1292       CI->setArgOperand(1, SplatRHS);
1293     } else if (!Arg0VecTy && Arg1VecTy) {
1294       Value *SplatLHS = B.CreateVectorSplat(Arg1VecTy->getElementCount(), Arg0);
1295       CI->setArgOperand(0, SplatLHS);
1296     }
1297   }
1298 
1299   CI->setCalledFunction(Intrinsic::getOrInsertDeclaration(
1300       CI->getModule(), IntrID, {CI->getType()}));
1301 }
1302 
1303 bool AMDGPULibCalls::tryReplaceLibcallWithSimpleIntrinsic(
1304     IRBuilder<> &B, CallInst *CI, Intrinsic::ID IntrID, bool AllowMinSizeF32,
1305     bool AllowF64, bool AllowStrictFP) {
1306   if (!shouldReplaceLibcallWithIntrinsic(CI, AllowMinSizeF32, AllowF64,
1307                                          AllowStrictFP))
1308     return false;
1309   replaceLibCallWithSimpleIntrinsic(B, CI, IntrID);
1310   return true;
1311 }
1312 
1313 std::tuple<Value *, Value *, Value *>
1314 AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B,
1315                              FunctionCallee Fsincos) {
1316   DebugLoc DL = B.getCurrentDebugLocation();
1317   Function *F = B.GetInsertBlock()->getParent();
1318   B.SetInsertPointPastAllocas(F);
1319 
1320   AllocaInst *Alloc = B.CreateAlloca(Arg->getType(), nullptr, "__sincos_");
1321 
1322   if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1323     // If the argument is an instruction, it must dominate all uses so put our
1324     // sincos call there. Otherwise, right after the allocas works well enough
1325     // if it's an argument or constant.
1326 
1327     B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1328 
1329     // SetInsertPoint unwelcomely always tries to set the debug loc.
1330     B.SetCurrentDebugLocation(DL);
1331   }
1332 
1333   Type *CosPtrTy = Fsincos.getFunctionType()->getParamType(1);
1334 
1335   // The allocaInst allocates the memory in private address space. This need
1336   // to be addrspacecasted to point to the address space of cos pointer type.
1337   // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1338   Value *CastAlloc = B.CreateAddrSpaceCast(Alloc, CosPtrTy);
1339 
1340   CallInst *SinCos = CreateCallEx2(B, Fsincos, Arg, CastAlloc);
1341 
1342   // TODO: Is it worth trying to preserve the location for the cos calls for the
1343   // load?
1344 
1345   LoadInst *LoadCos = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1346   return {SinCos, LoadCos, SinCos};
1347 }
1348 
1349 // fold sin, cos -> sincos.
1350 bool AMDGPULibCalls::fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B,
1351                                  const FuncInfo &fInfo) {
1352   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1353          fInfo.getId() == AMDGPULibFunc::EI_COS);
1354 
1355   if ((getArgType(fInfo) != AMDGPULibFunc::F32 &&
1356        getArgType(fInfo) != AMDGPULibFunc::F64) ||
1357       fInfo.getPrefix() != AMDGPULibFunc::NOPFX)
1358     return false;
1359 
1360   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1361 
1362   Value *CArgVal = FPOp->getOperand(0);
1363 
1364   // TODO: Constant fold the call
1365   if (isa<ConstantData>(CArgVal))
1366     return false;
1367 
1368   CallInst *CI = cast<CallInst>(FPOp);
1369 
1370   Function *F = B.GetInsertBlock()->getParent();
1371   Module *M = F->getParent();
1372 
1373   // Merge the sin and cos. For OpenCL 2.0, there may only be a generic pointer
1374   // implementation. Prefer the private form if available.
1375   AMDGPULibFunc SinCosLibFuncPrivate(AMDGPULibFunc::EI_SINCOS, fInfo);
1376   SinCosLibFuncPrivate.getLeads()[0].PtrKind =
1377       AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::PRIVATE_ADDRESS);
1378 
1379   AMDGPULibFunc SinCosLibFuncGeneric(AMDGPULibFunc::EI_SINCOS, fInfo);
1380   SinCosLibFuncGeneric.getLeads()[0].PtrKind =
1381       AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1382 
1383   FunctionCallee FSinCosPrivate = getFunction(M, SinCosLibFuncPrivate);
1384   FunctionCallee FSinCosGeneric = getFunction(M, SinCosLibFuncGeneric);
1385   FunctionCallee FSinCos = FSinCosPrivate ? FSinCosPrivate : FSinCosGeneric;
1386   if (!FSinCos)
1387     return false;
1388 
1389   SmallVector<CallInst *> SinCalls;
1390   SmallVector<CallInst *> CosCalls;
1391   SmallVector<CallInst *> SinCosCalls;
1392   FuncInfo PartnerInfo(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN,
1393                        fInfo);
1394   const std::string PairName = PartnerInfo.mangle();
1395 
1396   StringRef SinName = isSin ? CI->getCalledFunction()->getName() : PairName;
1397   StringRef CosName = isSin ? PairName : CI->getCalledFunction()->getName();
1398   const std::string SinCosPrivateName = SinCosLibFuncPrivate.mangle();
1399   const std::string SinCosGenericName = SinCosLibFuncGeneric.mangle();
1400 
1401   // Intersect the two sets of flags.
1402   FastMathFlags FMF = FPOp->getFastMathFlags();
1403   MDNode *FPMath = CI->getMetadata(LLVMContext::MD_fpmath);
1404 
1405   SmallVector<DILocation *> MergeDbgLocs = {CI->getDebugLoc()};
1406 
1407   for (User* U : CArgVal->users()) {
1408     CallInst *XI = dyn_cast<CallInst>(U);
1409     if (!XI || XI->getFunction() != F || XI->isNoBuiltin())
1410       continue;
1411 
1412     Function *UCallee = XI->getCalledFunction();
1413     if (!UCallee)
1414       continue;
1415 
1416     bool Handled = true;
1417 
1418     if (UCallee->getName() == SinName)
1419       SinCalls.push_back(XI);
1420     else if (UCallee->getName() == CosName)
1421       CosCalls.push_back(XI);
1422     else if (UCallee->getName() == SinCosPrivateName ||
1423              UCallee->getName() == SinCosGenericName)
1424       SinCosCalls.push_back(XI);
1425     else
1426       Handled = false;
1427 
1428     if (Handled) {
1429       MergeDbgLocs.push_back(XI->getDebugLoc());
1430       auto *OtherOp = cast<FPMathOperator>(XI);
1431       FMF &= OtherOp->getFastMathFlags();
1432       FPMath = MDNode::getMostGenericFPMath(
1433           FPMath, XI->getMetadata(LLVMContext::MD_fpmath));
1434     }
1435   }
1436 
1437   if (SinCalls.empty() || CosCalls.empty())
1438     return false;
1439 
1440   B.setFastMathFlags(FMF);
1441   B.setDefaultFPMathTag(FPMath);
1442   DILocation *DbgLoc = DILocation::getMergedLocations(MergeDbgLocs);
1443   B.SetCurrentDebugLocation(DbgLoc);
1444 
1445   auto [Sin, Cos, SinCos] = insertSinCos(CArgVal, FMF, B, FSinCos);
1446 
1447   auto replaceTrigInsts = [](ArrayRef<CallInst *> Calls, Value *Res) {
1448     for (CallInst *C : Calls)
1449       C->replaceAllUsesWith(Res);
1450 
1451     // Leave the other dead instructions to avoid clobbering iterators.
1452   };
1453 
1454   replaceTrigInsts(SinCalls, Sin);
1455   replaceTrigInsts(CosCalls, Cos);
1456   replaceTrigInsts(SinCosCalls, SinCos);
1457 
1458   // It's safe to delete the original now.
1459   CI->eraseFromParent();
1460   return true;
1461 }
1462 
1463 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0,
1464                                             double &Res1, Constant *copr0,
1465                                             Constant *copr1) {
1466   // By default, opr0/opr1/opr3 holds values of float/double type.
1467   // If they are not float/double, each function has to its
1468   // operand separately.
1469   double opr0 = 0.0, opr1 = 0.0;
1470   ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1471   ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1472   if (fpopr0) {
1473     opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1474              ? fpopr0->getValueAPF().convertToDouble()
1475              : (double)fpopr0->getValueAPF().convertToFloat();
1476   }
1477 
1478   if (fpopr1) {
1479     opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1480              ? fpopr1->getValueAPF().convertToDouble()
1481              : (double)fpopr1->getValueAPF().convertToFloat();
1482   }
1483 
1484   switch (FInfo.getId()) {
1485   default : return false;
1486 
1487   case AMDGPULibFunc::EI_ACOS:
1488     Res0 = acos(opr0);
1489     return true;
1490 
1491   case AMDGPULibFunc::EI_ACOSH:
1492     // acosh(x) == log(x + sqrt(x*x - 1))
1493     Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1494     return true;
1495 
1496   case AMDGPULibFunc::EI_ACOSPI:
1497     Res0 = acos(opr0) / MATH_PI;
1498     return true;
1499 
1500   case AMDGPULibFunc::EI_ASIN:
1501     Res0 = asin(opr0);
1502     return true;
1503 
1504   case AMDGPULibFunc::EI_ASINH:
1505     // asinh(x) == log(x + sqrt(x*x + 1))
1506     Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1507     return true;
1508 
1509   case AMDGPULibFunc::EI_ASINPI:
1510     Res0 = asin(opr0) / MATH_PI;
1511     return true;
1512 
1513   case AMDGPULibFunc::EI_ATAN:
1514     Res0 = atan(opr0);
1515     return true;
1516 
1517   case AMDGPULibFunc::EI_ATANH:
1518     // atanh(x) == (log(x+1) - log(x-1))/2;
1519     Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1520     return true;
1521 
1522   case AMDGPULibFunc::EI_ATANPI:
1523     Res0 = atan(opr0) / MATH_PI;
1524     return true;
1525 
1526   case AMDGPULibFunc::EI_CBRT:
1527     Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1528     return true;
1529 
1530   case AMDGPULibFunc::EI_COS:
1531     Res0 = cos(opr0);
1532     return true;
1533 
1534   case AMDGPULibFunc::EI_COSH:
1535     Res0 = cosh(opr0);
1536     return true;
1537 
1538   case AMDGPULibFunc::EI_COSPI:
1539     Res0 = cos(MATH_PI * opr0);
1540     return true;
1541 
1542   case AMDGPULibFunc::EI_EXP:
1543     Res0 = exp(opr0);
1544     return true;
1545 
1546   case AMDGPULibFunc::EI_EXP2:
1547     Res0 = pow(2.0, opr0);
1548     return true;
1549 
1550   case AMDGPULibFunc::EI_EXP10:
1551     Res0 = pow(10.0, opr0);
1552     return true;
1553 
1554   case AMDGPULibFunc::EI_LOG:
1555     Res0 = log(opr0);
1556     return true;
1557 
1558   case AMDGPULibFunc::EI_LOG2:
1559     Res0 = log(opr0) / log(2.0);
1560     return true;
1561 
1562   case AMDGPULibFunc::EI_LOG10:
1563     Res0 = log(opr0) / log(10.0);
1564     return true;
1565 
1566   case AMDGPULibFunc::EI_RSQRT:
1567     Res0 = 1.0 / sqrt(opr0);
1568     return true;
1569 
1570   case AMDGPULibFunc::EI_SIN:
1571     Res0 = sin(opr0);
1572     return true;
1573 
1574   case AMDGPULibFunc::EI_SINH:
1575     Res0 = sinh(opr0);
1576     return true;
1577 
1578   case AMDGPULibFunc::EI_SINPI:
1579     Res0 = sin(MATH_PI * opr0);
1580     return true;
1581 
1582   case AMDGPULibFunc::EI_TAN:
1583     Res0 = tan(opr0);
1584     return true;
1585 
1586   case AMDGPULibFunc::EI_TANH:
1587     Res0 = tanh(opr0);
1588     return true;
1589 
1590   case AMDGPULibFunc::EI_TANPI:
1591     Res0 = tan(MATH_PI * opr0);
1592     return true;
1593 
1594   // two-arg functions
1595   case AMDGPULibFunc::EI_POW:
1596   case AMDGPULibFunc::EI_POWR:
1597     Res0 = pow(opr0, opr1);
1598     return true;
1599 
1600   case AMDGPULibFunc::EI_POWN: {
1601     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1602       double val = (double)iopr1->getSExtValue();
1603       Res0 = pow(opr0, val);
1604       return true;
1605     }
1606     return false;
1607   }
1608 
1609   case AMDGPULibFunc::EI_ROOTN: {
1610     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1611       double val = (double)iopr1->getSExtValue();
1612       Res0 = pow(opr0, 1.0 / val);
1613       return true;
1614     }
1615     return false;
1616   }
1617 
1618   // with ptr arg
1619   case AMDGPULibFunc::EI_SINCOS:
1620     Res0 = sin(opr0);
1621     Res1 = cos(opr0);
1622     return true;
1623   }
1624 
1625   return false;
1626 }
1627 
1628 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) {
1629   int numArgs = (int)aCI->arg_size();
1630   if (numArgs > 3)
1631     return false;
1632 
1633   Constant *copr0 = nullptr;
1634   Constant *copr1 = nullptr;
1635   if (numArgs > 0) {
1636     if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1637       return false;
1638   }
1639 
1640   if (numArgs > 1) {
1641     if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1642       if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1643         return false;
1644     }
1645   }
1646 
1647   // At this point, all arguments to aCI are constants.
1648 
1649   // max vector size is 16, and sincos will generate two results.
1650   double DVal0[16], DVal1[16];
1651   int FuncVecSize = getVecSize(FInfo);
1652   bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1653   if (FuncVecSize == 1) {
1654     if (!evaluateScalarMathFunc(FInfo, DVal0[0], DVal1[0], copr0, copr1)) {
1655       return false;
1656     }
1657   } else {
1658     ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1659     ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1660     for (int i = 0; i < FuncVecSize; ++i) {
1661       Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1662       Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1663       if (!evaluateScalarMathFunc(FInfo, DVal0[i], DVal1[i], celt0, celt1)) {
1664         return false;
1665       }
1666     }
1667   }
1668 
1669   LLVMContext &context = aCI->getContext();
1670   Constant *nval0, *nval1;
1671   if (FuncVecSize == 1) {
1672     nval0 = ConstantFP::get(aCI->getType(), DVal0[0]);
1673     if (hasTwoResults)
1674       nval1 = ConstantFP::get(aCI->getType(), DVal1[0]);
1675   } else {
1676     if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1677       SmallVector <float, 0> FVal0, FVal1;
1678       for (int i = 0; i < FuncVecSize; ++i)
1679         FVal0.push_back((float)DVal0[i]);
1680       ArrayRef<float> tmp0(FVal0);
1681       nval0 = ConstantDataVector::get(context, tmp0);
1682       if (hasTwoResults) {
1683         for (int i = 0; i < FuncVecSize; ++i)
1684           FVal1.push_back((float)DVal1[i]);
1685         ArrayRef<float> tmp1(FVal1);
1686         nval1 = ConstantDataVector::get(context, tmp1);
1687       }
1688     } else {
1689       ArrayRef<double> tmp0(DVal0);
1690       nval0 = ConstantDataVector::get(context, tmp0);
1691       if (hasTwoResults) {
1692         ArrayRef<double> tmp1(DVal1);
1693         nval1 = ConstantDataVector::get(context, tmp1);
1694       }
1695     }
1696   }
1697 
1698   if (hasTwoResults) {
1699     // sincos
1700     assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1701            "math function with ptr arg not supported yet");
1702     new StoreInst(nval1, aCI->getArgOperand(1), aCI->getIterator());
1703   }
1704 
1705   replaceCall(aCI, nval0);
1706   return true;
1707 }
1708 
1709 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F,
1710                                                   FunctionAnalysisManager &AM) {
1711   AMDGPULibCalls Simplifier;
1712   Simplifier.initNativeFuncs();
1713   Simplifier.initFunction(F, AM);
1714 
1715   bool Changed = false;
1716 
1717   LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1718              F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1719 
1720   for (auto &BB : F) {
1721     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1722       // Ignore non-calls.
1723       CallInst *CI = dyn_cast<CallInst>(I);
1724       ++I;
1725 
1726       if (CI) {
1727         if (Simplifier.fold(CI))
1728           Changed = true;
1729       }
1730     }
1731   }
1732   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1733 }
1734 
1735 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F,
1736                                                 FunctionAnalysisManager &AM) {
1737   if (UseNative.empty())
1738     return PreservedAnalyses::all();
1739 
1740   AMDGPULibCalls Simplifier;
1741   Simplifier.initNativeFuncs();
1742   Simplifier.initFunction(F, AM);
1743 
1744   bool Changed = false;
1745   for (auto &BB : F) {
1746     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1747       // Ignore non-calls.
1748       CallInst *CI = dyn_cast<CallInst>(I);
1749       ++I;
1750       if (CI && Simplifier.useNative(CI))
1751         Changed = true;
1752     }
1753   }
1754   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1755 }
1756