xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp (revision ee795fd1cf6ceb4963bf264df0dcfad006d96381)
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   // sqrt
91   bool fold_sqrt(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
92 
93   /// Insert a value to sincos function \p Fsincos. Returns (value of sin, value
94   /// of cos, sincos call).
95   std::tuple<Value *, Value *, Value *> insertSinCos(Value *Arg,
96                                                      FastMathFlags FMF,
97                                                      IRBuilder<> &B,
98                                                      FunctionCallee Fsincos);
99 
100   // sin/cos
101   bool fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B, const FuncInfo &FInfo);
102 
103   // __read_pipe/__write_pipe
104   bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
105                             const FuncInfo &FInfo);
106 
107   // Get a scalar native builtin single argument FP function
108   FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo);
109 
110   /// Substitute a call to a known libcall with an intrinsic call. If \p
111   /// AllowMinSize is true, allow the replacement in a minsize function.
112   bool shouldReplaceLibcallWithIntrinsic(const CallInst *CI,
113                                          bool AllowMinSizeF32 = false,
114                                          bool AllowF64 = false,
115                                          bool AllowStrictFP = false);
116   void replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI,
117                                          Intrinsic::ID IntrID);
118 
119   bool tryReplaceLibcallWithSimpleIntrinsic(IRBuilder<> &B, CallInst *CI,
120                                             Intrinsic::ID IntrID,
121                                             bool AllowMinSizeF32 = false,
122                                             bool AllowF64 = false,
123                                             bool AllowStrictFP = false);
124 
125 protected:
126   bool isUnsafeMath(const FPMathOperator *FPOp) const;
127   bool isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const;
128 
129   bool canIncreasePrecisionOfConstantFold(const FPMathOperator *FPOp) const;
130 
131   static void replaceCall(Instruction *I, Value *With) {
132     I->replaceAllUsesWith(With);
133     I->eraseFromParent();
134   }
135 
136   static void replaceCall(FPMathOperator *I, Value *With) {
137     replaceCall(cast<Instruction>(I), With);
138   }
139 
140 public:
141   AMDGPULibCalls() {}
142 
143   bool fold(CallInst *CI);
144 
145   void initFunction(Function &F, FunctionAnalysisManager &FAM);
146   void initNativeFuncs();
147 
148   // Replace a normal math function call with that native version
149   bool useNative(CallInst *CI);
150 };
151 
152 } // end llvm namespace
153 
154 template <typename IRB>
155 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg,
156                               const Twine &Name = "") {
157   CallInst *R = B.CreateCall(Callee, Arg, Name);
158   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
159     R->setCallingConv(F->getCallingConv());
160   return R;
161 }
162 
163 template <typename IRB>
164 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1,
165                                Value *Arg2, const Twine &Name = "") {
166   CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
167   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
168     R->setCallingConv(F->getCallingConv());
169   return R;
170 }
171 
172 static FunctionType *getPownType(FunctionType *FT) {
173   Type *PowNExpTy = Type::getInt32Ty(FT->getContext());
174   if (VectorType *VecTy = dyn_cast<VectorType>(FT->getReturnType()))
175     PowNExpTy = VectorType::get(PowNExpTy, VecTy->getElementCount());
176 
177   return FunctionType::get(FT->getReturnType(),
178                            {FT->getParamType(0), PowNExpTy}, false);
179 }
180 
181 //  Data structures for table-driven optimizations.
182 //  FuncTbl works for both f32 and f64 functions with 1 input argument
183 
184 struct TableEntry {
185   double   result;
186   double   input;
187 };
188 
189 /* a list of {result, input} */
190 static const TableEntry tbl_acos[] = {
191   {MATH_PI / 2.0, 0.0},
192   {MATH_PI / 2.0, -0.0},
193   {0.0, 1.0},
194   {MATH_PI, -1.0}
195 };
196 static const TableEntry tbl_acosh[] = {
197   {0.0, 1.0}
198 };
199 static const TableEntry tbl_acospi[] = {
200   {0.5, 0.0},
201   {0.5, -0.0},
202   {0.0, 1.0},
203   {1.0, -1.0}
204 };
205 static const TableEntry tbl_asin[] = {
206   {0.0, 0.0},
207   {-0.0, -0.0},
208   {MATH_PI / 2.0, 1.0},
209   {-MATH_PI / 2.0, -1.0}
210 };
211 static const TableEntry tbl_asinh[] = {
212   {0.0, 0.0},
213   {-0.0, -0.0}
214 };
215 static const TableEntry tbl_asinpi[] = {
216   {0.0, 0.0},
217   {-0.0, -0.0},
218   {0.5, 1.0},
219   {-0.5, -1.0}
220 };
221 static const TableEntry tbl_atan[] = {
222   {0.0, 0.0},
223   {-0.0, -0.0},
224   {MATH_PI / 4.0, 1.0},
225   {-MATH_PI / 4.0, -1.0}
226 };
227 static const TableEntry tbl_atanh[] = {
228   {0.0, 0.0},
229   {-0.0, -0.0}
230 };
231 static const TableEntry tbl_atanpi[] = {
232   {0.0, 0.0},
233   {-0.0, -0.0},
234   {0.25, 1.0},
235   {-0.25, -1.0}
236 };
237 static const TableEntry tbl_cbrt[] = {
238   {0.0, 0.0},
239   {-0.0, -0.0},
240   {1.0, 1.0},
241   {-1.0, -1.0},
242 };
243 static const TableEntry tbl_cos[] = {
244   {1.0, 0.0},
245   {1.0, -0.0}
246 };
247 static const TableEntry tbl_cosh[] = {
248   {1.0, 0.0},
249   {1.0, -0.0}
250 };
251 static const TableEntry tbl_cospi[] = {
252   {1.0, 0.0},
253   {1.0, -0.0}
254 };
255 static const TableEntry tbl_erfc[] = {
256   {1.0, 0.0},
257   {1.0, -0.0}
258 };
259 static const TableEntry tbl_erf[] = {
260   {0.0, 0.0},
261   {-0.0, -0.0}
262 };
263 static const TableEntry tbl_exp[] = {
264   {1.0, 0.0},
265   {1.0, -0.0},
266   {MATH_E, 1.0}
267 };
268 static const TableEntry tbl_exp2[] = {
269   {1.0, 0.0},
270   {1.0, -0.0},
271   {2.0, 1.0}
272 };
273 static const TableEntry tbl_exp10[] = {
274   {1.0, 0.0},
275   {1.0, -0.0},
276   {10.0, 1.0}
277 };
278 static const TableEntry tbl_expm1[] = {
279   {0.0, 0.0},
280   {-0.0, -0.0}
281 };
282 static const TableEntry tbl_log[] = {
283   {0.0, 1.0},
284   {1.0, MATH_E}
285 };
286 static const TableEntry tbl_log2[] = {
287   {0.0, 1.0},
288   {1.0, 2.0}
289 };
290 static const TableEntry tbl_log10[] = {
291   {0.0, 1.0},
292   {1.0, 10.0}
293 };
294 static const TableEntry tbl_rsqrt[] = {
295   {1.0, 1.0},
296   {MATH_SQRT1_2, 2.0}
297 };
298 static const TableEntry tbl_sin[] = {
299   {0.0, 0.0},
300   {-0.0, -0.0}
301 };
302 static const TableEntry tbl_sinh[] = {
303   {0.0, 0.0},
304   {-0.0, -0.0}
305 };
306 static const TableEntry tbl_sinpi[] = {
307   {0.0, 0.0},
308   {-0.0, -0.0}
309 };
310 static const TableEntry tbl_sqrt[] = {
311   {0.0, 0.0},
312   {1.0, 1.0},
313   {MATH_SQRT2, 2.0}
314 };
315 static const TableEntry tbl_tan[] = {
316   {0.0, 0.0},
317   {-0.0, -0.0}
318 };
319 static const TableEntry tbl_tanh[] = {
320   {0.0, 0.0},
321   {-0.0, -0.0}
322 };
323 static const TableEntry tbl_tanpi[] = {
324   {0.0, 0.0},
325   {-0.0, -0.0}
326 };
327 static const TableEntry tbl_tgamma[] = {
328   {1.0, 1.0},
329   {1.0, 2.0},
330   {2.0, 3.0},
331   {6.0, 4.0}
332 };
333 
334 static bool HasNative(AMDGPULibFunc::EFuncId id) {
335   switch(id) {
336   case AMDGPULibFunc::EI_DIVIDE:
337   case AMDGPULibFunc::EI_COS:
338   case AMDGPULibFunc::EI_EXP:
339   case AMDGPULibFunc::EI_EXP2:
340   case AMDGPULibFunc::EI_EXP10:
341   case AMDGPULibFunc::EI_LOG:
342   case AMDGPULibFunc::EI_LOG2:
343   case AMDGPULibFunc::EI_LOG10:
344   case AMDGPULibFunc::EI_POWR:
345   case AMDGPULibFunc::EI_RECIP:
346   case AMDGPULibFunc::EI_RSQRT:
347   case AMDGPULibFunc::EI_SIN:
348   case AMDGPULibFunc::EI_SINCOS:
349   case AMDGPULibFunc::EI_SQRT:
350   case AMDGPULibFunc::EI_TAN:
351     return true;
352   default:;
353   }
354   return false;
355 }
356 
357 using TableRef = ArrayRef<TableEntry>;
358 
359 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
360   switch(id) {
361   case AMDGPULibFunc::EI_ACOS:    return TableRef(tbl_acos);
362   case AMDGPULibFunc::EI_ACOSH:   return TableRef(tbl_acosh);
363   case AMDGPULibFunc::EI_ACOSPI:  return TableRef(tbl_acospi);
364   case AMDGPULibFunc::EI_ASIN:    return TableRef(tbl_asin);
365   case AMDGPULibFunc::EI_ASINH:   return TableRef(tbl_asinh);
366   case AMDGPULibFunc::EI_ASINPI:  return TableRef(tbl_asinpi);
367   case AMDGPULibFunc::EI_ATAN:    return TableRef(tbl_atan);
368   case AMDGPULibFunc::EI_ATANH:   return TableRef(tbl_atanh);
369   case AMDGPULibFunc::EI_ATANPI:  return TableRef(tbl_atanpi);
370   case AMDGPULibFunc::EI_CBRT:    return TableRef(tbl_cbrt);
371   case AMDGPULibFunc::EI_NCOS:
372   case AMDGPULibFunc::EI_COS:     return TableRef(tbl_cos);
373   case AMDGPULibFunc::EI_COSH:    return TableRef(tbl_cosh);
374   case AMDGPULibFunc::EI_COSPI:   return TableRef(tbl_cospi);
375   case AMDGPULibFunc::EI_ERFC:    return TableRef(tbl_erfc);
376   case AMDGPULibFunc::EI_ERF:     return TableRef(tbl_erf);
377   case AMDGPULibFunc::EI_EXP:     return TableRef(tbl_exp);
378   case AMDGPULibFunc::EI_NEXP2:
379   case AMDGPULibFunc::EI_EXP2:    return TableRef(tbl_exp2);
380   case AMDGPULibFunc::EI_EXP10:   return TableRef(tbl_exp10);
381   case AMDGPULibFunc::EI_EXPM1:   return TableRef(tbl_expm1);
382   case AMDGPULibFunc::EI_LOG:     return TableRef(tbl_log);
383   case AMDGPULibFunc::EI_NLOG2:
384   case AMDGPULibFunc::EI_LOG2:    return TableRef(tbl_log2);
385   case AMDGPULibFunc::EI_LOG10:   return TableRef(tbl_log10);
386   case AMDGPULibFunc::EI_NRSQRT:
387   case AMDGPULibFunc::EI_RSQRT:   return TableRef(tbl_rsqrt);
388   case AMDGPULibFunc::EI_NSIN:
389   case AMDGPULibFunc::EI_SIN:     return TableRef(tbl_sin);
390   case AMDGPULibFunc::EI_SINH:    return TableRef(tbl_sinh);
391   case AMDGPULibFunc::EI_SINPI:   return TableRef(tbl_sinpi);
392   case AMDGPULibFunc::EI_NSQRT:
393   case AMDGPULibFunc::EI_SQRT:    return TableRef(tbl_sqrt);
394   case AMDGPULibFunc::EI_TAN:     return TableRef(tbl_tan);
395   case AMDGPULibFunc::EI_TANH:    return TableRef(tbl_tanh);
396   case AMDGPULibFunc::EI_TANPI:   return TableRef(tbl_tanpi);
397   case AMDGPULibFunc::EI_TGAMMA:  return TableRef(tbl_tgamma);
398   default:;
399   }
400   return TableRef();
401 }
402 
403 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
404   return FInfo.getLeads()[0].VectorSize;
405 }
406 
407 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
408   return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
409 }
410 
411 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) {
412   // If we are doing PreLinkOpt, the function is external. So it is safe to
413   // use getOrInsertFunction() at this stage.
414 
415   return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
416                        : AMDGPULibFunc::getFunction(M, fInfo);
417 }
418 
419 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName,
420                                        FuncInfo &FInfo) {
421   return AMDGPULibFunc::parse(FMangledName, FInfo);
422 }
423 
424 bool AMDGPULibCalls::isUnsafeMath(const FPMathOperator *FPOp) const {
425   return UnsafeFPMath || FPOp->isFast();
426 }
427 
428 bool AMDGPULibCalls::isUnsafeFiniteOnlyMath(const FPMathOperator *FPOp) const {
429   return UnsafeFPMath ||
430          (FPOp->hasApproxFunc() && FPOp->hasNoNaNs() && FPOp->hasNoInfs());
431 }
432 
433 bool AMDGPULibCalls::canIncreasePrecisionOfConstantFold(
434     const FPMathOperator *FPOp) const {
435   // TODO: Refine to approxFunc or contract
436   return isUnsafeMath(FPOp);
437 }
438 
439 void AMDGPULibCalls::initFunction(Function &F, FunctionAnalysisManager &FAM) {
440   UnsafeFPMath = F.getFnAttribute("unsafe-fp-math").getValueAsBool();
441   AC = &FAM.getResult<AssumptionAnalysis>(F);
442   TLInfo = &FAM.getResult<TargetLibraryAnalysis>(F);
443   DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
444 }
445 
446 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
447   return AllNative || llvm::is_contained(UseNative, F);
448 }
449 
450 void AMDGPULibCalls::initNativeFuncs() {
451   AllNative = useNativeFunc("all") ||
452               (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
453                UseNative.begin()->empty());
454 }
455 
456 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
457   bool native_sin = useNativeFunc("sin");
458   bool native_cos = useNativeFunc("cos");
459 
460   if (native_sin && native_cos) {
461     Module *M = aCI->getModule();
462     Value *opr0 = aCI->getArgOperand(0);
463 
464     AMDGPULibFunc nf;
465     nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
466     nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
467 
468     nf.setPrefix(AMDGPULibFunc::NATIVE);
469     nf.setId(AMDGPULibFunc::EI_SIN);
470     FunctionCallee sinExpr = getFunction(M, nf);
471 
472     nf.setPrefix(AMDGPULibFunc::NATIVE);
473     nf.setId(AMDGPULibFunc::EI_COS);
474     FunctionCallee cosExpr = getFunction(M, nf);
475     if (sinExpr && cosExpr) {
476       Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
477       Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
478       new StoreInst(cosval, aCI->getArgOperand(1), aCI);
479 
480       DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
481                                           << " with native version of sin/cos");
482 
483       replaceCall(aCI, sinval);
484       return true;
485     }
486   }
487   return false;
488 }
489 
490 bool AMDGPULibCalls::useNative(CallInst *aCI) {
491   Function *Callee = aCI->getCalledFunction();
492   if (!Callee || aCI->isNoBuiltin())
493     return false;
494 
495   FuncInfo FInfo;
496   if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() ||
497       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
498       getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
499       !(AllNative || useNativeFunc(FInfo.getName()))) {
500     return false;
501   }
502 
503   if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
504     return sincosUseNative(aCI, FInfo);
505 
506   FInfo.setPrefix(AMDGPULibFunc::NATIVE);
507   FunctionCallee F = getFunction(aCI->getModule(), FInfo);
508   if (!F)
509     return false;
510 
511   aCI->setCalledFunction(F);
512   DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
513                                       << " with native version");
514   return true;
515 }
516 
517 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
518 // builtin, with appended type size and alignment arguments, where 2 or 4
519 // indicates the original number of arguments. The library has optimized version
520 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
521 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
522 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
523 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
524 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
525                                           const FuncInfo &FInfo) {
526   auto *Callee = CI->getCalledFunction();
527   if (!Callee->isDeclaration())
528     return false;
529 
530   assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
531   auto *M = Callee->getParent();
532   std::string Name = std::string(Callee->getName());
533   auto NumArg = CI->arg_size();
534   if (NumArg != 4 && NumArg != 6)
535     return false;
536   ConstantInt *PacketSize =
537       dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 2));
538   ConstantInt *PacketAlign =
539       dyn_cast<ConstantInt>(CI->getArgOperand(NumArg - 1));
540   if (!PacketSize || !PacketAlign)
541     return false;
542 
543   unsigned Size = PacketSize->getZExtValue();
544   Align Alignment = PacketAlign->getAlignValue();
545   if (Alignment != Size)
546     return false;
547 
548   unsigned PtrArgLoc = CI->arg_size() - 3;
549   Value *PtrArg = CI->getArgOperand(PtrArgLoc);
550   Type *PtrTy = PtrArg->getType();
551 
552   SmallVector<llvm::Type *, 6> ArgTys;
553   for (unsigned I = 0; I != PtrArgLoc; ++I)
554     ArgTys.push_back(CI->getArgOperand(I)->getType());
555   ArgTys.push_back(PtrTy);
556 
557   Name = Name + "_" + std::to_string(Size);
558   auto *FTy = FunctionType::get(Callee->getReturnType(),
559                                 ArrayRef<Type *>(ArgTys), false);
560   AMDGPULibFunc NewLibFunc(Name, FTy);
561   FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
562   if (!F)
563     return false;
564 
565   auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
566   SmallVector<Value *, 6> Args;
567   for (unsigned I = 0; I != PtrArgLoc; ++I)
568     Args.push_back(CI->getArgOperand(I));
569   Args.push_back(BCast);
570 
571   auto *NCI = B.CreateCall(F, Args);
572   NCI->setAttributes(CI->getAttributes());
573   CI->replaceAllUsesWith(NCI);
574   CI->dropAllReferences();
575   CI->eraseFromParent();
576 
577   return true;
578 }
579 
580 static bool isKnownIntegral(const Value *V, const DataLayout &DL,
581                             FastMathFlags FMF) {
582   if (isa<UndefValue>(V))
583     return true;
584 
585   if (const ConstantFP *CF = dyn_cast<ConstantFP>(V))
586     return CF->getValueAPF().isInteger();
587 
588   if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) {
589     for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) {
590       Constant *ConstElt = CDV->getElementAsConstant(i);
591       if (isa<UndefValue>(ConstElt))
592         continue;
593       const ConstantFP *CFP = dyn_cast<ConstantFP>(ConstElt);
594       if (!CFP || !CFP->getValue().isInteger())
595         return false;
596     }
597 
598     return true;
599   }
600 
601   const Instruction *I = dyn_cast<Instruction>(V);
602   if (!I)
603     return false;
604 
605   switch (I->getOpcode()) {
606   case Instruction::SIToFP:
607   case Instruction::UIToFP:
608     // TODO: Could check nofpclass(inf) on incoming argument
609     if (FMF.noInfs())
610       return true;
611 
612     // Need to check int size cannot produce infinity, which computeKnownFPClass
613     // knows how to do already.
614     return isKnownNeverInfinity(I, DL);
615   case Instruction::Call: {
616     const CallInst *CI = cast<CallInst>(I);
617     switch (CI->getIntrinsicID()) {
618     case Intrinsic::trunc:
619     case Intrinsic::floor:
620     case Intrinsic::ceil:
621     case Intrinsic::rint:
622     case Intrinsic::nearbyint:
623     case Intrinsic::round:
624     case Intrinsic::roundeven:
625       return (FMF.noInfs() && FMF.noNaNs()) ||
626              isKnownNeverInfOrNaN(I, DL, nullptr);
627     default:
628       break;
629     }
630 
631     break;
632   }
633   default:
634     break;
635   }
636 
637   return false;
638 }
639 
640 // This function returns false if no change; return true otherwise.
641 bool AMDGPULibCalls::fold(CallInst *CI) {
642   Function *Callee = CI->getCalledFunction();
643   // Ignore indirect calls.
644   if (!Callee || Callee->isIntrinsic() || CI->isNoBuiltin())
645     return false;
646 
647   FuncInfo FInfo;
648   if (!parseFunctionName(Callee->getName(), FInfo))
649     return false;
650 
651   // Further check the number of arguments to see if they match.
652   // TODO: Check calling convention matches too
653   if (!FInfo.isCompatibleSignature(CI->getFunctionType()))
654     return false;
655 
656   LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << '\n');
657 
658   if (TDOFold(CI, FInfo))
659     return true;
660 
661   IRBuilder<> B(CI);
662 
663   if (FPMathOperator *FPOp = dyn_cast<FPMathOperator>(CI)) {
664     // Under unsafe-math, evaluate calls if possible.
665     // According to Brian Sumner, we can do this for all f32 function calls
666     // using host's double function calls.
667     if (canIncreasePrecisionOfConstantFold(FPOp) && evaluateCall(CI, FInfo))
668       return true;
669 
670     // Copy fast flags from the original call.
671     FastMathFlags FMF = FPOp->getFastMathFlags();
672     B.setFastMathFlags(FMF);
673 
674     // Specialized optimizations for each function call.
675     //
676     // TODO: Handle other simple intrinsic wrappers. Sqrt.
677     //
678     // TODO: Handle native functions
679     switch (FInfo.getId()) {
680     case AMDGPULibFunc::EI_EXP:
681       if (FMF.none())
682         return false;
683       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp,
684                                                   FMF.approxFunc());
685     case AMDGPULibFunc::EI_EXP2:
686       if (FMF.none())
687         return false;
688       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::exp2,
689                                                   FMF.approxFunc());
690     case AMDGPULibFunc::EI_LOG:
691       if (FMF.none())
692         return false;
693       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log,
694                                                   FMF.approxFunc());
695     case AMDGPULibFunc::EI_LOG2:
696       if (FMF.none())
697         return false;
698       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log2,
699                                                   FMF.approxFunc());
700     case AMDGPULibFunc::EI_LOG10:
701       if (FMF.none())
702         return false;
703       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::log10,
704                                                   FMF.approxFunc());
705     case AMDGPULibFunc::EI_FMIN:
706       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::minnum,
707                                                   true, true);
708     case AMDGPULibFunc::EI_FMAX:
709       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::maxnum,
710                                                   true, true);
711     case AMDGPULibFunc::EI_FMA:
712       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fma, true,
713                                                   true);
714     case AMDGPULibFunc::EI_MAD:
715       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fmuladd,
716                                                   true, true);
717     case AMDGPULibFunc::EI_FABS:
718       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::fabs, true,
719                                                   true, true);
720     case AMDGPULibFunc::EI_COPYSIGN:
721       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::copysign,
722                                                   true, true, true);
723     case AMDGPULibFunc::EI_FLOOR:
724       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::floor, true,
725                                                   true);
726     case AMDGPULibFunc::EI_CEIL:
727       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::ceil, true,
728                                                   true);
729     case AMDGPULibFunc::EI_TRUNC:
730       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::trunc, true,
731                                                   true);
732     case AMDGPULibFunc::EI_RINT:
733       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::rint, true,
734                                                   true);
735     case AMDGPULibFunc::EI_ROUND:
736       return tryReplaceLibcallWithSimpleIntrinsic(B, CI, Intrinsic::round, true,
737                                                   true);
738     case AMDGPULibFunc::EI_LDEXP: {
739       if (!shouldReplaceLibcallWithIntrinsic(CI, true, true))
740         return false;
741 
742       Value *Arg1 = CI->getArgOperand(1);
743       if (VectorType *VecTy = dyn_cast<VectorType>(CI->getType());
744           VecTy && !isa<VectorType>(Arg1->getType())) {
745         Value *SplatArg1 = B.CreateVectorSplat(VecTy->getElementCount(), Arg1);
746         CI->setArgOperand(1, SplatArg1);
747       }
748 
749       CI->setCalledFunction(Intrinsic::getDeclaration(
750           CI->getModule(), Intrinsic::ldexp,
751           {CI->getType(), CI->getArgOperand(1)->getType()}));
752       return true;
753     }
754     case AMDGPULibFunc::EI_POW: {
755       Module *M = Callee->getParent();
756       AMDGPULibFunc PowrInfo(AMDGPULibFunc::EI_POWR, FInfo);
757       FunctionCallee PowrFunc = getFunction(M, PowrInfo);
758       CallInst *Call = cast<CallInst>(FPOp);
759 
760       // pow(x, y) -> powr(x, y) for x >= -0.0
761       // TODO: Account for flags on current call
762       if (PowrFunc &&
763           cannotBeOrderedLessThanZero(FPOp->getOperand(0), M->getDataLayout(),
764                                       TLInfo, 0, AC, Call, DT)) {
765         Call->setCalledFunction(PowrFunc);
766         return fold_pow(FPOp, B, PowrInfo) || true;
767       }
768 
769       // pow(x, y) -> pown(x, y) for known integral y
770       if (isKnownIntegral(FPOp->getOperand(1), M->getDataLayout(),
771                           FPOp->getFastMathFlags())) {
772         FunctionType *PownType = getPownType(CI->getFunctionType());
773         AMDGPULibFunc PownInfo(AMDGPULibFunc::EI_POWN, PownType, true);
774         FunctionCallee PownFunc = getFunction(M, PownInfo);
775         if (PownFunc) {
776           // TODO: If the incoming integral value is an sitofp/uitofp, it won't
777           // fold out without a known range. We can probably take the source
778           // value directly.
779           Value *CastedArg =
780               B.CreateFPToSI(FPOp->getOperand(1), PownType->getParamType(1));
781           // Have to drop any nofpclass attributes on the original call site.
782           Call->removeParamAttrs(
783               1, AttributeFuncs::typeIncompatible(CastedArg->getType()));
784           Call->setCalledFunction(PownFunc);
785           Call->setArgOperand(1, CastedArg);
786           return fold_pow(FPOp, B, PownInfo) || true;
787         }
788       }
789 
790       return fold_pow(FPOp, B, FInfo);
791     }
792     case AMDGPULibFunc::EI_POWR:
793     case AMDGPULibFunc::EI_POWN:
794       return fold_pow(FPOp, B, FInfo);
795     case AMDGPULibFunc::EI_ROOTN:
796       return fold_rootn(FPOp, B, FInfo);
797     case AMDGPULibFunc::EI_SQRT:
798       return fold_sqrt(FPOp, B, FInfo);
799     case AMDGPULibFunc::EI_COS:
800     case AMDGPULibFunc::EI_SIN:
801       return fold_sincos(FPOp, B, FInfo);
802     default:
803       break;
804     }
805   } else {
806     // Specialized optimizations for each function call
807     switch (FInfo.getId()) {
808     case AMDGPULibFunc::EI_READ_PIPE_2:
809     case AMDGPULibFunc::EI_READ_PIPE_4:
810     case AMDGPULibFunc::EI_WRITE_PIPE_2:
811     case AMDGPULibFunc::EI_WRITE_PIPE_4:
812       return fold_read_write_pipe(CI, B, FInfo);
813     default:
814       break;
815     }
816   }
817 
818   return false;
819 }
820 
821 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
822   // Table-Driven optimization
823   const TableRef tr = getOptTable(FInfo.getId());
824   if (tr.empty())
825     return false;
826 
827   int const sz = (int)tr.size();
828   Value *opr0 = CI->getArgOperand(0);
829 
830   if (getVecSize(FInfo) > 1) {
831     if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
832       SmallVector<double, 0> DVal;
833       for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
834         ConstantFP *eltval = dyn_cast<ConstantFP>(
835                                CV->getElementAsConstant((unsigned)eltNo));
836         assert(eltval && "Non-FP arguments in math function!");
837         bool found = false;
838         for (int i=0; i < sz; ++i) {
839           if (eltval->isExactlyValue(tr[i].input)) {
840             DVal.push_back(tr[i].result);
841             found = true;
842             break;
843           }
844         }
845         if (!found) {
846           // This vector constants not handled yet.
847           return false;
848         }
849       }
850       LLVMContext &context = CI->getParent()->getParent()->getContext();
851       Constant *nval;
852       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
853         SmallVector<float, 0> FVal;
854         for (unsigned i = 0; i < DVal.size(); ++i) {
855           FVal.push_back((float)DVal[i]);
856         }
857         ArrayRef<float> tmp(FVal);
858         nval = ConstantDataVector::get(context, tmp);
859       } else { // F64
860         ArrayRef<double> tmp(DVal);
861         nval = ConstantDataVector::get(context, tmp);
862       }
863       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
864       replaceCall(CI, nval);
865       return true;
866     }
867   } else {
868     // Scalar version
869     if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
870       for (int i = 0; i < sz; ++i) {
871         if (CF->isExactlyValue(tr[i].input)) {
872           Value *nval = ConstantFP::get(CF->getType(), tr[i].result);
873           LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
874           replaceCall(CI, nval);
875           return true;
876         }
877       }
878     }
879   }
880 
881   return false;
882 }
883 
884 namespace llvm {
885 static double log2(double V) {
886 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L
887   return ::log2(V);
888 #else
889   return log(V) / numbers::ln2;
890 #endif
891 }
892 }
893 
894 bool AMDGPULibCalls::fold_pow(FPMathOperator *FPOp, IRBuilder<> &B,
895                               const FuncInfo &FInfo) {
896   assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
897           FInfo.getId() == AMDGPULibFunc::EI_POWR ||
898           FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
899          "fold_pow: encounter a wrong function call");
900 
901   Module *M = B.GetInsertBlock()->getModule();
902   Type *eltType = FPOp->getType()->getScalarType();
903   Value *opr0 = FPOp->getOperand(0);
904   Value *opr1 = FPOp->getOperand(1);
905 
906   const APFloat *CF = nullptr;
907   const APInt *CINT = nullptr;
908   if (!match(opr1, m_APFloatAllowUndef(CF)))
909     match(opr1, m_APIntAllowUndef(CINT));
910 
911   // 0x1111111 means that we don't do anything for this call.
912   int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
913 
914   if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0)) {
915     //  pow/powr/pown(x, 0) == 1
916     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1\n");
917     Constant *cnval = ConstantFP::get(eltType, 1.0);
918     if (getVecSize(FInfo) > 1) {
919       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
920     }
921     replaceCall(FPOp, cnval);
922     return true;
923   }
924   if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
925     // pow/powr/pown(x, 1.0) = x
926     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n");
927     replaceCall(FPOp, opr0);
928     return true;
929   }
930   if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
931     // pow/powr/pown(x, 2.0) = x*x
932     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << " * "
933                       << *opr0 << "\n");
934     Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
935     replaceCall(FPOp, nval);
936     return true;
937   }
938   if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
939     // pow/powr/pown(x, -1.0) = 1.0/x
940     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1 / " << *opr0 << "\n");
941     Constant *cnval = ConstantFP::get(eltType, 1.0);
942     if (getVecSize(FInfo) > 1) {
943       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
944     }
945     Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
946     replaceCall(FPOp, nval);
947     return true;
948   }
949 
950   if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
951     // pow[r](x, [-]0.5) = sqrt(x)
952     bool issqrt = CF->isExactlyValue(0.5);
953     if (FunctionCallee FPExpr =
954             getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
955                                                 : AMDGPULibFunc::EI_RSQRT,
956                                          FInfo))) {
957       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << FInfo.getName()
958                         << '(' << *opr0 << ")\n");
959       Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
960                                                         : "__pow2rsqrt");
961       replaceCall(FPOp, nval);
962       return true;
963     }
964   }
965 
966   if (!isUnsafeFiniteOnlyMath(FPOp))
967     return false;
968 
969   // Unsafe Math optimization
970 
971   // Remember that ci_opr1 is set if opr1 is integral
972   if (CF) {
973     double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
974                       ? (double)CF->convertToFloat()
975                       : CF->convertToDouble();
976     int ival = (int)dval;
977     if ((double)ival == dval) {
978       ci_opr1 = ival;
979     } else
980       ci_opr1 = 0x11111111;
981   }
982 
983   // pow/powr/pown(x, c) = [1/](x*x*..x); where
984   //   trunc(c) == c && the number of x == c && |c| <= 12
985   unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
986   if (abs_opr1 <= 12) {
987     Constant *cnval;
988     Value *nval;
989     if (abs_opr1 == 0) {
990       cnval = ConstantFP::get(eltType, 1.0);
991       if (getVecSize(FInfo) > 1) {
992         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
993       }
994       nval = cnval;
995     } else {
996       Value *valx2 = nullptr;
997       nval = nullptr;
998       while (abs_opr1 > 0) {
999         valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
1000         if (abs_opr1 & 1) {
1001           nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
1002         }
1003         abs_opr1 >>= 1;
1004       }
1005     }
1006 
1007     if (ci_opr1 < 0) {
1008       cnval = ConstantFP::get(eltType, 1.0);
1009       if (getVecSize(FInfo) > 1) {
1010         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
1011       }
1012       nval = B.CreateFDiv(cnval, nval, "__1powprod");
1013     }
1014     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1015                       << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
1016                       << ")\n");
1017     replaceCall(FPOp, nval);
1018     return true;
1019   }
1020 
1021   // If we should use the generic intrinsic instead of emitting a libcall
1022   const bool ShouldUseIntrinsic = eltType->isFloatTy() || eltType->isHalfTy();
1023 
1024   // powr ---> exp2(y * log2(x))
1025   // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
1026   FunctionCallee ExpExpr;
1027   if (ShouldUseIntrinsic)
1028     ExpExpr = Intrinsic::getDeclaration(M, Intrinsic::exp2, {FPOp->getType()});
1029   else {
1030     ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo));
1031     if (!ExpExpr)
1032       return false;
1033   }
1034 
1035   bool needlog = false;
1036   bool needabs = false;
1037   bool needcopysign = false;
1038   Constant *cnval = nullptr;
1039   if (getVecSize(FInfo) == 1) {
1040     CF = nullptr;
1041     match(opr0, m_APFloatAllowUndef(CF));
1042 
1043     if (CF) {
1044       double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1045                      ? (double)CF->convertToFloat()
1046                      : CF->convertToDouble();
1047 
1048       V = log2(std::abs(V));
1049       cnval = ConstantFP::get(eltType, V);
1050       needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
1051                      CF->isNegative();
1052     } else {
1053       needlog = true;
1054       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR &&
1055                                (!CF || CF->isNegative());
1056     }
1057   } else {
1058     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
1059 
1060     if (!CDV) {
1061       needlog = true;
1062       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1063     } else {
1064       assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
1065               "Wrong vector size detected");
1066 
1067       SmallVector<double, 0> DVal;
1068       for (int i=0; i < getVecSize(FInfo); ++i) {
1069         double V = CDV->getElementAsAPFloat(i).convertToDouble();
1070         if (V < 0.0) needcopysign = true;
1071         V = log2(std::abs(V));
1072         DVal.push_back(V);
1073       }
1074       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1075         SmallVector<float, 0> FVal;
1076         for (unsigned i=0; i < DVal.size(); ++i) {
1077           FVal.push_back((float)DVal[i]);
1078         }
1079         ArrayRef<float> tmp(FVal);
1080         cnval = ConstantDataVector::get(M->getContext(), tmp);
1081       } else {
1082         ArrayRef<double> tmp(DVal);
1083         cnval = ConstantDataVector::get(M->getContext(), tmp);
1084       }
1085     }
1086   }
1087 
1088   if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1089     // We cannot handle corner cases for a general pow() function, give up
1090     // unless y is a constant integral value. Then proceed as if it were pown.
1091     if (!isKnownIntegral(opr1, M->getDataLayout(), FPOp->getFastMathFlags()))
1092       return false;
1093   }
1094 
1095   Value *nval;
1096   if (needabs) {
1097     nval = B.CreateUnaryIntrinsic(Intrinsic::fabs, opr0, nullptr, "__fabs");
1098   } else {
1099     nval = cnval ? cnval : opr0;
1100   }
1101   if (needlog) {
1102     FunctionCallee LogExpr;
1103     if (ShouldUseIntrinsic) {
1104       LogExpr =
1105           Intrinsic::getDeclaration(M, Intrinsic::log2, {FPOp->getType()});
1106     } else {
1107       LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo));
1108       if (!LogExpr)
1109         return false;
1110     }
1111 
1112     nval = CreateCallEx(B,LogExpr, nval, "__log2");
1113   }
1114 
1115   if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1116     // convert int(32) to fp(f32 or f64)
1117     opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1118   }
1119   nval = B.CreateFMul(opr1, nval, "__ylogx");
1120   nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1121 
1122   if (needcopysign) {
1123     Value *opr_n;
1124     Type* rTy = opr0->getType();
1125     Type* nTyS = B.getIntNTy(eltType->getPrimitiveSizeInBits());
1126     Type *nTy = nTyS;
1127     if (const auto *vTy = dyn_cast<FixedVectorType>(rTy))
1128       nTy = FixedVectorType::get(nTyS, vTy);
1129     unsigned size = nTy->getScalarSizeInBits();
1130     opr_n = FPOp->getOperand(1);
1131     if (opr_n->getType()->isIntegerTy())
1132       opr_n = B.CreateZExtOrTrunc(opr_n, nTy, "__ytou");
1133     else
1134       opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1135 
1136     Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1137     sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1138     nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1139     nval = B.CreateBitCast(nval, opr0->getType());
1140   }
1141 
1142   LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1143                     << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1144   replaceCall(FPOp, nval);
1145 
1146   return true;
1147 }
1148 
1149 bool AMDGPULibCalls::fold_rootn(FPMathOperator *FPOp, IRBuilder<> &B,
1150                                 const FuncInfo &FInfo) {
1151   // skip vector function
1152   if (getVecSize(FInfo) != 1)
1153     return false;
1154 
1155   Value *opr0 = FPOp->getOperand(0);
1156   Value *opr1 = FPOp->getOperand(1);
1157 
1158   ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1159   if (!CINT) {
1160     return false;
1161   }
1162   int ci_opr1 = (int)CINT->getSExtValue();
1163   if (ci_opr1 == 1) {  // rootn(x, 1) = x
1164     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> " << *opr0 << "\n");
1165     replaceCall(FPOp, opr0);
1166     return true;
1167   }
1168 
1169   Module *M = B.GetInsertBlock()->getModule();
1170   if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x)
1171     if (FunctionCallee FPExpr =
1172             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1173       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> sqrt(" << *opr0
1174                         << ")\n");
1175       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1176       replaceCall(FPOp, nval);
1177       return true;
1178     }
1179   } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1180     if (FunctionCallee FPExpr =
1181             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1182       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> cbrt(" << *opr0
1183                         << ")\n");
1184       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1185       replaceCall(FPOp, nval);
1186       return true;
1187     }
1188   } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1189     LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> 1.0 / " << *opr0 << "\n");
1190     Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1191                                opr0,
1192                                "__rootn2div");
1193     replaceCall(FPOp, nval);
1194     return true;
1195   } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x)
1196     if (FunctionCallee FPExpr =
1197             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) {
1198       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> rsqrt(" << *opr0
1199                         << ")\n");
1200       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1201       replaceCall(FPOp, nval);
1202       return true;
1203     }
1204   }
1205   return false;
1206 }
1207 
1208 // Get a scalar native builtin single argument FP function
1209 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1210                                                  const FuncInfo &FInfo) {
1211   if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1212     return nullptr;
1213   FuncInfo nf = FInfo;
1214   nf.setPrefix(AMDGPULibFunc::NATIVE);
1215   return getFunction(M, nf);
1216 }
1217 
1218 // Some library calls are just wrappers around llvm intrinsics, but compiled
1219 // conservatively. Preserve the flags from the original call site by
1220 // substituting them with direct calls with all the flags.
1221 bool AMDGPULibCalls::shouldReplaceLibcallWithIntrinsic(const CallInst *CI,
1222                                                        bool AllowMinSizeF32,
1223                                                        bool AllowF64,
1224                                                        bool AllowStrictFP) {
1225   Type *FltTy = CI->getType()->getScalarType();
1226   const bool IsF32 = FltTy->isFloatTy();
1227 
1228   // f64 intrinsics aren't implemented for most operations.
1229   if (!IsF32 && !FltTy->isHalfTy() && (!AllowF64 || !FltTy->isDoubleTy()))
1230     return false;
1231 
1232   // We're implicitly inlining by replacing the libcall with the intrinsic, so
1233   // don't do it for noinline call sites.
1234   if (CI->isNoInline())
1235     return false;
1236 
1237   const Function *ParentF = CI->getFunction();
1238   // TODO: Handle strictfp
1239   if (!AllowStrictFP && ParentF->hasFnAttribute(Attribute::StrictFP))
1240     return false;
1241 
1242   if (IsF32 && !AllowMinSizeF32 && ParentF->hasMinSize())
1243     return false;
1244   return true;
1245 }
1246 
1247 void AMDGPULibCalls::replaceLibCallWithSimpleIntrinsic(IRBuilder<> &B,
1248                                                        CallInst *CI,
1249                                                        Intrinsic::ID IntrID) {
1250   if (CI->arg_size() == 2) {
1251     Value *Arg0 = CI->getArgOperand(0);
1252     Value *Arg1 = CI->getArgOperand(1);
1253     VectorType *Arg0VecTy = dyn_cast<VectorType>(Arg0->getType());
1254     VectorType *Arg1VecTy = dyn_cast<VectorType>(Arg1->getType());
1255     if (Arg0VecTy && !Arg1VecTy) {
1256       Value *SplatRHS = B.CreateVectorSplat(Arg0VecTy->getElementCount(), Arg1);
1257       CI->setArgOperand(1, SplatRHS);
1258     } else if (!Arg0VecTy && Arg1VecTy) {
1259       Value *SplatLHS = B.CreateVectorSplat(Arg1VecTy->getElementCount(), Arg0);
1260       CI->setArgOperand(0, SplatLHS);
1261     }
1262   }
1263 
1264   CI->setCalledFunction(
1265       Intrinsic::getDeclaration(CI->getModule(), IntrID, {CI->getType()}));
1266 }
1267 
1268 bool AMDGPULibCalls::tryReplaceLibcallWithSimpleIntrinsic(
1269     IRBuilder<> &B, CallInst *CI, Intrinsic::ID IntrID, bool AllowMinSizeF32,
1270     bool AllowF64, bool AllowStrictFP) {
1271   if (!shouldReplaceLibcallWithIntrinsic(CI, AllowMinSizeF32, AllowF64,
1272                                          AllowStrictFP))
1273     return false;
1274   replaceLibCallWithSimpleIntrinsic(B, CI, IntrID);
1275   return true;
1276 }
1277 
1278 // fold sqrt -> native_sqrt (x)
1279 bool AMDGPULibCalls::fold_sqrt(FPMathOperator *FPOp, IRBuilder<> &B,
1280                                const FuncInfo &FInfo) {
1281   if (!isUnsafeMath(FPOp))
1282     return false;
1283 
1284   if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1285       (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1286     Module *M = B.GetInsertBlock()->getModule();
1287 
1288     if (FunctionCallee FPExpr = getNativeFunction(
1289             M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1290       Value *opr0 = FPOp->getOperand(0);
1291       LLVM_DEBUG(errs() << "AMDIC: " << *FPOp << " ---> "
1292                         << "sqrt(" << *opr0 << ")\n");
1293       Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1294       replaceCall(FPOp, nval);
1295       return true;
1296     }
1297   }
1298   return false;
1299 }
1300 
1301 std::tuple<Value *, Value *, Value *>
1302 AMDGPULibCalls::insertSinCos(Value *Arg, FastMathFlags FMF, IRBuilder<> &B,
1303                              FunctionCallee Fsincos) {
1304   DebugLoc DL = B.getCurrentDebugLocation();
1305   Function *F = B.GetInsertBlock()->getParent();
1306   B.SetInsertPointPastAllocas(F);
1307 
1308   AllocaInst *Alloc = B.CreateAlloca(Arg->getType(), nullptr, "__sincos_");
1309 
1310   if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) {
1311     // If the argument is an instruction, it must dominate all uses so put our
1312     // sincos call there. Otherwise, right after the allocas works well enough
1313     // if it's an argument or constant.
1314 
1315     B.SetInsertPoint(ArgInst->getParent(), ++ArgInst->getIterator());
1316 
1317     // SetInsertPoint unwelcomely always tries to set the debug loc.
1318     B.SetCurrentDebugLocation(DL);
1319   }
1320 
1321   Type *CosPtrTy = Fsincos.getFunctionType()->getParamType(1);
1322 
1323   // The allocaInst allocates the memory in private address space. This need
1324   // to be addrspacecasted to point to the address space of cos pointer type.
1325   // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1326   Value *CastAlloc = B.CreateAddrSpaceCast(Alloc, CosPtrTy);
1327 
1328   CallInst *SinCos = CreateCallEx2(B, Fsincos, Arg, CastAlloc);
1329 
1330   // TODO: Is it worth trying to preserve the location for the cos calls for the
1331   // load?
1332 
1333   LoadInst *LoadCos = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1334   return {SinCos, LoadCos, SinCos};
1335 }
1336 
1337 // fold sin, cos -> sincos.
1338 bool AMDGPULibCalls::fold_sincos(FPMathOperator *FPOp, IRBuilder<> &B,
1339                                  const FuncInfo &fInfo) {
1340   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1341          fInfo.getId() == AMDGPULibFunc::EI_COS);
1342 
1343   if ((getArgType(fInfo) != AMDGPULibFunc::F32 &&
1344        getArgType(fInfo) != AMDGPULibFunc::F64) ||
1345       fInfo.getPrefix() != AMDGPULibFunc::NOPFX)
1346     return false;
1347 
1348   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1349 
1350   Value *CArgVal = FPOp->getOperand(0);
1351   CallInst *CI = cast<CallInst>(FPOp);
1352 
1353   Function *F = B.GetInsertBlock()->getParent();
1354   Module *M = F->getParent();
1355 
1356   // Merge the sin and cos. For OpenCL 2.0, there may only be a generic pointer
1357   // implementation. Prefer the private form if available.
1358   AMDGPULibFunc SinCosLibFuncPrivate(AMDGPULibFunc::EI_SINCOS, fInfo);
1359   SinCosLibFuncPrivate.getLeads()[0].PtrKind =
1360       AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::PRIVATE_ADDRESS);
1361 
1362   AMDGPULibFunc SinCosLibFuncGeneric(AMDGPULibFunc::EI_SINCOS, fInfo);
1363   SinCosLibFuncGeneric.getLeads()[0].PtrKind =
1364       AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1365 
1366   FunctionCallee FSinCosPrivate = getFunction(M, SinCosLibFuncPrivate);
1367   FunctionCallee FSinCosGeneric = getFunction(M, SinCosLibFuncGeneric);
1368   FunctionCallee FSinCos = FSinCosPrivate ? FSinCosPrivate : FSinCosGeneric;
1369   if (!FSinCos)
1370     return false;
1371 
1372   SmallVector<CallInst *> SinCalls;
1373   SmallVector<CallInst *> CosCalls;
1374   SmallVector<CallInst *> SinCosCalls;
1375   FuncInfo PartnerInfo(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN,
1376                        fInfo);
1377   const std::string PairName = PartnerInfo.mangle();
1378 
1379   StringRef SinName = isSin ? CI->getCalledFunction()->getName() : PairName;
1380   StringRef CosName = isSin ? PairName : CI->getCalledFunction()->getName();
1381   const std::string SinCosPrivateName = SinCosLibFuncPrivate.mangle();
1382   const std::string SinCosGenericName = SinCosLibFuncGeneric.mangle();
1383 
1384   // Intersect the two sets of flags.
1385   FastMathFlags FMF = FPOp->getFastMathFlags();
1386   MDNode *FPMath = CI->getMetadata(LLVMContext::MD_fpmath);
1387 
1388   SmallVector<DILocation *> MergeDbgLocs = {CI->getDebugLoc()};
1389 
1390   for (User* U : CArgVal->users()) {
1391     CallInst *XI = dyn_cast<CallInst>(U);
1392     if (!XI || XI->getFunction() != F || XI->isNoBuiltin())
1393       continue;
1394 
1395     Function *UCallee = XI->getCalledFunction();
1396     if (!UCallee)
1397       continue;
1398 
1399     bool Handled = true;
1400 
1401     if (UCallee->getName() == SinName)
1402       SinCalls.push_back(XI);
1403     else if (UCallee->getName() == CosName)
1404       CosCalls.push_back(XI);
1405     else if (UCallee->getName() == SinCosPrivateName ||
1406              UCallee->getName() == SinCosGenericName)
1407       SinCosCalls.push_back(XI);
1408     else
1409       Handled = false;
1410 
1411     if (Handled) {
1412       MergeDbgLocs.push_back(XI->getDebugLoc());
1413       auto *OtherOp = cast<FPMathOperator>(XI);
1414       FMF &= OtherOp->getFastMathFlags();
1415       FPMath = MDNode::getMostGenericFPMath(
1416           FPMath, XI->getMetadata(LLVMContext::MD_fpmath));
1417     }
1418   }
1419 
1420   if (SinCalls.empty() || CosCalls.empty())
1421     return false;
1422 
1423   B.setFastMathFlags(FMF);
1424   B.setDefaultFPMathTag(FPMath);
1425   DILocation *DbgLoc = DILocation::getMergedLocations(MergeDbgLocs);
1426   B.SetCurrentDebugLocation(DbgLoc);
1427 
1428   auto [Sin, Cos, SinCos] = insertSinCos(CArgVal, FMF, B, FSinCos);
1429 
1430   auto replaceTrigInsts = [](ArrayRef<CallInst *> Calls, Value *Res) {
1431     for (CallInst *C : Calls)
1432       C->replaceAllUsesWith(Res);
1433 
1434     // Leave the other dead instructions to avoid clobbering iterators.
1435   };
1436 
1437   replaceTrigInsts(SinCalls, Sin);
1438   replaceTrigInsts(CosCalls, Cos);
1439   replaceTrigInsts(SinCosCalls, SinCos);
1440 
1441   // It's safe to delete the original now.
1442   CI->eraseFromParent();
1443   return true;
1444 }
1445 
1446 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo, double &Res0,
1447                                             double &Res1, Constant *copr0,
1448                                             Constant *copr1) {
1449   // By default, opr0/opr1/opr3 holds values of float/double type.
1450   // If they are not float/double, each function has to its
1451   // operand separately.
1452   double opr0 = 0.0, opr1 = 0.0;
1453   ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1454   ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1455   if (fpopr0) {
1456     opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1457              ? fpopr0->getValueAPF().convertToDouble()
1458              : (double)fpopr0->getValueAPF().convertToFloat();
1459   }
1460 
1461   if (fpopr1) {
1462     opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1463              ? fpopr1->getValueAPF().convertToDouble()
1464              : (double)fpopr1->getValueAPF().convertToFloat();
1465   }
1466 
1467   switch (FInfo.getId()) {
1468   default : return false;
1469 
1470   case AMDGPULibFunc::EI_ACOS:
1471     Res0 = acos(opr0);
1472     return true;
1473 
1474   case AMDGPULibFunc::EI_ACOSH:
1475     // acosh(x) == log(x + sqrt(x*x - 1))
1476     Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1477     return true;
1478 
1479   case AMDGPULibFunc::EI_ACOSPI:
1480     Res0 = acos(opr0) / MATH_PI;
1481     return true;
1482 
1483   case AMDGPULibFunc::EI_ASIN:
1484     Res0 = asin(opr0);
1485     return true;
1486 
1487   case AMDGPULibFunc::EI_ASINH:
1488     // asinh(x) == log(x + sqrt(x*x + 1))
1489     Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1490     return true;
1491 
1492   case AMDGPULibFunc::EI_ASINPI:
1493     Res0 = asin(opr0) / MATH_PI;
1494     return true;
1495 
1496   case AMDGPULibFunc::EI_ATAN:
1497     Res0 = atan(opr0);
1498     return true;
1499 
1500   case AMDGPULibFunc::EI_ATANH:
1501     // atanh(x) == (log(x+1) - log(x-1))/2;
1502     Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1503     return true;
1504 
1505   case AMDGPULibFunc::EI_ATANPI:
1506     Res0 = atan(opr0) / MATH_PI;
1507     return true;
1508 
1509   case AMDGPULibFunc::EI_CBRT:
1510     Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1511     return true;
1512 
1513   case AMDGPULibFunc::EI_COS:
1514     Res0 = cos(opr0);
1515     return true;
1516 
1517   case AMDGPULibFunc::EI_COSH:
1518     Res0 = cosh(opr0);
1519     return true;
1520 
1521   case AMDGPULibFunc::EI_COSPI:
1522     Res0 = cos(MATH_PI * opr0);
1523     return true;
1524 
1525   case AMDGPULibFunc::EI_EXP:
1526     Res0 = exp(opr0);
1527     return true;
1528 
1529   case AMDGPULibFunc::EI_EXP2:
1530     Res0 = pow(2.0, opr0);
1531     return true;
1532 
1533   case AMDGPULibFunc::EI_EXP10:
1534     Res0 = pow(10.0, opr0);
1535     return true;
1536 
1537   case AMDGPULibFunc::EI_LOG:
1538     Res0 = log(opr0);
1539     return true;
1540 
1541   case AMDGPULibFunc::EI_LOG2:
1542     Res0 = log(opr0) / log(2.0);
1543     return true;
1544 
1545   case AMDGPULibFunc::EI_LOG10:
1546     Res0 = log(opr0) / log(10.0);
1547     return true;
1548 
1549   case AMDGPULibFunc::EI_RSQRT:
1550     Res0 = 1.0 / sqrt(opr0);
1551     return true;
1552 
1553   case AMDGPULibFunc::EI_SIN:
1554     Res0 = sin(opr0);
1555     return true;
1556 
1557   case AMDGPULibFunc::EI_SINH:
1558     Res0 = sinh(opr0);
1559     return true;
1560 
1561   case AMDGPULibFunc::EI_SINPI:
1562     Res0 = sin(MATH_PI * opr0);
1563     return true;
1564 
1565   case AMDGPULibFunc::EI_TAN:
1566     Res0 = tan(opr0);
1567     return true;
1568 
1569   case AMDGPULibFunc::EI_TANH:
1570     Res0 = tanh(opr0);
1571     return true;
1572 
1573   case AMDGPULibFunc::EI_TANPI:
1574     Res0 = tan(MATH_PI * opr0);
1575     return true;
1576 
1577   // two-arg functions
1578   case AMDGPULibFunc::EI_POW:
1579   case AMDGPULibFunc::EI_POWR:
1580     Res0 = pow(opr0, opr1);
1581     return true;
1582 
1583   case AMDGPULibFunc::EI_POWN: {
1584     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1585       double val = (double)iopr1->getSExtValue();
1586       Res0 = pow(opr0, val);
1587       return true;
1588     }
1589     return false;
1590   }
1591 
1592   case AMDGPULibFunc::EI_ROOTN: {
1593     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1594       double val = (double)iopr1->getSExtValue();
1595       Res0 = pow(opr0, 1.0 / val);
1596       return true;
1597     }
1598     return false;
1599   }
1600 
1601   // with ptr arg
1602   case AMDGPULibFunc::EI_SINCOS:
1603     Res0 = sin(opr0);
1604     Res1 = cos(opr0);
1605     return true;
1606   }
1607 
1608   return false;
1609 }
1610 
1611 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) {
1612   int numArgs = (int)aCI->arg_size();
1613   if (numArgs > 3)
1614     return false;
1615 
1616   Constant *copr0 = nullptr;
1617   Constant *copr1 = nullptr;
1618   if (numArgs > 0) {
1619     if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1620       return false;
1621   }
1622 
1623   if (numArgs > 1) {
1624     if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1625       if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1626         return false;
1627     }
1628   }
1629 
1630   // At this point, all arguments to aCI are constants.
1631 
1632   // max vector size is 16, and sincos will generate two results.
1633   double DVal0[16], DVal1[16];
1634   int FuncVecSize = getVecSize(FInfo);
1635   bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1636   if (FuncVecSize == 1) {
1637     if (!evaluateScalarMathFunc(FInfo, DVal0[0], DVal1[0], copr0, copr1)) {
1638       return false;
1639     }
1640   } else {
1641     ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1642     ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1643     for (int i = 0; i < FuncVecSize; ++i) {
1644       Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1645       Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1646       if (!evaluateScalarMathFunc(FInfo, DVal0[i], DVal1[i], celt0, celt1)) {
1647         return false;
1648       }
1649     }
1650   }
1651 
1652   LLVMContext &context = aCI->getContext();
1653   Constant *nval0, *nval1;
1654   if (FuncVecSize == 1) {
1655     nval0 = ConstantFP::get(aCI->getType(), DVal0[0]);
1656     if (hasTwoResults)
1657       nval1 = ConstantFP::get(aCI->getType(), DVal1[0]);
1658   } else {
1659     if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1660       SmallVector <float, 0> FVal0, FVal1;
1661       for (int i = 0; i < FuncVecSize; ++i)
1662         FVal0.push_back((float)DVal0[i]);
1663       ArrayRef<float> tmp0(FVal0);
1664       nval0 = ConstantDataVector::get(context, tmp0);
1665       if (hasTwoResults) {
1666         for (int i = 0; i < FuncVecSize; ++i)
1667           FVal1.push_back((float)DVal1[i]);
1668         ArrayRef<float> tmp1(FVal1);
1669         nval1 = ConstantDataVector::get(context, tmp1);
1670       }
1671     } else {
1672       ArrayRef<double> tmp0(DVal0);
1673       nval0 = ConstantDataVector::get(context, tmp0);
1674       if (hasTwoResults) {
1675         ArrayRef<double> tmp1(DVal1);
1676         nval1 = ConstantDataVector::get(context, tmp1);
1677       }
1678     }
1679   }
1680 
1681   if (hasTwoResults) {
1682     // sincos
1683     assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1684            "math function with ptr arg not supported yet");
1685     new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1686   }
1687 
1688   replaceCall(aCI, nval0);
1689   return true;
1690 }
1691 
1692 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F,
1693                                                   FunctionAnalysisManager &AM) {
1694   AMDGPULibCalls Simplifier;
1695   Simplifier.initNativeFuncs();
1696   Simplifier.initFunction(F, AM);
1697 
1698   bool Changed = false;
1699 
1700   LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1701              F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1702 
1703   for (auto &BB : F) {
1704     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1705       // Ignore non-calls.
1706       CallInst *CI = dyn_cast<CallInst>(I);
1707       ++I;
1708 
1709       if (CI) {
1710         if (Simplifier.fold(CI))
1711           Changed = true;
1712       }
1713     }
1714   }
1715   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1716 }
1717 
1718 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F,
1719                                                 FunctionAnalysisManager &AM) {
1720   if (UseNative.empty())
1721     return PreservedAnalyses::all();
1722 
1723   AMDGPULibCalls Simplifier;
1724   Simplifier.initNativeFuncs();
1725   Simplifier.initFunction(F, AM);
1726 
1727   bool Changed = false;
1728   for (auto &BB : F) {
1729     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1730       // Ignore non-calls.
1731       CallInst *CI = dyn_cast<CallInst>(I);
1732       ++I;
1733       if (CI && Simplifier.useNative(CI))
1734         Changed = true;
1735     }
1736   }
1737   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1738 }
1739