xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp (revision fc5121a722cb7d6ce2f10dcb698ff64dbad2eaa3)
1 //===- AMDGPULibCalls.cpp -------------------------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief This file does AMD library function optimizations.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #define DEBUG_TYPE "amdgpu-simplifylib"
16 
17 #include "AMDGPU.h"
18 #include "AMDGPULibFunc.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/Loads.h"
21 #include "llvm/ADT/StringSet.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/ValueSymbolTable.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include <vector>
34 #include <cmath>
35 
36 using namespace llvm;
37 
38 static cl::opt<bool> EnablePreLink("amdgpu-prelink",
39   cl::desc("Enable pre-link mode optimizations"),
40   cl::init(false),
41   cl::Hidden);
42 
43 static cl::list<std::string> UseNative("amdgpu-use-native",
44   cl::desc("Comma separated list of functions to replace with native, or all"),
45   cl::CommaSeparated, cl::ValueOptional,
46   cl::Hidden);
47 
48 #define MATH_PI     3.14159265358979323846264338327950288419716939937511
49 #define MATH_E      2.71828182845904523536028747135266249775724709369996
50 #define MATH_SQRT2  1.41421356237309504880168872420969807856967187537695
51 
52 #define MATH_LOG2E     1.4426950408889634073599246810018921374266459541529859
53 #define MATH_LOG10E    0.4342944819032518276511289189166050822943970058036665
54 // Value of log2(10)
55 #define MATH_LOG2_10   3.3219280948873623478703194294893901758648313930245806
56 // Value of 1 / log2(10)
57 #define MATH_RLOG2_10  0.3010299956639811952137388947244930267681898814621085
58 // Value of 1 / M_LOG2E_F = 1 / log2(e)
59 #define MATH_RLOG2_E   0.6931471805599453094172321214581765680755001343602552
60 
61 namespace llvm {
62 
63 class AMDGPULibCalls {
64 private:
65 
66   typedef llvm::AMDGPULibFunc FuncInfo;
67 
68   // -fuse-native.
69   bool AllNative = false;
70 
71   bool useNativeFunc(const StringRef F) const;
72 
73   // Return a pointer (pointer expr) to the function if function defintion with
74   // "FuncName" exists. It may create a new function prototype in pre-link mode.
75   Constant *getFunction(Module *M, const FuncInfo& fInfo);
76 
77   // Replace a normal function with its native version.
78   bool replaceWithNative(CallInst *CI, const FuncInfo &FInfo);
79 
80   bool parseFunctionName(const StringRef& FMangledName,
81                          FuncInfo *FInfo=nullptr /*out*/);
82 
83   bool TDOFold(CallInst *CI, const FuncInfo &FInfo);
84 
85   /* Specialized optimizations */
86 
87   // recip (half or native)
88   bool fold_recip(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
89 
90   // divide (half or native)
91   bool fold_divide(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
92 
93   // pow/powr/pown
94   bool fold_pow(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
95 
96   // rootn
97   bool fold_rootn(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
98 
99   // fma/mad
100   bool fold_fma_mad(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
101 
102   // -fuse-native for sincos
103   bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo);
104 
105   // evaluate calls if calls' arguments are constants.
106   bool evaluateScalarMathFunc(FuncInfo &FInfo, double& Res0,
107     double& Res1, Constant *copr0, Constant *copr1, Constant *copr2);
108   bool evaluateCall(CallInst *aCI, FuncInfo &FInfo);
109 
110   // exp
111   bool fold_exp(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
112 
113   // exp2
114   bool fold_exp2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
115 
116   // exp10
117   bool fold_exp10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
118 
119   // log
120   bool fold_log(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
121 
122   // log2
123   bool fold_log2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
124 
125   // log10
126   bool fold_log10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
127 
128   // sqrt
129   bool fold_sqrt(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
130 
131   // sin/cos
132   bool fold_sincos(CallInst * CI, IRBuilder<> &B, AliasAnalysis * AA);
133 
134   // __read_pipe/__write_pipe
135   bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, FuncInfo &FInfo);
136 
137   // Get insertion point at entry.
138   BasicBlock::iterator getEntryIns(CallInst * UI);
139   // Insert an Alloc instruction.
140   AllocaInst* insertAlloca(CallInst * UI, IRBuilder<> &B, const char *prefix);
141   // Get a scalar native builtin signle argument FP function
142   Constant* getNativeFunction(Module* M, const FuncInfo &FInfo);
143 
144 protected:
145   CallInst *CI;
146 
147   bool isUnsafeMath(const CallInst *CI) const;
148 
149   void replaceCall(Value *With) {
150     CI->replaceAllUsesWith(With);
151     CI->eraseFromParent();
152   }
153 
154 public:
155   bool fold(CallInst *CI, AliasAnalysis *AA = nullptr);
156 
157   void initNativeFuncs();
158 
159   // Replace a normal math function call with that native version
160   bool useNative(CallInst *CI);
161 };
162 
163 } // end llvm namespace
164 
165 namespace {
166 
167   class AMDGPUSimplifyLibCalls : public FunctionPass {
168 
169   AMDGPULibCalls Simplifier;
170 
171   public:
172     static char ID; // Pass identification
173 
174     AMDGPUSimplifyLibCalls() : FunctionPass(ID) {
175       initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
176     }
177 
178     void getAnalysisUsage(AnalysisUsage &AU) const override {
179       AU.addRequired<AAResultsWrapperPass>();
180     }
181 
182     bool runOnFunction(Function &M) override;
183   };
184 
185   class AMDGPUUseNativeCalls : public FunctionPass {
186 
187   AMDGPULibCalls Simplifier;
188 
189   public:
190     static char ID; // Pass identification
191 
192     AMDGPUUseNativeCalls() : FunctionPass(ID) {
193       initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry());
194       Simplifier.initNativeFuncs();
195     }
196 
197     bool runOnFunction(Function &F) override;
198   };
199 
200 } // end anonymous namespace.
201 
202 char AMDGPUSimplifyLibCalls::ID = 0;
203 char AMDGPUUseNativeCalls::ID = 0;
204 
205 INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
206                       "Simplify well-known AMD library calls", false, false)
207 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
208 INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
209                     "Simplify well-known AMD library calls", false, false)
210 
211 INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative",
212                 "Replace builtin math calls with that native versions.",
213                 false, false)
214 
215 template <typename IRB>
216 CallInst *CreateCallEx(IRB &B, Value *Callee, Value *Arg, const Twine &Name="")
217 {
218   CallInst *R = B.CreateCall(Callee, Arg, Name);
219   if (Function* F = dyn_cast<Function>(Callee))
220     R->setCallingConv(F->getCallingConv());
221   return R;
222 }
223 
224 template <typename IRB>
225 CallInst *CreateCallEx2(IRB &B, Value *Callee, Value *Arg1, Value *Arg2,
226                         const Twine &Name="") {
227   CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
228   if (Function* F = dyn_cast<Function>(Callee))
229     R->setCallingConv(F->getCallingConv());
230   return R;
231 }
232 
233 //  Data structures for table-driven optimizations.
234 //  FuncTbl works for both f32 and f64 functions with 1 input argument
235 
236 struct TableEntry {
237   double   result;
238   double   input;
239 };
240 
241 /* a list of {result, input} */
242 static const TableEntry tbl_acos[] = {
243   {MATH_PI/2.0, 0.0},
244   {MATH_PI/2.0, -0.0},
245   {0.0, 1.0},
246   {MATH_PI, -1.0}
247 };
248 static const TableEntry tbl_acosh[] = {
249   {0.0, 1.0}
250 };
251 static const TableEntry tbl_acospi[] = {
252   {0.5, 0.0},
253   {0.5, -0.0},
254   {0.0, 1.0},
255   {1.0, -1.0}
256 };
257 static const TableEntry tbl_asin[] = {
258   {0.0, 0.0},
259   {-0.0, -0.0},
260   {MATH_PI/2.0, 1.0},
261   {-MATH_PI/2.0, -1.0}
262 };
263 static const TableEntry tbl_asinh[] = {
264   {0.0, 0.0},
265   {-0.0, -0.0}
266 };
267 static const TableEntry tbl_asinpi[] = {
268   {0.0, 0.0},
269   {-0.0, -0.0},
270   {0.5, 1.0},
271   {-0.5, -1.0}
272 };
273 static const TableEntry tbl_atan[] = {
274   {0.0, 0.0},
275   {-0.0, -0.0},
276   {MATH_PI/4.0, 1.0},
277   {-MATH_PI/4.0, -1.0}
278 };
279 static const TableEntry tbl_atanh[] = {
280   {0.0, 0.0},
281   {-0.0, -0.0}
282 };
283 static const TableEntry tbl_atanpi[] = {
284   {0.0, 0.0},
285   {-0.0, -0.0},
286   {0.25, 1.0},
287   {-0.25, -1.0}
288 };
289 static const TableEntry tbl_cbrt[] = {
290   {0.0, 0.0},
291   {-0.0, -0.0},
292   {1.0, 1.0},
293   {-1.0, -1.0},
294 };
295 static const TableEntry tbl_cos[] = {
296   {1.0, 0.0},
297   {1.0, -0.0}
298 };
299 static const TableEntry tbl_cosh[] = {
300   {1.0, 0.0},
301   {1.0, -0.0}
302 };
303 static const TableEntry tbl_cospi[] = {
304   {1.0, 0.0},
305   {1.0, -0.0}
306 };
307 static const TableEntry tbl_erfc[] = {
308   {1.0, 0.0},
309   {1.0, -0.0}
310 };
311 static const TableEntry tbl_erf[] = {
312   {0.0, 0.0},
313   {-0.0, -0.0}
314 };
315 static const TableEntry tbl_exp[] = {
316   {1.0, 0.0},
317   {1.0, -0.0},
318   {MATH_E, 1.0}
319 };
320 static const TableEntry tbl_exp2[] = {
321   {1.0, 0.0},
322   {1.0, -0.0},
323   {2.0, 1.0}
324 };
325 static const TableEntry tbl_exp10[] = {
326   {1.0, 0.0},
327   {1.0, -0.0},
328   {10.0, 1.0}
329 };
330 static const TableEntry tbl_expm1[] = {
331   {0.0, 0.0},
332   {-0.0, -0.0}
333 };
334 static const TableEntry tbl_log[] = {
335   {0.0, 1.0},
336   {1.0, MATH_E}
337 };
338 static const TableEntry tbl_log2[] = {
339   {0.0, 1.0},
340   {1.0, 2.0}
341 };
342 static const TableEntry tbl_log10[] = {
343   {0.0, 1.0},
344   {1.0, 10.0}
345 };
346 static const TableEntry tbl_rsqrt[] = {
347   {1.0, 1.0},
348   {1.0/MATH_SQRT2, 2.0}
349 };
350 static const TableEntry tbl_sin[] = {
351   {0.0, 0.0},
352   {-0.0, -0.0}
353 };
354 static const TableEntry tbl_sinh[] = {
355   {0.0, 0.0},
356   {-0.0, -0.0}
357 };
358 static const TableEntry tbl_sinpi[] = {
359   {0.0, 0.0},
360   {-0.0, -0.0}
361 };
362 static const TableEntry tbl_sqrt[] = {
363   {0.0, 0.0},
364   {1.0, 1.0},
365   {MATH_SQRT2, 2.0}
366 };
367 static const TableEntry tbl_tan[] = {
368   {0.0, 0.0},
369   {-0.0, -0.0}
370 };
371 static const TableEntry tbl_tanh[] = {
372   {0.0, 0.0},
373   {-0.0, -0.0}
374 };
375 static const TableEntry tbl_tanpi[] = {
376   {0.0, 0.0},
377   {-0.0, -0.0}
378 };
379 static const TableEntry tbl_tgamma[] = {
380   {1.0, 1.0},
381   {1.0, 2.0},
382   {2.0, 3.0},
383   {6.0, 4.0}
384 };
385 
386 static bool HasNative(AMDGPULibFunc::EFuncId id) {
387   switch(id) {
388   case AMDGPULibFunc::EI_DIVIDE:
389   case AMDGPULibFunc::EI_COS:
390   case AMDGPULibFunc::EI_EXP:
391   case AMDGPULibFunc::EI_EXP2:
392   case AMDGPULibFunc::EI_EXP10:
393   case AMDGPULibFunc::EI_LOG:
394   case AMDGPULibFunc::EI_LOG2:
395   case AMDGPULibFunc::EI_LOG10:
396   case AMDGPULibFunc::EI_POWR:
397   case AMDGPULibFunc::EI_RECIP:
398   case AMDGPULibFunc::EI_RSQRT:
399   case AMDGPULibFunc::EI_SIN:
400   case AMDGPULibFunc::EI_SINCOS:
401   case AMDGPULibFunc::EI_SQRT:
402   case AMDGPULibFunc::EI_TAN:
403     return true;
404   default:;
405   }
406   return false;
407 }
408 
409 struct TableRef {
410   size_t size;
411   const TableEntry *table; // variable size: from 0 to (size - 1)
412 
413   TableRef() : size(0), table(nullptr) {}
414 
415   template <size_t N>
416   TableRef(const TableEntry (&tbl)[N]) : size(N), table(&tbl[0]) {}
417 };
418 
419 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
420   switch(id) {
421   case AMDGPULibFunc::EI_ACOS:    return TableRef(tbl_acos);
422   case AMDGPULibFunc::EI_ACOSH:   return TableRef(tbl_acosh);
423   case AMDGPULibFunc::EI_ACOSPI:  return TableRef(tbl_acospi);
424   case AMDGPULibFunc::EI_ASIN:    return TableRef(tbl_asin);
425   case AMDGPULibFunc::EI_ASINH:   return TableRef(tbl_asinh);
426   case AMDGPULibFunc::EI_ASINPI:  return TableRef(tbl_asinpi);
427   case AMDGPULibFunc::EI_ATAN:    return TableRef(tbl_atan);
428   case AMDGPULibFunc::EI_ATANH:   return TableRef(tbl_atanh);
429   case AMDGPULibFunc::EI_ATANPI:  return TableRef(tbl_atanpi);
430   case AMDGPULibFunc::EI_CBRT:    return TableRef(tbl_cbrt);
431   case AMDGPULibFunc::EI_NCOS:
432   case AMDGPULibFunc::EI_COS:     return TableRef(tbl_cos);
433   case AMDGPULibFunc::EI_COSH:    return TableRef(tbl_cosh);
434   case AMDGPULibFunc::EI_COSPI:   return TableRef(tbl_cospi);
435   case AMDGPULibFunc::EI_ERFC:    return TableRef(tbl_erfc);
436   case AMDGPULibFunc::EI_ERF:     return TableRef(tbl_erf);
437   case AMDGPULibFunc::EI_EXP:     return TableRef(tbl_exp);
438   case AMDGPULibFunc::EI_NEXP2:
439   case AMDGPULibFunc::EI_EXP2:    return TableRef(tbl_exp2);
440   case AMDGPULibFunc::EI_EXP10:   return TableRef(tbl_exp10);
441   case AMDGPULibFunc::EI_EXPM1:   return TableRef(tbl_expm1);
442   case AMDGPULibFunc::EI_LOG:     return TableRef(tbl_log);
443   case AMDGPULibFunc::EI_NLOG2:
444   case AMDGPULibFunc::EI_LOG2:    return TableRef(tbl_log2);
445   case AMDGPULibFunc::EI_LOG10:   return TableRef(tbl_log10);
446   case AMDGPULibFunc::EI_NRSQRT:
447   case AMDGPULibFunc::EI_RSQRT:   return TableRef(tbl_rsqrt);
448   case AMDGPULibFunc::EI_NSIN:
449   case AMDGPULibFunc::EI_SIN:     return TableRef(tbl_sin);
450   case AMDGPULibFunc::EI_SINH:    return TableRef(tbl_sinh);
451   case AMDGPULibFunc::EI_SINPI:   return TableRef(tbl_sinpi);
452   case AMDGPULibFunc::EI_NSQRT:
453   case AMDGPULibFunc::EI_SQRT:    return TableRef(tbl_sqrt);
454   case AMDGPULibFunc::EI_TAN:     return TableRef(tbl_tan);
455   case AMDGPULibFunc::EI_TANH:    return TableRef(tbl_tanh);
456   case AMDGPULibFunc::EI_TANPI:   return TableRef(tbl_tanpi);
457   case AMDGPULibFunc::EI_TGAMMA:  return TableRef(tbl_tgamma);
458   default:;
459   }
460   return TableRef();
461 }
462 
463 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
464   return FInfo.getLeads()[0].VectorSize;
465 }
466 
467 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
468   return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
469 }
470 
471 Constant *AMDGPULibCalls::getFunction(Module *M, const FuncInfo& fInfo) {
472   // If we are doing PreLinkOpt, the function is external. So it is safe to
473   // use getOrInsertFunction() at this stage.
474 
475   return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
476                        : AMDGPULibFunc::getFunction(M, fInfo);
477 }
478 
479 bool AMDGPULibCalls::parseFunctionName(const StringRef& FMangledName,
480                                     FuncInfo *FInfo) {
481   return AMDGPULibFunc::parse(FMangledName, *FInfo);
482 }
483 
484 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const {
485   if (auto Op = dyn_cast<FPMathOperator>(CI))
486     if (Op->hasUnsafeAlgebra())
487       return true;
488   const Function *F = CI->getParent()->getParent();
489   Attribute Attr = F->getFnAttribute("unsafe-fp-math");
490   return Attr.getValueAsString() == "true";
491 }
492 
493 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
494   return AllNative ||
495          std::find(UseNative.begin(), UseNative.end(), F) != UseNative.end();
496 }
497 
498 void AMDGPULibCalls::initNativeFuncs() {
499   AllNative = useNativeFunc("all") ||
500               (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
501                UseNative.begin()->empty());
502 }
503 
504 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
505   bool native_sin = useNativeFunc("sin");
506   bool native_cos = useNativeFunc("cos");
507 
508   if (native_sin && native_cos) {
509     Module *M = aCI->getModule();
510     Value *opr0 = aCI->getArgOperand(0);
511 
512     AMDGPULibFunc nf;
513     nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
514     nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
515 
516     nf.setPrefix(AMDGPULibFunc::NATIVE);
517     nf.setId(AMDGPULibFunc::EI_SIN);
518     Constant *sinExpr = getFunction(M, nf);
519 
520     nf.setPrefix(AMDGPULibFunc::NATIVE);
521     nf.setId(AMDGPULibFunc::EI_COS);
522     Constant *cosExpr = getFunction(M, nf);
523     if (sinExpr && cosExpr) {
524       Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
525       Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
526       new StoreInst(cosval, aCI->getArgOperand(1), aCI);
527 
528       DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
529                                           << " with native version of sin/cos");
530 
531       replaceCall(sinval);
532       return true;
533     }
534   }
535   return false;
536 }
537 
538 bool AMDGPULibCalls::useNative(CallInst *aCI) {
539   CI = aCI;
540   Function *Callee = aCI->getCalledFunction();
541 
542   FuncInfo FInfo;
543   if (!parseFunctionName(Callee->getName(), &FInfo) || !FInfo.isMangled() ||
544       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
545       getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
546       !(AllNative || useNativeFunc(FInfo.getName()))) {
547     return false;
548   }
549 
550   if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
551     return sincosUseNative(aCI, FInfo);
552 
553   FInfo.setPrefix(AMDGPULibFunc::NATIVE);
554   Constant *F = getFunction(aCI->getModule(), FInfo);
555   if (!F)
556     return false;
557 
558   aCI->setCalledFunction(F);
559   DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
560                                       << " with native version");
561   return true;
562 }
563 
564 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
565 // builtin, with appended type size and alignment arguments, where 2 or 4
566 // indicates the original number of arguments. The library has optimized version
567 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
568 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
569 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
570 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
571 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
572                                           FuncInfo &FInfo) {
573   auto *Callee = CI->getCalledFunction();
574   if (!Callee->isDeclaration())
575     return false;
576 
577   assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
578   auto *M = Callee->getParent();
579   auto &Ctx = M->getContext();
580   std::string Name = Callee->getName();
581   auto NumArg = CI->getNumArgOperands();
582   if (NumArg != 4 && NumArg != 6)
583     return false;
584   auto *PacketSize = CI->getArgOperand(NumArg - 2);
585   auto *PacketAlign = CI->getArgOperand(NumArg - 1);
586   if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign))
587     return false;
588   unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue();
589   unsigned Align = cast<ConstantInt>(PacketAlign)->getZExtValue();
590   if (Size != Align || !isPowerOf2_32(Size))
591     return false;
592 
593   Type *PtrElemTy;
594   if (Size <= 8)
595     PtrElemTy = Type::getIntNTy(Ctx, Size * 8);
596   else
597     PtrElemTy = VectorType::get(Type::getInt64Ty(Ctx), Size / 8);
598   unsigned PtrArgLoc = CI->getNumArgOperands() - 3;
599   auto PtrArg = CI->getArgOperand(PtrArgLoc);
600   unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace();
601   auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS);
602 
603   SmallVector<llvm::Type *, 6> ArgTys;
604   for (unsigned I = 0; I != PtrArgLoc; ++I)
605     ArgTys.push_back(CI->getArgOperand(I)->getType());
606   ArgTys.push_back(PtrTy);
607 
608   Name = Name + "_" + std::to_string(Size);
609   auto *FTy = FunctionType::get(Callee->getReturnType(),
610                                 ArrayRef<Type *>(ArgTys), false);
611   AMDGPULibFunc NewLibFunc(Name, FTy);
612   auto *F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
613   if (!F)
614     return false;
615 
616   auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
617   SmallVector<Value *, 6> Args;
618   for (unsigned I = 0; I != PtrArgLoc; ++I)
619     Args.push_back(CI->getArgOperand(I));
620   Args.push_back(BCast);
621 
622   auto *NCI = B.CreateCall(F, Args);
623   NCI->setAttributes(CI->getAttributes());
624   CI->replaceAllUsesWith(NCI);
625   CI->dropAllReferences();
626   CI->eraseFromParent();
627 
628   return true;
629 }
630 
631 // This function returns false if no change; return true otherwise.
632 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) {
633   this->CI = CI;
634   Function *Callee = CI->getCalledFunction();
635 
636   // Ignore indirect calls.
637   if (Callee == 0) return false;
638 
639   FuncInfo FInfo;
640   if (!parseFunctionName(Callee->getName(), &FInfo))
641     return false;
642 
643   // Further check the number of arguments to see if they match.
644   if (CI->getNumArgOperands() != FInfo.getNumArgs())
645     return false;
646 
647   BasicBlock *BB = CI->getParent();
648   LLVMContext &Context = CI->getParent()->getContext();
649   IRBuilder<> B(Context);
650 
651   // Set the builder to the instruction after the call.
652   B.SetInsertPoint(BB, CI->getIterator());
653 
654   // Copy fast flags from the original call.
655   if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI))
656     B.setFastMathFlags(FPOp->getFastMathFlags());
657 
658   if (TDOFold(CI, FInfo))
659     return true;
660 
661   // Under unsafe-math, evaluate calls if possible.
662   // According to Brian Sumner, we can do this for all f32 function calls
663   // using host's double function calls.
664   if (isUnsafeMath(CI) && evaluateCall(CI, FInfo))
665     return true;
666 
667   // Specilized optimizations for each function call
668   switch (FInfo.getId()) {
669   case AMDGPULibFunc::EI_RECIP:
670     // skip vector function
671     assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
672              FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
673             "recip must be an either native or half function");
674     return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo);
675 
676   case AMDGPULibFunc::EI_DIVIDE:
677     // skip vector function
678     assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
679              FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
680             "divide must be an either native or half function");
681     return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo);
682 
683   case AMDGPULibFunc::EI_POW:
684   case AMDGPULibFunc::EI_POWR:
685   case AMDGPULibFunc::EI_POWN:
686     return fold_pow(CI, B, FInfo);
687 
688   case AMDGPULibFunc::EI_ROOTN:
689     // skip vector function
690     return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo);
691 
692   case AMDGPULibFunc::EI_FMA:
693   case AMDGPULibFunc::EI_MAD:
694   case AMDGPULibFunc::EI_NFMA:
695     // skip vector function
696     return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo);
697 
698   case AMDGPULibFunc::EI_SQRT:
699     return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo);
700   case AMDGPULibFunc::EI_COS:
701   case AMDGPULibFunc::EI_SIN:
702     if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
703          getArgType(FInfo) == AMDGPULibFunc::F64)
704         && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX))
705       return fold_sincos(CI, B, AA);
706 
707     break;
708   case AMDGPULibFunc::EI_READ_PIPE_2:
709   case AMDGPULibFunc::EI_READ_PIPE_4:
710   case AMDGPULibFunc::EI_WRITE_PIPE_2:
711   case AMDGPULibFunc::EI_WRITE_PIPE_4:
712     return fold_read_write_pipe(CI, B, FInfo);
713 
714   default:
715     break;
716   }
717 
718   return false;
719 }
720 
721 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
722   // Table-Driven optimization
723   const TableRef tr = getOptTable(FInfo.getId());
724   if (tr.size==0)
725     return false;
726 
727   int const sz = (int)tr.size;
728   const TableEntry * const ftbl = tr.table;
729   Value *opr0 = CI->getArgOperand(0);
730 
731   if (getVecSize(FInfo) > 1) {
732     if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
733       SmallVector<double, 0> DVal;
734       for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
735         ConstantFP *eltval = dyn_cast<ConstantFP>(
736                                CV->getElementAsConstant((unsigned)eltNo));
737         assert(eltval && "Non-FP arguments in math function!");
738         bool found = false;
739         for (int i=0; i < sz; ++i) {
740           if (eltval->isExactlyValue(ftbl[i].input)) {
741             DVal.push_back(ftbl[i].result);
742             found = true;
743             break;
744           }
745         }
746         if (!found) {
747           // This vector constants not handled yet.
748           return false;
749         }
750       }
751       LLVMContext &context = CI->getParent()->getParent()->getContext();
752       Constant *nval;
753       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
754         SmallVector<float, 0> FVal;
755         for (unsigned i = 0; i < DVal.size(); ++i) {
756           FVal.push_back((float)DVal[i]);
757         }
758         ArrayRef<float> tmp(FVal);
759         nval = ConstantDataVector::get(context, tmp);
760       } else { // F64
761         ArrayRef<double> tmp(DVal);
762         nval = ConstantDataVector::get(context, tmp);
763       }
764       DEBUG(errs() << "AMDIC: " << *CI
765                    << " ---> " << *nval << "\n");
766       replaceCall(nval);
767       return true;
768     }
769   } else {
770     // Scalar version
771     if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
772       for (int i = 0; i < sz; ++i) {
773         if (CF->isExactlyValue(ftbl[i].input)) {
774           Value *nval = ConstantFP::get(CF->getType(), ftbl[i].result);
775           DEBUG(errs() << "AMDIC: " << *CI
776                        << " ---> " << *nval << "\n");
777           replaceCall(nval);
778           return true;
779         }
780       }
781     }
782   }
783 
784   return false;
785 }
786 
787 bool AMDGPULibCalls::replaceWithNative(CallInst *CI, const FuncInfo &FInfo) {
788   Module *M = CI->getModule();
789   if (getArgType(FInfo) != AMDGPULibFunc::F32 ||
790       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
791       !HasNative(FInfo.getId()))
792     return false;
793 
794   AMDGPULibFunc nf = FInfo;
795   nf.setPrefix(AMDGPULibFunc::NATIVE);
796   if (Constant *FPExpr = getFunction(M, nf)) {
797     DEBUG(dbgs() << "AMDIC: " << *CI << " ---> ");
798 
799     CI->setCalledFunction(FPExpr);
800 
801     DEBUG(dbgs() << *CI << '\n');
802 
803     return true;
804   }
805   return false;
806 }
807 
808 //  [native_]half_recip(c) ==> 1.0/c
809 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B,
810                                 const FuncInfo &FInfo) {
811   Value *opr0 = CI->getArgOperand(0);
812   if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
813     // Just create a normal div. Later, InstCombine will be able
814     // to compute the divide into a constant (avoid check float infinity
815     // or subnormal at this point).
816     Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0),
817                                opr0,
818                                "recip2div");
819     DEBUG(errs() << "AMDIC: " << *CI
820                  << " ---> " << *nval << "\n");
821     replaceCall(nval);
822     return true;
823   }
824   return false;
825 }
826 
827 //  [native_]half_divide(x, c) ==> x/c
828 bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B,
829                                  const FuncInfo &FInfo) {
830   Value *opr0 = CI->getArgOperand(0);
831   Value *opr1 = CI->getArgOperand(1);
832   ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
833   ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
834 
835   if ((CF0 && CF1) ||  // both are constants
836       (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32)))
837       // CF1 is constant && f32 divide
838   {
839     Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0),
840                                 opr1, "__div2recip");
841     Value *nval  = B.CreateFMul(opr0, nval1, "__div2mul");
842     replaceCall(nval);
843     return true;
844   }
845   return false;
846 }
847 
848 namespace llvm {
849 static double log2(double V) {
850 #if _XOPEN_SOURCE >= 600 || _ISOC99_SOURCE || _POSIX_C_SOURCE >= 200112L
851   return ::log2(V);
852 #else
853   return log(V) / 0.693147180559945309417;
854 #endif
855 }
856 }
857 
858 bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B,
859                               const FuncInfo &FInfo) {
860   assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
861           FInfo.getId() == AMDGPULibFunc::EI_POWR ||
862           FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
863          "fold_pow: encounter a wrong function call");
864 
865   Value *opr0, *opr1;
866   ConstantFP *CF;
867   ConstantInt *CINT;
868   ConstantAggregateZero *CZero;
869   Type *eltType;
870 
871   opr0 = CI->getArgOperand(0);
872   opr1 = CI->getArgOperand(1);
873   CZero = dyn_cast<ConstantAggregateZero>(opr1);
874   if (getVecSize(FInfo) == 1) {
875     eltType = opr0->getType();
876     CF = dyn_cast<ConstantFP>(opr1);
877     CINT = dyn_cast<ConstantInt>(opr1);
878   } else {
879     VectorType *VTy = dyn_cast<VectorType>(opr0->getType());
880     assert(VTy && "Oprand of vector function should be of vectortype");
881     eltType = VTy->getElementType();
882     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1);
883 
884     // Now, only Handle vector const whose elements have the same value.
885     CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr;
886     CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr;
887   }
888 
889   // No unsafe math , no constant argument, do nothing
890   if (!isUnsafeMath(CI) && !CF && !CINT && !CZero)
891     return false;
892 
893   // 0x1111111 means that we don't do anything for this call.
894   int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
895 
896   if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) {
897     //  pow/powr/pown(x, 0) == 1
898     DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n");
899     Constant *cnval = ConstantFP::get(eltType, 1.0);
900     if (getVecSize(FInfo) > 1) {
901       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
902     }
903     replaceCall(cnval);
904     return true;
905   }
906   if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
907     // pow/powr/pown(x, 1.0) = x
908     DEBUG(errs() << "AMDIC: " << *CI
909                  << " ---> " << *opr0 << "\n");
910     replaceCall(opr0);
911     return true;
912   }
913   if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
914     // pow/powr/pown(x, 2.0) = x*x
915     DEBUG(errs() << "AMDIC: " << *CI
916                  << " ---> " << *opr0 << " * " << *opr0 << "\n");
917     Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
918     replaceCall(nval);
919     return true;
920   }
921   if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
922     // pow/powr/pown(x, -1.0) = 1.0/x
923     DEBUG(errs() << "AMDIC: " << *CI
924                  << " ---> 1 / " << *opr0 << "\n");
925     Constant *cnval = ConstantFP::get(eltType, 1.0);
926     if (getVecSize(FInfo) > 1) {
927       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
928     }
929     Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
930     replaceCall(nval);
931     return true;
932   }
933 
934   Module *M = CI->getModule();
935   if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
936     // pow[r](x, [-]0.5) = sqrt(x)
937     bool issqrt = CF->isExactlyValue(0.5);
938     if (Constant *FPExpr = getFunction(M,
939         AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
940                              : AMDGPULibFunc::EI_RSQRT, FInfo))) {
941       DEBUG(errs() << "AMDIC: " << *CI << " ---> "
942                    << FInfo.getName().c_str() << "(" << *opr0 << ")\n");
943       Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
944                                                         : "__pow2rsqrt");
945       replaceCall(nval);
946       return true;
947     }
948   }
949 
950   if (!isUnsafeMath(CI))
951     return false;
952 
953   // Unsafe Math optimization
954 
955   // Remember that ci_opr1 is set if opr1 is integral
956   if (CF) {
957     double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
958                     ? (double)CF->getValueAPF().convertToFloat()
959                     : CF->getValueAPF().convertToDouble();
960     int ival = (int)dval;
961     if ((double)ival == dval) {
962       ci_opr1 = ival;
963     } else
964       ci_opr1 = 0x11111111;
965   }
966 
967   // pow/powr/pown(x, c) = [1/](x*x*..x); where
968   //   trunc(c) == c && the number of x == c && |c| <= 12
969   unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
970   if (abs_opr1 <= 12) {
971     Constant *cnval;
972     Value *nval;
973     if (abs_opr1 == 0) {
974       cnval = ConstantFP::get(eltType, 1.0);
975       if (getVecSize(FInfo) > 1) {
976         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
977       }
978       nval = cnval;
979     } else {
980       Value *valx2 = nullptr;
981       nval = nullptr;
982       while (abs_opr1 > 0) {
983         valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
984         if (abs_opr1 & 1) {
985           nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
986         }
987         abs_opr1 >>= 1;
988       }
989     }
990 
991     if (ci_opr1 < 0) {
992       cnval = ConstantFP::get(eltType, 1.0);
993       if (getVecSize(FInfo) > 1) {
994         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
995       }
996       nval = B.CreateFDiv(cnval, nval, "__1powprod");
997     }
998     DEBUG(errs() << "AMDIC: " << *CI << " ---> "
999                  <<  ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0 << ")\n");
1000     replaceCall(nval);
1001     return true;
1002   }
1003 
1004   // powr ---> exp2(y * log2(x))
1005   // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
1006   Constant *ExpExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2,
1007                                                    FInfo));
1008   if (!ExpExpr)
1009     return false;
1010 
1011   bool needlog = false;
1012   bool needabs = false;
1013   bool needcopysign = false;
1014   Constant *cnval = nullptr;
1015   if (getVecSize(FInfo) == 1) {
1016     CF = dyn_cast<ConstantFP>(opr0);
1017 
1018     if (CF) {
1019       double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1020                    ? (double)CF->getValueAPF().convertToFloat()
1021                    : CF->getValueAPF().convertToDouble();
1022 
1023       V = log2(std::abs(V));
1024       cnval = ConstantFP::get(eltType, V);
1025       needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
1026                      CF->isNegative();
1027     } else {
1028       needlog = true;
1029       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR &&
1030                                (!CF || CF->isNegative());
1031     }
1032   } else {
1033     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
1034 
1035     if (!CDV) {
1036       needlog = true;
1037       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1038     } else {
1039       assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
1040               "Wrong vector size detected");
1041 
1042       SmallVector<double, 0> DVal;
1043       for (int i=0; i < getVecSize(FInfo); ++i) {
1044         double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1045                      ? (double)CDV->getElementAsFloat(i)
1046                      : CDV->getElementAsDouble(i);
1047         if (V < 0.0) needcopysign = true;
1048         V = log2(std::abs(V));
1049         DVal.push_back(V);
1050       }
1051       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1052         SmallVector<float, 0> FVal;
1053         for (unsigned i=0; i < DVal.size(); ++i) {
1054           FVal.push_back((float)DVal[i]);
1055         }
1056         ArrayRef<float> tmp(FVal);
1057         cnval = ConstantDataVector::get(M->getContext(), tmp);
1058       } else {
1059         ArrayRef<double> tmp(DVal);
1060         cnval = ConstantDataVector::get(M->getContext(), tmp);
1061       }
1062     }
1063   }
1064 
1065   if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1066     // We cannot handle corner cases for a general pow() function, give up
1067     // unless y is a constant integral value. Then proceed as if it were pown.
1068     if (getVecSize(FInfo) == 1) {
1069       if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) {
1070         double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1071                    ? (double)CF->getValueAPF().convertToFloat()
1072                    : CF->getValueAPF().convertToDouble();
1073         if (y != (double)(int64_t)y)
1074           return false;
1075       } else
1076         return false;
1077     } else {
1078       if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) {
1079         for (int i=0; i < getVecSize(FInfo); ++i) {
1080           double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1081                      ? (double)CDV->getElementAsFloat(i)
1082                      : CDV->getElementAsDouble(i);
1083           if (y != (double)(int64_t)y)
1084             return false;
1085         }
1086       } else
1087         return false;
1088     }
1089   }
1090 
1091   Value *nval;
1092   if (needabs) {
1093     Constant *AbsExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS,
1094                                                      FInfo));
1095     if (!AbsExpr)
1096       return false;
1097     nval = CreateCallEx(B, AbsExpr, opr0, "__fabs");
1098   } else {
1099     nval = cnval ? cnval : opr0;
1100   }
1101   if (needlog) {
1102     Constant *LogExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2,
1103                                                      FInfo));
1104     if (!LogExpr)
1105       return false;
1106     nval = CreateCallEx(B,LogExpr, nval, "__log2");
1107   }
1108 
1109   if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1110     // convert int(32) to fp(f32 or f64)
1111     opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1112   }
1113   nval = B.CreateFMul(opr1, nval, "__ylogx");
1114   nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1115 
1116   if (needcopysign) {
1117     Value *opr_n;
1118     Type* rTy = opr0->getType();
1119     Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty();
1120     Type *nTy = nTyS;
1121     if (const VectorType *vTy = dyn_cast<VectorType>(rTy))
1122       nTy = VectorType::get(nTyS, vTy->getNumElements());
1123     unsigned size = nTy->getScalarSizeInBits();
1124     opr_n = CI->getArgOperand(1);
1125     if (opr_n->getType()->isIntegerTy())
1126       opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou");
1127     else
1128       opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1129 
1130     Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1131     sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1132     nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1133     nval = B.CreateBitCast(nval, opr0->getType());
1134   }
1135 
1136   DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1137                << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1138   replaceCall(nval);
1139 
1140   return true;
1141 }
1142 
1143 bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B,
1144                                 const FuncInfo &FInfo) {
1145   Value *opr0 = CI->getArgOperand(0);
1146   Value *opr1 = CI->getArgOperand(1);
1147 
1148   ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1149   if (!CINT) {
1150     return false;
1151   }
1152   int ci_opr1 = (int)CINT->getSExtValue();
1153   if (ci_opr1 == 1) {  // rootn(x, 1) = x
1154     DEBUG(errs() << "AMDIC: " << *CI
1155                  << " ---> " << *opr0 << "\n");
1156     replaceCall(opr0);
1157     return true;
1158   }
1159   if (ci_opr1 == 2) {  // rootn(x, 2) = sqrt(x)
1160     std::vector<const Type*> ParamsTys;
1161     ParamsTys.push_back(opr0->getType());
1162     Module *M = CI->getModule();
1163     if (Constant *FPExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT,
1164                                                         FInfo))) {
1165       DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n");
1166       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1167       replaceCall(nval);
1168       return true;
1169     }
1170   } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1171     Module *M = CI->getModule();
1172     if (Constant *FPExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT,
1173                                                         FInfo))) {
1174       DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n");
1175       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1176       replaceCall(nval);
1177       return true;
1178     }
1179   } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1180     DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n");
1181     Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1182                                opr0,
1183                                "__rootn2div");
1184     replaceCall(nval);
1185     return true;
1186   } else if (ci_opr1 == -2) {  // rootn(x, -2) = rsqrt(x)
1187     std::vector<const Type*> ParamsTys;
1188     ParamsTys.push_back(opr0->getType());
1189     Module *M = CI->getModule();
1190     if (Constant *FPExpr = getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT,
1191                                                         FInfo))) {
1192       DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0 << ")\n");
1193       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1194       replaceCall(nval);
1195       return true;
1196     }
1197   }
1198   return false;
1199 }
1200 
1201 bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B,
1202                                   const FuncInfo &FInfo) {
1203   Value *opr0 = CI->getArgOperand(0);
1204   Value *opr1 = CI->getArgOperand(1);
1205   Value *opr2 = CI->getArgOperand(2);
1206 
1207   ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
1208   ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
1209   if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) {
1210     // fma/mad(a, b, c) = c if a=0 || b=0
1211     DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n");
1212     replaceCall(opr2);
1213     return true;
1214   }
1215   if (CF0 && CF0->isExactlyValue(1.0f)) {
1216     // fma/mad(a, b, c) = b+c if a=1
1217     DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1218                  << *opr1 << " + " << *opr2 << "\n");
1219     Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd");
1220     replaceCall(nval);
1221     return true;
1222   }
1223   if (CF1 && CF1->isExactlyValue(1.0f)) {
1224     // fma/mad(a, b, c) = a+c if b=1
1225     DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1226                  << *opr0 << " + " << *opr2 << "\n");
1227     Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd");
1228     replaceCall(nval);
1229     return true;
1230   }
1231   if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) {
1232     if (CF->isZero()) {
1233       // fma/mad(a, b, c) = a*b if c=0
1234       DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1235                    << *opr0 << " * " << *opr1 << "\n");
1236       Value *nval = B.CreateFMul(opr0, opr1, "fmamul");
1237       replaceCall(nval);
1238       return true;
1239     }
1240   }
1241 
1242   return false;
1243 }
1244 
1245 // Get a scalar native builtin signle argument FP function
1246 Constant* AMDGPULibCalls::getNativeFunction(Module* M, const FuncInfo& FInfo) {
1247   if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1248     return nullptr;
1249   FuncInfo nf = FInfo;
1250   nf.setPrefix(AMDGPULibFunc::NATIVE);
1251   return getFunction(M, nf);
1252 }
1253 
1254 // fold sqrt -> native_sqrt (x)
1255 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
1256                                const FuncInfo &FInfo) {
1257   if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1258       (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1259     if (Constant *FPExpr = getNativeFunction(
1260         CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1261       Value *opr0 = CI->getArgOperand(0);
1262       DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1263                    << "sqrt(" << *opr0 << ")\n");
1264       Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1265       replaceCall(nval);
1266       return true;
1267     }
1268   }
1269   return false;
1270 }
1271 
1272 // fold sin, cos -> sincos.
1273 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
1274                                  AliasAnalysis *AA) {
1275   AMDGPULibFunc fInfo;
1276   if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo))
1277     return false;
1278 
1279   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1280          fInfo.getId() == AMDGPULibFunc::EI_COS);
1281   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1282 
1283   Value *CArgVal = CI->getArgOperand(0);
1284   BasicBlock * const CBB = CI->getParent();
1285 
1286   int const MaxScan = 30;
1287 
1288   { // fold in load value.
1289     LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
1290     if (LI && LI->getParent() == CBB) {
1291       BasicBlock::iterator BBI = LI->getIterator();
1292       Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
1293       if (AvailableVal) {
1294         CArgVal->replaceAllUsesWith(AvailableVal);
1295         if (CArgVal->getNumUses() == 0)
1296           LI->eraseFromParent();
1297         CArgVal = CI->getArgOperand(0);
1298       }
1299     }
1300   }
1301 
1302   Module *M = CI->getModule();
1303   fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN);
1304   std::string const PairName = fInfo.mangle();
1305 
1306   CallInst *UI = nullptr;
1307   for (User* U : CArgVal->users()) {
1308     CallInst *XI = dyn_cast_or_null<CallInst>(U);
1309     if (!XI || XI == CI || XI->getParent() != CBB)
1310       continue;
1311 
1312     Function *UCallee = XI->getCalledFunction();
1313     if (!UCallee || !UCallee->getName().equals(PairName))
1314       continue;
1315 
1316     BasicBlock::iterator BBI = CI->getIterator();
1317     if (BBI == CI->getParent()->begin())
1318       break;
1319     --BBI;
1320     for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
1321       if (cast<Instruction>(BBI) == XI) {
1322         UI = XI;
1323         break;
1324       }
1325     }
1326     if (UI) break;
1327   }
1328 
1329   if (!UI) return false;
1330 
1331   // Merge the sin and cos.
1332 
1333   // for OpenCL 2.0 we have only generic implementation of sincos
1334   // function.
1335   AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
1336   nf.getLeads()[0].PtrKind = AMDGPULibFunc::GENERIC;
1337   Function *Fsincos = dyn_cast_or_null<Function>(getFunction(M, nf));
1338   if (!Fsincos) return false;
1339 
1340   BasicBlock::iterator ItOld = B.GetInsertPoint();
1341   AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
1342   B.SetInsertPoint(UI);
1343 
1344   Value *P = Alloc;
1345   Type *PTy = Fsincos->getFunctionType()->getParamType(1);
1346   // The allocaInst allocates the memory in private address space. This need
1347   // to be bitcasted to point to the address space of cos pointer type.
1348   // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1349   const AMDGPUAS AS = AMDGPU::getAMDGPUAS(*M);
1350   if (PTy->getPointerAddressSpace() != AS.PRIVATE_ADDRESS)
1351     P = B.CreateAddrSpaceCast(Alloc, PTy);
1352   CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P);
1353 
1354   DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI
1355                << ") with " << *Call << "\n");
1356 
1357   if (!isSin) { // CI->cos, UI->sin
1358     B.SetInsertPoint(&*ItOld);
1359     UI->replaceAllUsesWith(&*Call);
1360     Instruction *Reload = B.CreateLoad(Alloc);
1361     CI->replaceAllUsesWith(Reload);
1362     UI->eraseFromParent();
1363     CI->eraseFromParent();
1364   } else { // CI->sin, UI->cos
1365     Instruction *Reload = B.CreateLoad(Alloc);
1366     UI->replaceAllUsesWith(Reload);
1367     CI->replaceAllUsesWith(Call);
1368     UI->eraseFromParent();
1369     CI->eraseFromParent();
1370   }
1371   return true;
1372 }
1373 
1374 // Get insertion point at entry.
1375 BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) {
1376   Function * Func = UI->getParent()->getParent();
1377   BasicBlock * BB = &Func->getEntryBlock();
1378   assert(BB && "Entry block not found!");
1379   BasicBlock::iterator ItNew = BB->begin();
1380   return ItNew;
1381 }
1382 
1383 // Insert a AllocsInst at the beginning of function entry block.
1384 AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B,
1385                                          const char *prefix) {
1386   BasicBlock::iterator ItNew = getEntryIns(UI);
1387   Function *UCallee = UI->getCalledFunction();
1388   Type *RetType = UCallee->getReturnType();
1389   B.SetInsertPoint(&*ItNew);
1390   AllocaInst *Alloc = B.CreateAlloca(RetType, 0,
1391     std::string(prefix) + UI->getName());
1392   Alloc->setAlignment(UCallee->getParent()->getDataLayout()
1393                        .getTypeAllocSize(RetType));
1394   return Alloc;
1395 }
1396 
1397 bool AMDGPULibCalls::evaluateScalarMathFunc(FuncInfo &FInfo,
1398                                             double& Res0, double& Res1,
1399                                             Constant *copr0, Constant *copr1,
1400                                             Constant *copr2) {
1401   // By default, opr0/opr1/opr3 holds values of float/double type.
1402   // If they are not float/double, each function has to its
1403   // operand separately.
1404   double opr0=0.0, opr1=0.0, opr2=0.0;
1405   ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1406   ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1407   ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2);
1408   if (fpopr0) {
1409     opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1410              ? fpopr0->getValueAPF().convertToDouble()
1411              : (double)fpopr0->getValueAPF().convertToFloat();
1412   }
1413 
1414   if (fpopr1) {
1415     opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1416              ? fpopr1->getValueAPF().convertToDouble()
1417              : (double)fpopr1->getValueAPF().convertToFloat();
1418   }
1419 
1420   if (fpopr2) {
1421     opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1422              ? fpopr2->getValueAPF().convertToDouble()
1423              : (double)fpopr2->getValueAPF().convertToFloat();
1424   }
1425 
1426   switch (FInfo.getId()) {
1427   default : return false;
1428 
1429   case AMDGPULibFunc::EI_ACOS:
1430     Res0 = acos(opr0);
1431     return true;
1432 
1433   case AMDGPULibFunc::EI_ACOSH:
1434     // acosh(x) == log(x + sqrt(x*x - 1))
1435     Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1436     return true;
1437 
1438   case AMDGPULibFunc::EI_ACOSPI:
1439     Res0 = acos(opr0) / MATH_PI;
1440     return true;
1441 
1442   case AMDGPULibFunc::EI_ASIN:
1443     Res0 = asin(opr0);
1444     return true;
1445 
1446   case AMDGPULibFunc::EI_ASINH:
1447     // asinh(x) == log(x + sqrt(x*x + 1))
1448     Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1449     return true;
1450 
1451   case AMDGPULibFunc::EI_ASINPI:
1452     Res0 = asin(opr0) / MATH_PI;
1453     return true;
1454 
1455   case AMDGPULibFunc::EI_ATAN:
1456     Res0 = atan(opr0);
1457     return true;
1458 
1459   case AMDGPULibFunc::EI_ATANH:
1460     // atanh(x) == (log(x+1) - log(x-1))/2;
1461     Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1462     return true;
1463 
1464   case AMDGPULibFunc::EI_ATANPI:
1465     Res0 = atan(opr0) / MATH_PI;
1466     return true;
1467 
1468   case AMDGPULibFunc::EI_CBRT:
1469     Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1470     return true;
1471 
1472   case AMDGPULibFunc::EI_COS:
1473     Res0 = cos(opr0);
1474     return true;
1475 
1476   case AMDGPULibFunc::EI_COSH:
1477     Res0 = cosh(opr0);
1478     return true;
1479 
1480   case AMDGPULibFunc::EI_COSPI:
1481     Res0 = cos(MATH_PI * opr0);
1482     return true;
1483 
1484   case AMDGPULibFunc::EI_EXP:
1485     Res0 = exp(opr0);
1486     return true;
1487 
1488   case AMDGPULibFunc::EI_EXP2:
1489     Res0 = pow(2.0, opr0);
1490     return true;
1491 
1492   case AMDGPULibFunc::EI_EXP10:
1493     Res0 = pow(10.0, opr0);
1494     return true;
1495 
1496   case AMDGPULibFunc::EI_EXPM1:
1497     Res0 = exp(opr0) - 1.0;
1498     return true;
1499 
1500   case AMDGPULibFunc::EI_LOG:
1501     Res0 = log(opr0);
1502     return true;
1503 
1504   case AMDGPULibFunc::EI_LOG2:
1505     Res0 = log(opr0) / log(2.0);
1506     return true;
1507 
1508   case AMDGPULibFunc::EI_LOG10:
1509     Res0 = log(opr0) / log(10.0);
1510     return true;
1511 
1512   case AMDGPULibFunc::EI_RSQRT:
1513     Res0 = 1.0 / sqrt(opr0);
1514     return true;
1515 
1516   case AMDGPULibFunc::EI_SIN:
1517     Res0 = sin(opr0);
1518     return true;
1519 
1520   case AMDGPULibFunc::EI_SINH:
1521     Res0 = sinh(opr0);
1522     return true;
1523 
1524   case AMDGPULibFunc::EI_SINPI:
1525     Res0 = sin(MATH_PI * opr0);
1526     return true;
1527 
1528   case AMDGPULibFunc::EI_SQRT:
1529     Res0 = sqrt(opr0);
1530     return true;
1531 
1532   case AMDGPULibFunc::EI_TAN:
1533     Res0 = tan(opr0);
1534     return true;
1535 
1536   case AMDGPULibFunc::EI_TANH:
1537     Res0 = tanh(opr0);
1538     return true;
1539 
1540   case AMDGPULibFunc::EI_TANPI:
1541     Res0 = tan(MATH_PI * opr0);
1542     return true;
1543 
1544   case AMDGPULibFunc::EI_RECIP:
1545     Res0 = 1.0 / opr0;
1546     return true;
1547 
1548   // two-arg functions
1549   case AMDGPULibFunc::EI_DIVIDE:
1550     Res0 = opr0 / opr1;
1551     return true;
1552 
1553   case AMDGPULibFunc::EI_POW:
1554   case AMDGPULibFunc::EI_POWR:
1555     Res0 = pow(opr0, opr1);
1556     return true;
1557 
1558   case AMDGPULibFunc::EI_POWN: {
1559     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1560       double val = (double)iopr1->getSExtValue();
1561       Res0 = pow(opr0, val);
1562       return true;
1563     }
1564     return false;
1565   }
1566 
1567   case AMDGPULibFunc::EI_ROOTN: {
1568     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1569       double val = (double)iopr1->getSExtValue();
1570       Res0 = pow(opr0, 1.0 / val);
1571       return true;
1572     }
1573     return false;
1574   }
1575 
1576   // with ptr arg
1577   case AMDGPULibFunc::EI_SINCOS:
1578     Res0 = sin(opr0);
1579     Res1 = cos(opr0);
1580     return true;
1581 
1582   // three-arg functions
1583   case AMDGPULibFunc::EI_FMA:
1584   case AMDGPULibFunc::EI_MAD:
1585     Res0 = opr0 * opr1 + opr2;
1586     return true;
1587   }
1588 
1589   return false;
1590 }
1591 
1592 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, FuncInfo &FInfo) {
1593   int numArgs = (int)aCI->getNumArgOperands();
1594   if (numArgs > 3)
1595     return false;
1596 
1597   Constant *copr0 = nullptr;
1598   Constant *copr1 = nullptr;
1599   Constant *copr2 = nullptr;
1600   if (numArgs > 0) {
1601     if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1602       return false;
1603   }
1604 
1605   if (numArgs > 1) {
1606     if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1607       if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1608         return false;
1609     }
1610   }
1611 
1612   if (numArgs > 2) {
1613     if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr)
1614       return false;
1615   }
1616 
1617   // At this point, all arguments to aCI are constants.
1618 
1619   // max vector size is 16, and sincos will generate two results.
1620   double DVal0[16], DVal1[16];
1621   bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1622   if (getVecSize(FInfo) == 1) {
1623     if (!evaluateScalarMathFunc(FInfo, DVal0[0],
1624                                 DVal1[0], copr0, copr1, copr2)) {
1625       return false;
1626     }
1627   } else {
1628     ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1629     ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1630     ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2);
1631     for (int i=0; i < getVecSize(FInfo); ++i) {
1632       Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1633       Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1634       Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr;
1635       if (!evaluateScalarMathFunc(FInfo, DVal0[i],
1636                                   DVal1[i], celt0, celt1, celt2)) {
1637         return false;
1638       }
1639     }
1640   }
1641 
1642   LLVMContext &context = CI->getParent()->getParent()->getContext();
1643   Constant *nval0, *nval1;
1644   if (getVecSize(FInfo) == 1) {
1645     nval0 = ConstantFP::get(CI->getType(), DVal0[0]);
1646     if (hasTwoResults)
1647       nval1 = ConstantFP::get(CI->getType(), DVal1[0]);
1648   } else {
1649     if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1650       SmallVector <float, 0> FVal0, FVal1;
1651       for (int i=0; i < getVecSize(FInfo); ++i)
1652         FVal0.push_back((float)DVal0[i]);
1653       ArrayRef<float> tmp0(FVal0);
1654       nval0 = ConstantDataVector::get(context, tmp0);
1655       if (hasTwoResults) {
1656         for (int i=0; i < getVecSize(FInfo); ++i)
1657           FVal1.push_back((float)DVal1[i]);
1658         ArrayRef<float> tmp1(FVal1);
1659         nval1 = ConstantDataVector::get(context, tmp1);
1660       }
1661     } else {
1662       ArrayRef<double> tmp0(DVal0);
1663       nval0 = ConstantDataVector::get(context, tmp0);
1664       if (hasTwoResults) {
1665         ArrayRef<double> tmp1(DVal1);
1666         nval1 = ConstantDataVector::get(context, tmp1);
1667       }
1668     }
1669   }
1670 
1671   if (hasTwoResults) {
1672     // sincos
1673     assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1674            "math function with ptr arg not supported yet");
1675     new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1676   }
1677 
1678   replaceCall(nval0);
1679   return true;
1680 }
1681 
1682 // Public interface to the Simplify LibCalls pass.
1683 FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass() {
1684   return new AMDGPUSimplifyLibCalls();
1685 }
1686 
1687 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
1688   return new AMDGPUUseNativeCalls();
1689 }
1690 
1691 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
1692   if (skipFunction(F))
1693     return false;
1694 
1695   bool Changed = false;
1696   auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1697 
1698   DEBUG(dbgs() << "AMDIC: process function ";
1699         F.printAsOperand(dbgs(), false, F.getParent());
1700         dbgs() << '\n';);
1701 
1702   for (auto &BB : F) {
1703     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1704       // Ignore non-calls.
1705       CallInst *CI = dyn_cast<CallInst>(I);
1706       ++I;
1707       if (!CI) continue;
1708 
1709       // Ignore indirect calls.
1710       Function *Callee = CI->getCalledFunction();
1711       if (Callee == 0) continue;
1712 
1713       DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1714             dbgs().flush());
1715       if(Simplifier.fold(CI, AA))
1716         Changed = true;
1717     }
1718   }
1719   return Changed;
1720 }
1721 
1722 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) {
1723   if (skipFunction(F) || UseNative.empty())
1724     return false;
1725 
1726   bool Changed = false;
1727   for (auto &BB : F) {
1728     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1729       // Ignore non-calls.
1730       CallInst *CI = dyn_cast<CallInst>(I);
1731       ++I;
1732       if (!CI) continue;
1733 
1734       // Ignore indirect calls.
1735       Function *Callee = CI->getCalledFunction();
1736       if (Callee == 0) continue;
1737 
1738       if(Simplifier.useNative(CI))
1739         Changed = true;
1740     }
1741   }
1742   return Changed;
1743 }
1744