xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp (revision 14359ef1b6a0610ac91df5f5a91c88a0b51c187c)
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 #define DEBUG_TYPE "amdgpu-simplifylib"
15 
16 #include "AMDGPU.h"
17 #include "AMDGPULibFunc.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/Loads.h"
20 #include "llvm/ADT/StringSet.h"
21 #include "llvm/ADT/StringRef.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/IR/ValueSymbolTable.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetOptions.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   FunctionCallee 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   FunctionCallee 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   const TargetOptions Options;
172 
173   public:
174     static char ID; // Pass identification
175 
176     AMDGPUSimplifyLibCalls(const TargetOptions &Opt = TargetOptions())
177       : FunctionPass(ID), Options(Opt) {
178       initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
179     }
180 
181     void getAnalysisUsage(AnalysisUsage &AU) const override {
182       AU.addRequired<AAResultsWrapperPass>();
183     }
184 
185     bool runOnFunction(Function &M) override;
186   };
187 
188   class AMDGPUUseNativeCalls : public FunctionPass {
189 
190   AMDGPULibCalls Simplifier;
191 
192   public:
193     static char ID; // Pass identification
194 
195     AMDGPUUseNativeCalls() : FunctionPass(ID) {
196       initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry());
197       Simplifier.initNativeFuncs();
198     }
199 
200     bool runOnFunction(Function &F) override;
201   };
202 
203 } // end anonymous namespace.
204 
205 char AMDGPUSimplifyLibCalls::ID = 0;
206 char AMDGPUUseNativeCalls::ID = 0;
207 
208 INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
209                       "Simplify well-known AMD library calls", false, false)
210 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
211 INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
212                     "Simplify well-known AMD library calls", false, false)
213 
214 INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative",
215                 "Replace builtin math calls with that native versions.",
216                 false, false)
217 
218 template <typename IRB>
219 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg,
220                               const Twine &Name = "") {
221   CallInst *R = B.CreateCall(Callee, Arg, Name);
222   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
223     R->setCallingConv(F->getCallingConv());
224   return R;
225 }
226 
227 template <typename IRB>
228 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1,
229                                Value *Arg2, const Twine &Name = "") {
230   CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
231   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
232     R->setCallingConv(F->getCallingConv());
233   return R;
234 }
235 
236 //  Data structures for table-driven optimizations.
237 //  FuncTbl works for both f32 and f64 functions with 1 input argument
238 
239 struct TableEntry {
240   double   result;
241   double   input;
242 };
243 
244 /* a list of {result, input} */
245 static const TableEntry tbl_acos[] = {
246   {MATH_PI/2.0, 0.0},
247   {MATH_PI/2.0, -0.0},
248   {0.0, 1.0},
249   {MATH_PI, -1.0}
250 };
251 static const TableEntry tbl_acosh[] = {
252   {0.0, 1.0}
253 };
254 static const TableEntry tbl_acospi[] = {
255   {0.5, 0.0},
256   {0.5, -0.0},
257   {0.0, 1.0},
258   {1.0, -1.0}
259 };
260 static const TableEntry tbl_asin[] = {
261   {0.0, 0.0},
262   {-0.0, -0.0},
263   {MATH_PI/2.0, 1.0},
264   {-MATH_PI/2.0, -1.0}
265 };
266 static const TableEntry tbl_asinh[] = {
267   {0.0, 0.0},
268   {-0.0, -0.0}
269 };
270 static const TableEntry tbl_asinpi[] = {
271   {0.0, 0.0},
272   {-0.0, -0.0},
273   {0.5, 1.0},
274   {-0.5, -1.0}
275 };
276 static const TableEntry tbl_atan[] = {
277   {0.0, 0.0},
278   {-0.0, -0.0},
279   {MATH_PI/4.0, 1.0},
280   {-MATH_PI/4.0, -1.0}
281 };
282 static const TableEntry tbl_atanh[] = {
283   {0.0, 0.0},
284   {-0.0, -0.0}
285 };
286 static const TableEntry tbl_atanpi[] = {
287   {0.0, 0.0},
288   {-0.0, -0.0},
289   {0.25, 1.0},
290   {-0.25, -1.0}
291 };
292 static const TableEntry tbl_cbrt[] = {
293   {0.0, 0.0},
294   {-0.0, -0.0},
295   {1.0, 1.0},
296   {-1.0, -1.0},
297 };
298 static const TableEntry tbl_cos[] = {
299   {1.0, 0.0},
300   {1.0, -0.0}
301 };
302 static const TableEntry tbl_cosh[] = {
303   {1.0, 0.0},
304   {1.0, -0.0}
305 };
306 static const TableEntry tbl_cospi[] = {
307   {1.0, 0.0},
308   {1.0, -0.0}
309 };
310 static const TableEntry tbl_erfc[] = {
311   {1.0, 0.0},
312   {1.0, -0.0}
313 };
314 static const TableEntry tbl_erf[] = {
315   {0.0, 0.0},
316   {-0.0, -0.0}
317 };
318 static const TableEntry tbl_exp[] = {
319   {1.0, 0.0},
320   {1.0, -0.0},
321   {MATH_E, 1.0}
322 };
323 static const TableEntry tbl_exp2[] = {
324   {1.0, 0.0},
325   {1.0, -0.0},
326   {2.0, 1.0}
327 };
328 static const TableEntry tbl_exp10[] = {
329   {1.0, 0.0},
330   {1.0, -0.0},
331   {10.0, 1.0}
332 };
333 static const TableEntry tbl_expm1[] = {
334   {0.0, 0.0},
335   {-0.0, -0.0}
336 };
337 static const TableEntry tbl_log[] = {
338   {0.0, 1.0},
339   {1.0, MATH_E}
340 };
341 static const TableEntry tbl_log2[] = {
342   {0.0, 1.0},
343   {1.0, 2.0}
344 };
345 static const TableEntry tbl_log10[] = {
346   {0.0, 1.0},
347   {1.0, 10.0}
348 };
349 static const TableEntry tbl_rsqrt[] = {
350   {1.0, 1.0},
351   {1.0/MATH_SQRT2, 2.0}
352 };
353 static const TableEntry tbl_sin[] = {
354   {0.0, 0.0},
355   {-0.0, -0.0}
356 };
357 static const TableEntry tbl_sinh[] = {
358   {0.0, 0.0},
359   {-0.0, -0.0}
360 };
361 static const TableEntry tbl_sinpi[] = {
362   {0.0, 0.0},
363   {-0.0, -0.0}
364 };
365 static const TableEntry tbl_sqrt[] = {
366   {0.0, 0.0},
367   {1.0, 1.0},
368   {MATH_SQRT2, 2.0}
369 };
370 static const TableEntry tbl_tan[] = {
371   {0.0, 0.0},
372   {-0.0, -0.0}
373 };
374 static const TableEntry tbl_tanh[] = {
375   {0.0, 0.0},
376   {-0.0, -0.0}
377 };
378 static const TableEntry tbl_tanpi[] = {
379   {0.0, 0.0},
380   {-0.0, -0.0}
381 };
382 static const TableEntry tbl_tgamma[] = {
383   {1.0, 1.0},
384   {1.0, 2.0},
385   {2.0, 3.0},
386   {6.0, 4.0}
387 };
388 
389 static bool HasNative(AMDGPULibFunc::EFuncId id) {
390   switch(id) {
391   case AMDGPULibFunc::EI_DIVIDE:
392   case AMDGPULibFunc::EI_COS:
393   case AMDGPULibFunc::EI_EXP:
394   case AMDGPULibFunc::EI_EXP2:
395   case AMDGPULibFunc::EI_EXP10:
396   case AMDGPULibFunc::EI_LOG:
397   case AMDGPULibFunc::EI_LOG2:
398   case AMDGPULibFunc::EI_LOG10:
399   case AMDGPULibFunc::EI_POWR:
400   case AMDGPULibFunc::EI_RECIP:
401   case AMDGPULibFunc::EI_RSQRT:
402   case AMDGPULibFunc::EI_SIN:
403   case AMDGPULibFunc::EI_SINCOS:
404   case AMDGPULibFunc::EI_SQRT:
405   case AMDGPULibFunc::EI_TAN:
406     return true;
407   default:;
408   }
409   return false;
410 }
411 
412 struct TableRef {
413   size_t size;
414   const TableEntry *table; // variable size: from 0 to (size - 1)
415 
416   TableRef() : size(0), table(nullptr) {}
417 
418   template <size_t N>
419   TableRef(const TableEntry (&tbl)[N]) : size(N), table(&tbl[0]) {}
420 };
421 
422 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
423   switch(id) {
424   case AMDGPULibFunc::EI_ACOS:    return TableRef(tbl_acos);
425   case AMDGPULibFunc::EI_ACOSH:   return TableRef(tbl_acosh);
426   case AMDGPULibFunc::EI_ACOSPI:  return TableRef(tbl_acospi);
427   case AMDGPULibFunc::EI_ASIN:    return TableRef(tbl_asin);
428   case AMDGPULibFunc::EI_ASINH:   return TableRef(tbl_asinh);
429   case AMDGPULibFunc::EI_ASINPI:  return TableRef(tbl_asinpi);
430   case AMDGPULibFunc::EI_ATAN:    return TableRef(tbl_atan);
431   case AMDGPULibFunc::EI_ATANH:   return TableRef(tbl_atanh);
432   case AMDGPULibFunc::EI_ATANPI:  return TableRef(tbl_atanpi);
433   case AMDGPULibFunc::EI_CBRT:    return TableRef(tbl_cbrt);
434   case AMDGPULibFunc::EI_NCOS:
435   case AMDGPULibFunc::EI_COS:     return TableRef(tbl_cos);
436   case AMDGPULibFunc::EI_COSH:    return TableRef(tbl_cosh);
437   case AMDGPULibFunc::EI_COSPI:   return TableRef(tbl_cospi);
438   case AMDGPULibFunc::EI_ERFC:    return TableRef(tbl_erfc);
439   case AMDGPULibFunc::EI_ERF:     return TableRef(tbl_erf);
440   case AMDGPULibFunc::EI_EXP:     return TableRef(tbl_exp);
441   case AMDGPULibFunc::EI_NEXP2:
442   case AMDGPULibFunc::EI_EXP2:    return TableRef(tbl_exp2);
443   case AMDGPULibFunc::EI_EXP10:   return TableRef(tbl_exp10);
444   case AMDGPULibFunc::EI_EXPM1:   return TableRef(tbl_expm1);
445   case AMDGPULibFunc::EI_LOG:     return TableRef(tbl_log);
446   case AMDGPULibFunc::EI_NLOG2:
447   case AMDGPULibFunc::EI_LOG2:    return TableRef(tbl_log2);
448   case AMDGPULibFunc::EI_LOG10:   return TableRef(tbl_log10);
449   case AMDGPULibFunc::EI_NRSQRT:
450   case AMDGPULibFunc::EI_RSQRT:   return TableRef(tbl_rsqrt);
451   case AMDGPULibFunc::EI_NSIN:
452   case AMDGPULibFunc::EI_SIN:     return TableRef(tbl_sin);
453   case AMDGPULibFunc::EI_SINH:    return TableRef(tbl_sinh);
454   case AMDGPULibFunc::EI_SINPI:   return TableRef(tbl_sinpi);
455   case AMDGPULibFunc::EI_NSQRT:
456   case AMDGPULibFunc::EI_SQRT:    return TableRef(tbl_sqrt);
457   case AMDGPULibFunc::EI_TAN:     return TableRef(tbl_tan);
458   case AMDGPULibFunc::EI_TANH:    return TableRef(tbl_tanh);
459   case AMDGPULibFunc::EI_TANPI:   return TableRef(tbl_tanpi);
460   case AMDGPULibFunc::EI_TGAMMA:  return TableRef(tbl_tgamma);
461   default:;
462   }
463   return TableRef();
464 }
465 
466 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
467   return FInfo.getLeads()[0].VectorSize;
468 }
469 
470 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
471   return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
472 }
473 
474 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) {
475   // If we are doing PreLinkOpt, the function is external. So it is safe to
476   // use getOrInsertFunction() at this stage.
477 
478   return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
479                        : AMDGPULibFunc::getFunction(M, fInfo);
480 }
481 
482 bool AMDGPULibCalls::parseFunctionName(const StringRef& FMangledName,
483                                     FuncInfo *FInfo) {
484   return AMDGPULibFunc::parse(FMangledName, *FInfo);
485 }
486 
487 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const {
488   if (auto Op = dyn_cast<FPMathOperator>(CI))
489     if (Op->isFast())
490       return true;
491   const Function *F = CI->getParent()->getParent();
492   Attribute Attr = F->getFnAttribute("unsafe-fp-math");
493   return Attr.getValueAsString() == "true";
494 }
495 
496 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
497   return AllNative ||
498          std::find(UseNative.begin(), UseNative.end(), F) != UseNative.end();
499 }
500 
501 void AMDGPULibCalls::initNativeFuncs() {
502   AllNative = useNativeFunc("all") ||
503               (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
504                UseNative.begin()->empty());
505 }
506 
507 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
508   bool native_sin = useNativeFunc("sin");
509   bool native_cos = useNativeFunc("cos");
510 
511   if (native_sin && native_cos) {
512     Module *M = aCI->getModule();
513     Value *opr0 = aCI->getArgOperand(0);
514 
515     AMDGPULibFunc nf;
516     nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
517     nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
518 
519     nf.setPrefix(AMDGPULibFunc::NATIVE);
520     nf.setId(AMDGPULibFunc::EI_SIN);
521     FunctionCallee sinExpr = getFunction(M, nf);
522 
523     nf.setPrefix(AMDGPULibFunc::NATIVE);
524     nf.setId(AMDGPULibFunc::EI_COS);
525     FunctionCallee cosExpr = getFunction(M, nf);
526     if (sinExpr && cosExpr) {
527       Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
528       Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
529       new StoreInst(cosval, aCI->getArgOperand(1), aCI);
530 
531       DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
532                                           << " with native version of sin/cos");
533 
534       replaceCall(sinval);
535       return true;
536     }
537   }
538   return false;
539 }
540 
541 bool AMDGPULibCalls::useNative(CallInst *aCI) {
542   CI = aCI;
543   Function *Callee = aCI->getCalledFunction();
544 
545   FuncInfo FInfo;
546   if (!parseFunctionName(Callee->getName(), &FInfo) || !FInfo.isMangled() ||
547       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
548       getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
549       !(AllNative || useNativeFunc(FInfo.getName()))) {
550     return false;
551   }
552 
553   if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
554     return sincosUseNative(aCI, FInfo);
555 
556   FInfo.setPrefix(AMDGPULibFunc::NATIVE);
557   FunctionCallee F = getFunction(aCI->getModule(), FInfo);
558   if (!F)
559     return false;
560 
561   aCI->setCalledFunction(F);
562   DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
563                                       << " with native version");
564   return true;
565 }
566 
567 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
568 // builtin, with appended type size and alignment arguments, where 2 or 4
569 // indicates the original number of arguments. The library has optimized version
570 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
571 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
572 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
573 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
574 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
575                                           FuncInfo &FInfo) {
576   auto *Callee = CI->getCalledFunction();
577   if (!Callee->isDeclaration())
578     return false;
579 
580   assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
581   auto *M = Callee->getParent();
582   auto &Ctx = M->getContext();
583   std::string Name = Callee->getName();
584   auto NumArg = CI->getNumArgOperands();
585   if (NumArg != 4 && NumArg != 6)
586     return false;
587   auto *PacketSize = CI->getArgOperand(NumArg - 2);
588   auto *PacketAlign = CI->getArgOperand(NumArg - 1);
589   if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign))
590     return false;
591   unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue();
592   unsigned Align = cast<ConstantInt>(PacketAlign)->getZExtValue();
593   if (Size != Align || !isPowerOf2_32(Size))
594     return false;
595 
596   Type *PtrElemTy;
597   if (Size <= 8)
598     PtrElemTy = Type::getIntNTy(Ctx, Size * 8);
599   else
600     PtrElemTy = VectorType::get(Type::getInt64Ty(Ctx), Size / 8);
601   unsigned PtrArgLoc = CI->getNumArgOperands() - 3;
602   auto PtrArg = CI->getArgOperand(PtrArgLoc);
603   unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace();
604   auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS);
605 
606   SmallVector<llvm::Type *, 6> ArgTys;
607   for (unsigned I = 0; I != PtrArgLoc; ++I)
608     ArgTys.push_back(CI->getArgOperand(I)->getType());
609   ArgTys.push_back(PtrTy);
610 
611   Name = Name + "_" + std::to_string(Size);
612   auto *FTy = FunctionType::get(Callee->getReturnType(),
613                                 ArrayRef<Type *>(ArgTys), false);
614   AMDGPULibFunc NewLibFunc(Name, FTy);
615   FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
616   if (!F)
617     return false;
618 
619   auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
620   SmallVector<Value *, 6> Args;
621   for (unsigned I = 0; I != PtrArgLoc; ++I)
622     Args.push_back(CI->getArgOperand(I));
623   Args.push_back(BCast);
624 
625   auto *NCI = B.CreateCall(F, Args);
626   NCI->setAttributes(CI->getAttributes());
627   CI->replaceAllUsesWith(NCI);
628   CI->dropAllReferences();
629   CI->eraseFromParent();
630 
631   return true;
632 }
633 
634 // This function returns false if no change; return true otherwise.
635 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) {
636   this->CI = CI;
637   Function *Callee = CI->getCalledFunction();
638 
639   // Ignore indirect calls.
640   if (Callee == 0) return false;
641 
642   FuncInfo FInfo;
643   if (!parseFunctionName(Callee->getName(), &FInfo))
644     return false;
645 
646   // Further check the number of arguments to see if they match.
647   if (CI->getNumArgOperands() != FInfo.getNumArgs())
648     return false;
649 
650   BasicBlock *BB = CI->getParent();
651   LLVMContext &Context = CI->getParent()->getContext();
652   IRBuilder<> B(Context);
653 
654   // Set the builder to the instruction after the call.
655   B.SetInsertPoint(BB, CI->getIterator());
656 
657   // Copy fast flags from the original call.
658   if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI))
659     B.setFastMathFlags(FPOp->getFastMathFlags());
660 
661   if (TDOFold(CI, FInfo))
662     return true;
663 
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 (isUnsafeMath(CI) && evaluateCall(CI, FInfo))
668     return true;
669 
670   // Specilized optimizations for each function call
671   switch (FInfo.getId()) {
672   case AMDGPULibFunc::EI_RECIP:
673     // skip vector function
674     assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
675              FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
676             "recip must be an either native or half function");
677     return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo);
678 
679   case AMDGPULibFunc::EI_DIVIDE:
680     // skip vector function
681     assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
682              FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
683             "divide must be an either native or half function");
684     return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo);
685 
686   case AMDGPULibFunc::EI_POW:
687   case AMDGPULibFunc::EI_POWR:
688   case AMDGPULibFunc::EI_POWN:
689     return fold_pow(CI, B, FInfo);
690 
691   case AMDGPULibFunc::EI_ROOTN:
692     // skip vector function
693     return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo);
694 
695   case AMDGPULibFunc::EI_FMA:
696   case AMDGPULibFunc::EI_MAD:
697   case AMDGPULibFunc::EI_NFMA:
698     // skip vector function
699     return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo);
700 
701   case AMDGPULibFunc::EI_SQRT:
702     return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo);
703   case AMDGPULibFunc::EI_COS:
704   case AMDGPULibFunc::EI_SIN:
705     if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
706          getArgType(FInfo) == AMDGPULibFunc::F64)
707         && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX))
708       return fold_sincos(CI, B, AA);
709 
710     break;
711   case AMDGPULibFunc::EI_READ_PIPE_2:
712   case AMDGPULibFunc::EI_READ_PIPE_4:
713   case AMDGPULibFunc::EI_WRITE_PIPE_2:
714   case AMDGPULibFunc::EI_WRITE_PIPE_4:
715     return fold_read_write_pipe(CI, B, FInfo);
716 
717   default:
718     break;
719   }
720 
721   return false;
722 }
723 
724 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
725   // Table-Driven optimization
726   const TableRef tr = getOptTable(FInfo.getId());
727   if (tr.size==0)
728     return false;
729 
730   int const sz = (int)tr.size;
731   const TableEntry * const ftbl = tr.table;
732   Value *opr0 = CI->getArgOperand(0);
733 
734   if (getVecSize(FInfo) > 1) {
735     if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
736       SmallVector<double, 0> DVal;
737       for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
738         ConstantFP *eltval = dyn_cast<ConstantFP>(
739                                CV->getElementAsConstant((unsigned)eltNo));
740         assert(eltval && "Non-FP arguments in math function!");
741         bool found = false;
742         for (int i=0; i < sz; ++i) {
743           if (eltval->isExactlyValue(ftbl[i].input)) {
744             DVal.push_back(ftbl[i].result);
745             found = true;
746             break;
747           }
748         }
749         if (!found) {
750           // This vector constants not handled yet.
751           return false;
752         }
753       }
754       LLVMContext &context = CI->getParent()->getParent()->getContext();
755       Constant *nval;
756       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
757         SmallVector<float, 0> FVal;
758         for (unsigned i = 0; i < DVal.size(); ++i) {
759           FVal.push_back((float)DVal[i]);
760         }
761         ArrayRef<float> tmp(FVal);
762         nval = ConstantDataVector::get(context, tmp);
763       } else { // F64
764         ArrayRef<double> tmp(DVal);
765         nval = ConstantDataVector::get(context, tmp);
766       }
767       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
768       replaceCall(nval);
769       return true;
770     }
771   } else {
772     // Scalar version
773     if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
774       for (int i = 0; i < sz; ++i) {
775         if (CF->isExactlyValue(ftbl[i].input)) {
776           Value *nval = ConstantFP::get(CF->getType(), ftbl[i].result);
777           LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
778           replaceCall(nval);
779           return true;
780         }
781       }
782     }
783   }
784 
785   return false;
786 }
787 
788 bool AMDGPULibCalls::replaceWithNative(CallInst *CI, const FuncInfo &FInfo) {
789   Module *M = CI->getModule();
790   if (getArgType(FInfo) != AMDGPULibFunc::F32 ||
791       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
792       !HasNative(FInfo.getId()))
793     return false;
794 
795   AMDGPULibFunc nf = FInfo;
796   nf.setPrefix(AMDGPULibFunc::NATIVE);
797   if (FunctionCallee FPExpr = getFunction(M, nf)) {
798     LLVM_DEBUG(dbgs() << "AMDIC: " << *CI << " ---> ");
799 
800     CI->setCalledFunction(FPExpr);
801 
802     LLVM_DEBUG(dbgs() << *CI << '\n');
803 
804     return true;
805   }
806   return false;
807 }
808 
809 //  [native_]half_recip(c) ==> 1.0/c
810 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B,
811                                 const FuncInfo &FInfo) {
812   Value *opr0 = CI->getArgOperand(0);
813   if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
814     // Just create a normal div. Later, InstCombine will be able
815     // to compute the divide into a constant (avoid check float infinity
816     // or subnormal at this point).
817     Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0),
818                                opr0,
819                                "recip2div");
820     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *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     LLVM_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     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
909     replaceCall(opr0);
910     return true;
911   }
912   if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
913     // pow/powr/pown(x, 2.0) = x*x
914     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0
915                       << "\n");
916     Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
917     replaceCall(nval);
918     return true;
919   }
920   if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
921     // pow/powr/pown(x, -1.0) = 1.0/x
922     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n");
923     Constant *cnval = ConstantFP::get(eltType, 1.0);
924     if (getVecSize(FInfo) > 1) {
925       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
926     }
927     Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
928     replaceCall(nval);
929     return true;
930   }
931 
932   Module *M = CI->getModule();
933   if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
934     // pow[r](x, [-]0.5) = sqrt(x)
935     bool issqrt = CF->isExactlyValue(0.5);
936     if (FunctionCallee FPExpr =
937             getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
938                                                 : AMDGPULibFunc::EI_RSQRT,
939                                          FInfo))) {
940       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
941                         << FInfo.getName().c_str() << "(" << *opr0 << ")\n");
942       Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
943                                                         : "__pow2rsqrt");
944       replaceCall(nval);
945       return true;
946     }
947   }
948 
949   if (!isUnsafeMath(CI))
950     return false;
951 
952   // Unsafe Math optimization
953 
954   // Remember that ci_opr1 is set if opr1 is integral
955   if (CF) {
956     double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
957                     ? (double)CF->getValueAPF().convertToFloat()
958                     : CF->getValueAPF().convertToDouble();
959     int ival = (int)dval;
960     if ((double)ival == dval) {
961       ci_opr1 = ival;
962     } else
963       ci_opr1 = 0x11111111;
964   }
965 
966   // pow/powr/pown(x, c) = [1/](x*x*..x); where
967   //   trunc(c) == c && the number of x == c && |c| <= 12
968   unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
969   if (abs_opr1 <= 12) {
970     Constant *cnval;
971     Value *nval;
972     if (abs_opr1 == 0) {
973       cnval = ConstantFP::get(eltType, 1.0);
974       if (getVecSize(FInfo) > 1) {
975         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
976       }
977       nval = cnval;
978     } else {
979       Value *valx2 = nullptr;
980       nval = nullptr;
981       while (abs_opr1 > 0) {
982         valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
983         if (abs_opr1 & 1) {
984           nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
985         }
986         abs_opr1 >>= 1;
987       }
988     }
989 
990     if (ci_opr1 < 0) {
991       cnval = ConstantFP::get(eltType, 1.0);
992       if (getVecSize(FInfo) > 1) {
993         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
994       }
995       nval = B.CreateFDiv(cnval, nval, "__1powprod");
996     }
997     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
998                       << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
999                       << ")\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   FunctionCallee ExpExpr =
1007       getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, 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     FunctionCallee AbsExpr =
1094         getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS, 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     FunctionCallee LogExpr =
1103         getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, 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   LLVM_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     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
1155     replaceCall(opr0);
1156     return true;
1157   }
1158   if (ci_opr1 == 2) {  // rootn(x, 2) = sqrt(x)
1159     std::vector<const Type*> ParamsTys;
1160     ParamsTys.push_back(opr0->getType());
1161     Module *M = CI->getModule();
1162     if (FunctionCallee FPExpr =
1163             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1164       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n");
1165       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1166       replaceCall(nval);
1167       return true;
1168     }
1169   } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1170     Module *M = CI->getModule();
1171     if (FunctionCallee FPExpr =
1172             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1173       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n");
1174       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1175       replaceCall(nval);
1176       return true;
1177     }
1178   } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1179     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n");
1180     Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1181                                opr0,
1182                                "__rootn2div");
1183     replaceCall(nval);
1184     return true;
1185   } else if (ci_opr1 == -2) {  // rootn(x, -2) = rsqrt(x)
1186     std::vector<const Type*> ParamsTys;
1187     ParamsTys.push_back(opr0->getType());
1188     Module *M = CI->getModule();
1189     if (FunctionCallee FPExpr =
1190             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) {
1191       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0
1192                         << ")\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     LLVM_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     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2
1218                       << "\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     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2
1226                       << "\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       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * "
1235                         << *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 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1247                                                  const FuncInfo &FInfo) {
1248   if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1249     return nullptr;
1250   FuncInfo nf = FInfo;
1251   nf.setPrefix(AMDGPULibFunc::NATIVE);
1252   return getFunction(M, nf);
1253 }
1254 
1255 // fold sqrt -> native_sqrt (x)
1256 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
1257                                const FuncInfo &FInfo) {
1258   if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1259       (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1260     if (FunctionCallee FPExpr = getNativeFunction(
1261             CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1262       Value *opr0 = CI->getArgOperand(0);
1263       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1264                         << "sqrt(" << *opr0 << ")\n");
1265       Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1266       replaceCall(nval);
1267       return true;
1268     }
1269   }
1270   return false;
1271 }
1272 
1273 // fold sin, cos -> sincos.
1274 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
1275                                  AliasAnalysis *AA) {
1276   AMDGPULibFunc fInfo;
1277   if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo))
1278     return false;
1279 
1280   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1281          fInfo.getId() == AMDGPULibFunc::EI_COS);
1282   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1283 
1284   Value *CArgVal = CI->getArgOperand(0);
1285   BasicBlock * const CBB = CI->getParent();
1286 
1287   int const MaxScan = 30;
1288 
1289   { // fold in load value.
1290     LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
1291     if (LI && LI->getParent() == CBB) {
1292       BasicBlock::iterator BBI = LI->getIterator();
1293       Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
1294       if (AvailableVal) {
1295         CArgVal->replaceAllUsesWith(AvailableVal);
1296         if (CArgVal->getNumUses() == 0)
1297           LI->eraseFromParent();
1298         CArgVal = CI->getArgOperand(0);
1299       }
1300     }
1301   }
1302 
1303   Module *M = CI->getModule();
1304   fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN);
1305   std::string const PairName = fInfo.mangle();
1306 
1307   CallInst *UI = nullptr;
1308   for (User* U : CArgVal->users()) {
1309     CallInst *XI = dyn_cast_or_null<CallInst>(U);
1310     if (!XI || XI == CI || XI->getParent() != CBB)
1311       continue;
1312 
1313     Function *UCallee = XI->getCalledFunction();
1314     if (!UCallee || !UCallee->getName().equals(PairName))
1315       continue;
1316 
1317     BasicBlock::iterator BBI = CI->getIterator();
1318     if (BBI == CI->getParent()->begin())
1319       break;
1320     --BBI;
1321     for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
1322       if (cast<Instruction>(BBI) == XI) {
1323         UI = XI;
1324         break;
1325       }
1326     }
1327     if (UI) break;
1328   }
1329 
1330   if (!UI) return false;
1331 
1332   // Merge the sin and cos.
1333 
1334   // for OpenCL 2.0 we have only generic implementation of sincos
1335   // function.
1336   AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
1337   nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1338   FunctionCallee Fsincos = getFunction(M, nf);
1339   if (!Fsincos) return false;
1340 
1341   BasicBlock::iterator ItOld = B.GetInsertPoint();
1342   AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
1343   B.SetInsertPoint(UI);
1344 
1345   Value *P = Alloc;
1346   Type *PTy = Fsincos.getFunctionType()->getParamType(1);
1347   // The allocaInst allocates the memory in private address space. This need
1348   // to be bitcasted to point to the address space of cos pointer type.
1349   // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1350   if (PTy->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
1351     P = B.CreateAddrSpaceCast(Alloc, PTy);
1352   CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P);
1353 
1354   LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with "
1355                     << *Call << "\n");
1356 
1357   if (!isSin) { // CI->cos, UI->sin
1358     B.SetInsertPoint(&*ItOld);
1359     UI->replaceAllUsesWith(&*Call);
1360     Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1361     CI->replaceAllUsesWith(Reload);
1362     UI->eraseFromParent();
1363     CI->eraseFromParent();
1364   } else { // CI->sin, UI->cos
1365     Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), 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(const TargetOptions &Opt) {
1684   return new AMDGPUSimplifyLibCalls(Opt);
1685 }
1686 
1687 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
1688   return new AMDGPUUseNativeCalls();
1689 }
1690 
1691 static bool setFastFlags(Function &F, const TargetOptions &Options) {
1692   AttrBuilder B;
1693 
1694   if (Options.UnsafeFPMath || Options.NoInfsFPMath)
1695     B.addAttribute("no-infs-fp-math", "true");
1696   if (Options.UnsafeFPMath || Options.NoNaNsFPMath)
1697     B.addAttribute("no-nans-fp-math", "true");
1698   if (Options.UnsafeFPMath) {
1699     B.addAttribute("less-precise-fpmad", "true");
1700     B.addAttribute("unsafe-fp-math", "true");
1701   }
1702 
1703   if (!B.hasAttributes())
1704     return false;
1705 
1706   F.addAttributes(AttributeList::FunctionIndex, B);
1707 
1708   return true;
1709 }
1710 
1711 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
1712   if (skipFunction(F))
1713     return false;
1714 
1715   bool Changed = false;
1716   auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1717 
1718   LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1719              F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1720 
1721   if (!EnablePreLink)
1722     Changed |= setFastFlags(F, Options);
1723 
1724   for (auto &BB : F) {
1725     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1726       // Ignore non-calls.
1727       CallInst *CI = dyn_cast<CallInst>(I);
1728       ++I;
1729       if (!CI) continue;
1730 
1731       // Ignore indirect calls.
1732       Function *Callee = CI->getCalledFunction();
1733       if (Callee == 0) continue;
1734 
1735       LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1736                  dbgs().flush());
1737       if(Simplifier.fold(CI, AA))
1738         Changed = true;
1739     }
1740   }
1741   return Changed;
1742 }
1743 
1744 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) {
1745   if (skipFunction(F) || UseNative.empty())
1746     return false;
1747 
1748   bool Changed = false;
1749   for (auto &BB : F) {
1750     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1751       // Ignore non-calls.
1752       CallInst *CI = dyn_cast<CallInst>(I);
1753       ++I;
1754       if (!CI) continue;
1755 
1756       // Ignore indirect calls.
1757       Function *Callee = CI->getCalledFunction();
1758       if (Callee == 0) continue;
1759 
1760       if(Simplifier.useNative(CI))
1761         Changed = true;
1762     }
1763   }
1764   return Changed;
1765 }
1766