xref: /llvm-project/llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp (revision 360a5d5612d913134ed643bf85b40a7518164146)
1 //===- AMDGPULibCalls.cpp -------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file does AMD library function optimizations.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "AMDGPU.h"
15 #include "AMDGPULibFunc.h"
16 #include "GCNSubtarget.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/Loads.h"
19 #include "llvm/IR/IRBuilder.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/IR/IntrinsicsAMDGPU.h"
22 #include "llvm/InitializePasses.h"
23 #include "llvm/Target/TargetMachine.h"
24 #include <cmath>
25 
26 #define DEBUG_TYPE "amdgpu-simplifylib"
27 
28 using namespace llvm;
29 
30 static cl::opt<bool> EnablePreLink("amdgpu-prelink",
31   cl::desc("Enable pre-link mode optimizations"),
32   cl::init(false),
33   cl::Hidden);
34 
35 static cl::list<std::string> UseNative("amdgpu-use-native",
36   cl::desc("Comma separated list of functions to replace with native, or all"),
37   cl::CommaSeparated, cl::ValueOptional,
38   cl::Hidden);
39 
40 #define MATH_PI      numbers::pi
41 #define MATH_E       numbers::e
42 #define MATH_SQRT2   numbers::sqrt2
43 #define MATH_SQRT1_2 numbers::inv_sqrt2
44 
45 namespace llvm {
46 
47 class AMDGPULibCalls {
48 private:
49 
50   typedef llvm::AMDGPULibFunc FuncInfo;
51 
52   const TargetMachine *TM;
53 
54   // -fuse-native.
55   bool AllNative = false;
56 
57   bool useNativeFunc(const StringRef F) const;
58 
59   // Return a pointer (pointer expr) to the function if function definition with
60   // "FuncName" exists. It may create a new function prototype in pre-link mode.
61   FunctionCallee getFunction(Module *M, const FuncInfo &fInfo);
62 
63   bool parseFunctionName(const StringRef &FMangledName, FuncInfo &FInfo);
64 
65   bool TDOFold(CallInst *CI, const FuncInfo &FInfo);
66 
67   /* Specialized optimizations */
68 
69   // recip (half or native)
70   bool fold_recip(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
71 
72   // divide (half or native)
73   bool fold_divide(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
74 
75   // pow/powr/pown
76   bool fold_pow(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
77 
78   // rootn
79   bool fold_rootn(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
80 
81   // fma/mad
82   bool fold_fma_mad(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
83 
84   // -fuse-native for sincos
85   bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo);
86 
87   // evaluate calls if calls' arguments are constants.
88   bool evaluateScalarMathFunc(const FuncInfo &FInfo, double& Res0,
89     double& Res1, Constant *copr0, Constant *copr1, Constant *copr2);
90   bool evaluateCall(CallInst *aCI, const FuncInfo &FInfo);
91 
92   // sqrt
93   bool fold_sqrt(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
94 
95   // sin/cos
96   bool fold_sincos(CallInst * CI, IRBuilder<> &B, AliasAnalysis * AA);
97 
98   // __read_pipe/__write_pipe
99   bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
100                             const FuncInfo &FInfo);
101 
102   // llvm.amdgcn.wavefrontsize
103   bool fold_wavefrontsize(CallInst *CI, IRBuilder<> &B);
104 
105   // Get insertion point at entry.
106   BasicBlock::iterator getEntryIns(CallInst * UI);
107   // Insert an Alloc instruction.
108   AllocaInst* insertAlloca(CallInst * UI, IRBuilder<> &B, const char *prefix);
109   // Get a scalar native builtin single argument FP function
110   FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo);
111 
112 protected:
113   CallInst *CI;
114 
115   bool isUnsafeMath(const CallInst *CI) const;
116 
117   void replaceCall(Value *With) {
118     CI->replaceAllUsesWith(With);
119     CI->eraseFromParent();
120   }
121 
122 public:
123   AMDGPULibCalls(const TargetMachine *TM_ = nullptr) : TM(TM_) {}
124 
125   bool fold(CallInst *CI, AliasAnalysis *AA = nullptr);
126 
127   void initNativeFuncs();
128 
129   // Replace a normal math function call with that native version
130   bool useNative(CallInst *CI);
131 };
132 
133 } // end llvm namespace
134 
135 namespace {
136 
137   class AMDGPUSimplifyLibCalls : public FunctionPass {
138 
139   AMDGPULibCalls Simplifier;
140 
141   public:
142     static char ID; // Pass identification
143 
144     AMDGPUSimplifyLibCalls(const TargetMachine *TM = nullptr)
145       : FunctionPass(ID), Simplifier(TM) {
146       initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
147     }
148 
149     void getAnalysisUsage(AnalysisUsage &AU) const override {
150       AU.addRequired<AAResultsWrapperPass>();
151     }
152 
153     bool runOnFunction(Function &M) override;
154   };
155 
156   class AMDGPUUseNativeCalls : public FunctionPass {
157 
158   AMDGPULibCalls Simplifier;
159 
160   public:
161     static char ID; // Pass identification
162 
163     AMDGPUUseNativeCalls() : FunctionPass(ID) {
164       initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry());
165       Simplifier.initNativeFuncs();
166     }
167 
168     bool runOnFunction(Function &F) override;
169   };
170 
171 } // end anonymous namespace.
172 
173 char AMDGPUSimplifyLibCalls::ID = 0;
174 char AMDGPUUseNativeCalls::ID = 0;
175 
176 INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
177                       "Simplify well-known AMD library calls", false, false)
178 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
179 INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",
180                     "Simplify well-known AMD library calls", false, false)
181 
182 INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative",
183                 "Replace builtin math calls with that native versions.",
184                 false, false)
185 
186 template <typename IRB>
187 static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg,
188                               const Twine &Name = "") {
189   CallInst *R = B.CreateCall(Callee, Arg, Name);
190   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
191     R->setCallingConv(F->getCallingConv());
192   return R;
193 }
194 
195 template <typename IRB>
196 static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1,
197                                Value *Arg2, const Twine &Name = "") {
198   CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
199   if (Function *F = dyn_cast<Function>(Callee.getCallee()))
200     R->setCallingConv(F->getCallingConv());
201   return R;
202 }
203 
204 //  Data structures for table-driven optimizations.
205 //  FuncTbl works for both f32 and f64 functions with 1 input argument
206 
207 struct TableEntry {
208   double   result;
209   double   input;
210 };
211 
212 /* a list of {result, input} */
213 static const TableEntry tbl_acos[] = {
214   {MATH_PI / 2.0, 0.0},
215   {MATH_PI / 2.0, -0.0},
216   {0.0, 1.0},
217   {MATH_PI, -1.0}
218 };
219 static const TableEntry tbl_acosh[] = {
220   {0.0, 1.0}
221 };
222 static const TableEntry tbl_acospi[] = {
223   {0.5, 0.0},
224   {0.5, -0.0},
225   {0.0, 1.0},
226   {1.0, -1.0}
227 };
228 static const TableEntry tbl_asin[] = {
229   {0.0, 0.0},
230   {-0.0, -0.0},
231   {MATH_PI / 2.0, 1.0},
232   {-MATH_PI / 2.0, -1.0}
233 };
234 static const TableEntry tbl_asinh[] = {
235   {0.0, 0.0},
236   {-0.0, -0.0}
237 };
238 static const TableEntry tbl_asinpi[] = {
239   {0.0, 0.0},
240   {-0.0, -0.0},
241   {0.5, 1.0},
242   {-0.5, -1.0}
243 };
244 static const TableEntry tbl_atan[] = {
245   {0.0, 0.0},
246   {-0.0, -0.0},
247   {MATH_PI / 4.0, 1.0},
248   {-MATH_PI / 4.0, -1.0}
249 };
250 static const TableEntry tbl_atanh[] = {
251   {0.0, 0.0},
252   {-0.0, -0.0}
253 };
254 static const TableEntry tbl_atanpi[] = {
255   {0.0, 0.0},
256   {-0.0, -0.0},
257   {0.25, 1.0},
258   {-0.25, -1.0}
259 };
260 static const TableEntry tbl_cbrt[] = {
261   {0.0, 0.0},
262   {-0.0, -0.0},
263   {1.0, 1.0},
264   {-1.0, -1.0},
265 };
266 static const TableEntry tbl_cos[] = {
267   {1.0, 0.0},
268   {1.0, -0.0}
269 };
270 static const TableEntry tbl_cosh[] = {
271   {1.0, 0.0},
272   {1.0, -0.0}
273 };
274 static const TableEntry tbl_cospi[] = {
275   {1.0, 0.0},
276   {1.0, -0.0}
277 };
278 static const TableEntry tbl_erfc[] = {
279   {1.0, 0.0},
280   {1.0, -0.0}
281 };
282 static const TableEntry tbl_erf[] = {
283   {0.0, 0.0},
284   {-0.0, -0.0}
285 };
286 static const TableEntry tbl_exp[] = {
287   {1.0, 0.0},
288   {1.0, -0.0},
289   {MATH_E, 1.0}
290 };
291 static const TableEntry tbl_exp2[] = {
292   {1.0, 0.0},
293   {1.0, -0.0},
294   {2.0, 1.0}
295 };
296 static const TableEntry tbl_exp10[] = {
297   {1.0, 0.0},
298   {1.0, -0.0},
299   {10.0, 1.0}
300 };
301 static const TableEntry tbl_expm1[] = {
302   {0.0, 0.0},
303   {-0.0, -0.0}
304 };
305 static const TableEntry tbl_log[] = {
306   {0.0, 1.0},
307   {1.0, MATH_E}
308 };
309 static const TableEntry tbl_log2[] = {
310   {0.0, 1.0},
311   {1.0, 2.0}
312 };
313 static const TableEntry tbl_log10[] = {
314   {0.0, 1.0},
315   {1.0, 10.0}
316 };
317 static const TableEntry tbl_rsqrt[] = {
318   {1.0, 1.0},
319   {MATH_SQRT1_2, 2.0}
320 };
321 static const TableEntry tbl_sin[] = {
322   {0.0, 0.0},
323   {-0.0, -0.0}
324 };
325 static const TableEntry tbl_sinh[] = {
326   {0.0, 0.0},
327   {-0.0, -0.0}
328 };
329 static const TableEntry tbl_sinpi[] = {
330   {0.0, 0.0},
331   {-0.0, -0.0}
332 };
333 static const TableEntry tbl_sqrt[] = {
334   {0.0, 0.0},
335   {1.0, 1.0},
336   {MATH_SQRT2, 2.0}
337 };
338 static const TableEntry tbl_tan[] = {
339   {0.0, 0.0},
340   {-0.0, -0.0}
341 };
342 static const TableEntry tbl_tanh[] = {
343   {0.0, 0.0},
344   {-0.0, -0.0}
345 };
346 static const TableEntry tbl_tanpi[] = {
347   {0.0, 0.0},
348   {-0.0, -0.0}
349 };
350 static const TableEntry tbl_tgamma[] = {
351   {1.0, 1.0},
352   {1.0, 2.0},
353   {2.0, 3.0},
354   {6.0, 4.0}
355 };
356 
357 static bool HasNative(AMDGPULibFunc::EFuncId id) {
358   switch(id) {
359   case AMDGPULibFunc::EI_DIVIDE:
360   case AMDGPULibFunc::EI_COS:
361   case AMDGPULibFunc::EI_EXP:
362   case AMDGPULibFunc::EI_EXP2:
363   case AMDGPULibFunc::EI_EXP10:
364   case AMDGPULibFunc::EI_LOG:
365   case AMDGPULibFunc::EI_LOG2:
366   case AMDGPULibFunc::EI_LOG10:
367   case AMDGPULibFunc::EI_POWR:
368   case AMDGPULibFunc::EI_RECIP:
369   case AMDGPULibFunc::EI_RSQRT:
370   case AMDGPULibFunc::EI_SIN:
371   case AMDGPULibFunc::EI_SINCOS:
372   case AMDGPULibFunc::EI_SQRT:
373   case AMDGPULibFunc::EI_TAN:
374     return true;
375   default:;
376   }
377   return false;
378 }
379 
380 using TableRef = ArrayRef<TableEntry>;
381 
382 static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
383   switch(id) {
384   case AMDGPULibFunc::EI_ACOS:    return TableRef(tbl_acos);
385   case AMDGPULibFunc::EI_ACOSH:   return TableRef(tbl_acosh);
386   case AMDGPULibFunc::EI_ACOSPI:  return TableRef(tbl_acospi);
387   case AMDGPULibFunc::EI_ASIN:    return TableRef(tbl_asin);
388   case AMDGPULibFunc::EI_ASINH:   return TableRef(tbl_asinh);
389   case AMDGPULibFunc::EI_ASINPI:  return TableRef(tbl_asinpi);
390   case AMDGPULibFunc::EI_ATAN:    return TableRef(tbl_atan);
391   case AMDGPULibFunc::EI_ATANH:   return TableRef(tbl_atanh);
392   case AMDGPULibFunc::EI_ATANPI:  return TableRef(tbl_atanpi);
393   case AMDGPULibFunc::EI_CBRT:    return TableRef(tbl_cbrt);
394   case AMDGPULibFunc::EI_NCOS:
395   case AMDGPULibFunc::EI_COS:     return TableRef(tbl_cos);
396   case AMDGPULibFunc::EI_COSH:    return TableRef(tbl_cosh);
397   case AMDGPULibFunc::EI_COSPI:   return TableRef(tbl_cospi);
398   case AMDGPULibFunc::EI_ERFC:    return TableRef(tbl_erfc);
399   case AMDGPULibFunc::EI_ERF:     return TableRef(tbl_erf);
400   case AMDGPULibFunc::EI_EXP:     return TableRef(tbl_exp);
401   case AMDGPULibFunc::EI_NEXP2:
402   case AMDGPULibFunc::EI_EXP2:    return TableRef(tbl_exp2);
403   case AMDGPULibFunc::EI_EXP10:   return TableRef(tbl_exp10);
404   case AMDGPULibFunc::EI_EXPM1:   return TableRef(tbl_expm1);
405   case AMDGPULibFunc::EI_LOG:     return TableRef(tbl_log);
406   case AMDGPULibFunc::EI_NLOG2:
407   case AMDGPULibFunc::EI_LOG2:    return TableRef(tbl_log2);
408   case AMDGPULibFunc::EI_LOG10:   return TableRef(tbl_log10);
409   case AMDGPULibFunc::EI_NRSQRT:
410   case AMDGPULibFunc::EI_RSQRT:   return TableRef(tbl_rsqrt);
411   case AMDGPULibFunc::EI_NSIN:
412   case AMDGPULibFunc::EI_SIN:     return TableRef(tbl_sin);
413   case AMDGPULibFunc::EI_SINH:    return TableRef(tbl_sinh);
414   case AMDGPULibFunc::EI_SINPI:   return TableRef(tbl_sinpi);
415   case AMDGPULibFunc::EI_NSQRT:
416   case AMDGPULibFunc::EI_SQRT:    return TableRef(tbl_sqrt);
417   case AMDGPULibFunc::EI_TAN:     return TableRef(tbl_tan);
418   case AMDGPULibFunc::EI_TANH:    return TableRef(tbl_tanh);
419   case AMDGPULibFunc::EI_TANPI:   return TableRef(tbl_tanpi);
420   case AMDGPULibFunc::EI_TGAMMA:  return TableRef(tbl_tgamma);
421   default:;
422   }
423   return TableRef();
424 }
425 
426 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
427   return FInfo.getLeads()[0].VectorSize;
428 }
429 
430 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
431   return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
432 }
433 
434 FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) {
435   // If we are doing PreLinkOpt, the function is external. So it is safe to
436   // use getOrInsertFunction() at this stage.
437 
438   return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
439                        : AMDGPULibFunc::getFunction(M, fInfo);
440 }
441 
442 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName,
443                                        FuncInfo &FInfo) {
444   return AMDGPULibFunc::parse(FMangledName, FInfo);
445 }
446 
447 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const {
448   if (auto Op = dyn_cast<FPMathOperator>(CI))
449     if (Op->isFast())
450       return true;
451   const Function *F = CI->getParent()->getParent();
452   Attribute Attr = F->getFnAttribute("unsafe-fp-math");
453   return Attr.getValueAsBool();
454 }
455 
456 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
457   return AllNative || llvm::is_contained(UseNative, F);
458 }
459 
460 void AMDGPULibCalls::initNativeFuncs() {
461   AllNative = useNativeFunc("all") ||
462               (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
463                UseNative.begin()->empty());
464 }
465 
466 bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
467   bool native_sin = useNativeFunc("sin");
468   bool native_cos = useNativeFunc("cos");
469 
470   if (native_sin && native_cos) {
471     Module *M = aCI->getModule();
472     Value *opr0 = aCI->getArgOperand(0);
473 
474     AMDGPULibFunc nf;
475     nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
476     nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
477 
478     nf.setPrefix(AMDGPULibFunc::NATIVE);
479     nf.setId(AMDGPULibFunc::EI_SIN);
480     FunctionCallee sinExpr = getFunction(M, nf);
481 
482     nf.setPrefix(AMDGPULibFunc::NATIVE);
483     nf.setId(AMDGPULibFunc::EI_COS);
484     FunctionCallee cosExpr = getFunction(M, nf);
485     if (sinExpr && cosExpr) {
486       Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
487       Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
488       new StoreInst(cosval, aCI->getArgOperand(1), aCI);
489 
490       DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
491                                           << " with native version of sin/cos");
492 
493       replaceCall(sinval);
494       return true;
495     }
496   }
497   return false;
498 }
499 
500 bool AMDGPULibCalls::useNative(CallInst *aCI) {
501   CI = aCI;
502   Function *Callee = aCI->getCalledFunction();
503 
504   FuncInfo FInfo;
505   if (!parseFunctionName(Callee->getName(), FInfo) || !FInfo.isMangled() ||
506       FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
507       getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
508       !(AllNative || useNativeFunc(FInfo.getName()))) {
509     return false;
510   }
511 
512   if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
513     return sincosUseNative(aCI, FInfo);
514 
515   FInfo.setPrefix(AMDGPULibFunc::NATIVE);
516   FunctionCallee F = getFunction(aCI->getModule(), FInfo);
517   if (!F)
518     return false;
519 
520   aCI->setCalledFunction(F);
521   DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCI
522                                       << " with native version");
523   return true;
524 }
525 
526 // Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
527 // builtin, with appended type size and alignment arguments, where 2 or 4
528 // indicates the original number of arguments. The library has optimized version
529 // of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
530 // power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
531 // for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
532 // 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
533 bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
534                                           const FuncInfo &FInfo) {
535   auto *Callee = CI->getCalledFunction();
536   if (!Callee->isDeclaration())
537     return false;
538 
539   assert(Callee->hasName() && "Invalid read_pipe/write_pipe function");
540   auto *M = Callee->getParent();
541   std::string Name = std::string(Callee->getName());
542   auto NumArg = CI->arg_size();
543   if (NumArg != 4 && NumArg != 6)
544     return false;
545   auto *PacketSize = CI->getArgOperand(NumArg - 2);
546   auto *PacketAlign = CI->getArgOperand(NumArg - 1);
547   if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign))
548     return false;
549   unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue();
550   Align Alignment = cast<ConstantInt>(PacketAlign)->getAlignValue();
551   if (Alignment != Size)
552     return false;
553 
554   unsigned PtrArgLoc = CI->arg_size() - 3;
555   Value *PtrArg = CI->getArgOperand(PtrArgLoc);
556   Type *PtrTy = PtrArg->getType();
557 
558   SmallVector<llvm::Type *, 6> ArgTys;
559   for (unsigned I = 0; I != PtrArgLoc; ++I)
560     ArgTys.push_back(CI->getArgOperand(I)->getType());
561   ArgTys.push_back(PtrTy);
562 
563   Name = Name + "_" + std::to_string(Size);
564   auto *FTy = FunctionType::get(Callee->getReturnType(),
565                                 ArrayRef<Type *>(ArgTys), false);
566   AMDGPULibFunc NewLibFunc(Name, FTy);
567   FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
568   if (!F)
569     return false;
570 
571   auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
572   SmallVector<Value *, 6> Args;
573   for (unsigned I = 0; I != PtrArgLoc; ++I)
574     Args.push_back(CI->getArgOperand(I));
575   Args.push_back(BCast);
576 
577   auto *NCI = B.CreateCall(F, Args);
578   NCI->setAttributes(CI->getAttributes());
579   CI->replaceAllUsesWith(NCI);
580   CI->dropAllReferences();
581   CI->eraseFromParent();
582 
583   return true;
584 }
585 
586 // This function returns false if no change; return true otherwise.
587 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) {
588   this->CI = CI;
589   Function *Callee = CI->getCalledFunction();
590 
591   // Ignore indirect calls.
592   if (Callee == nullptr)
593     return false;
594 
595   BasicBlock *BB = CI->getParent();
596   LLVMContext &Context = CI->getParent()->getContext();
597   IRBuilder<> B(Context);
598 
599   // Set the builder to the instruction after the call.
600   B.SetInsertPoint(BB, CI->getIterator());
601 
602   // Copy fast flags from the original call.
603   if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI))
604     B.setFastMathFlags(FPOp->getFastMathFlags());
605 
606   switch (Callee->getIntrinsicID()) {
607   default:
608     break;
609   case Intrinsic::amdgcn_wavefrontsize:
610     return !EnablePreLink && fold_wavefrontsize(CI, B);
611   }
612 
613   FuncInfo FInfo;
614   if (!parseFunctionName(Callee->getName(), FInfo))
615     return false;
616 
617   // Further check the number of arguments to see if they match.
618   if (CI->arg_size() != FInfo.getNumArgs())
619     return false;
620 
621   if (TDOFold(CI, FInfo))
622     return true;
623 
624   // Under unsafe-math, evaluate calls if possible.
625   // According to Brian Sumner, we can do this for all f32 function calls
626   // using host's double function calls.
627   if (isUnsafeMath(CI) && evaluateCall(CI, FInfo))
628     return true;
629 
630   // Specialized optimizations for each function call
631   switch (FInfo.getId()) {
632   case AMDGPULibFunc::EI_RECIP:
633     // skip vector function
634     assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
635              FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
636             "recip must be an either native or half function");
637     return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo);
638 
639   case AMDGPULibFunc::EI_DIVIDE:
640     // skip vector function
641     assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
642              FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
643             "divide must be an either native or half function");
644     return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo);
645 
646   case AMDGPULibFunc::EI_POW:
647   case AMDGPULibFunc::EI_POWR:
648   case AMDGPULibFunc::EI_POWN:
649     return fold_pow(CI, B, FInfo);
650 
651   case AMDGPULibFunc::EI_ROOTN:
652     // skip vector function
653     return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo);
654 
655   case AMDGPULibFunc::EI_FMA:
656   case AMDGPULibFunc::EI_MAD:
657   case AMDGPULibFunc::EI_NFMA:
658     // skip vector function
659     return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo);
660 
661   case AMDGPULibFunc::EI_SQRT:
662     return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo);
663   case AMDGPULibFunc::EI_COS:
664   case AMDGPULibFunc::EI_SIN:
665     if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
666          getArgType(FInfo) == AMDGPULibFunc::F64)
667         && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX))
668       return fold_sincos(CI, B, AA);
669 
670     break;
671   case AMDGPULibFunc::EI_READ_PIPE_2:
672   case AMDGPULibFunc::EI_READ_PIPE_4:
673   case AMDGPULibFunc::EI_WRITE_PIPE_2:
674   case AMDGPULibFunc::EI_WRITE_PIPE_4:
675     return fold_read_write_pipe(CI, B, FInfo);
676 
677   default:
678     break;
679   }
680 
681   return false;
682 }
683 
684 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
685   // Table-Driven optimization
686   const TableRef tr = getOptTable(FInfo.getId());
687   if (tr.empty())
688     return false;
689 
690   int const sz = (int)tr.size();
691   Value *opr0 = CI->getArgOperand(0);
692 
693   if (getVecSize(FInfo) > 1) {
694     if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
695       SmallVector<double, 0> DVal;
696       for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
697         ConstantFP *eltval = dyn_cast<ConstantFP>(
698                                CV->getElementAsConstant((unsigned)eltNo));
699         assert(eltval && "Non-FP arguments in math function!");
700         bool found = false;
701         for (int i=0; i < sz; ++i) {
702           if (eltval->isExactlyValue(tr[i].input)) {
703             DVal.push_back(tr[i].result);
704             found = true;
705             break;
706           }
707         }
708         if (!found) {
709           // This vector constants not handled yet.
710           return false;
711         }
712       }
713       LLVMContext &context = CI->getParent()->getParent()->getContext();
714       Constant *nval;
715       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
716         SmallVector<float, 0> FVal;
717         for (unsigned i = 0; i < DVal.size(); ++i) {
718           FVal.push_back((float)DVal[i]);
719         }
720         ArrayRef<float> tmp(FVal);
721         nval = ConstantDataVector::get(context, tmp);
722       } else { // F64
723         ArrayRef<double> tmp(DVal);
724         nval = ConstantDataVector::get(context, tmp);
725       }
726       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
727       replaceCall(nval);
728       return true;
729     }
730   } else {
731     // Scalar version
732     if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
733       for (int i = 0; i < sz; ++i) {
734         if (CF->isExactlyValue(tr[i].input)) {
735           Value *nval = ConstantFP::get(CF->getType(), tr[i].result);
736           LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
737           replaceCall(nval);
738           return true;
739         }
740       }
741     }
742   }
743 
744   return false;
745 }
746 
747 //  [native_]half_recip(c) ==> 1.0/c
748 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B,
749                                 const FuncInfo &FInfo) {
750   Value *opr0 = CI->getArgOperand(0);
751   if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
752     // Just create a normal div. Later, InstCombine will be able
753     // to compute the divide into a constant (avoid check float infinity
754     // or subnormal at this point).
755     Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0),
756                                opr0,
757                                "recip2div");
758     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
759     replaceCall(nval);
760     return true;
761   }
762   return false;
763 }
764 
765 //  [native_]half_divide(x, c) ==> x/c
766 bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B,
767                                  const FuncInfo &FInfo) {
768   Value *opr0 = CI->getArgOperand(0);
769   Value *opr1 = CI->getArgOperand(1);
770   ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
771   ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
772 
773   if ((CF0 && CF1) ||  // both are constants
774       (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32)))
775       // CF1 is constant && f32 divide
776   {
777     Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0),
778                                 opr1, "__div2recip");
779     Value *nval  = B.CreateFMul(opr0, nval1, "__div2mul");
780     replaceCall(nval);
781     return true;
782   }
783   return false;
784 }
785 
786 namespace llvm {
787 static double log2(double V) {
788 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L
789   return ::log2(V);
790 #else
791   return log(V) / numbers::ln2;
792 #endif
793 }
794 }
795 
796 bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B,
797                               const FuncInfo &FInfo) {
798   assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
799           FInfo.getId() == AMDGPULibFunc::EI_POWR ||
800           FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
801          "fold_pow: encounter a wrong function call");
802 
803   Value *opr0, *opr1;
804   ConstantFP *CF;
805   ConstantInt *CINT;
806   ConstantAggregateZero *CZero;
807   Type *eltType;
808 
809   opr0 = CI->getArgOperand(0);
810   opr1 = CI->getArgOperand(1);
811   CZero = dyn_cast<ConstantAggregateZero>(opr1);
812   if (getVecSize(FInfo) == 1) {
813     eltType = opr0->getType();
814     CF = dyn_cast<ConstantFP>(opr1);
815     CINT = dyn_cast<ConstantInt>(opr1);
816   } else {
817     VectorType *VTy = dyn_cast<VectorType>(opr0->getType());
818     assert(VTy && "Oprand of vector function should be of vectortype");
819     eltType = VTy->getElementType();
820     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1);
821 
822     // Now, only Handle vector const whose elements have the same value.
823     CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr;
824     CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr;
825   }
826 
827   // No unsafe math , no constant argument, do nothing
828   if (!isUnsafeMath(CI) && !CF && !CINT && !CZero)
829     return false;
830 
831   // 0x1111111 means that we don't do anything for this call.
832   int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
833 
834   if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) {
835     //  pow/powr/pown(x, 0) == 1
836     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n");
837     Constant *cnval = ConstantFP::get(eltType, 1.0);
838     if (getVecSize(FInfo) > 1) {
839       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
840     }
841     replaceCall(cnval);
842     return true;
843   }
844   if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
845     // pow/powr/pown(x, 1.0) = x
846     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
847     replaceCall(opr0);
848     return true;
849   }
850   if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
851     // pow/powr/pown(x, 2.0) = x*x
852     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0
853                       << "\n");
854     Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
855     replaceCall(nval);
856     return true;
857   }
858   if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
859     // pow/powr/pown(x, -1.0) = 1.0/x
860     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n");
861     Constant *cnval = ConstantFP::get(eltType, 1.0);
862     if (getVecSize(FInfo) > 1) {
863       cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
864     }
865     Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
866     replaceCall(nval);
867     return true;
868   }
869 
870   Module *M = CI->getModule();
871   if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
872     // pow[r](x, [-]0.5) = sqrt(x)
873     bool issqrt = CF->isExactlyValue(0.5);
874     if (FunctionCallee FPExpr =
875             getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
876                                                 : AMDGPULibFunc::EI_RSQRT,
877                                          FInfo))) {
878       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
879                         << FInfo.getName().c_str() << "(" << *opr0 << ")\n");
880       Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
881                                                         : "__pow2rsqrt");
882       replaceCall(nval);
883       return true;
884     }
885   }
886 
887   if (!isUnsafeMath(CI))
888     return false;
889 
890   // Unsafe Math optimization
891 
892   // Remember that ci_opr1 is set if opr1 is integral
893   if (CF) {
894     double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
895                     ? (double)CF->getValueAPF().convertToFloat()
896                     : CF->getValueAPF().convertToDouble();
897     int ival = (int)dval;
898     if ((double)ival == dval) {
899       ci_opr1 = ival;
900     } else
901       ci_opr1 = 0x11111111;
902   }
903 
904   // pow/powr/pown(x, c) = [1/](x*x*..x); where
905   //   trunc(c) == c && the number of x == c && |c| <= 12
906   unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
907   if (abs_opr1 <= 12) {
908     Constant *cnval;
909     Value *nval;
910     if (abs_opr1 == 0) {
911       cnval = ConstantFP::get(eltType, 1.0);
912       if (getVecSize(FInfo) > 1) {
913         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
914       }
915       nval = cnval;
916     } else {
917       Value *valx2 = nullptr;
918       nval = nullptr;
919       while (abs_opr1 > 0) {
920         valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
921         if (abs_opr1 & 1) {
922           nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
923         }
924         abs_opr1 >>= 1;
925       }
926     }
927 
928     if (ci_opr1 < 0) {
929       cnval = ConstantFP::get(eltType, 1.0);
930       if (getVecSize(FInfo) > 1) {
931         cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
932       }
933       nval = B.CreateFDiv(cnval, nval, "__1powprod");
934     }
935     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
936                       << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
937                       << ")\n");
938     replaceCall(nval);
939     return true;
940   }
941 
942   // powr ---> exp2(y * log2(x))
943   // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
944   FunctionCallee ExpExpr =
945       getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo));
946   if (!ExpExpr)
947     return false;
948 
949   bool needlog = false;
950   bool needabs = false;
951   bool needcopysign = false;
952   Constant *cnval = nullptr;
953   if (getVecSize(FInfo) == 1) {
954     CF = dyn_cast<ConstantFP>(opr0);
955 
956     if (CF) {
957       double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
958                    ? (double)CF->getValueAPF().convertToFloat()
959                    : CF->getValueAPF().convertToDouble();
960 
961       V = log2(std::abs(V));
962       cnval = ConstantFP::get(eltType, V);
963       needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
964                      CF->isNegative();
965     } else {
966       needlog = true;
967       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR &&
968                                (!CF || CF->isNegative());
969     }
970   } else {
971     ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
972 
973     if (!CDV) {
974       needlog = true;
975       needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
976     } else {
977       assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
978               "Wrong vector size detected");
979 
980       SmallVector<double, 0> DVal;
981       for (int i=0; i < getVecSize(FInfo); ++i) {
982         double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
983                      ? (double)CDV->getElementAsFloat(i)
984                      : CDV->getElementAsDouble(i);
985         if (V < 0.0) needcopysign = true;
986         V = log2(std::abs(V));
987         DVal.push_back(V);
988       }
989       if (getArgType(FInfo) == AMDGPULibFunc::F32) {
990         SmallVector<float, 0> FVal;
991         for (unsigned i=0; i < DVal.size(); ++i) {
992           FVal.push_back((float)DVal[i]);
993         }
994         ArrayRef<float> tmp(FVal);
995         cnval = ConstantDataVector::get(M->getContext(), tmp);
996       } else {
997         ArrayRef<double> tmp(DVal);
998         cnval = ConstantDataVector::get(M->getContext(), tmp);
999       }
1000     }
1001   }
1002 
1003   if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1004     // We cannot handle corner cases for a general pow() function, give up
1005     // unless y is a constant integral value. Then proceed as if it were pown.
1006     if (getVecSize(FInfo) == 1) {
1007       if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) {
1008         double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1009                    ? (double)CF->getValueAPF().convertToFloat()
1010                    : CF->getValueAPF().convertToDouble();
1011         if (y != (double)(int64_t)y)
1012           return false;
1013       } else
1014         return false;
1015     } else {
1016       if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) {
1017         for (int i=0; i < getVecSize(FInfo); ++i) {
1018           double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1019                      ? (double)CDV->getElementAsFloat(i)
1020                      : CDV->getElementAsDouble(i);
1021           if (y != (double)(int64_t)y)
1022             return false;
1023         }
1024       } else
1025         return false;
1026     }
1027   }
1028 
1029   Value *nval;
1030   if (needabs) {
1031     FunctionCallee AbsExpr =
1032         getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS, FInfo));
1033     if (!AbsExpr)
1034       return false;
1035     nval = CreateCallEx(B, AbsExpr, opr0, "__fabs");
1036   } else {
1037     nval = cnval ? cnval : opr0;
1038   }
1039   if (needlog) {
1040     FunctionCallee LogExpr =
1041         getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo));
1042     if (!LogExpr)
1043       return false;
1044     nval = CreateCallEx(B,LogExpr, nval, "__log2");
1045   }
1046 
1047   if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1048     // convert int(32) to fp(f32 or f64)
1049     opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1050   }
1051   nval = B.CreateFMul(opr1, nval, "__ylogx");
1052   nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1053 
1054   if (needcopysign) {
1055     Value *opr_n;
1056     Type* rTy = opr0->getType();
1057     Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty();
1058     Type *nTy = nTyS;
1059     if (const auto *vTy = dyn_cast<FixedVectorType>(rTy))
1060       nTy = FixedVectorType::get(nTyS, vTy);
1061     unsigned size = nTy->getScalarSizeInBits();
1062     opr_n = CI->getArgOperand(1);
1063     if (opr_n->getType()->isIntegerTy())
1064       opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou");
1065     else
1066       opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1067 
1068     Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1069     sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1070     nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1071     nval = B.CreateBitCast(nval, opr0->getType());
1072   }
1073 
1074   LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1075                     << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1076   replaceCall(nval);
1077 
1078   return true;
1079 }
1080 
1081 bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B,
1082                                 const FuncInfo &FInfo) {
1083   Value *opr0 = CI->getArgOperand(0);
1084   Value *opr1 = CI->getArgOperand(1);
1085 
1086   ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1087   if (!CINT) {
1088     return false;
1089   }
1090   int ci_opr1 = (int)CINT->getSExtValue();
1091   if (ci_opr1 == 1) {  // rootn(x, 1) = x
1092     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
1093     replaceCall(opr0);
1094     return true;
1095   }
1096   if (ci_opr1 == 2) {  // rootn(x, 2) = sqrt(x)
1097     Module *M = CI->getModule();
1098     if (FunctionCallee FPExpr =
1099             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1100       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n");
1101       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1102       replaceCall(nval);
1103       return true;
1104     }
1105   } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1106     Module *M = CI->getModule();
1107     if (FunctionCallee FPExpr =
1108             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1109       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n");
1110       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1111       replaceCall(nval);
1112       return true;
1113     }
1114   } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1115     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n");
1116     Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1117                                opr0,
1118                                "__rootn2div");
1119     replaceCall(nval);
1120     return true;
1121   } else if (ci_opr1 == -2) {  // rootn(x, -2) = rsqrt(x)
1122     Module *M = CI->getModule();
1123     if (FunctionCallee FPExpr =
1124             getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) {
1125       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0
1126                         << ")\n");
1127       Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1128       replaceCall(nval);
1129       return true;
1130     }
1131   }
1132   return false;
1133 }
1134 
1135 bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B,
1136                                   const FuncInfo &FInfo) {
1137   Value *opr0 = CI->getArgOperand(0);
1138   Value *opr1 = CI->getArgOperand(1);
1139   Value *opr2 = CI->getArgOperand(2);
1140 
1141   ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
1142   ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
1143   if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) {
1144     // fma/mad(a, b, c) = c if a=0 || b=0
1145     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n");
1146     replaceCall(opr2);
1147     return true;
1148   }
1149   if (CF0 && CF0->isExactlyValue(1.0f)) {
1150     // fma/mad(a, b, c) = b+c if a=1
1151     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2
1152                       << "\n");
1153     Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd");
1154     replaceCall(nval);
1155     return true;
1156   }
1157   if (CF1 && CF1->isExactlyValue(1.0f)) {
1158     // fma/mad(a, b, c) = a+c if b=1
1159     LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2
1160                       << "\n");
1161     Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd");
1162     replaceCall(nval);
1163     return true;
1164   }
1165   if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) {
1166     if (CF->isZero()) {
1167       // fma/mad(a, b, c) = a*b if c=0
1168       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * "
1169                         << *opr1 << "\n");
1170       Value *nval = B.CreateFMul(opr0, opr1, "fmamul");
1171       replaceCall(nval);
1172       return true;
1173     }
1174   }
1175 
1176   return false;
1177 }
1178 
1179 // Get a scalar native builtin single argument FP function
1180 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1181                                                  const FuncInfo &FInfo) {
1182   if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1183     return nullptr;
1184   FuncInfo nf = FInfo;
1185   nf.setPrefix(AMDGPULibFunc::NATIVE);
1186   return getFunction(M, nf);
1187 }
1188 
1189 // fold sqrt -> native_sqrt (x)
1190 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
1191                                const FuncInfo &FInfo) {
1192   if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1193       (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1194     if (FunctionCallee FPExpr = getNativeFunction(
1195             CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1196       Value *opr0 = CI->getArgOperand(0);
1197       LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1198                         << "sqrt(" << *opr0 << ")\n");
1199       Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1200       replaceCall(nval);
1201       return true;
1202     }
1203   }
1204   return false;
1205 }
1206 
1207 // fold sin, cos -> sincos.
1208 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
1209                                  AliasAnalysis *AA) {
1210   AMDGPULibFunc fInfo;
1211   if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo))
1212     return false;
1213 
1214   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1215          fInfo.getId() == AMDGPULibFunc::EI_COS);
1216   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1217 
1218   Value *CArgVal = CI->getArgOperand(0);
1219   BasicBlock * const CBB = CI->getParent();
1220 
1221   int const MaxScan = 30;
1222   bool Changed = false;
1223 
1224   { // fold in load value.
1225     LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
1226     if (LI && LI->getParent() == CBB) {
1227       BasicBlock::iterator BBI = LI->getIterator();
1228       Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
1229       if (AvailableVal) {
1230         Changed = true;
1231         CArgVal->replaceAllUsesWith(AvailableVal);
1232         if (CArgVal->getNumUses() == 0)
1233           LI->eraseFromParent();
1234         CArgVal = CI->getArgOperand(0);
1235       }
1236     }
1237   }
1238 
1239   Module *M = CI->getModule();
1240   fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN);
1241   std::string const PairName = fInfo.mangle();
1242 
1243   CallInst *UI = nullptr;
1244   for (User* U : CArgVal->users()) {
1245     CallInst *XI = dyn_cast_or_null<CallInst>(U);
1246     if (!XI || XI == CI || XI->getParent() != CBB)
1247       continue;
1248 
1249     Function *UCallee = XI->getCalledFunction();
1250     if (!UCallee || !UCallee->getName().equals(PairName))
1251       continue;
1252 
1253     BasicBlock::iterator BBI = CI->getIterator();
1254     if (BBI == CI->getParent()->begin())
1255       break;
1256     --BBI;
1257     for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
1258       if (cast<Instruction>(BBI) == XI) {
1259         UI = XI;
1260         break;
1261       }
1262     }
1263     if (UI) break;
1264   }
1265 
1266   if (!UI)
1267     return Changed;
1268 
1269   // Merge the sin and cos.
1270 
1271   // for OpenCL 2.0 we have only generic implementation of sincos
1272   // function.
1273   AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
1274   nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1275   FunctionCallee Fsincos = getFunction(M, nf);
1276   if (!Fsincos)
1277     return Changed;
1278 
1279   BasicBlock::iterator ItOld = B.GetInsertPoint();
1280   AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
1281   B.SetInsertPoint(UI);
1282 
1283   Value *P = Alloc;
1284   Type *PTy = Fsincos.getFunctionType()->getParamType(1);
1285   // The allocaInst allocates the memory in private address space. This need
1286   // to be bitcasted to point to the address space of cos pointer type.
1287   // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1288   if (PTy->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
1289     P = B.CreateAddrSpaceCast(Alloc, PTy);
1290   CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P);
1291 
1292   LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with "
1293                     << *Call << "\n");
1294 
1295   if (!isSin) { // CI->cos, UI->sin
1296     B.SetInsertPoint(&*ItOld);
1297     UI->replaceAllUsesWith(&*Call);
1298     Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1299     CI->replaceAllUsesWith(Reload);
1300     UI->eraseFromParent();
1301     CI->eraseFromParent();
1302   } else { // CI->sin, UI->cos
1303     Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1304     UI->replaceAllUsesWith(Reload);
1305     CI->replaceAllUsesWith(Call);
1306     UI->eraseFromParent();
1307     CI->eraseFromParent();
1308   }
1309   return true;
1310 }
1311 
1312 bool AMDGPULibCalls::fold_wavefrontsize(CallInst *CI, IRBuilder<> &B) {
1313   if (!TM)
1314     return false;
1315 
1316   StringRef CPU = TM->getTargetCPU();
1317   StringRef Features = TM->getTargetFeatureString();
1318   if ((CPU.empty() || CPU.equals_insensitive("generic")) &&
1319       (Features.empty() || !Features.contains_insensitive("wavefrontsize")))
1320     return false;
1321 
1322   Function *F = CI->getParent()->getParent();
1323   const GCNSubtarget &ST = TM->getSubtarget<GCNSubtarget>(*F);
1324   unsigned N = ST.getWavefrontSize();
1325 
1326   LLVM_DEBUG(errs() << "AMDIC: fold_wavefrontsize (" << *CI << ") with "
1327                << N << "\n");
1328 
1329   CI->replaceAllUsesWith(ConstantInt::get(B.getInt32Ty(), N));
1330   CI->eraseFromParent();
1331   return true;
1332 }
1333 
1334 // Get insertion point at entry.
1335 BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) {
1336   Function * Func = UI->getParent()->getParent();
1337   BasicBlock * BB = &Func->getEntryBlock();
1338   assert(BB && "Entry block not found!");
1339   BasicBlock::iterator ItNew = BB->begin();
1340   return ItNew;
1341 }
1342 
1343 // Insert a AllocsInst at the beginning of function entry block.
1344 AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B,
1345                                          const char *prefix) {
1346   BasicBlock::iterator ItNew = getEntryIns(UI);
1347   Function *UCallee = UI->getCalledFunction();
1348   Type *RetType = UCallee->getReturnType();
1349   B.SetInsertPoint(&*ItNew);
1350   AllocaInst *Alloc =
1351       B.CreateAlloca(RetType, nullptr, std::string(prefix) + UI->getName());
1352   Alloc->setAlignment(
1353       Align(UCallee->getParent()->getDataLayout().getTypeAllocSize(RetType)));
1354   return Alloc;
1355 }
1356 
1357 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo,
1358                                             double& Res0, double& Res1,
1359                                             Constant *copr0, Constant *copr1,
1360                                             Constant *copr2) {
1361   // By default, opr0/opr1/opr3 holds values of float/double type.
1362   // If they are not float/double, each function has to its
1363   // operand separately.
1364   double opr0=0.0, opr1=0.0, opr2=0.0;
1365   ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1366   ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1367   ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2);
1368   if (fpopr0) {
1369     opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1370              ? fpopr0->getValueAPF().convertToDouble()
1371              : (double)fpopr0->getValueAPF().convertToFloat();
1372   }
1373 
1374   if (fpopr1) {
1375     opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1376              ? fpopr1->getValueAPF().convertToDouble()
1377              : (double)fpopr1->getValueAPF().convertToFloat();
1378   }
1379 
1380   if (fpopr2) {
1381     opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1382              ? fpopr2->getValueAPF().convertToDouble()
1383              : (double)fpopr2->getValueAPF().convertToFloat();
1384   }
1385 
1386   switch (FInfo.getId()) {
1387   default : return false;
1388 
1389   case AMDGPULibFunc::EI_ACOS:
1390     Res0 = acos(opr0);
1391     return true;
1392 
1393   case AMDGPULibFunc::EI_ACOSH:
1394     // acosh(x) == log(x + sqrt(x*x - 1))
1395     Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1396     return true;
1397 
1398   case AMDGPULibFunc::EI_ACOSPI:
1399     Res0 = acos(opr0) / MATH_PI;
1400     return true;
1401 
1402   case AMDGPULibFunc::EI_ASIN:
1403     Res0 = asin(opr0);
1404     return true;
1405 
1406   case AMDGPULibFunc::EI_ASINH:
1407     // asinh(x) == log(x + sqrt(x*x + 1))
1408     Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1409     return true;
1410 
1411   case AMDGPULibFunc::EI_ASINPI:
1412     Res0 = asin(opr0) / MATH_PI;
1413     return true;
1414 
1415   case AMDGPULibFunc::EI_ATAN:
1416     Res0 = atan(opr0);
1417     return true;
1418 
1419   case AMDGPULibFunc::EI_ATANH:
1420     // atanh(x) == (log(x+1) - log(x-1))/2;
1421     Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1422     return true;
1423 
1424   case AMDGPULibFunc::EI_ATANPI:
1425     Res0 = atan(opr0) / MATH_PI;
1426     return true;
1427 
1428   case AMDGPULibFunc::EI_CBRT:
1429     Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1430     return true;
1431 
1432   case AMDGPULibFunc::EI_COS:
1433     Res0 = cos(opr0);
1434     return true;
1435 
1436   case AMDGPULibFunc::EI_COSH:
1437     Res0 = cosh(opr0);
1438     return true;
1439 
1440   case AMDGPULibFunc::EI_COSPI:
1441     Res0 = cos(MATH_PI * opr0);
1442     return true;
1443 
1444   case AMDGPULibFunc::EI_EXP:
1445     Res0 = exp(opr0);
1446     return true;
1447 
1448   case AMDGPULibFunc::EI_EXP2:
1449     Res0 = pow(2.0, opr0);
1450     return true;
1451 
1452   case AMDGPULibFunc::EI_EXP10:
1453     Res0 = pow(10.0, opr0);
1454     return true;
1455 
1456   case AMDGPULibFunc::EI_EXPM1:
1457     Res0 = exp(opr0) - 1.0;
1458     return true;
1459 
1460   case AMDGPULibFunc::EI_LOG:
1461     Res0 = log(opr0);
1462     return true;
1463 
1464   case AMDGPULibFunc::EI_LOG2:
1465     Res0 = log(opr0) / log(2.0);
1466     return true;
1467 
1468   case AMDGPULibFunc::EI_LOG10:
1469     Res0 = log(opr0) / log(10.0);
1470     return true;
1471 
1472   case AMDGPULibFunc::EI_RSQRT:
1473     Res0 = 1.0 / sqrt(opr0);
1474     return true;
1475 
1476   case AMDGPULibFunc::EI_SIN:
1477     Res0 = sin(opr0);
1478     return true;
1479 
1480   case AMDGPULibFunc::EI_SINH:
1481     Res0 = sinh(opr0);
1482     return true;
1483 
1484   case AMDGPULibFunc::EI_SINPI:
1485     Res0 = sin(MATH_PI * opr0);
1486     return true;
1487 
1488   case AMDGPULibFunc::EI_SQRT:
1489     Res0 = sqrt(opr0);
1490     return true;
1491 
1492   case AMDGPULibFunc::EI_TAN:
1493     Res0 = tan(opr0);
1494     return true;
1495 
1496   case AMDGPULibFunc::EI_TANH:
1497     Res0 = tanh(opr0);
1498     return true;
1499 
1500   case AMDGPULibFunc::EI_TANPI:
1501     Res0 = tan(MATH_PI * opr0);
1502     return true;
1503 
1504   case AMDGPULibFunc::EI_RECIP:
1505     Res0 = 1.0 / opr0;
1506     return true;
1507 
1508   // two-arg functions
1509   case AMDGPULibFunc::EI_DIVIDE:
1510     Res0 = opr0 / opr1;
1511     return true;
1512 
1513   case AMDGPULibFunc::EI_POW:
1514   case AMDGPULibFunc::EI_POWR:
1515     Res0 = pow(opr0, opr1);
1516     return true;
1517 
1518   case AMDGPULibFunc::EI_POWN: {
1519     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1520       double val = (double)iopr1->getSExtValue();
1521       Res0 = pow(opr0, val);
1522       return true;
1523     }
1524     return false;
1525   }
1526 
1527   case AMDGPULibFunc::EI_ROOTN: {
1528     if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1529       double val = (double)iopr1->getSExtValue();
1530       Res0 = pow(opr0, 1.0 / val);
1531       return true;
1532     }
1533     return false;
1534   }
1535 
1536   // with ptr arg
1537   case AMDGPULibFunc::EI_SINCOS:
1538     Res0 = sin(opr0);
1539     Res1 = cos(opr0);
1540     return true;
1541 
1542   // three-arg functions
1543   case AMDGPULibFunc::EI_FMA:
1544   case AMDGPULibFunc::EI_MAD:
1545     Res0 = opr0 * opr1 + opr2;
1546     return true;
1547   }
1548 
1549   return false;
1550 }
1551 
1552 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) {
1553   int numArgs = (int)aCI->arg_size();
1554   if (numArgs > 3)
1555     return false;
1556 
1557   Constant *copr0 = nullptr;
1558   Constant *copr1 = nullptr;
1559   Constant *copr2 = nullptr;
1560   if (numArgs > 0) {
1561     if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1562       return false;
1563   }
1564 
1565   if (numArgs > 1) {
1566     if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1567       if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1568         return false;
1569     }
1570   }
1571 
1572   if (numArgs > 2) {
1573     if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr)
1574       return false;
1575   }
1576 
1577   // At this point, all arguments to aCI are constants.
1578 
1579   // max vector size is 16, and sincos will generate two results.
1580   double DVal0[16], DVal1[16];
1581   int FuncVecSize = getVecSize(FInfo);
1582   bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1583   if (FuncVecSize == 1) {
1584     if (!evaluateScalarMathFunc(FInfo, DVal0[0],
1585                                 DVal1[0], copr0, copr1, copr2)) {
1586       return false;
1587     }
1588   } else {
1589     ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1590     ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1591     ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2);
1592     for (int i = 0; i < FuncVecSize; ++i) {
1593       Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1594       Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1595       Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr;
1596       if (!evaluateScalarMathFunc(FInfo, DVal0[i],
1597                                   DVal1[i], celt0, celt1, celt2)) {
1598         return false;
1599       }
1600     }
1601   }
1602 
1603   LLVMContext &context = CI->getParent()->getParent()->getContext();
1604   Constant *nval0, *nval1;
1605   if (FuncVecSize == 1) {
1606     nval0 = ConstantFP::get(CI->getType(), DVal0[0]);
1607     if (hasTwoResults)
1608       nval1 = ConstantFP::get(CI->getType(), DVal1[0]);
1609   } else {
1610     if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1611       SmallVector <float, 0> FVal0, FVal1;
1612       for (int i = 0; i < FuncVecSize; ++i)
1613         FVal0.push_back((float)DVal0[i]);
1614       ArrayRef<float> tmp0(FVal0);
1615       nval0 = ConstantDataVector::get(context, tmp0);
1616       if (hasTwoResults) {
1617         for (int i = 0; i < FuncVecSize; ++i)
1618           FVal1.push_back((float)DVal1[i]);
1619         ArrayRef<float> tmp1(FVal1);
1620         nval1 = ConstantDataVector::get(context, tmp1);
1621       }
1622     } else {
1623       ArrayRef<double> tmp0(DVal0);
1624       nval0 = ConstantDataVector::get(context, tmp0);
1625       if (hasTwoResults) {
1626         ArrayRef<double> tmp1(DVal1);
1627         nval1 = ConstantDataVector::get(context, tmp1);
1628       }
1629     }
1630   }
1631 
1632   if (hasTwoResults) {
1633     // sincos
1634     assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1635            "math function with ptr arg not supported yet");
1636     new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1637   }
1638 
1639   replaceCall(nval0);
1640   return true;
1641 }
1642 
1643 // Public interface to the Simplify LibCalls pass.
1644 FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetMachine *TM) {
1645   return new AMDGPUSimplifyLibCalls(TM);
1646 }
1647 
1648 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
1649   return new AMDGPUUseNativeCalls();
1650 }
1651 
1652 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
1653   if (skipFunction(F))
1654     return false;
1655 
1656   bool Changed = false;
1657   auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1658 
1659   LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1660              F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1661 
1662   for (auto &BB : F) {
1663     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1664       // Ignore non-calls.
1665       CallInst *CI = dyn_cast<CallInst>(I);
1666       ++I;
1667       // Ignore intrinsics that do not become real instructions.
1668       if (!CI || isa<DbgInfoIntrinsic>(CI) || CI->isLifetimeStartOrEnd())
1669         continue;
1670 
1671       // Ignore indirect calls.
1672       Function *Callee = CI->getCalledFunction();
1673       if (Callee == nullptr)
1674         continue;
1675 
1676       LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1677                  dbgs().flush());
1678       if(Simplifier.fold(CI, AA))
1679         Changed = true;
1680     }
1681   }
1682   return Changed;
1683 }
1684 
1685 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F,
1686                                                   FunctionAnalysisManager &AM) {
1687   AMDGPULibCalls Simplifier(&TM);
1688   Simplifier.initNativeFuncs();
1689 
1690   bool Changed = false;
1691   auto AA = &AM.getResult<AAManager>(F);
1692 
1693   LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1694              F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1695 
1696   for (auto &BB : F) {
1697     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1698       // Ignore non-calls.
1699       CallInst *CI = dyn_cast<CallInst>(I);
1700       ++I;
1701       // Ignore intrinsics that do not become real instructions.
1702       if (!CI || isa<DbgInfoIntrinsic>(CI) || CI->isLifetimeStartOrEnd())
1703         continue;
1704 
1705       // Ignore indirect calls.
1706       Function *Callee = CI->getCalledFunction();
1707       if (Callee == nullptr)
1708         continue;
1709 
1710       LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1711                  dbgs().flush());
1712       if (Simplifier.fold(CI, AA))
1713         Changed = true;
1714     }
1715   }
1716   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1717 }
1718 
1719 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) {
1720   if (skipFunction(F) || UseNative.empty())
1721     return false;
1722 
1723   bool Changed = false;
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 == nullptr)
1734         continue;
1735 
1736       if (Simplifier.useNative(CI))
1737         Changed = true;
1738     }
1739   }
1740   return Changed;
1741 }
1742 
1743 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F,
1744                                                 FunctionAnalysisManager &AM) {
1745   if (UseNative.empty())
1746     return PreservedAnalyses::all();
1747 
1748   AMDGPULibCalls Simplifier;
1749   Simplifier.initNativeFuncs();
1750 
1751   bool Changed = false;
1752   for (auto &BB : F) {
1753     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1754       // Ignore non-calls.
1755       CallInst *CI = dyn_cast<CallInst>(I);
1756       ++I;
1757       if (!CI)
1758         continue;
1759 
1760       // Ignore indirect calls.
1761       Function *Callee = CI->getCalledFunction();
1762       if (Callee == nullptr)
1763         continue;
1764 
1765       if (Simplifier.useNative(CI))
1766         Changed = true;
1767     }
1768   }
1769   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1770 }
1771