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
replaceCall(Value * With)117 void replaceCall(Value *With) {
118 CI->replaceAllUsesWith(With);
119 CI->eraseFromParent();
120 }
121
122 public:
AMDGPULibCalls(const TargetMachine * TM_=nullptr)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
AMDGPUSimplifyLibCalls(const TargetMachine * TM=nullptr)144 AMDGPUSimplifyLibCalls(const TargetMachine *TM = nullptr)
145 : FunctionPass(ID), Simplifier(TM) {
146 initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
147 }
148
getAnalysisUsage(AnalysisUsage & AU) const149 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
AMDGPUUseNativeCalls()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)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)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>
CreateCallEx2(IRB & B,FunctionCallee Callee,Value * Arg1,Value * Arg2,const Twine & Name="")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
HasNative(AMDGPULibFunc::EFuncId id)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
getOptTable(AMDGPULibFunc::EFuncId id)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
getVecSize(const AMDGPULibFunc & FInfo)426 static inline int getVecSize(const AMDGPULibFunc& FInfo) {
427 return FInfo.getLeads()[0].VectorSize;
428 }
429
getArgType(const AMDGPULibFunc & FInfo)430 static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
431 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
432 }
433
getFunction(Module * M,const FuncInfo & fInfo)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
parseFunctionName(const StringRef & FMangledName,FuncInfo & FInfo)442 bool AMDGPULibCalls::parseFunctionName(const StringRef &FMangledName,
443 FuncInfo &FInfo) {
444 return AMDGPULibFunc::parse(FMangledName, FInfo);
445 }
446
isUnsafeMath(const CallInst * CI) const447 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
useNativeFunc(const StringRef F) const456 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
457 return AllNative || llvm::is_contained(UseNative, F);
458 }
459
initNativeFuncs()460 void AMDGPULibCalls::initNativeFuncs() {
461 AllNative = useNativeFunc("all") ||
462 (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
463 UseNative.begin()->empty());
464 }
465
sincosUseNative(CallInst * aCI,const FuncInfo & FInfo)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
useNative(CallInst * aCI)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.
fold_read_write_pipe(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)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 auto &Ctx = M->getContext();
542 std::string Name = std::string(Callee->getName());
543 auto NumArg = CI->arg_size();
544 if (NumArg != 4 && NumArg != 6)
545 return false;
546 auto *PacketSize = CI->getArgOperand(NumArg - 2);
547 auto *PacketAlign = CI->getArgOperand(NumArg - 1);
548 if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign))
549 return false;
550 unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue();
551 Align Alignment = cast<ConstantInt>(PacketAlign)->getAlignValue();
552 if (Alignment != Size)
553 return false;
554
555 Type *PtrElemTy;
556 if (Size <= 8)
557 PtrElemTy = Type::getIntNTy(Ctx, Size * 8);
558 else
559 PtrElemTy = FixedVectorType::get(Type::getInt64Ty(Ctx), Size / 8);
560 unsigned PtrArgLoc = CI->arg_size() - 3;
561 auto PtrArg = CI->getArgOperand(PtrArgLoc);
562 unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace();
563 auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS);
564
565 SmallVector<llvm::Type *, 6> ArgTys;
566 for (unsigned I = 0; I != PtrArgLoc; ++I)
567 ArgTys.push_back(CI->getArgOperand(I)->getType());
568 ArgTys.push_back(PtrTy);
569
570 Name = Name + "_" + std::to_string(Size);
571 auto *FTy = FunctionType::get(Callee->getReturnType(),
572 ArrayRef<Type *>(ArgTys), false);
573 AMDGPULibFunc NewLibFunc(Name, FTy);
574 FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
575 if (!F)
576 return false;
577
578 auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
579 SmallVector<Value *, 6> Args;
580 for (unsigned I = 0; I != PtrArgLoc; ++I)
581 Args.push_back(CI->getArgOperand(I));
582 Args.push_back(BCast);
583
584 auto *NCI = B.CreateCall(F, Args);
585 NCI->setAttributes(CI->getAttributes());
586 CI->replaceAllUsesWith(NCI);
587 CI->dropAllReferences();
588 CI->eraseFromParent();
589
590 return true;
591 }
592
593 // This function returns false if no change; return true otherwise.
fold(CallInst * CI,AliasAnalysis * AA)594 bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) {
595 this->CI = CI;
596 Function *Callee = CI->getCalledFunction();
597
598 // Ignore indirect calls.
599 if (Callee == nullptr)
600 return false;
601
602 BasicBlock *BB = CI->getParent();
603 LLVMContext &Context = CI->getParent()->getContext();
604 IRBuilder<> B(Context);
605
606 // Set the builder to the instruction after the call.
607 B.SetInsertPoint(BB, CI->getIterator());
608
609 // Copy fast flags from the original call.
610 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(CI))
611 B.setFastMathFlags(FPOp->getFastMathFlags());
612
613 switch (Callee->getIntrinsicID()) {
614 default:
615 break;
616 case Intrinsic::amdgcn_wavefrontsize:
617 return !EnablePreLink && fold_wavefrontsize(CI, B);
618 }
619
620 FuncInfo FInfo;
621 if (!parseFunctionName(Callee->getName(), FInfo))
622 return false;
623
624 // Further check the number of arguments to see if they match.
625 if (CI->arg_size() != FInfo.getNumArgs())
626 return false;
627
628 if (TDOFold(CI, FInfo))
629 return true;
630
631 // Under unsafe-math, evaluate calls if possible.
632 // According to Brian Sumner, we can do this for all f32 function calls
633 // using host's double function calls.
634 if (isUnsafeMath(CI) && evaluateCall(CI, FInfo))
635 return true;
636
637 // Specialized optimizations for each function call
638 switch (FInfo.getId()) {
639 case AMDGPULibFunc::EI_RECIP:
640 // skip vector function
641 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
642 FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
643 "recip must be an either native or half function");
644 return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo);
645
646 case AMDGPULibFunc::EI_DIVIDE:
647 // skip vector function
648 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||
649 FInfo.getPrefix() == AMDGPULibFunc::HALF) &&
650 "divide must be an either native or half function");
651 return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo);
652
653 case AMDGPULibFunc::EI_POW:
654 case AMDGPULibFunc::EI_POWR:
655 case AMDGPULibFunc::EI_POWN:
656 return fold_pow(CI, B, FInfo);
657
658 case AMDGPULibFunc::EI_ROOTN:
659 // skip vector function
660 return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo);
661
662 case AMDGPULibFunc::EI_FMA:
663 case AMDGPULibFunc::EI_MAD:
664 case AMDGPULibFunc::EI_NFMA:
665 // skip vector function
666 return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo);
667
668 case AMDGPULibFunc::EI_SQRT:
669 return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo);
670 case AMDGPULibFunc::EI_COS:
671 case AMDGPULibFunc::EI_SIN:
672 if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
673 getArgType(FInfo) == AMDGPULibFunc::F64)
674 && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX))
675 return fold_sincos(CI, B, AA);
676
677 break;
678 case AMDGPULibFunc::EI_READ_PIPE_2:
679 case AMDGPULibFunc::EI_READ_PIPE_4:
680 case AMDGPULibFunc::EI_WRITE_PIPE_2:
681 case AMDGPULibFunc::EI_WRITE_PIPE_4:
682 return fold_read_write_pipe(CI, B, FInfo);
683
684 default:
685 break;
686 }
687
688 return false;
689 }
690
TDOFold(CallInst * CI,const FuncInfo & FInfo)691 bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
692 // Table-Driven optimization
693 const TableRef tr = getOptTable(FInfo.getId());
694 if (tr.empty())
695 return false;
696
697 int const sz = (int)tr.size();
698 Value *opr0 = CI->getArgOperand(0);
699
700 if (getVecSize(FInfo) > 1) {
701 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
702 SmallVector<double, 0> DVal;
703 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
704 ConstantFP *eltval = dyn_cast<ConstantFP>(
705 CV->getElementAsConstant((unsigned)eltNo));
706 assert(eltval && "Non-FP arguments in math function!");
707 bool found = false;
708 for (int i=0; i < sz; ++i) {
709 if (eltval->isExactlyValue(tr[i].input)) {
710 DVal.push_back(tr[i].result);
711 found = true;
712 break;
713 }
714 }
715 if (!found) {
716 // This vector constants not handled yet.
717 return false;
718 }
719 }
720 LLVMContext &context = CI->getParent()->getParent()->getContext();
721 Constant *nval;
722 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
723 SmallVector<float, 0> FVal;
724 for (unsigned i = 0; i < DVal.size(); ++i) {
725 FVal.push_back((float)DVal[i]);
726 }
727 ArrayRef<float> tmp(FVal);
728 nval = ConstantDataVector::get(context, tmp);
729 } else { // F64
730 ArrayRef<double> tmp(DVal);
731 nval = ConstantDataVector::get(context, tmp);
732 }
733 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
734 replaceCall(nval);
735 return true;
736 }
737 } else {
738 // Scalar version
739 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
740 for (int i = 0; i < sz; ++i) {
741 if (CF->isExactlyValue(tr[i].input)) {
742 Value *nval = ConstantFP::get(CF->getType(), tr[i].result);
743 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
744 replaceCall(nval);
745 return true;
746 }
747 }
748 }
749 }
750
751 return false;
752 }
753
754 // [native_]half_recip(c) ==> 1.0/c
fold_recip(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)755 bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B,
756 const FuncInfo &FInfo) {
757 Value *opr0 = CI->getArgOperand(0);
758 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
759 // Just create a normal div. Later, InstCombine will be able
760 // to compute the divide into a constant (avoid check float infinity
761 // or subnormal at this point).
762 Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0),
763 opr0,
764 "recip2div");
765 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n");
766 replaceCall(nval);
767 return true;
768 }
769 return false;
770 }
771
772 // [native_]half_divide(x, c) ==> x/c
fold_divide(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)773 bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B,
774 const FuncInfo &FInfo) {
775 Value *opr0 = CI->getArgOperand(0);
776 Value *opr1 = CI->getArgOperand(1);
777 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
778 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
779
780 if ((CF0 && CF1) || // both are constants
781 (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32)))
782 // CF1 is constant && f32 divide
783 {
784 Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0),
785 opr1, "__div2recip");
786 Value *nval = B.CreateFMul(opr0, nval1, "__div2mul");
787 replaceCall(nval);
788 return true;
789 }
790 return false;
791 }
792
793 namespace llvm {
log2(double V)794 static double log2(double V) {
795 #if _XOPEN_SOURCE >= 600 || defined(_ISOC99_SOURCE) || _POSIX_C_SOURCE >= 200112L
796 return ::log2(V);
797 #else
798 return log(V) / numbers::ln2;
799 #endif
800 }
801 }
802
fold_pow(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)803 bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B,
804 const FuncInfo &FInfo) {
805 assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||
806 FInfo.getId() == AMDGPULibFunc::EI_POWR ||
807 FInfo.getId() == AMDGPULibFunc::EI_POWN) &&
808 "fold_pow: encounter a wrong function call");
809
810 Value *opr0, *opr1;
811 ConstantFP *CF;
812 ConstantInt *CINT;
813 ConstantAggregateZero *CZero;
814 Type *eltType;
815
816 opr0 = CI->getArgOperand(0);
817 opr1 = CI->getArgOperand(1);
818 CZero = dyn_cast<ConstantAggregateZero>(opr1);
819 if (getVecSize(FInfo) == 1) {
820 eltType = opr0->getType();
821 CF = dyn_cast<ConstantFP>(opr1);
822 CINT = dyn_cast<ConstantInt>(opr1);
823 } else {
824 VectorType *VTy = dyn_cast<VectorType>(opr0->getType());
825 assert(VTy && "Oprand of vector function should be of vectortype");
826 eltType = VTy->getElementType();
827 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1);
828
829 // Now, only Handle vector const whose elements have the same value.
830 CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr;
831 CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr;
832 }
833
834 // No unsafe math , no constant argument, do nothing
835 if (!isUnsafeMath(CI) && !CF && !CINT && !CZero)
836 return false;
837
838 // 0x1111111 means that we don't do anything for this call.
839 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
840
841 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) {
842 // pow/powr/pown(x, 0) == 1
843 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n");
844 Constant *cnval = ConstantFP::get(eltType, 1.0);
845 if (getVecSize(FInfo) > 1) {
846 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
847 }
848 replaceCall(cnval);
849 return true;
850 }
851 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
852 // pow/powr/pown(x, 1.0) = x
853 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
854 replaceCall(opr0);
855 return true;
856 }
857 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
858 // pow/powr/pown(x, 2.0) = x*x
859 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0
860 << "\n");
861 Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
862 replaceCall(nval);
863 return true;
864 }
865 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
866 // pow/powr/pown(x, -1.0) = 1.0/x
867 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n");
868 Constant *cnval = ConstantFP::get(eltType, 1.0);
869 if (getVecSize(FInfo) > 1) {
870 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
871 }
872 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
873 replaceCall(nval);
874 return true;
875 }
876
877 Module *M = CI->getModule();
878 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
879 // pow[r](x, [-]0.5) = sqrt(x)
880 bool issqrt = CF->isExactlyValue(0.5);
881 if (FunctionCallee FPExpr =
882 getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
883 : AMDGPULibFunc::EI_RSQRT,
884 FInfo))) {
885 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
886 << FInfo.getName().c_str() << "(" << *opr0 << ")\n");
887 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
888 : "__pow2rsqrt");
889 replaceCall(nval);
890 return true;
891 }
892 }
893
894 if (!isUnsafeMath(CI))
895 return false;
896
897 // Unsafe Math optimization
898
899 // Remember that ci_opr1 is set if opr1 is integral
900 if (CF) {
901 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
902 ? (double)CF->getValueAPF().convertToFloat()
903 : CF->getValueAPF().convertToDouble();
904 int ival = (int)dval;
905 if ((double)ival == dval) {
906 ci_opr1 = ival;
907 } else
908 ci_opr1 = 0x11111111;
909 }
910
911 // pow/powr/pown(x, c) = [1/](x*x*..x); where
912 // trunc(c) == c && the number of x == c && |c| <= 12
913 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
914 if (abs_opr1 <= 12) {
915 Constant *cnval;
916 Value *nval;
917 if (abs_opr1 == 0) {
918 cnval = ConstantFP::get(eltType, 1.0);
919 if (getVecSize(FInfo) > 1) {
920 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
921 }
922 nval = cnval;
923 } else {
924 Value *valx2 = nullptr;
925 nval = nullptr;
926 while (abs_opr1 > 0) {
927 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
928 if (abs_opr1 & 1) {
929 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
930 }
931 abs_opr1 >>= 1;
932 }
933 }
934
935 if (ci_opr1 < 0) {
936 cnval = ConstantFP::get(eltType, 1.0);
937 if (getVecSize(FInfo) > 1) {
938 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
939 }
940 nval = B.CreateFDiv(cnval, nval, "__1powprod");
941 }
942 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
943 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0
944 << ")\n");
945 replaceCall(nval);
946 return true;
947 }
948
949 // powr ---> exp2(y * log2(x))
950 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
951 FunctionCallee ExpExpr =
952 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo));
953 if (!ExpExpr)
954 return false;
955
956 bool needlog = false;
957 bool needabs = false;
958 bool needcopysign = false;
959 Constant *cnval = nullptr;
960 if (getVecSize(FInfo) == 1) {
961 CF = dyn_cast<ConstantFP>(opr0);
962
963 if (CF) {
964 double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
965 ? (double)CF->getValueAPF().convertToFloat()
966 : CF->getValueAPF().convertToDouble();
967
968 V = log2(std::abs(V));
969 cnval = ConstantFP::get(eltType, V);
970 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
971 CF->isNegative();
972 } else {
973 needlog = true;
974 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR &&
975 (!CF || CF->isNegative());
976 }
977 } else {
978 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
979
980 if (!CDV) {
981 needlog = true;
982 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
983 } else {
984 assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&
985 "Wrong vector size detected");
986
987 SmallVector<double, 0> DVal;
988 for (int i=0; i < getVecSize(FInfo); ++i) {
989 double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
990 ? (double)CDV->getElementAsFloat(i)
991 : CDV->getElementAsDouble(i);
992 if (V < 0.0) needcopysign = true;
993 V = log2(std::abs(V));
994 DVal.push_back(V);
995 }
996 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
997 SmallVector<float, 0> FVal;
998 for (unsigned i=0; i < DVal.size(); ++i) {
999 FVal.push_back((float)DVal[i]);
1000 }
1001 ArrayRef<float> tmp(FVal);
1002 cnval = ConstantDataVector::get(M->getContext(), tmp);
1003 } else {
1004 ArrayRef<double> tmp(DVal);
1005 cnval = ConstantDataVector::get(M->getContext(), tmp);
1006 }
1007 }
1008 }
1009
1010 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1011 // We cannot handle corner cases for a general pow() function, give up
1012 // unless y is a constant integral value. Then proceed as if it were pown.
1013 if (getVecSize(FInfo) == 1) {
1014 if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) {
1015 double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1016 ? (double)CF->getValueAPF().convertToFloat()
1017 : CF->getValueAPF().convertToDouble();
1018 if (y != (double)(int64_t)y)
1019 return false;
1020 } else
1021 return false;
1022 } else {
1023 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) {
1024 for (int i=0; i < getVecSize(FInfo); ++i) {
1025 double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1026 ? (double)CDV->getElementAsFloat(i)
1027 : CDV->getElementAsDouble(i);
1028 if (y != (double)(int64_t)y)
1029 return false;
1030 }
1031 } else
1032 return false;
1033 }
1034 }
1035
1036 Value *nval;
1037 if (needabs) {
1038 FunctionCallee AbsExpr =
1039 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS, FInfo));
1040 if (!AbsExpr)
1041 return false;
1042 nval = CreateCallEx(B, AbsExpr, opr0, "__fabs");
1043 } else {
1044 nval = cnval ? cnval : opr0;
1045 }
1046 if (needlog) {
1047 FunctionCallee LogExpr =
1048 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo));
1049 if (!LogExpr)
1050 return false;
1051 nval = CreateCallEx(B,LogExpr, nval, "__log2");
1052 }
1053
1054 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1055 // convert int(32) to fp(f32 or f64)
1056 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1057 }
1058 nval = B.CreateFMul(opr1, nval, "__ylogx");
1059 nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1060
1061 if (needcopysign) {
1062 Value *opr_n;
1063 Type* rTy = opr0->getType();
1064 Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty();
1065 Type *nTy = nTyS;
1066 if (const auto *vTy = dyn_cast<FixedVectorType>(rTy))
1067 nTy = FixedVectorType::get(nTyS, vTy);
1068 unsigned size = nTy->getScalarSizeInBits();
1069 opr_n = CI->getArgOperand(1);
1070 if (opr_n->getType()->isIntegerTy())
1071 opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou");
1072 else
1073 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1074
1075 Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1076 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1077 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1078 nval = B.CreateBitCast(nval, opr0->getType());
1079 }
1080
1081 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1082 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n");
1083 replaceCall(nval);
1084
1085 return true;
1086 }
1087
fold_rootn(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)1088 bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B,
1089 const FuncInfo &FInfo) {
1090 Value *opr0 = CI->getArgOperand(0);
1091 Value *opr1 = CI->getArgOperand(1);
1092
1093 ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1094 if (!CINT) {
1095 return false;
1096 }
1097 int ci_opr1 = (int)CINT->getSExtValue();
1098 if (ci_opr1 == 1) { // rootn(x, 1) = x
1099 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n");
1100 replaceCall(opr0);
1101 return true;
1102 }
1103 if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x)
1104 Module *M = CI->getModule();
1105 if (FunctionCallee FPExpr =
1106 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1107 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n");
1108 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1109 replaceCall(nval);
1110 return true;
1111 }
1112 } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1113 Module *M = CI->getModule();
1114 if (FunctionCallee FPExpr =
1115 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1116 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n");
1117 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1118 replaceCall(nval);
1119 return true;
1120 }
1121 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1122 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n");
1123 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1124 opr0,
1125 "__rootn2div");
1126 replaceCall(nval);
1127 return true;
1128 } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x)
1129 Module *M = CI->getModule();
1130 if (FunctionCallee FPExpr =
1131 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) {
1132 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0
1133 << ")\n");
1134 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1135 replaceCall(nval);
1136 return true;
1137 }
1138 }
1139 return false;
1140 }
1141
fold_fma_mad(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)1142 bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B,
1143 const FuncInfo &FInfo) {
1144 Value *opr0 = CI->getArgOperand(0);
1145 Value *opr1 = CI->getArgOperand(1);
1146 Value *opr2 = CI->getArgOperand(2);
1147
1148 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
1149 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
1150 if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) {
1151 // fma/mad(a, b, c) = c if a=0 || b=0
1152 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n");
1153 replaceCall(opr2);
1154 return true;
1155 }
1156 if (CF0 && CF0->isExactlyValue(1.0f)) {
1157 // fma/mad(a, b, c) = b+c if a=1
1158 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2
1159 << "\n");
1160 Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd");
1161 replaceCall(nval);
1162 return true;
1163 }
1164 if (CF1 && CF1->isExactlyValue(1.0f)) {
1165 // fma/mad(a, b, c) = a+c if b=1
1166 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2
1167 << "\n");
1168 Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd");
1169 replaceCall(nval);
1170 return true;
1171 }
1172 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) {
1173 if (CF->isZero()) {
1174 // fma/mad(a, b, c) = a*b if c=0
1175 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * "
1176 << *opr1 << "\n");
1177 Value *nval = B.CreateFMul(opr0, opr1, "fmamul");
1178 replaceCall(nval);
1179 return true;
1180 }
1181 }
1182
1183 return false;
1184 }
1185
1186 // Get a scalar native builtin single argument FP function
getNativeFunction(Module * M,const FuncInfo & FInfo)1187 FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1188 const FuncInfo &FInfo) {
1189 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1190 return nullptr;
1191 FuncInfo nf = FInfo;
1192 nf.setPrefix(AMDGPULibFunc::NATIVE);
1193 return getFunction(M, nf);
1194 }
1195
1196 // fold sqrt -> native_sqrt (x)
fold_sqrt(CallInst * CI,IRBuilder<> & B,const FuncInfo & FInfo)1197 bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
1198 const FuncInfo &FInfo) {
1199 if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1200 (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1201 if (FunctionCallee FPExpr = getNativeFunction(
1202 CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1203 Value *opr0 = CI->getArgOperand(0);
1204 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "
1205 << "sqrt(" << *opr0 << ")\n");
1206 Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1207 replaceCall(nval);
1208 return true;
1209 }
1210 }
1211 return false;
1212 }
1213
1214 // fold sin, cos -> sincos.
fold_sincos(CallInst * CI,IRBuilder<> & B,AliasAnalysis * AA)1215 bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
1216 AliasAnalysis *AA) {
1217 AMDGPULibFunc fInfo;
1218 if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo))
1219 return false;
1220
1221 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
1222 fInfo.getId() == AMDGPULibFunc::EI_COS);
1223 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1224
1225 Value *CArgVal = CI->getArgOperand(0);
1226 BasicBlock * const CBB = CI->getParent();
1227
1228 int const MaxScan = 30;
1229 bool Changed = false;
1230
1231 { // fold in load value.
1232 LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
1233 if (LI && LI->getParent() == CBB) {
1234 BasicBlock::iterator BBI = LI->getIterator();
1235 Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
1236 if (AvailableVal) {
1237 Changed = true;
1238 CArgVal->replaceAllUsesWith(AvailableVal);
1239 if (CArgVal->getNumUses() == 0)
1240 LI->eraseFromParent();
1241 CArgVal = CI->getArgOperand(0);
1242 }
1243 }
1244 }
1245
1246 Module *M = CI->getModule();
1247 fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN);
1248 std::string const PairName = fInfo.mangle();
1249
1250 CallInst *UI = nullptr;
1251 for (User* U : CArgVal->users()) {
1252 CallInst *XI = dyn_cast_or_null<CallInst>(U);
1253 if (!XI || XI == CI || XI->getParent() != CBB)
1254 continue;
1255
1256 Function *UCallee = XI->getCalledFunction();
1257 if (!UCallee || !UCallee->getName().equals(PairName))
1258 continue;
1259
1260 BasicBlock::iterator BBI = CI->getIterator();
1261 if (BBI == CI->getParent()->begin())
1262 break;
1263 --BBI;
1264 for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
1265 if (cast<Instruction>(BBI) == XI) {
1266 UI = XI;
1267 break;
1268 }
1269 }
1270 if (UI) break;
1271 }
1272
1273 if (!UI)
1274 return Changed;
1275
1276 // Merge the sin and cos.
1277
1278 // for OpenCL 2.0 we have only generic implementation of sincos
1279 // function.
1280 AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
1281 nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1282 FunctionCallee Fsincos = getFunction(M, nf);
1283 if (!Fsincos)
1284 return Changed;
1285
1286 BasicBlock::iterator ItOld = B.GetInsertPoint();
1287 AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
1288 B.SetInsertPoint(UI);
1289
1290 Value *P = Alloc;
1291 Type *PTy = Fsincos.getFunctionType()->getParamType(1);
1292 // The allocaInst allocates the memory in private address space. This need
1293 // to be bitcasted to point to the address space of cos pointer type.
1294 // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1295 if (PTy->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
1296 P = B.CreateAddrSpaceCast(Alloc, PTy);
1297 CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P);
1298
1299 LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with "
1300 << *Call << "\n");
1301
1302 if (!isSin) { // CI->cos, UI->sin
1303 B.SetInsertPoint(&*ItOld);
1304 UI->replaceAllUsesWith(&*Call);
1305 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1306 CI->replaceAllUsesWith(Reload);
1307 UI->eraseFromParent();
1308 CI->eraseFromParent();
1309 } else { // CI->sin, UI->cos
1310 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1311 UI->replaceAllUsesWith(Reload);
1312 CI->replaceAllUsesWith(Call);
1313 UI->eraseFromParent();
1314 CI->eraseFromParent();
1315 }
1316 return true;
1317 }
1318
fold_wavefrontsize(CallInst * CI,IRBuilder<> & B)1319 bool AMDGPULibCalls::fold_wavefrontsize(CallInst *CI, IRBuilder<> &B) {
1320 if (!TM)
1321 return false;
1322
1323 StringRef CPU = TM->getTargetCPU();
1324 StringRef Features = TM->getTargetFeatureString();
1325 if ((CPU.empty() || CPU.equals_insensitive("generic")) &&
1326 (Features.empty() || !Features.contains_insensitive("wavefrontsize")))
1327 return false;
1328
1329 Function *F = CI->getParent()->getParent();
1330 const GCNSubtarget &ST = TM->getSubtarget<GCNSubtarget>(*F);
1331 unsigned N = ST.getWavefrontSize();
1332
1333 LLVM_DEBUG(errs() << "AMDIC: fold_wavefrontsize (" << *CI << ") with "
1334 << N << "\n");
1335
1336 CI->replaceAllUsesWith(ConstantInt::get(B.getInt32Ty(), N));
1337 CI->eraseFromParent();
1338 return true;
1339 }
1340
1341 // Get insertion point at entry.
getEntryIns(CallInst * UI)1342 BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) {
1343 Function * Func = UI->getParent()->getParent();
1344 BasicBlock * BB = &Func->getEntryBlock();
1345 assert(BB && "Entry block not found!");
1346 BasicBlock::iterator ItNew = BB->begin();
1347 return ItNew;
1348 }
1349
1350 // Insert a AllocsInst at the beginning of function entry block.
insertAlloca(CallInst * UI,IRBuilder<> & B,const char * prefix)1351 AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B,
1352 const char *prefix) {
1353 BasicBlock::iterator ItNew = getEntryIns(UI);
1354 Function *UCallee = UI->getCalledFunction();
1355 Type *RetType = UCallee->getReturnType();
1356 B.SetInsertPoint(&*ItNew);
1357 AllocaInst *Alloc =
1358 B.CreateAlloca(RetType, nullptr, std::string(prefix) + UI->getName());
1359 Alloc->setAlignment(
1360 Align(UCallee->getParent()->getDataLayout().getTypeAllocSize(RetType)));
1361 return Alloc;
1362 }
1363
evaluateScalarMathFunc(const FuncInfo & FInfo,double & Res0,double & Res1,Constant * copr0,Constant * copr1,Constant * copr2)1364 bool AMDGPULibCalls::evaluateScalarMathFunc(const FuncInfo &FInfo,
1365 double& Res0, double& Res1,
1366 Constant *copr0, Constant *copr1,
1367 Constant *copr2) {
1368 // By default, opr0/opr1/opr3 holds values of float/double type.
1369 // If they are not float/double, each function has to its
1370 // operand separately.
1371 double opr0=0.0, opr1=0.0, opr2=0.0;
1372 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1373 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1374 ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2);
1375 if (fpopr0) {
1376 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1377 ? fpopr0->getValueAPF().convertToDouble()
1378 : (double)fpopr0->getValueAPF().convertToFloat();
1379 }
1380
1381 if (fpopr1) {
1382 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1383 ? fpopr1->getValueAPF().convertToDouble()
1384 : (double)fpopr1->getValueAPF().convertToFloat();
1385 }
1386
1387 if (fpopr2) {
1388 opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1389 ? fpopr2->getValueAPF().convertToDouble()
1390 : (double)fpopr2->getValueAPF().convertToFloat();
1391 }
1392
1393 switch (FInfo.getId()) {
1394 default : return false;
1395
1396 case AMDGPULibFunc::EI_ACOS:
1397 Res0 = acos(opr0);
1398 return true;
1399
1400 case AMDGPULibFunc::EI_ACOSH:
1401 // acosh(x) == log(x + sqrt(x*x - 1))
1402 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1403 return true;
1404
1405 case AMDGPULibFunc::EI_ACOSPI:
1406 Res0 = acos(opr0) / MATH_PI;
1407 return true;
1408
1409 case AMDGPULibFunc::EI_ASIN:
1410 Res0 = asin(opr0);
1411 return true;
1412
1413 case AMDGPULibFunc::EI_ASINH:
1414 // asinh(x) == log(x + sqrt(x*x + 1))
1415 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1416 return true;
1417
1418 case AMDGPULibFunc::EI_ASINPI:
1419 Res0 = asin(opr0) / MATH_PI;
1420 return true;
1421
1422 case AMDGPULibFunc::EI_ATAN:
1423 Res0 = atan(opr0);
1424 return true;
1425
1426 case AMDGPULibFunc::EI_ATANH:
1427 // atanh(x) == (log(x+1) - log(x-1))/2;
1428 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1429 return true;
1430
1431 case AMDGPULibFunc::EI_ATANPI:
1432 Res0 = atan(opr0) / MATH_PI;
1433 return true;
1434
1435 case AMDGPULibFunc::EI_CBRT:
1436 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1437 return true;
1438
1439 case AMDGPULibFunc::EI_COS:
1440 Res0 = cos(opr0);
1441 return true;
1442
1443 case AMDGPULibFunc::EI_COSH:
1444 Res0 = cosh(opr0);
1445 return true;
1446
1447 case AMDGPULibFunc::EI_COSPI:
1448 Res0 = cos(MATH_PI * opr0);
1449 return true;
1450
1451 case AMDGPULibFunc::EI_EXP:
1452 Res0 = exp(opr0);
1453 return true;
1454
1455 case AMDGPULibFunc::EI_EXP2:
1456 Res0 = pow(2.0, opr0);
1457 return true;
1458
1459 case AMDGPULibFunc::EI_EXP10:
1460 Res0 = pow(10.0, opr0);
1461 return true;
1462
1463 case AMDGPULibFunc::EI_EXPM1:
1464 Res0 = exp(opr0) - 1.0;
1465 return true;
1466
1467 case AMDGPULibFunc::EI_LOG:
1468 Res0 = log(opr0);
1469 return true;
1470
1471 case AMDGPULibFunc::EI_LOG2:
1472 Res0 = log(opr0) / log(2.0);
1473 return true;
1474
1475 case AMDGPULibFunc::EI_LOG10:
1476 Res0 = log(opr0) / log(10.0);
1477 return true;
1478
1479 case AMDGPULibFunc::EI_RSQRT:
1480 Res0 = 1.0 / sqrt(opr0);
1481 return true;
1482
1483 case AMDGPULibFunc::EI_SIN:
1484 Res0 = sin(opr0);
1485 return true;
1486
1487 case AMDGPULibFunc::EI_SINH:
1488 Res0 = sinh(opr0);
1489 return true;
1490
1491 case AMDGPULibFunc::EI_SINPI:
1492 Res0 = sin(MATH_PI * opr0);
1493 return true;
1494
1495 case AMDGPULibFunc::EI_SQRT:
1496 Res0 = sqrt(opr0);
1497 return true;
1498
1499 case AMDGPULibFunc::EI_TAN:
1500 Res0 = tan(opr0);
1501 return true;
1502
1503 case AMDGPULibFunc::EI_TANH:
1504 Res0 = tanh(opr0);
1505 return true;
1506
1507 case AMDGPULibFunc::EI_TANPI:
1508 Res0 = tan(MATH_PI * opr0);
1509 return true;
1510
1511 case AMDGPULibFunc::EI_RECIP:
1512 Res0 = 1.0 / opr0;
1513 return true;
1514
1515 // two-arg functions
1516 case AMDGPULibFunc::EI_DIVIDE:
1517 Res0 = opr0 / opr1;
1518 return true;
1519
1520 case AMDGPULibFunc::EI_POW:
1521 case AMDGPULibFunc::EI_POWR:
1522 Res0 = pow(opr0, opr1);
1523 return true;
1524
1525 case AMDGPULibFunc::EI_POWN: {
1526 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1527 double val = (double)iopr1->getSExtValue();
1528 Res0 = pow(opr0, val);
1529 return true;
1530 }
1531 return false;
1532 }
1533
1534 case AMDGPULibFunc::EI_ROOTN: {
1535 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1536 double val = (double)iopr1->getSExtValue();
1537 Res0 = pow(opr0, 1.0 / val);
1538 return true;
1539 }
1540 return false;
1541 }
1542
1543 // with ptr arg
1544 case AMDGPULibFunc::EI_SINCOS:
1545 Res0 = sin(opr0);
1546 Res1 = cos(opr0);
1547 return true;
1548
1549 // three-arg functions
1550 case AMDGPULibFunc::EI_FMA:
1551 case AMDGPULibFunc::EI_MAD:
1552 Res0 = opr0 * opr1 + opr2;
1553 return true;
1554 }
1555
1556 return false;
1557 }
1558
evaluateCall(CallInst * aCI,const FuncInfo & FInfo)1559 bool AMDGPULibCalls::evaluateCall(CallInst *aCI, const FuncInfo &FInfo) {
1560 int numArgs = (int)aCI->arg_size();
1561 if (numArgs > 3)
1562 return false;
1563
1564 Constant *copr0 = nullptr;
1565 Constant *copr1 = nullptr;
1566 Constant *copr2 = nullptr;
1567 if (numArgs > 0) {
1568 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1569 return false;
1570 }
1571
1572 if (numArgs > 1) {
1573 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1574 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1575 return false;
1576 }
1577 }
1578
1579 if (numArgs > 2) {
1580 if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr)
1581 return false;
1582 }
1583
1584 // At this point, all arguments to aCI are constants.
1585
1586 // max vector size is 16, and sincos will generate two results.
1587 double DVal0[16], DVal1[16];
1588 int FuncVecSize = getVecSize(FInfo);
1589 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
1590 if (FuncVecSize == 1) {
1591 if (!evaluateScalarMathFunc(FInfo, DVal0[0],
1592 DVal1[0], copr0, copr1, copr2)) {
1593 return false;
1594 }
1595 } else {
1596 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
1597 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
1598 ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2);
1599 for (int i = 0; i < FuncVecSize; ++i) {
1600 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1601 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1602 Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr;
1603 if (!evaluateScalarMathFunc(FInfo, DVal0[i],
1604 DVal1[i], celt0, celt1, celt2)) {
1605 return false;
1606 }
1607 }
1608 }
1609
1610 LLVMContext &context = CI->getParent()->getParent()->getContext();
1611 Constant *nval0, *nval1;
1612 if (FuncVecSize == 1) {
1613 nval0 = ConstantFP::get(CI->getType(), DVal0[0]);
1614 if (hasTwoResults)
1615 nval1 = ConstantFP::get(CI->getType(), DVal1[0]);
1616 } else {
1617 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1618 SmallVector <float, 0> FVal0, FVal1;
1619 for (int i = 0; i < FuncVecSize; ++i)
1620 FVal0.push_back((float)DVal0[i]);
1621 ArrayRef<float> tmp0(FVal0);
1622 nval0 = ConstantDataVector::get(context, tmp0);
1623 if (hasTwoResults) {
1624 for (int i = 0; i < FuncVecSize; ++i)
1625 FVal1.push_back((float)DVal1[i]);
1626 ArrayRef<float> tmp1(FVal1);
1627 nval1 = ConstantDataVector::get(context, tmp1);
1628 }
1629 } else {
1630 ArrayRef<double> tmp0(DVal0);
1631 nval0 = ConstantDataVector::get(context, tmp0);
1632 if (hasTwoResults) {
1633 ArrayRef<double> tmp1(DVal1);
1634 nval1 = ConstantDataVector::get(context, tmp1);
1635 }
1636 }
1637 }
1638
1639 if (hasTwoResults) {
1640 // sincos
1641 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&
1642 "math function with ptr arg not supported yet");
1643 new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1644 }
1645
1646 replaceCall(nval0);
1647 return true;
1648 }
1649
1650 // Public interface to the Simplify LibCalls pass.
createAMDGPUSimplifyLibCallsPass(const TargetMachine * TM)1651 FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetMachine *TM) {
1652 return new AMDGPUSimplifyLibCalls(TM);
1653 }
1654
createAMDGPUUseNativeCallsPass()1655 FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
1656 return new AMDGPUUseNativeCalls();
1657 }
1658
runOnFunction(Function & F)1659 bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
1660 if (skipFunction(F))
1661 return false;
1662
1663 bool Changed = false;
1664 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1665
1666 LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1667 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1668
1669 for (auto &BB : F) {
1670 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1671 // Ignore non-calls.
1672 CallInst *CI = dyn_cast<CallInst>(I);
1673 ++I;
1674 // Ignore intrinsics that do not become real instructions.
1675 if (!CI || isa<DbgInfoIntrinsic>(CI) || CI->isLifetimeStartOrEnd())
1676 continue;
1677
1678 // Ignore indirect calls.
1679 Function *Callee = CI->getCalledFunction();
1680 if (Callee == nullptr)
1681 continue;
1682
1683 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1684 dbgs().flush());
1685 if(Simplifier.fold(CI, AA))
1686 Changed = true;
1687 }
1688 }
1689 return Changed;
1690 }
1691
run(Function & F,FunctionAnalysisManager & AM)1692 PreservedAnalyses AMDGPUSimplifyLibCallsPass::run(Function &F,
1693 FunctionAnalysisManager &AM) {
1694 AMDGPULibCalls Simplifier(&TM);
1695 Simplifier.initNativeFuncs();
1696
1697 bool Changed = false;
1698 auto AA = &AM.getResult<AAManager>(F);
1699
1700 LLVM_DEBUG(dbgs() << "AMDIC: process function ";
1701 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';);
1702
1703 for (auto &BB : F) {
1704 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1705 // Ignore non-calls.
1706 CallInst *CI = dyn_cast<CallInst>(I);
1707 ++I;
1708 // Ignore intrinsics that do not become real instructions.
1709 if (!CI || isa<DbgInfoIntrinsic>(CI) || CI->isLifetimeStartOrEnd())
1710 continue;
1711
1712 // Ignore indirect calls.
1713 Function *Callee = CI->getCalledFunction();
1714 if (Callee == nullptr)
1715 continue;
1716
1717 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";
1718 dbgs().flush());
1719 if (Simplifier.fold(CI, AA))
1720 Changed = true;
1721 }
1722 }
1723 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1724 }
1725
runOnFunction(Function & F)1726 bool AMDGPUUseNativeCalls::runOnFunction(Function &F) {
1727 if (skipFunction(F) || UseNative.empty())
1728 return false;
1729
1730 bool Changed = false;
1731 for (auto &BB : F) {
1732 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1733 // Ignore non-calls.
1734 CallInst *CI = dyn_cast<CallInst>(I);
1735 ++I;
1736 if (!CI) continue;
1737
1738 // Ignore indirect calls.
1739 Function *Callee = CI->getCalledFunction();
1740 if (Callee == nullptr)
1741 continue;
1742
1743 if (Simplifier.useNative(CI))
1744 Changed = true;
1745 }
1746 }
1747 return Changed;
1748 }
1749
run(Function & F,FunctionAnalysisManager & AM)1750 PreservedAnalyses AMDGPUUseNativeCallsPass::run(Function &F,
1751 FunctionAnalysisManager &AM) {
1752 if (UseNative.empty())
1753 return PreservedAnalyses::all();
1754
1755 AMDGPULibCalls Simplifier;
1756 Simplifier.initNativeFuncs();
1757
1758 bool Changed = false;
1759 for (auto &BB : F) {
1760 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E;) {
1761 // Ignore non-calls.
1762 CallInst *CI = dyn_cast<CallInst>(I);
1763 ++I;
1764 if (!CI)
1765 continue;
1766
1767 // Ignore indirect calls.
1768 Function *Callee = CI->getCalledFunction();
1769 if (Callee == nullptr)
1770 continue;
1771
1772 if (Simplifier.useNative(CI))
1773 Changed = true;
1774 }
1775 }
1776 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1777 }
1778