Bug Summary

File:lib/Target/AMDGPU/AMDGPULibCalls.cpp
Warning:line 1679, column 13
2nd function call argument is an uninitialized value

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name AMDGPULibCalls.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -mframe-pointer=none -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~svn374877/build-llvm/lib/Target/AMDGPU -I /build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU -I /build/llvm-toolchain-snapshot-10~svn374877/build-llvm/include -I /build/llvm-toolchain-snapshot-10~svn374877/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~svn374877/build-llvm/lib/Target/AMDGPU -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~svn374877=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2019-10-15-233810-7101-1 -x c++ /build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp
1//===- AMDGPULibCalls.cpp -------------------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9/// \file
10/// This file does AMD library function optimizations.
11//
12//===----------------------------------------------------------------------===//
13
14#define DEBUG_TYPE"amdgpu-simplifylib" "amdgpu-simplifylib"
15
16#include "AMDGPU.h"
17#include "AMDGPULibFunc.h"
18#include "AMDGPUSubtarget.h"
19#include "llvm/Analysis/AliasAnalysis.h"
20#include "llvm/Analysis/Loads.h"
21#include "llvm/ADT/StringSet.h"
22#include "llvm/ADT/StringRef.h"
23#include "llvm/IR/Constants.h"
24#include "llvm/IR/DerivedTypes.h"
25#include "llvm/IR/Instructions.h"
26#include "llvm/IR/Intrinsics.h"
27#include "llvm/IR/IRBuilder.h"
28#include "llvm/IR/Function.h"
29#include "llvm/IR/LLVMContext.h"
30#include "llvm/IR/Module.h"
31#include "llvm/IR/ValueSymbolTable.h"
32#include "llvm/Support/Debug.h"
33#include "llvm/Support/MathExtras.h"
34#include "llvm/Support/raw_ostream.h"
35#include "llvm/Target/TargetMachine.h"
36#include "llvm/Target/TargetOptions.h"
37#include <vector>
38#include <cmath>
39
40using namespace llvm;
41
42static cl::opt<bool> EnablePreLink("amdgpu-prelink",
43 cl::desc("Enable pre-link mode optimizations"),
44 cl::init(false),
45 cl::Hidden);
46
47static cl::list<std::string> UseNative("amdgpu-use-native",
48 cl::desc("Comma separated list of functions to replace with native, or all"),
49 cl::CommaSeparated, cl::ValueOptional,
50 cl::Hidden);
51
52#define MATH_PInumbers::pi numbers::pi
53#define MATH_Enumbers::e numbers::e
54#define MATH_SQRT2numbers::sqrt2 numbers::sqrt2
55#define MATH_SQRT1_2numbers::inv_sqrt2 numbers::inv_sqrt2
56
57namespace llvm {
58
59class AMDGPULibCalls {
60private:
61
62 typedef llvm::AMDGPULibFunc FuncInfo;
63
64 const TargetMachine *TM;
65
66 // -fuse-native.
67 bool AllNative = false;
68
69 bool useNativeFunc(const StringRef F) const;
70
71 // Return a pointer (pointer expr) to the function if function defintion with
72 // "FuncName" exists. It may create a new function prototype in pre-link mode.
73 FunctionCallee getFunction(Module *M, const FuncInfo &fInfo);
74
75 // Replace a normal function with its native version.
76 bool replaceWithNative(CallInst *CI, const FuncInfo &FInfo);
77
78 bool parseFunctionName(const StringRef& FMangledName,
79 FuncInfo *FInfo=nullptr /*out*/);
80
81 bool TDOFold(CallInst *CI, const FuncInfo &FInfo);
82
83 /* Specialized optimizations */
84
85 // recip (half or native)
86 bool fold_recip(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
87
88 // divide (half or native)
89 bool fold_divide(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
90
91 // pow/powr/pown
92 bool fold_pow(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
93
94 // rootn
95 bool fold_rootn(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
96
97 // fma/mad
98 bool fold_fma_mad(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
99
100 // -fuse-native for sincos
101 bool sincosUseNative(CallInst *aCI, const FuncInfo &FInfo);
102
103 // evaluate calls if calls' arguments are constants.
104 bool evaluateScalarMathFunc(FuncInfo &FInfo, double& Res0,
105 double& Res1, Constant *copr0, Constant *copr1, Constant *copr2);
106 bool evaluateCall(CallInst *aCI, FuncInfo &FInfo);
107
108 // exp
109 bool fold_exp(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
110
111 // exp2
112 bool fold_exp2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
113
114 // exp10
115 bool fold_exp10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
116
117 // log
118 bool fold_log(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
119
120 // log2
121 bool fold_log2(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
122
123 // log10
124 bool fold_log10(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
125
126 // sqrt
127 bool fold_sqrt(CallInst *CI, IRBuilder<> &B, const FuncInfo &FInfo);
128
129 // sin/cos
130 bool fold_sincos(CallInst * CI, IRBuilder<> &B, AliasAnalysis * AA);
131
132 // __read_pipe/__write_pipe
133 bool fold_read_write_pipe(CallInst *CI, IRBuilder<> &B, FuncInfo &FInfo);
134
135 // llvm.amdgcn.wavefrontsize
136 bool fold_wavefrontsize(CallInst *CI, IRBuilder<> &B);
137
138 // Get insertion point at entry.
139 BasicBlock::iterator getEntryIns(CallInst * UI);
140 // Insert an Alloc instruction.
141 AllocaInst* insertAlloca(CallInst * UI, IRBuilder<> &B, const char *prefix);
142 // Get a scalar native builtin signle argument FP function
143 FunctionCallee getNativeFunction(Module *M, const FuncInfo &FInfo);
144
145protected:
146 CallInst *CI;
147
148 bool isUnsafeMath(const CallInst *CI) const;
149
150 void replaceCall(Value *With) {
151 CI->replaceAllUsesWith(With);
152 CI->eraseFromParent();
153 }
154
155public:
156 AMDGPULibCalls(const TargetMachine *TM_ = nullptr) : TM(TM_) {}
157
158 bool fold(CallInst *CI, AliasAnalysis *AA = nullptr);
159
160 void initNativeFuncs();
161
162 // Replace a normal math function call with that native version
163 bool useNative(CallInst *CI);
164};
165
166} // end llvm namespace
167
168namespace {
169
170 class AMDGPUSimplifyLibCalls : public FunctionPass {
171
172 const TargetOptions Options;
173
174 AMDGPULibCalls Simplifier;
175
176 public:
177 static char ID; // Pass identification
178
179 AMDGPUSimplifyLibCalls(const TargetOptions &Opt = TargetOptions(),
180 const TargetMachine *TM = nullptr)
181 : FunctionPass(ID), Options(Opt), Simplifier(TM) {
182 initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
183 }
184
185 void getAnalysisUsage(AnalysisUsage &AU) const override {
186 AU.addRequired<AAResultsWrapperPass>();
187 }
188
189 bool runOnFunction(Function &M) override;
190 };
191
192 class AMDGPUUseNativeCalls : public FunctionPass {
193
194 AMDGPULibCalls Simplifier;
195
196 public:
197 static char ID; // Pass identification
198
199 AMDGPUUseNativeCalls() : FunctionPass(ID) {
200 initializeAMDGPUUseNativeCallsPass(*PassRegistry::getPassRegistry());
201 Simplifier.initNativeFuncs();
202 }
203
204 bool runOnFunction(Function &F) override;
205 };
206
207} // end anonymous namespace.
208
209char AMDGPUSimplifyLibCalls::ID = 0;
210char AMDGPUUseNativeCalls::ID = 0;
211
212INITIALIZE_PASS_BEGIN(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",static void *initializeAMDGPUSimplifyLibCallsPassOnce(PassRegistry
&Registry) {
213 "Simplify well-known AMD library calls", false, false)static void *initializeAMDGPUSimplifyLibCallsPassOnce(PassRegistry
&Registry) {
214INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
215INITIALIZE_PASS_END(AMDGPUSimplifyLibCalls, "amdgpu-simplifylib",PassInfo *PI = new PassInfo( "Simplify well-known AMD library calls"
, "amdgpu-simplifylib", &AMDGPUSimplifyLibCalls::ID, PassInfo
::NormalCtor_t(callDefaultCtor<AMDGPUSimplifyLibCalls>)
, false, false); Registry.registerPass(*PI, true); return PI;
} static llvm::once_flag InitializeAMDGPUSimplifyLibCallsPassFlag
; void llvm::initializeAMDGPUSimplifyLibCallsPass(PassRegistry
&Registry) { llvm::call_once(InitializeAMDGPUSimplifyLibCallsPassFlag
, initializeAMDGPUSimplifyLibCallsPassOnce, std::ref(Registry
)); }
216 "Simplify well-known AMD library calls", false, false)PassInfo *PI = new PassInfo( "Simplify well-known AMD library calls"
, "amdgpu-simplifylib", &AMDGPUSimplifyLibCalls::ID, PassInfo
::NormalCtor_t(callDefaultCtor<AMDGPUSimplifyLibCalls>)
, false, false); Registry.registerPass(*PI, true); return PI;
} static llvm::once_flag InitializeAMDGPUSimplifyLibCallsPassFlag
; void llvm::initializeAMDGPUSimplifyLibCallsPass(PassRegistry
&Registry) { llvm::call_once(InitializeAMDGPUSimplifyLibCallsPassFlag
, initializeAMDGPUSimplifyLibCallsPassOnce, std::ref(Registry
)); }
217
218INITIALIZE_PASS(AMDGPUUseNativeCalls, "amdgpu-usenative",static void *initializeAMDGPUUseNativeCallsPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "Replace builtin math calls with that native versions."
, "amdgpu-usenative", &AMDGPUUseNativeCalls::ID, PassInfo
::NormalCtor_t(callDefaultCtor<AMDGPUUseNativeCalls>), false
, false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeAMDGPUUseNativeCallsPassFlag; void
llvm::initializeAMDGPUUseNativeCallsPass(PassRegistry &Registry
) { llvm::call_once(InitializeAMDGPUUseNativeCallsPassFlag, initializeAMDGPUUseNativeCallsPassOnce
, std::ref(Registry)); }
219 "Replace builtin math calls with that native versions.",static void *initializeAMDGPUUseNativeCallsPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "Replace builtin math calls with that native versions."
, "amdgpu-usenative", &AMDGPUUseNativeCalls::ID, PassInfo
::NormalCtor_t(callDefaultCtor<AMDGPUUseNativeCalls>), false
, false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeAMDGPUUseNativeCallsPassFlag; void
llvm::initializeAMDGPUUseNativeCallsPass(PassRegistry &Registry
) { llvm::call_once(InitializeAMDGPUUseNativeCallsPassFlag, initializeAMDGPUUseNativeCallsPassOnce
, std::ref(Registry)); }
220 false, false)static void *initializeAMDGPUUseNativeCallsPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "Replace builtin math calls with that native versions."
, "amdgpu-usenative", &AMDGPUUseNativeCalls::ID, PassInfo
::NormalCtor_t(callDefaultCtor<AMDGPUUseNativeCalls>), false
, false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeAMDGPUUseNativeCallsPassFlag; void
llvm::initializeAMDGPUUseNativeCallsPass(PassRegistry &Registry
) { llvm::call_once(InitializeAMDGPUUseNativeCallsPassFlag, initializeAMDGPUUseNativeCallsPassOnce
, std::ref(Registry)); }
221
222template <typename IRB>
223static CallInst *CreateCallEx(IRB &B, FunctionCallee Callee, Value *Arg,
224 const Twine &Name = "") {
225 CallInst *R = B.CreateCall(Callee, Arg, Name);
226 if (Function *F = dyn_cast<Function>(Callee.getCallee()))
227 R->setCallingConv(F->getCallingConv());
228 return R;
229}
230
231template <typename IRB>
232static CallInst *CreateCallEx2(IRB &B, FunctionCallee Callee, Value *Arg1,
233 Value *Arg2, const Twine &Name = "") {
234 CallInst *R = B.CreateCall(Callee, {Arg1, Arg2}, Name);
235 if (Function *F = dyn_cast<Function>(Callee.getCallee()))
236 R->setCallingConv(F->getCallingConv());
237 return R;
238}
239
240// Data structures for table-driven optimizations.
241// FuncTbl works for both f32 and f64 functions with 1 input argument
242
243struct TableEntry {
244 double result;
245 double input;
246};
247
248/* a list of {result, input} */
249static const TableEntry tbl_acos[] = {
250 {MATH_PInumbers::pi / 2.0, 0.0},
251 {MATH_PInumbers::pi / 2.0, -0.0},
252 {0.0, 1.0},
253 {MATH_PInumbers::pi, -1.0}
254};
255static const TableEntry tbl_acosh[] = {
256 {0.0, 1.0}
257};
258static const TableEntry tbl_acospi[] = {
259 {0.5, 0.0},
260 {0.5, -0.0},
261 {0.0, 1.0},
262 {1.0, -1.0}
263};
264static const TableEntry tbl_asin[] = {
265 {0.0, 0.0},
266 {-0.0, -0.0},
267 {MATH_PInumbers::pi / 2.0, 1.0},
268 {-MATH_PInumbers::pi / 2.0, -1.0}
269};
270static const TableEntry tbl_asinh[] = {
271 {0.0, 0.0},
272 {-0.0, -0.0}
273};
274static const TableEntry tbl_asinpi[] = {
275 {0.0, 0.0},
276 {-0.0, -0.0},
277 {0.5, 1.0},
278 {-0.5, -1.0}
279};
280static const TableEntry tbl_atan[] = {
281 {0.0, 0.0},
282 {-0.0, -0.0},
283 {MATH_PInumbers::pi / 4.0, 1.0},
284 {-MATH_PInumbers::pi / 4.0, -1.0}
285};
286static const TableEntry tbl_atanh[] = {
287 {0.0, 0.0},
288 {-0.0, -0.0}
289};
290static const TableEntry tbl_atanpi[] = {
291 {0.0, 0.0},
292 {-0.0, -0.0},
293 {0.25, 1.0},
294 {-0.25, -1.0}
295};
296static const TableEntry tbl_cbrt[] = {
297 {0.0, 0.0},
298 {-0.0, -0.0},
299 {1.0, 1.0},
300 {-1.0, -1.0},
301};
302static const TableEntry tbl_cos[] = {
303 {1.0, 0.0},
304 {1.0, -0.0}
305};
306static const TableEntry tbl_cosh[] = {
307 {1.0, 0.0},
308 {1.0, -0.0}
309};
310static const TableEntry tbl_cospi[] = {
311 {1.0, 0.0},
312 {1.0, -0.0}
313};
314static const TableEntry tbl_erfc[] = {
315 {1.0, 0.0},
316 {1.0, -0.0}
317};
318static const TableEntry tbl_erf[] = {
319 {0.0, 0.0},
320 {-0.0, -0.0}
321};
322static const TableEntry tbl_exp[] = {
323 {1.0, 0.0},
324 {1.0, -0.0},
325 {MATH_Enumbers::e, 1.0}
326};
327static const TableEntry tbl_exp2[] = {
328 {1.0, 0.0},
329 {1.0, -0.0},
330 {2.0, 1.0}
331};
332static const TableEntry tbl_exp10[] = {
333 {1.0, 0.0},
334 {1.0, -0.0},
335 {10.0, 1.0}
336};
337static const TableEntry tbl_expm1[] = {
338 {0.0, 0.0},
339 {-0.0, -0.0}
340};
341static const TableEntry tbl_log[] = {
342 {0.0, 1.0},
343 {1.0, MATH_Enumbers::e}
344};
345static const TableEntry tbl_log2[] = {
346 {0.0, 1.0},
347 {1.0, 2.0}
348};
349static const TableEntry tbl_log10[] = {
350 {0.0, 1.0},
351 {1.0, 10.0}
352};
353static const TableEntry tbl_rsqrt[] = {
354 {1.0, 1.0},
355 {MATH_SQRT1_2numbers::inv_sqrt2, 2.0}
356};
357static const TableEntry tbl_sin[] = {
358 {0.0, 0.0},
359 {-0.0, -0.0}
360};
361static const TableEntry tbl_sinh[] = {
362 {0.0, 0.0},
363 {-0.0, -0.0}
364};
365static const TableEntry tbl_sinpi[] = {
366 {0.0, 0.0},
367 {-0.0, -0.0}
368};
369static const TableEntry tbl_sqrt[] = {
370 {0.0, 0.0},
371 {1.0, 1.0},
372 {MATH_SQRT2numbers::sqrt2, 2.0}
373};
374static const TableEntry tbl_tan[] = {
375 {0.0, 0.0},
376 {-0.0, -0.0}
377};
378static const TableEntry tbl_tanh[] = {
379 {0.0, 0.0},
380 {-0.0, -0.0}
381};
382static const TableEntry tbl_tanpi[] = {
383 {0.0, 0.0},
384 {-0.0, -0.0}
385};
386static const TableEntry tbl_tgamma[] = {
387 {1.0, 1.0},
388 {1.0, 2.0},
389 {2.0, 3.0},
390 {6.0, 4.0}
391};
392
393static bool HasNative(AMDGPULibFunc::EFuncId id) {
394 switch(id) {
395 case AMDGPULibFunc::EI_DIVIDE:
396 case AMDGPULibFunc::EI_COS:
397 case AMDGPULibFunc::EI_EXP:
398 case AMDGPULibFunc::EI_EXP2:
399 case AMDGPULibFunc::EI_EXP10:
400 case AMDGPULibFunc::EI_LOG:
401 case AMDGPULibFunc::EI_LOG2:
402 case AMDGPULibFunc::EI_LOG10:
403 case AMDGPULibFunc::EI_POWR:
404 case AMDGPULibFunc::EI_RECIP:
405 case AMDGPULibFunc::EI_RSQRT:
406 case AMDGPULibFunc::EI_SIN:
407 case AMDGPULibFunc::EI_SINCOS:
408 case AMDGPULibFunc::EI_SQRT:
409 case AMDGPULibFunc::EI_TAN:
410 return true;
411 default:;
412 }
413 return false;
414}
415
416struct TableRef {
417 size_t size;
418 const TableEntry *table; // variable size: from 0 to (size - 1)
419
420 TableRef() : size(0), table(nullptr) {}
421
422 template <size_t N>
423 TableRef(const TableEntry (&tbl)[N]) : size(N), table(&tbl[0]) {}
424};
425
426static TableRef getOptTable(AMDGPULibFunc::EFuncId id) {
427 switch(id) {
428 case AMDGPULibFunc::EI_ACOS: return TableRef(tbl_acos);
429 case AMDGPULibFunc::EI_ACOSH: return TableRef(tbl_acosh);
430 case AMDGPULibFunc::EI_ACOSPI: return TableRef(tbl_acospi);
431 case AMDGPULibFunc::EI_ASIN: return TableRef(tbl_asin);
432 case AMDGPULibFunc::EI_ASINH: return TableRef(tbl_asinh);
433 case AMDGPULibFunc::EI_ASINPI: return TableRef(tbl_asinpi);
434 case AMDGPULibFunc::EI_ATAN: return TableRef(tbl_atan);
435 case AMDGPULibFunc::EI_ATANH: return TableRef(tbl_atanh);
436 case AMDGPULibFunc::EI_ATANPI: return TableRef(tbl_atanpi);
437 case AMDGPULibFunc::EI_CBRT: return TableRef(tbl_cbrt);
438 case AMDGPULibFunc::EI_NCOS:
439 case AMDGPULibFunc::EI_COS: return TableRef(tbl_cos);
440 case AMDGPULibFunc::EI_COSH: return TableRef(tbl_cosh);
441 case AMDGPULibFunc::EI_COSPI: return TableRef(tbl_cospi);
442 case AMDGPULibFunc::EI_ERFC: return TableRef(tbl_erfc);
443 case AMDGPULibFunc::EI_ERF: return TableRef(tbl_erf);
444 case AMDGPULibFunc::EI_EXP: return TableRef(tbl_exp);
445 case AMDGPULibFunc::EI_NEXP2:
446 case AMDGPULibFunc::EI_EXP2: return TableRef(tbl_exp2);
447 case AMDGPULibFunc::EI_EXP10: return TableRef(tbl_exp10);
448 case AMDGPULibFunc::EI_EXPM1: return TableRef(tbl_expm1);
449 case AMDGPULibFunc::EI_LOG: return TableRef(tbl_log);
450 case AMDGPULibFunc::EI_NLOG2:
451 case AMDGPULibFunc::EI_LOG2: return TableRef(tbl_log2);
452 case AMDGPULibFunc::EI_LOG10: return TableRef(tbl_log10);
453 case AMDGPULibFunc::EI_NRSQRT:
454 case AMDGPULibFunc::EI_RSQRT: return TableRef(tbl_rsqrt);
455 case AMDGPULibFunc::EI_NSIN:
456 case AMDGPULibFunc::EI_SIN: return TableRef(tbl_sin);
457 case AMDGPULibFunc::EI_SINH: return TableRef(tbl_sinh);
458 case AMDGPULibFunc::EI_SINPI: return TableRef(tbl_sinpi);
459 case AMDGPULibFunc::EI_NSQRT:
460 case AMDGPULibFunc::EI_SQRT: return TableRef(tbl_sqrt);
461 case AMDGPULibFunc::EI_TAN: return TableRef(tbl_tan);
462 case AMDGPULibFunc::EI_TANH: return TableRef(tbl_tanh);
463 case AMDGPULibFunc::EI_TANPI: return TableRef(tbl_tanpi);
464 case AMDGPULibFunc::EI_TGAMMA: return TableRef(tbl_tgamma);
465 default:;
466 }
467 return TableRef();
468}
469
470static inline int getVecSize(const AMDGPULibFunc& FInfo) {
471 return FInfo.getLeads()[0].VectorSize;
472}
473
474static inline AMDGPULibFunc::EType getArgType(const AMDGPULibFunc& FInfo) {
475 return (AMDGPULibFunc::EType)FInfo.getLeads()[0].ArgType;
476}
477
478FunctionCallee AMDGPULibCalls::getFunction(Module *M, const FuncInfo &fInfo) {
479 // If we are doing PreLinkOpt, the function is external. So it is safe to
480 // use getOrInsertFunction() at this stage.
481
482 return EnablePreLink ? AMDGPULibFunc::getOrInsertFunction(M, fInfo)
483 : AMDGPULibFunc::getFunction(M, fInfo);
484}
485
486bool AMDGPULibCalls::parseFunctionName(const StringRef& FMangledName,
487 FuncInfo *FInfo) {
488 return AMDGPULibFunc::parse(FMangledName, *FInfo);
489}
490
491bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const {
492 if (auto Op = dyn_cast<FPMathOperator>(CI))
493 if (Op->isFast())
494 return true;
495 const Function *F = CI->getParent()->getParent();
496 Attribute Attr = F->getFnAttribute("unsafe-fp-math");
497 return Attr.getValueAsString() == "true";
498}
499
500bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
501 return AllNative ||
502 std::find(UseNative.begin(), UseNative.end(), F) != UseNative.end();
503}
504
505void AMDGPULibCalls::initNativeFuncs() {
506 AllNative = useNativeFunc("all") ||
507 (UseNative.getNumOccurrences() && UseNative.size() == 1 &&
508 UseNative.begin()->empty());
509}
510
511bool AMDGPULibCalls::sincosUseNative(CallInst *aCI, const FuncInfo &FInfo) {
512 bool native_sin = useNativeFunc("sin");
513 bool native_cos = useNativeFunc("cos");
514
515 if (native_sin && native_cos) {
516 Module *M = aCI->getModule();
517 Value *opr0 = aCI->getArgOperand(0);
518
519 AMDGPULibFunc nf;
520 nf.getLeads()[0].ArgType = FInfo.getLeads()[0].ArgType;
521 nf.getLeads()[0].VectorSize = FInfo.getLeads()[0].VectorSize;
522
523 nf.setPrefix(AMDGPULibFunc::NATIVE);
524 nf.setId(AMDGPULibFunc::EI_SIN);
525 FunctionCallee sinExpr = getFunction(M, nf);
526
527 nf.setPrefix(AMDGPULibFunc::NATIVE);
528 nf.setId(AMDGPULibFunc::EI_COS);
529 FunctionCallee cosExpr = getFunction(M, nf);
530 if (sinExpr && cosExpr) {
531 Value *sinval = CallInst::Create(sinExpr, opr0, "splitsin", aCI);
532 Value *cosval = CallInst::Create(cosExpr, opr0, "splitcos", aCI);
533 new StoreInst(cosval, aCI->getArgOperand(1), aCI);
534
535 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("usenative")) { dbgs() << "<useNative> replace "
<< *aCI << " with native version of sin/cos"; } }
while (false)
536 << " with native version of sin/cos")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("usenative")) { dbgs() << "<useNative> replace "
<< *aCI << " with native version of sin/cos"; } }
while (false)
;
537
538 replaceCall(sinval);
539 return true;
540 }
541 }
542 return false;
543}
544
545bool AMDGPULibCalls::useNative(CallInst *aCI) {
546 CI = aCI;
547 Function *Callee = aCI->getCalledFunction();
548
549 FuncInfo FInfo;
550 if (!parseFunctionName(Callee->getName(), &FInfo) || !FInfo.isMangled() ||
551 FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
552 getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()) ||
553 !(AllNative || useNativeFunc(FInfo.getName()))) {
554 return false;
555 }
556
557 if (FInfo.getId() == AMDGPULibFunc::EI_SINCOS)
558 return sincosUseNative(aCI, FInfo);
559
560 FInfo.setPrefix(AMDGPULibFunc::NATIVE);
561 FunctionCallee F = getFunction(aCI->getModule(), FInfo);
562 if (!F)
563 return false;
564
565 aCI->setCalledFunction(F);
566 DEBUG_WITH_TYPE("usenative", dbgs() << "<useNative> replace " << *aCIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("usenative")) { dbgs() << "<useNative> replace "
<< *aCI << " with native version"; } } while (false
)
567 << " with native version")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("usenative")) { dbgs() << "<useNative> replace "
<< *aCI << " with native version"; } } while (false
)
;
568 return true;
569}
570
571// Clang emits call of __read_pipe_2 or __read_pipe_4 for OpenCL read_pipe
572// builtin, with appended type size and alignment arguments, where 2 or 4
573// indicates the original number of arguments. The library has optimized version
574// of __read_pipe_2/__read_pipe_4 when the type size and alignment has the same
575// power of 2 value. This function transforms __read_pipe_2 to __read_pipe_2_N
576// for such cases where N is the size in bytes of the type (N = 1, 2, 4, 8, ...,
577// 128). The same for __read_pipe_4, write_pipe_2, and write_pipe_4.
578bool AMDGPULibCalls::fold_read_write_pipe(CallInst *CI, IRBuilder<> &B,
579 FuncInfo &FInfo) {
580 auto *Callee = CI->getCalledFunction();
581 if (!Callee->isDeclaration())
582 return false;
583
584 assert(Callee->hasName() && "Invalid read_pipe/write_pipe function")((Callee->hasName() && "Invalid read_pipe/write_pipe function"
) ? static_cast<void> (0) : __assert_fail ("Callee->hasName() && \"Invalid read_pipe/write_pipe function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 584, __PRETTY_FUNCTION__))
;
585 auto *M = Callee->getParent();
586 auto &Ctx = M->getContext();
587 std::string Name = Callee->getName();
588 auto NumArg = CI->getNumArgOperands();
589 if (NumArg != 4 && NumArg != 6)
590 return false;
591 auto *PacketSize = CI->getArgOperand(NumArg - 2);
592 auto *PacketAlign = CI->getArgOperand(NumArg - 1);
593 if (!isa<ConstantInt>(PacketSize) || !isa<ConstantInt>(PacketAlign))
594 return false;
595 unsigned Size = cast<ConstantInt>(PacketSize)->getZExtValue();
596 unsigned Align = cast<ConstantInt>(PacketAlign)->getZExtValue();
597 if (Size != Align || !isPowerOf2_32(Size))
598 return false;
599
600 Type *PtrElemTy;
601 if (Size <= 8)
602 PtrElemTy = Type::getIntNTy(Ctx, Size * 8);
603 else
604 PtrElemTy = VectorType::get(Type::getInt64Ty(Ctx), Size / 8);
605 unsigned PtrArgLoc = CI->getNumArgOperands() - 3;
606 auto PtrArg = CI->getArgOperand(PtrArgLoc);
607 unsigned PtrArgAS = PtrArg->getType()->getPointerAddressSpace();
608 auto *PtrTy = llvm::PointerType::get(PtrElemTy, PtrArgAS);
609
610 SmallVector<llvm::Type *, 6> ArgTys;
611 for (unsigned I = 0; I != PtrArgLoc; ++I)
612 ArgTys.push_back(CI->getArgOperand(I)->getType());
613 ArgTys.push_back(PtrTy);
614
615 Name = Name + "_" + std::to_string(Size);
616 auto *FTy = FunctionType::get(Callee->getReturnType(),
617 ArrayRef<Type *>(ArgTys), false);
618 AMDGPULibFunc NewLibFunc(Name, FTy);
619 FunctionCallee F = AMDGPULibFunc::getOrInsertFunction(M, NewLibFunc);
620 if (!F)
621 return false;
622
623 auto *BCast = B.CreatePointerCast(PtrArg, PtrTy);
624 SmallVector<Value *, 6> Args;
625 for (unsigned I = 0; I != PtrArgLoc; ++I)
626 Args.push_back(CI->getArgOperand(I));
627 Args.push_back(BCast);
628
629 auto *NCI = B.CreateCall(F, Args);
630 NCI->setAttributes(CI->getAttributes());
631 CI->replaceAllUsesWith(NCI);
632 CI->dropAllReferences();
633 CI->eraseFromParent();
634
635 return true;
636}
637
638// This function returns false if no change; return true otherwise.
639bool AMDGPULibCalls::fold(CallInst *CI, AliasAnalysis *AA) {
640 this->CI = CI;
641 Function *Callee = CI->getCalledFunction();
642
643 // Ignore indirect calls.
644 if (Callee
13.1
'Callee' is not equal to null
== 0) return false;
14
Taking false branch
645
646 BasicBlock *BB = CI->getParent();
647 LLVMContext &Context = CI->getParent()->getContext();
648 IRBuilder<> B(Context);
649
650 // Set the builder to the instruction after the call.
651 B.SetInsertPoint(BB, CI->getIterator());
652
653 // Copy fast flags from the original call.
654 if (const FPMathOperator *FPOp
15.1
'FPOp' is null
= dyn_cast<const FPMathOperator>(CI))
15
Assuming 'CI' is not a 'FPMathOperator'
16
Taking false branch
655 B.setFastMathFlags(FPOp->getFastMathFlags());
656
657 switch (Callee->getIntrinsicID()) {
17
Control jumps to the 'default' case at line 658
658 default:
659 break;
18
Execution continues on line 664
660 case Intrinsic::amdgcn_wavefrontsize:
661 return !EnablePreLink && fold_wavefrontsize(CI, B);
662 }
663
664 FuncInfo FInfo;
665 if (!parseFunctionName(Callee->getName(), &FInfo))
19
Assuming the condition is false
20
Taking false branch
666 return false;
667
668 // Further check the number of arguments to see if they match.
669 if (CI->getNumArgOperands() != FInfo.getNumArgs())
21
Assuming the condition is false
22
Taking false branch
670 return false;
671
672 if (TDOFold(CI, FInfo))
23
Taking false branch
673 return true;
674
675 // Under unsafe-math, evaluate calls if possible.
676 // According to Brian Sumner, we can do this for all f32 function calls
677 // using host's double function calls.
678 if (isUnsafeMath(CI) && evaluateCall(CI, FInfo))
24
Assuming the condition is true
25
Calling 'AMDGPULibCalls::evaluateCall'
679 return true;
680
681 // Specilized optimizations for each function call
682 switch (FInfo.getId()) {
683 case AMDGPULibFunc::EI_RECIP:
684 // skip vector function
685 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||(((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix
() == AMDGPULibFunc::HALF) && "recip must be an either native or half function"
) ? static_cast<void> (0) : __assert_fail ("(FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix() == AMDGPULibFunc::HALF) && \"recip must be an either native or half function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 687, __PRETTY_FUNCTION__))
686 FInfo.getPrefix() == AMDGPULibFunc::HALF) &&(((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix
() == AMDGPULibFunc::HALF) && "recip must be an either native or half function"
) ? static_cast<void> (0) : __assert_fail ("(FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix() == AMDGPULibFunc::HALF) && \"recip must be an either native or half function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 687, __PRETTY_FUNCTION__))
687 "recip must be an either native or half function")(((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix
() == AMDGPULibFunc::HALF) && "recip must be an either native or half function"
) ? static_cast<void> (0) : __assert_fail ("(FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix() == AMDGPULibFunc::HALF) && \"recip must be an either native or half function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 687, __PRETTY_FUNCTION__))
;
688 return (getVecSize(FInfo) != 1) ? false : fold_recip(CI, B, FInfo);
689
690 case AMDGPULibFunc::EI_DIVIDE:
691 // skip vector function
692 assert ((FInfo.getPrefix() == AMDGPULibFunc::NATIVE ||(((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix
() == AMDGPULibFunc::HALF) && "divide must be an either native or half function"
) ? static_cast<void> (0) : __assert_fail ("(FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix() == AMDGPULibFunc::HALF) && \"divide must be an either native or half function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 694, __PRETTY_FUNCTION__))
693 FInfo.getPrefix() == AMDGPULibFunc::HALF) &&(((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix
() == AMDGPULibFunc::HALF) && "divide must be an either native or half function"
) ? static_cast<void> (0) : __assert_fail ("(FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix() == AMDGPULibFunc::HALF) && \"divide must be an either native or half function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 694, __PRETTY_FUNCTION__))
694 "divide must be an either native or half function")(((FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix
() == AMDGPULibFunc::HALF) && "divide must be an either native or half function"
) ? static_cast<void> (0) : __assert_fail ("(FInfo.getPrefix() == AMDGPULibFunc::NATIVE || FInfo.getPrefix() == AMDGPULibFunc::HALF) && \"divide must be an either native or half function\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 694, __PRETTY_FUNCTION__))
;
695 return (getVecSize(FInfo) != 1) ? false : fold_divide(CI, B, FInfo);
696
697 case AMDGPULibFunc::EI_POW:
698 case AMDGPULibFunc::EI_POWR:
699 case AMDGPULibFunc::EI_POWN:
700 return fold_pow(CI, B, FInfo);
701
702 case AMDGPULibFunc::EI_ROOTN:
703 // skip vector function
704 return (getVecSize(FInfo) != 1) ? false : fold_rootn(CI, B, FInfo);
705
706 case AMDGPULibFunc::EI_FMA:
707 case AMDGPULibFunc::EI_MAD:
708 case AMDGPULibFunc::EI_NFMA:
709 // skip vector function
710 return (getVecSize(FInfo) != 1) ? false : fold_fma_mad(CI, B, FInfo);
711
712 case AMDGPULibFunc::EI_SQRT:
713 return isUnsafeMath(CI) && fold_sqrt(CI, B, FInfo);
714 case AMDGPULibFunc::EI_COS:
715 case AMDGPULibFunc::EI_SIN:
716 if ((getArgType(FInfo) == AMDGPULibFunc::F32 ||
717 getArgType(FInfo) == AMDGPULibFunc::F64)
718 && (FInfo.getPrefix() == AMDGPULibFunc::NOPFX))
719 return fold_sincos(CI, B, AA);
720
721 break;
722 case AMDGPULibFunc::EI_READ_PIPE_2:
723 case AMDGPULibFunc::EI_READ_PIPE_4:
724 case AMDGPULibFunc::EI_WRITE_PIPE_2:
725 case AMDGPULibFunc::EI_WRITE_PIPE_4:
726 return fold_read_write_pipe(CI, B, FInfo);
727
728 default:
729 break;
730 }
731
732 return false;
733}
734
735bool AMDGPULibCalls::TDOFold(CallInst *CI, const FuncInfo &FInfo) {
736 // Table-Driven optimization
737 const TableRef tr = getOptTable(FInfo.getId());
738 if (tr.size==0)
739 return false;
740
741 int const sz = (int)tr.size;
742 const TableEntry * const ftbl = tr.table;
743 Value *opr0 = CI->getArgOperand(0);
744
745 if (getVecSize(FInfo) > 1) {
746 if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(opr0)) {
747 SmallVector<double, 0> DVal;
748 for (int eltNo = 0; eltNo < getVecSize(FInfo); ++eltNo) {
749 ConstantFP *eltval = dyn_cast<ConstantFP>(
750 CV->getElementAsConstant((unsigned)eltNo));
751 assert(eltval && "Non-FP arguments in math function!")((eltval && "Non-FP arguments in math function!") ? static_cast
<void> (0) : __assert_fail ("eltval && \"Non-FP arguments in math function!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 751, __PRETTY_FUNCTION__))
;
752 bool found = false;
753 for (int i=0; i < sz; ++i) {
754 if (eltval->isExactlyValue(ftbl[i].input)) {
755 DVal.push_back(ftbl[i].result);
756 found = true;
757 break;
758 }
759 }
760 if (!found) {
761 // This vector constants not handled yet.
762 return false;
763 }
764 }
765 LLVMContext &context = CI->getParent()->getParent()->getContext();
766 Constant *nval;
767 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
768 SmallVector<float, 0> FVal;
769 for (unsigned i = 0; i < DVal.size(); ++i) {
770 FVal.push_back((float)DVal[i]);
771 }
772 ArrayRef<float> tmp(FVal);
773 nval = ConstantDataVector::get(context, tmp);
774 } else { // F64
775 ArrayRef<double> tmp(DVal);
776 nval = ConstantDataVector::get(context, tmp);
777 }
778 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *nval << "\n"; } } while
(false)
;
779 replaceCall(nval);
780 return true;
781 }
782 } else {
783 // Scalar version
784 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
785 for (int i = 0; i < sz; ++i) {
786 if (CF->isExactlyValue(ftbl[i].input)) {
787 Value *nval = ConstantFP::get(CF->getType(), ftbl[i].result);
788 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *nval << "\n"; } } while
(false)
;
789 replaceCall(nval);
790 return true;
791 }
792 }
793 }
794 }
795
796 return false;
797}
798
799bool AMDGPULibCalls::replaceWithNative(CallInst *CI, const FuncInfo &FInfo) {
800 Module *M = CI->getModule();
801 if (getArgType(FInfo) != AMDGPULibFunc::F32 ||
802 FInfo.getPrefix() != AMDGPULibFunc::NOPFX ||
803 !HasNative(FInfo.getId()))
804 return false;
805
806 AMDGPULibFunc nf = FInfo;
807 nf.setPrefix(AMDGPULibFunc::NATIVE);
808 if (FunctionCallee FPExpr = getFunction(M, nf)) {
809 LLVM_DEBUG(dbgs() << "AMDIC: " << *CI << " ---> ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { dbgs() << "AMDIC: " << *
CI << " ---> "; } } while (false)
;
810
811 CI->setCalledFunction(FPExpr);
812
813 LLVM_DEBUG(dbgs() << *CI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { dbgs() << *CI << '\n'; }
} while (false)
;
814
815 return true;
816 }
817 return false;
818}
819
820// [native_]half_recip(c) ==> 1.0/c
821bool AMDGPULibCalls::fold_recip(CallInst *CI, IRBuilder<> &B,
822 const FuncInfo &FInfo) {
823 Value *opr0 = CI->getArgOperand(0);
824 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr0)) {
825 // Just create a normal div. Later, InstCombine will be able
826 // to compute the divide into a constant (avoid check float infinity
827 // or subnormal at this point).
828 Value *nval = B.CreateFDiv(ConstantFP::get(CF->getType(), 1.0),
829 opr0,
830 "recip2div");
831 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *nval << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *nval << "\n"; } } while
(false)
;
832 replaceCall(nval);
833 return true;
834 }
835 return false;
836}
837
838// [native_]half_divide(x, c) ==> x/c
839bool AMDGPULibCalls::fold_divide(CallInst *CI, IRBuilder<> &B,
840 const FuncInfo &FInfo) {
841 Value *opr0 = CI->getArgOperand(0);
842 Value *opr1 = CI->getArgOperand(1);
843 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
844 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
845
846 if ((CF0 && CF1) || // both are constants
847 (CF1 && (getArgType(FInfo) == AMDGPULibFunc::F32)))
848 // CF1 is constant && f32 divide
849 {
850 Value *nval1 = B.CreateFDiv(ConstantFP::get(opr1->getType(), 1.0),
851 opr1, "__div2recip");
852 Value *nval = B.CreateFMul(opr0, nval1, "__div2mul");
853 replaceCall(nval);
854 return true;
855 }
856 return false;
857}
858
859namespace llvm {
860static double log2(double V) {
861#if _XOPEN_SOURCE700 >= 600 || defined(_ISOC99_SOURCE1) || _POSIX_C_SOURCE200809L >= 200112L
862 return ::log2(V);
863#else
864 return log(V) / numbers::ln2;
865#endif
866}
867}
868
869bool AMDGPULibCalls::fold_pow(CallInst *CI, IRBuilder<> &B,
870 const FuncInfo &FInfo) {
871 assert((FInfo.getId() == AMDGPULibFunc::EI_POW ||(((FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() ==
AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN
) && "fold_pow: encounter a wrong function call") ? static_cast
<void> (0) : __assert_fail ("(FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() == AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN) && \"fold_pow: encounter a wrong function call\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 874, __PRETTY_FUNCTION__))
872 FInfo.getId() == AMDGPULibFunc::EI_POWR ||(((FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() ==
AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN
) && "fold_pow: encounter a wrong function call") ? static_cast
<void> (0) : __assert_fail ("(FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() == AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN) && \"fold_pow: encounter a wrong function call\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 874, __PRETTY_FUNCTION__))
873 FInfo.getId() == AMDGPULibFunc::EI_POWN) &&(((FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() ==
AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN
) && "fold_pow: encounter a wrong function call") ? static_cast
<void> (0) : __assert_fail ("(FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() == AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN) && \"fold_pow: encounter a wrong function call\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 874, __PRETTY_FUNCTION__))
874 "fold_pow: encounter a wrong function call")(((FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() ==
AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN
) && "fold_pow: encounter a wrong function call") ? static_cast
<void> (0) : __assert_fail ("(FInfo.getId() == AMDGPULibFunc::EI_POW || FInfo.getId() == AMDGPULibFunc::EI_POWR || FInfo.getId() == AMDGPULibFunc::EI_POWN) && \"fold_pow: encounter a wrong function call\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 874, __PRETTY_FUNCTION__))
;
875
876 Value *opr0, *opr1;
877 ConstantFP *CF;
878 ConstantInt *CINT;
879 ConstantAggregateZero *CZero;
880 Type *eltType;
881
882 opr0 = CI->getArgOperand(0);
883 opr1 = CI->getArgOperand(1);
884 CZero = dyn_cast<ConstantAggregateZero>(opr1);
885 if (getVecSize(FInfo) == 1) {
886 eltType = opr0->getType();
887 CF = dyn_cast<ConstantFP>(opr1);
888 CINT = dyn_cast<ConstantInt>(opr1);
889 } else {
890 VectorType *VTy = dyn_cast<VectorType>(opr0->getType());
891 assert(VTy && "Oprand of vector function should be of vectortype")((VTy && "Oprand of vector function should be of vectortype"
) ? static_cast<void> (0) : __assert_fail ("VTy && \"Oprand of vector function should be of vectortype\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 891, __PRETTY_FUNCTION__))
;
892 eltType = VTy->getElementType();
893 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1);
894
895 // Now, only Handle vector const whose elements have the same value.
896 CF = CDV ? dyn_cast_or_null<ConstantFP>(CDV->getSplatValue()) : nullptr;
897 CINT = CDV ? dyn_cast_or_null<ConstantInt>(CDV->getSplatValue()) : nullptr;
898 }
899
900 // No unsafe math , no constant argument, do nothing
901 if (!isUnsafeMath(CI) && !CF && !CINT && !CZero)
902 return false;
903
904 // 0x1111111 means that we don't do anything for this call.
905 int ci_opr1 = (CINT ? (int)CINT->getSExtValue() : 0x1111111);
906
907 if ((CF && CF->isZero()) || (CINT && ci_opr1 == 0) || CZero) {
908 // pow/powr/pown(x, 0) == 1
909 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> 1\n"; } } while (false)
;
910 Constant *cnval = ConstantFP::get(eltType, 1.0);
911 if (getVecSize(FInfo) > 1) {
912 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
913 }
914 replaceCall(cnval);
915 return true;
916 }
917 if ((CF && CF->isExactlyValue(1.0)) || (CINT && ci_opr1 == 1)) {
918 // pow/powr/pown(x, 1.0) = x
919 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << "\n"; } } while
(false)
;
920 replaceCall(opr0);
921 return true;
922 }
923 if ((CF && CF->isExactlyValue(2.0)) || (CINT && ci_opr1 == 2)) {
924 // pow/powr/pown(x, 2.0) = x*x
925 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * " << *opr0do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << " * " <<
*opr0 << "\n"; } } while (false)
926 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << " * " <<
*opr0 << "\n"; } } while (false)
;
927 Value *nval = B.CreateFMul(opr0, opr0, "__pow2");
928 replaceCall(nval);
929 return true;
930 }
931 if ((CF && CF->isExactlyValue(-1.0)) || (CINT && ci_opr1 == -1)) {
932 // pow/powr/pown(x, -1.0) = 1.0/x
933 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1 / " << *opr0 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> 1 / " << *opr0 << "\n"; } }
while (false)
;
934 Constant *cnval = ConstantFP::get(eltType, 1.0);
935 if (getVecSize(FInfo) > 1) {
936 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
937 }
938 Value *nval = B.CreateFDiv(cnval, opr0, "__powrecip");
939 replaceCall(nval);
940 return true;
941 }
942
943 Module *M = CI->getModule();
944 if (CF && (CF->isExactlyValue(0.5) || CF->isExactlyValue(-0.5))) {
945 // pow[r](x, [-]0.5) = sqrt(x)
946 bool issqrt = CF->isExactlyValue(0.5);
947 if (FunctionCallee FPExpr =
948 getFunction(M, AMDGPULibFunc(issqrt ? AMDGPULibFunc::EI_SQRT
949 : AMDGPULibFunc::EI_RSQRT,
950 FInfo))) {
951 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << FInfo.getName().c_str() <<
"(" << *opr0 << ")\n"; } } while (false)
952 << FInfo.getName().c_str() << "(" << *opr0 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << FInfo.getName().c_str() <<
"(" << *opr0 << ")\n"; } } while (false)
;
953 Value *nval = CreateCallEx(B,FPExpr, opr0, issqrt ? "__pow2sqrt"
954 : "__pow2rsqrt");
955 replaceCall(nval);
956 return true;
957 }
958 }
959
960 if (!isUnsafeMath(CI))
961 return false;
962
963 // Unsafe Math optimization
964
965 // Remember that ci_opr1 is set if opr1 is integral
966 if (CF) {
967 double dval = (getArgType(FInfo) == AMDGPULibFunc::F32)
968 ? (double)CF->getValueAPF().convertToFloat()
969 : CF->getValueAPF().convertToDouble();
970 int ival = (int)dval;
971 if ((double)ival == dval) {
972 ci_opr1 = ival;
973 } else
974 ci_opr1 = 0x11111111;
975 }
976
977 // pow/powr/pown(x, c) = [1/](x*x*..x); where
978 // trunc(c) == c && the number of x == c && |c| <= 12
979 unsigned abs_opr1 = (ci_opr1 < 0) ? -ci_opr1 : ci_opr1;
980 if (abs_opr1 <= 12) {
981 Constant *cnval;
982 Value *nval;
983 if (abs_opr1 == 0) {
984 cnval = ConstantFP::get(eltType, 1.0);
985 if (getVecSize(FInfo) > 1) {
986 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
987 }
988 nval = cnval;
989 } else {
990 Value *valx2 = nullptr;
991 nval = nullptr;
992 while (abs_opr1 > 0) {
993 valx2 = valx2 ? B.CreateFMul(valx2, valx2, "__powx2") : opr0;
994 if (abs_opr1 & 1) {
995 nval = nval ? B.CreateFMul(nval, valx2, "__powprod") : valx2;
996 }
997 abs_opr1 >>= 1;
998 }
999 }
1000
1001 if (ci_opr1 < 0) {
1002 cnval = ConstantFP::get(eltType, 1.0);
1003 if (getVecSize(FInfo) > 1) {
1004 cnval = ConstantDataVector::getSplat(getVecSize(FInfo), cnval);
1005 }
1006 nval = B.CreateFDiv(cnval, nval, "__1powprod");
1007 }
1008 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << ((ci_opr1 < 0) ? "1/prod("
: "prod(") << *opr0 << ")\n"; } } while (false)
1009 << ((ci_opr1 < 0) ? "1/prod(" : "prod(") << *opr0do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << ((ci_opr1 < 0) ? "1/prod("
: "prod(") << *opr0 << ")\n"; } } while (false)
1010 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << ((ci_opr1 < 0) ? "1/prod("
: "prod(") << *opr0 << ")\n"; } } while (false)
;
1011 replaceCall(nval);
1012 return true;
1013 }
1014
1015 // powr ---> exp2(y * log2(x))
1016 // pown/pow ---> powr(fabs(x), y) | (x & ((int)y << 31))
1017 FunctionCallee ExpExpr =
1018 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_EXP2, FInfo));
1019 if (!ExpExpr)
1020 return false;
1021
1022 bool needlog = false;
1023 bool needabs = false;
1024 bool needcopysign = false;
1025 Constant *cnval = nullptr;
1026 if (getVecSize(FInfo) == 1) {
1027 CF = dyn_cast<ConstantFP>(opr0);
1028
1029 if (CF) {
1030 double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1031 ? (double)CF->getValueAPF().convertToFloat()
1032 : CF->getValueAPF().convertToDouble();
1033
1034 V = log2(std::abs(V));
1035 cnval = ConstantFP::get(eltType, V);
1036 needcopysign = (FInfo.getId() != AMDGPULibFunc::EI_POWR) &&
1037 CF->isNegative();
1038 } else {
1039 needlog = true;
1040 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR &&
1041 (!CF || CF->isNegative());
1042 }
1043 } else {
1044 ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr0);
1045
1046 if (!CDV) {
1047 needlog = true;
1048 needcopysign = needabs = FInfo.getId() != AMDGPULibFunc::EI_POWR;
1049 } else {
1050 assert ((int)CDV->getNumElements() == getVecSize(FInfo) &&(((int)CDV->getNumElements() == getVecSize(FInfo) &&
"Wrong vector size detected") ? static_cast<void> (0) :
__assert_fail ("(int)CDV->getNumElements() == getVecSize(FInfo) && \"Wrong vector size detected\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1051, __PRETTY_FUNCTION__))
1051 "Wrong vector size detected")(((int)CDV->getNumElements() == getVecSize(FInfo) &&
"Wrong vector size detected") ? static_cast<void> (0) :
__assert_fail ("(int)CDV->getNumElements() == getVecSize(FInfo) && \"Wrong vector size detected\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1051, __PRETTY_FUNCTION__))
;
1052
1053 SmallVector<double, 0> DVal;
1054 for (int i=0; i < getVecSize(FInfo); ++i) {
1055 double V = (getArgType(FInfo) == AMDGPULibFunc::F32)
1056 ? (double)CDV->getElementAsFloat(i)
1057 : CDV->getElementAsDouble(i);
1058 if (V < 0.0) needcopysign = true;
1059 V = log2(std::abs(V));
1060 DVal.push_back(V);
1061 }
1062 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1063 SmallVector<float, 0> FVal;
1064 for (unsigned i=0; i < DVal.size(); ++i) {
1065 FVal.push_back((float)DVal[i]);
1066 }
1067 ArrayRef<float> tmp(FVal);
1068 cnval = ConstantDataVector::get(M->getContext(), tmp);
1069 } else {
1070 ArrayRef<double> tmp(DVal);
1071 cnval = ConstantDataVector::get(M->getContext(), tmp);
1072 }
1073 }
1074 }
1075
1076 if (needcopysign && (FInfo.getId() == AMDGPULibFunc::EI_POW)) {
1077 // We cannot handle corner cases for a general pow() function, give up
1078 // unless y is a constant integral value. Then proceed as if it were pown.
1079 if (getVecSize(FInfo) == 1) {
1080 if (const ConstantFP *CF = dyn_cast<ConstantFP>(opr1)) {
1081 double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1082 ? (double)CF->getValueAPF().convertToFloat()
1083 : CF->getValueAPF().convertToDouble();
1084 if (y != (double)(int64_t)y)
1085 return false;
1086 } else
1087 return false;
1088 } else {
1089 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(opr1)) {
1090 for (int i=0; i < getVecSize(FInfo); ++i) {
1091 double y = (getArgType(FInfo) == AMDGPULibFunc::F32)
1092 ? (double)CDV->getElementAsFloat(i)
1093 : CDV->getElementAsDouble(i);
1094 if (y != (double)(int64_t)y)
1095 return false;
1096 }
1097 } else
1098 return false;
1099 }
1100 }
1101
1102 Value *nval;
1103 if (needabs) {
1104 FunctionCallee AbsExpr =
1105 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_FABS, FInfo));
1106 if (!AbsExpr)
1107 return false;
1108 nval = CreateCallEx(B, AbsExpr, opr0, "__fabs");
1109 } else {
1110 nval = cnval ? cnval : opr0;
1111 }
1112 if (needlog) {
1113 FunctionCallee LogExpr =
1114 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_LOG2, FInfo));
1115 if (!LogExpr)
1116 return false;
1117 nval = CreateCallEx(B,LogExpr, nval, "__log2");
1118 }
1119
1120 if (FInfo.getId() == AMDGPULibFunc::EI_POWN) {
1121 // convert int(32) to fp(f32 or f64)
1122 opr1 = B.CreateSIToFP(opr1, nval->getType(), "pownI2F");
1123 }
1124 nval = B.CreateFMul(opr1, nval, "__ylogx");
1125 nval = CreateCallEx(B,ExpExpr, nval, "__exp2");
1126
1127 if (needcopysign) {
1128 Value *opr_n;
1129 Type* rTy = opr0->getType();
1130 Type* nTyS = eltType->isDoubleTy() ? B.getInt64Ty() : B.getInt32Ty();
1131 Type *nTy = nTyS;
1132 if (const VectorType *vTy = dyn_cast<VectorType>(rTy))
1133 nTy = VectorType::get(nTyS, vTy->getNumElements());
1134 unsigned size = nTy->getScalarSizeInBits();
1135 opr_n = CI->getArgOperand(1);
1136 if (opr_n->getType()->isIntegerTy())
1137 opr_n = B.CreateZExtOrBitCast(opr_n, nTy, "__ytou");
1138 else
1139 opr_n = B.CreateFPToSI(opr1, nTy, "__ytou");
1140
1141 Value *sign = B.CreateShl(opr_n, size-1, "__yeven");
1142 sign = B.CreateAnd(B.CreateBitCast(opr0, nTy), sign, "__pow_sign");
1143 nval = B.CreateOr(B.CreateBitCast(nval, nTy), sign);
1144 nval = B.CreateBitCast(nval, opr0->getType());
1145 }
1146
1147 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << "exp2(" << *opr1 <<
" * log2(" << *opr0 << "))\n"; } } while (false)
1148 << "exp2(" << *opr1 << " * log2(" << *opr0 << "))\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << "exp2(" << *opr1 <<
" * log2(" << *opr0 << "))\n"; } } while (false)
;
1149 replaceCall(nval);
1150
1151 return true;
1152}
1153
1154bool AMDGPULibCalls::fold_rootn(CallInst *CI, IRBuilder<> &B,
1155 const FuncInfo &FInfo) {
1156 Value *opr0 = CI->getArgOperand(0);
1157 Value *opr1 = CI->getArgOperand(1);
1158
1159 ConstantInt *CINT = dyn_cast<ConstantInt>(opr1);
1160 if (!CINT) {
1161 return false;
1162 }
1163 int ci_opr1 = (int)CINT->getSExtValue();
1164 if (ci_opr1 == 1) { // rootn(x, 1) = x
1165 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << "\n"; } } while
(false)
;
1166 replaceCall(opr0);
1167 return true;
1168 }
1169 if (ci_opr1 == 2) { // rootn(x, 2) = sqrt(x)
1170 std::vector<const Type*> ParamsTys;
1171 ParamsTys.push_back(opr0->getType());
1172 Module *M = CI->getModule();
1173 if (FunctionCallee FPExpr =
1174 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1175 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> sqrt(" << *opr0 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> sqrt(" << *opr0 << ")\n"; }
} while (false)
;
1176 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2sqrt");
1177 replaceCall(nval);
1178 return true;
1179 }
1180 } else if (ci_opr1 == 3) { // rootn(x, 3) = cbrt(x)
1181 Module *M = CI->getModule();
1182 if (FunctionCallee FPExpr =
1183 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_CBRT, FInfo))) {
1184 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> cbrt(" << *opr0 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> cbrt(" << *opr0 << ")\n"; }
} while (false)
;
1185 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2cbrt");
1186 replaceCall(nval);
1187 return true;
1188 }
1189 } else if (ci_opr1 == -1) { // rootn(x, -1) = 1.0/x
1190 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> 1.0 / " << *opr0 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> 1.0 / " << *opr0 << "\n"; }
} while (false)
;
1191 Value *nval = B.CreateFDiv(ConstantFP::get(opr0->getType(), 1.0),
1192 opr0,
1193 "__rootn2div");
1194 replaceCall(nval);
1195 return true;
1196 } else if (ci_opr1 == -2) { // rootn(x, -2) = rsqrt(x)
1197 std::vector<const Type*> ParamsTys;
1198 ParamsTys.push_back(opr0->getType());
1199 Module *M = CI->getModule();
1200 if (FunctionCallee FPExpr =
1201 getFunction(M, AMDGPULibFunc(AMDGPULibFunc::EI_RSQRT, FInfo))) {
1202 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> rsqrt(" << *opr0do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> rsqrt(" << *opr0 << ")\n"; }
} while (false)
1203 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> rsqrt(" << *opr0 << ")\n"; }
} while (false)
;
1204 Value *nval = CreateCallEx(B,FPExpr, opr0, "__rootn2rsqrt");
1205 replaceCall(nval);
1206 return true;
1207 }
1208 }
1209 return false;
1210}
1211
1212bool AMDGPULibCalls::fold_fma_mad(CallInst *CI, IRBuilder<> &B,
1213 const FuncInfo &FInfo) {
1214 Value *opr0 = CI->getArgOperand(0);
1215 Value *opr1 = CI->getArgOperand(1);
1216 Value *opr2 = CI->getArgOperand(2);
1217
1218 ConstantFP *CF0 = dyn_cast<ConstantFP>(opr0);
1219 ConstantFP *CF1 = dyn_cast<ConstantFP>(opr1);
1220 if ((CF0 && CF0->isZero()) || (CF1 && CF1->isZero())) {
1221 // fma/mad(a, b, c) = c if a=0 || b=0
1222 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr2 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr2 << "\n"; } } while
(false)
;
1223 replaceCall(opr2);
1224 return true;
1225 }
1226 if (CF0 && CF0->isExactlyValue(1.0f)) {
1227 // fma/mad(a, b, c) = b+c if a=1
1228 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr1 << " + " << *opr2do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr1 << " + " <<
*opr2 << "\n"; } } while (false)
1229 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr1 << " + " <<
*opr2 << "\n"; } } while (false)
;
1230 Value *nval = B.CreateFAdd(opr1, opr2, "fmaadd");
1231 replaceCall(nval);
1232 return true;
1233 }
1234 if (CF1 && CF1->isExactlyValue(1.0f)) {
1235 // fma/mad(a, b, c) = a+c if b=1
1236 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " + " << *opr2do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << " + " <<
*opr2 << "\n"; } } while (false)
1237 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << " + " <<
*opr2 << "\n"; } } while (false)
;
1238 Value *nval = B.CreateFAdd(opr0, opr2, "fmaadd");
1239 replaceCall(nval);
1240 return true;
1241 }
1242 if (ConstantFP *CF = dyn_cast<ConstantFP>(opr2)) {
1243 if (CF->isZero()) {
1244 // fma/mad(a, b, c) = a*b if c=0
1245 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> " << *opr0 << " * "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << " * " <<
*opr1 << "\n"; } } while (false)
1246 << *opr1 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << *opr0 << " * " <<
*opr1 << "\n"; } } while (false)
;
1247 Value *nval = B.CreateFMul(opr0, opr1, "fmamul");
1248 replaceCall(nval);
1249 return true;
1250 }
1251 }
1252
1253 return false;
1254}
1255
1256// Get a scalar native builtin signle argument FP function
1257FunctionCallee AMDGPULibCalls::getNativeFunction(Module *M,
1258 const FuncInfo &FInfo) {
1259 if (getArgType(FInfo) == AMDGPULibFunc::F64 || !HasNative(FInfo.getId()))
1260 return nullptr;
1261 FuncInfo nf = FInfo;
1262 nf.setPrefix(AMDGPULibFunc::NATIVE);
1263 return getFunction(M, nf);
1264}
1265
1266// fold sqrt -> native_sqrt (x)
1267bool AMDGPULibCalls::fold_sqrt(CallInst *CI, IRBuilder<> &B,
1268 const FuncInfo &FInfo) {
1269 if (getArgType(FInfo) == AMDGPULibFunc::F32 && (getVecSize(FInfo) == 1) &&
1270 (FInfo.getPrefix() != AMDGPULibFunc::NATIVE)) {
1271 if (FunctionCallee FPExpr = getNativeFunction(
1272 CI->getModule(), AMDGPULibFunc(AMDGPULibFunc::EI_SQRT, FInfo))) {
1273 Value *opr0 = CI->getArgOperand(0);
1274 LLVM_DEBUG(errs() << "AMDIC: " << *CI << " ---> "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << "sqrt(" << *opr0 <<
")\n"; } } while (false)
1275 << "sqrt(" << *opr0 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: " << *
CI << " ---> " << "sqrt(" << *opr0 <<
")\n"; } } while (false)
;
1276 Value *nval = CreateCallEx(B,FPExpr, opr0, "__sqrt");
1277 replaceCall(nval);
1278 return true;
1279 }
1280 }
1281 return false;
1282}
1283
1284// fold sin, cos -> sincos.
1285bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
1286 AliasAnalysis *AA) {
1287 AMDGPULibFunc fInfo;
1288 if (!AMDGPULibFunc::parse(CI->getCalledFunction()->getName(), fInfo))
1289 return false;
1290
1291 assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||((fInfo.getId() == AMDGPULibFunc::EI_SIN || fInfo.getId() == AMDGPULibFunc
::EI_COS) ? static_cast<void> (0) : __assert_fail ("fInfo.getId() == AMDGPULibFunc::EI_SIN || fInfo.getId() == AMDGPULibFunc::EI_COS"
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1292, __PRETTY_FUNCTION__))
1292 fInfo.getId() == AMDGPULibFunc::EI_COS)((fInfo.getId() == AMDGPULibFunc::EI_SIN || fInfo.getId() == AMDGPULibFunc
::EI_COS) ? static_cast<void> (0) : __assert_fail ("fInfo.getId() == AMDGPULibFunc::EI_SIN || fInfo.getId() == AMDGPULibFunc::EI_COS"
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1292, __PRETTY_FUNCTION__))
;
1293 bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
1294
1295 Value *CArgVal = CI->getArgOperand(0);
1296 BasicBlock * const CBB = CI->getParent();
1297
1298 int const MaxScan = 30;
1299
1300 { // fold in load value.
1301 LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
1302 if (LI && LI->getParent() == CBB) {
1303 BasicBlock::iterator BBI = LI->getIterator();
1304 Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
1305 if (AvailableVal) {
1306 CArgVal->replaceAllUsesWith(AvailableVal);
1307 if (CArgVal->getNumUses() == 0)
1308 LI->eraseFromParent();
1309 CArgVal = CI->getArgOperand(0);
1310 }
1311 }
1312 }
1313
1314 Module *M = CI->getModule();
1315 fInfo.setId(isSin ? AMDGPULibFunc::EI_COS : AMDGPULibFunc::EI_SIN);
1316 std::string const PairName = fInfo.mangle();
1317
1318 CallInst *UI = nullptr;
1319 for (User* U : CArgVal->users()) {
1320 CallInst *XI = dyn_cast_or_null<CallInst>(U);
1321 if (!XI || XI == CI || XI->getParent() != CBB)
1322 continue;
1323
1324 Function *UCallee = XI->getCalledFunction();
1325 if (!UCallee || !UCallee->getName().equals(PairName))
1326 continue;
1327
1328 BasicBlock::iterator BBI = CI->getIterator();
1329 if (BBI == CI->getParent()->begin())
1330 break;
1331 --BBI;
1332 for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
1333 if (cast<Instruction>(BBI) == XI) {
1334 UI = XI;
1335 break;
1336 }
1337 }
1338 if (UI) break;
1339 }
1340
1341 if (!UI) return false;
1342
1343 // Merge the sin and cos.
1344
1345 // for OpenCL 2.0 we have only generic implementation of sincos
1346 // function.
1347 AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
1348 nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
1349 FunctionCallee Fsincos = getFunction(M, nf);
1350 if (!Fsincos) return false;
1351
1352 BasicBlock::iterator ItOld = B.GetInsertPoint();
1353 AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
1354 B.SetInsertPoint(UI);
1355
1356 Value *P = Alloc;
1357 Type *PTy = Fsincos.getFunctionType()->getParamType(1);
1358 // The allocaInst allocates the memory in private address space. This need
1359 // to be bitcasted to point to the address space of cos pointer type.
1360 // In OpenCL 2.0 this is generic, while in 1.2 that is private.
1361 if (PTy->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
1362 P = B.CreateAddrSpaceCast(Alloc, PTy);
1363 CallInst *Call = CreateCallEx2(B, Fsincos, UI->getArgOperand(0), P);
1364
1365 LLVM_DEBUG(errs() << "AMDIC: fold_sincos (" << *CI << ", " << *UI << ") with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: fold_sincos ("
<< *CI << ", " << *UI << ") with " <<
*Call << "\n"; } } while (false)
1366 << *Call << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: fold_sincos ("
<< *CI << ", " << *UI << ") with " <<
*Call << "\n"; } } while (false)
;
1367
1368 if (!isSin) { // CI->cos, UI->sin
1369 B.SetInsertPoint(&*ItOld);
1370 UI->replaceAllUsesWith(&*Call);
1371 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1372 CI->replaceAllUsesWith(Reload);
1373 UI->eraseFromParent();
1374 CI->eraseFromParent();
1375 } else { // CI->sin, UI->cos
1376 Instruction *Reload = B.CreateLoad(Alloc->getAllocatedType(), Alloc);
1377 UI->replaceAllUsesWith(Reload);
1378 CI->replaceAllUsesWith(Call);
1379 UI->eraseFromParent();
1380 CI->eraseFromParent();
1381 }
1382 return true;
1383}
1384
1385bool AMDGPULibCalls::fold_wavefrontsize(CallInst *CI, IRBuilder<> &B) {
1386 if (!TM)
1387 return false;
1388
1389 StringRef CPU = TM->getTargetCPU();
1390 StringRef Features = TM->getTargetFeatureString();
1391 if ((CPU.empty() || CPU.equals_lower("generic")) &&
1392 (Features.empty() ||
1393 Features.find_lower("wavefrontsize") == StringRef::npos))
1394 return false;
1395
1396 Function *F = CI->getParent()->getParent();
1397 const GCNSubtarget &ST = TM->getSubtarget<GCNSubtarget>(*F);
1398 unsigned N = ST.getWavefrontSize();
1399
1400 LLVM_DEBUG(errs() << "AMDIC: fold_wavefrontsize (" << *CI << ") with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: fold_wavefrontsize ("
<< *CI << ") with " << N << "\n"; } }
while (false)
1401 << N << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { errs() << "AMDIC: fold_wavefrontsize ("
<< *CI << ") with " << N << "\n"; } }
while (false)
;
1402
1403 CI->replaceAllUsesWith(ConstantInt::get(B.getInt32Ty(), N));
1404 CI->eraseFromParent();
1405 return true;
1406}
1407
1408// Get insertion point at entry.
1409BasicBlock::iterator AMDGPULibCalls::getEntryIns(CallInst * UI) {
1410 Function * Func = UI->getParent()->getParent();
1411 BasicBlock * BB = &Func->getEntryBlock();
1412 assert(BB && "Entry block not found!")((BB && "Entry block not found!") ? static_cast<void
> (0) : __assert_fail ("BB && \"Entry block not found!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1412, __PRETTY_FUNCTION__))
;
1413 BasicBlock::iterator ItNew = BB->begin();
1414 return ItNew;
1415}
1416
1417// Insert a AllocsInst at the beginning of function entry block.
1418AllocaInst* AMDGPULibCalls::insertAlloca(CallInst *UI, IRBuilder<> &B,
1419 const char *prefix) {
1420 BasicBlock::iterator ItNew = getEntryIns(UI);
1421 Function *UCallee = UI->getCalledFunction();
1422 Type *RetType = UCallee->getReturnType();
1423 B.SetInsertPoint(&*ItNew);
1424 AllocaInst *Alloc = B.CreateAlloca(RetType, 0,
1425 std::string(prefix) + UI->getName());
1426 Alloc->setAlignment(MaybeAlign(
1427 UCallee->getParent()->getDataLayout().getTypeAllocSize(RetType)));
1428 return Alloc;
1429}
1430
1431bool AMDGPULibCalls::evaluateScalarMathFunc(FuncInfo &FInfo,
1432 double& Res0, double& Res1,
1433 Constant *copr0, Constant *copr1,
1434 Constant *copr2) {
1435 // By default, opr0/opr1/opr3 holds values of float/double type.
1436 // If they are not float/double, each function has to its
1437 // operand separately.
1438 double opr0=0.0, opr1=0.0, opr2=0.0;
1439 ConstantFP *fpopr0 = dyn_cast_or_null<ConstantFP>(copr0);
1440 ConstantFP *fpopr1 = dyn_cast_or_null<ConstantFP>(copr1);
1441 ConstantFP *fpopr2 = dyn_cast_or_null<ConstantFP>(copr2);
1442 if (fpopr0) {
1443 opr0 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1444 ? fpopr0->getValueAPF().convertToDouble()
1445 : (double)fpopr0->getValueAPF().convertToFloat();
1446 }
1447
1448 if (fpopr1) {
1449 opr1 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1450 ? fpopr1->getValueAPF().convertToDouble()
1451 : (double)fpopr1->getValueAPF().convertToFloat();
1452 }
1453
1454 if (fpopr2) {
1455 opr2 = (getArgType(FInfo) == AMDGPULibFunc::F64)
1456 ? fpopr2->getValueAPF().convertToDouble()
1457 : (double)fpopr2->getValueAPF().convertToFloat();
1458 }
1459
1460 switch (FInfo.getId()) {
1461 default : return false;
1462
1463 case AMDGPULibFunc::EI_ACOS:
1464 Res0 = acos(opr0);
1465 return true;
1466
1467 case AMDGPULibFunc::EI_ACOSH:
1468 // acosh(x) == log(x + sqrt(x*x - 1))
1469 Res0 = log(opr0 + sqrt(opr0*opr0 - 1.0));
1470 return true;
1471
1472 case AMDGPULibFunc::EI_ACOSPI:
1473 Res0 = acos(opr0) / MATH_PInumbers::pi;
1474 return true;
1475
1476 case AMDGPULibFunc::EI_ASIN:
1477 Res0 = asin(opr0);
1478 return true;
1479
1480 case AMDGPULibFunc::EI_ASINH:
1481 // asinh(x) == log(x + sqrt(x*x + 1))
1482 Res0 = log(opr0 + sqrt(opr0*opr0 + 1.0));
1483 return true;
1484
1485 case AMDGPULibFunc::EI_ASINPI:
1486 Res0 = asin(opr0) / MATH_PInumbers::pi;
1487 return true;
1488
1489 case AMDGPULibFunc::EI_ATAN:
1490 Res0 = atan(opr0);
1491 return true;
1492
1493 case AMDGPULibFunc::EI_ATANH:
1494 // atanh(x) == (log(x+1) - log(x-1))/2;
1495 Res0 = (log(opr0 + 1.0) - log(opr0 - 1.0))/2.0;
1496 return true;
1497
1498 case AMDGPULibFunc::EI_ATANPI:
1499 Res0 = atan(opr0) / MATH_PInumbers::pi;
1500 return true;
1501
1502 case AMDGPULibFunc::EI_CBRT:
1503 Res0 = (opr0 < 0.0) ? -pow(-opr0, 1.0/3.0) : pow(opr0, 1.0/3.0);
1504 return true;
1505
1506 case AMDGPULibFunc::EI_COS:
1507 Res0 = cos(opr0);
1508 return true;
1509
1510 case AMDGPULibFunc::EI_COSH:
1511 Res0 = cosh(opr0);
1512 return true;
1513
1514 case AMDGPULibFunc::EI_COSPI:
1515 Res0 = cos(MATH_PInumbers::pi * opr0);
1516 return true;
1517
1518 case AMDGPULibFunc::EI_EXP:
1519 Res0 = exp(opr0);
1520 return true;
1521
1522 case AMDGPULibFunc::EI_EXP2:
1523 Res0 = pow(2.0, opr0);
1524 return true;
1525
1526 case AMDGPULibFunc::EI_EXP10:
1527 Res0 = pow(10.0, opr0);
1528 return true;
1529
1530 case AMDGPULibFunc::EI_EXPM1:
1531 Res0 = exp(opr0) - 1.0;
1532 return true;
1533
1534 case AMDGPULibFunc::EI_LOG:
1535 Res0 = log(opr0);
1536 return true;
1537
1538 case AMDGPULibFunc::EI_LOG2:
1539 Res0 = log(opr0) / log(2.0);
1540 return true;
1541
1542 case AMDGPULibFunc::EI_LOG10:
1543 Res0 = log(opr0) / log(10.0);
1544 return true;
1545
1546 case AMDGPULibFunc::EI_RSQRT:
1547 Res0 = 1.0 / sqrt(opr0);
1548 return true;
1549
1550 case AMDGPULibFunc::EI_SIN:
1551 Res0 = sin(opr0);
1552 return true;
1553
1554 case AMDGPULibFunc::EI_SINH:
1555 Res0 = sinh(opr0);
1556 return true;
1557
1558 case AMDGPULibFunc::EI_SINPI:
1559 Res0 = sin(MATH_PInumbers::pi * opr0);
1560 return true;
1561
1562 case AMDGPULibFunc::EI_SQRT:
1563 Res0 = sqrt(opr0);
1564 return true;
1565
1566 case AMDGPULibFunc::EI_TAN:
1567 Res0 = tan(opr0);
1568 return true;
1569
1570 case AMDGPULibFunc::EI_TANH:
1571 Res0 = tanh(opr0);
1572 return true;
1573
1574 case AMDGPULibFunc::EI_TANPI:
1575 Res0 = tan(MATH_PInumbers::pi * opr0);
1576 return true;
1577
1578 case AMDGPULibFunc::EI_RECIP:
1579 Res0 = 1.0 / opr0;
1580 return true;
1581
1582 // two-arg functions
1583 case AMDGPULibFunc::EI_DIVIDE:
1584 Res0 = opr0 / opr1;
1585 return true;
1586
1587 case AMDGPULibFunc::EI_POW:
1588 case AMDGPULibFunc::EI_POWR:
1589 Res0 = pow(opr0, opr1);
1590 return true;
1591
1592 case AMDGPULibFunc::EI_POWN: {
1593 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1594 double val = (double)iopr1->getSExtValue();
1595 Res0 = pow(opr0, val);
1596 return true;
1597 }
1598 return false;
1599 }
1600
1601 case AMDGPULibFunc::EI_ROOTN: {
1602 if (ConstantInt *iopr1 = dyn_cast_or_null<ConstantInt>(copr1)) {
1603 double val = (double)iopr1->getSExtValue();
1604 Res0 = pow(opr0, 1.0 / val);
1605 return true;
1606 }
1607 return false;
1608 }
1609
1610 // with ptr arg
1611 case AMDGPULibFunc::EI_SINCOS:
1612 Res0 = sin(opr0);
1613 Res1 = cos(opr0);
1614 return true;
1615
1616 // three-arg functions
1617 case AMDGPULibFunc::EI_FMA:
1618 case AMDGPULibFunc::EI_MAD:
1619 Res0 = opr0 * opr1 + opr2;
1620 return true;
1621 }
1622
1623 return false;
1624}
1625
1626bool AMDGPULibCalls::evaluateCall(CallInst *aCI, FuncInfo &FInfo) {
1627 int numArgs = (int)aCI->getNumArgOperands();
1628 if (numArgs > 3)
26
Assuming 'numArgs' is <= 3
27
Taking false branch
1629 return false;
1630
1631 Constant *copr0 = nullptr;
1632 Constant *copr1 = nullptr;
1633 Constant *copr2 = nullptr;
1634 if (numArgs > 0) {
28
Assuming 'numArgs' is <= 0
29
Taking false branch
1635 if ((copr0 = dyn_cast<Constant>(aCI->getArgOperand(0))) == nullptr)
1636 return false;
1637 }
1638
1639 if (numArgs
29.1
'numArgs' is <= 1
> 1) {
30
Taking false branch
1640 if ((copr1 = dyn_cast<Constant>(aCI->getArgOperand(1))) == nullptr) {
1641 if (FInfo.getId() != AMDGPULibFunc::EI_SINCOS)
1642 return false;
1643 }
1644 }
1645
1646 if (numArgs
30.1
'numArgs' is <= 2
> 2) {
31
Taking false branch
1647 if ((copr2 = dyn_cast<Constant>(aCI->getArgOperand(2))) == nullptr)
1648 return false;
1649 }
1650
1651 // At this point, all arguments to aCI are constants.
1652
1653 // max vector size is 16, and sincos will generate two results.
1654 double DVal0[16], DVal1[16];
1655 bool hasTwoResults = (FInfo.getId() == AMDGPULibFunc::EI_SINCOS);
32
Assuming the condition is false
1656 if (getVecSize(FInfo) == 1) {
33
Assuming the condition is false
34
Taking false branch
1657 if (!evaluateScalarMathFunc(FInfo, DVal0[0],
1658 DVal1[0], copr0, copr1, copr2)) {
1659 return false;
1660 }
1661 } else {
1662 ConstantDataVector *CDV0 = dyn_cast_or_null<ConstantDataVector>(copr0);
35
Assuming null pointer is passed into cast
1663 ConstantDataVector *CDV1 = dyn_cast_or_null<ConstantDataVector>(copr1);
36
Assuming null pointer is passed into cast
1664 ConstantDataVector *CDV2 = dyn_cast_or_null<ConstantDataVector>(copr2);
37
Assuming null pointer is passed into cast
1665 for (int i=0; i < getVecSize(FInfo); ++i) {
38
Assuming the condition is false
39
Loop condition is false. Execution continues on line 1676
1666 Constant *celt0 = CDV0 ? CDV0->getElementAsConstant(i) : nullptr;
1667 Constant *celt1 = CDV1 ? CDV1->getElementAsConstant(i) : nullptr;
1668 Constant *celt2 = CDV2 ? CDV2->getElementAsConstant(i) : nullptr;
1669 if (!evaluateScalarMathFunc(FInfo, DVal0[i],
1670 DVal1[i], celt0, celt1, celt2)) {
1671 return false;
1672 }
1673 }
1674 }
1675
1676 LLVMContext &context = CI->getParent()->getParent()->getContext();
1677 Constant *nval0, *nval1;
1678 if (getVecSize(FInfo) == 1) {
40
Assuming the condition is true
41
Taking true branch
1679 nval0 = ConstantFP::get(CI->getType(), DVal0[0]);
42
2nd function call argument is an uninitialized value
1680 if (hasTwoResults)
1681 nval1 = ConstantFP::get(CI->getType(), DVal1[0]);
1682 } else {
1683 if (getArgType(FInfo) == AMDGPULibFunc::F32) {
1684 SmallVector <float, 0> FVal0, FVal1;
1685 for (int i=0; i < getVecSize(FInfo); ++i)
1686 FVal0.push_back((float)DVal0[i]);
1687 ArrayRef<float> tmp0(FVal0);
1688 nval0 = ConstantDataVector::get(context, tmp0);
1689 if (hasTwoResults) {
1690 for (int i=0; i < getVecSize(FInfo); ++i)
1691 FVal1.push_back((float)DVal1[i]);
1692 ArrayRef<float> tmp1(FVal1);
1693 nval1 = ConstantDataVector::get(context, tmp1);
1694 }
1695 } else {
1696 ArrayRef<double> tmp0(DVal0);
1697 nval0 = ConstantDataVector::get(context, tmp0);
1698 if (hasTwoResults) {
1699 ArrayRef<double> tmp1(DVal1);
1700 nval1 = ConstantDataVector::get(context, tmp1);
1701 }
1702 }
1703 }
1704
1705 if (hasTwoResults) {
1706 // sincos
1707 assert(FInfo.getId() == AMDGPULibFunc::EI_SINCOS &&((FInfo.getId() == AMDGPULibFunc::EI_SINCOS && "math function with ptr arg not supported yet"
) ? static_cast<void> (0) : __assert_fail ("FInfo.getId() == AMDGPULibFunc::EI_SINCOS && \"math function with ptr arg not supported yet\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1708, __PRETTY_FUNCTION__))
1708 "math function with ptr arg not supported yet")((FInfo.getId() == AMDGPULibFunc::EI_SINCOS && "math function with ptr arg not supported yet"
) ? static_cast<void> (0) : __assert_fail ("FInfo.getId() == AMDGPULibFunc::EI_SINCOS && \"math function with ptr arg not supported yet\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Target/AMDGPU/AMDGPULibCalls.cpp"
, 1708, __PRETTY_FUNCTION__))
;
1709 new StoreInst(nval1, aCI->getArgOperand(1), aCI);
1710 }
1711
1712 replaceCall(nval0);
1713 return true;
1714}
1715
1716// Public interface to the Simplify LibCalls pass.
1717FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetOptions &Opt,
1718 const TargetMachine *TM) {
1719 return new AMDGPUSimplifyLibCalls(Opt, TM);
1720}
1721
1722FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
1723 return new AMDGPUUseNativeCalls();
1724}
1725
1726static bool setFastFlags(Function &F, const TargetOptions &Options) {
1727 AttrBuilder B;
1728
1729 if (Options.UnsafeFPMath || Options.NoInfsFPMath)
1730 B.addAttribute("no-infs-fp-math", "true");
1731 if (Options.UnsafeFPMath || Options.NoNaNsFPMath)
1732 B.addAttribute("no-nans-fp-math", "true");
1733 if (Options.UnsafeFPMath) {
1734 B.addAttribute("less-precise-fpmad", "true");
1735 B.addAttribute("unsafe-fp-math", "true");
1736 }
1737
1738 if (!B.hasAttributes())
1739 return false;
1740
1741 F.addAttributes(AttributeList::FunctionIndex, B);
1742
1743 return true;
1744}
1745
1746bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
1747 if (skipFunction(F))
1
Assuming the condition is false
2
Taking false branch
1748 return false;
1749
1750 bool Changed = false;
1751 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1752
1753 LLVM_DEBUG(dbgs() << "AMDIC: process function ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { dbgs() << "AMDIC: process function "
; F.printAsOperand(dbgs(), false, F.getParent()); dbgs() <<
'\n';; } } while (false)
3
Assuming 'DebugFlag' is false
4
Loop condition is false. Exiting loop
1754 F.printAsOperand(dbgs(), false, F.getParent()); dbgs() << '\n';)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { dbgs() << "AMDIC: process function "
; F.printAsOperand(dbgs(), false, F.getParent()); dbgs() <<
'\n';; } } while (false)
;
1755
1756 if (!EnablePreLink)
5
Assuming the condition is false
6
Taking false branch
1757 Changed |= setFastFlags(F, Options);
1758
1759 for (auto &BB : F) {
1760 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
7
Loop condition is true. Entering loop body
1761 // Ignore non-calls.
1762 CallInst *CI = dyn_cast<CallInst>(I);
1763 ++I;
1764 if (!CI) continue;
8
Assuming 'CI' is non-null
9
Taking false branch
1765
1766 // Ignore indirect calls.
1767 Function *Callee = CI->getCalledFunction();
1768 if (Callee
9.1
'Callee' is not equal to null
== 0) continue;
10
Taking false branch
1769
1770 LLVM_DEBUG(dbgs() << "AMDIC: try folding " << *CI << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { dbgs() << "AMDIC: try folding "
<< *CI << "\n"; dbgs().flush(); } } while (false
)
11
Assuming 'DebugFlag' is false
12
Loop condition is false. Exiting loop
1771 dbgs().flush())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("amdgpu-simplifylib")) { dbgs() << "AMDIC: try folding "
<< *CI << "\n"; dbgs().flush(); } } while (false
)
;
1772 if(Simplifier.fold(CI, AA))
13
Calling 'AMDGPULibCalls::fold'
1773 Changed = true;
1774 }
1775 }
1776 return Changed;
1777}
1778
1779bool AMDGPUUseNativeCalls::runOnFunction(Function &F) {
1780 if (skipFunction(F) || UseNative.empty())
1781 return false;
1782
1783 bool Changed = false;
1784 for (auto &BB : F) {
1785 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
1786 // Ignore non-calls.
1787 CallInst *CI = dyn_cast<CallInst>(I);
1788 ++I;
1789 if (!CI) continue;
1790
1791 // Ignore indirect calls.
1792 Function *Callee = CI->getCalledFunction();
1793 if (Callee == 0) continue;
1794
1795 if(Simplifier.useNative(CI))
1796 Changed = true;
1797 }
1798 }
1799 return Changed;
1800}