/home/runner/work/DirectXShaderCompiler/DirectXShaderCompiler/lib/Transforms/Utils/SimplifyLibCalls.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===------ SimplifyLibCalls.cpp - Library calls simplifier ---------------===// |
2 | | // |
3 | | // The LLVM Compiler Infrastructure |
4 | | // |
5 | | // This file is distributed under the University of Illinois Open Source |
6 | | // License. See LICENSE.TXT for details. |
7 | | // |
8 | | //===----------------------------------------------------------------------===// |
9 | | // |
10 | | // This is a utility pass used for testing the InstructionSimplify analysis. |
11 | | // The analysis is applied to every instruction, and if it simplifies then the |
12 | | // instruction is replaced by the simplification. If you are looking for a pass |
13 | | // that performs serious instruction folding, use the instcombine pass instead. |
14 | | // |
15 | | //===----------------------------------------------------------------------===// |
16 | | |
17 | | #include "llvm/Transforms/Utils/SimplifyLibCalls.h" |
18 | | #include "llvm/ADT/SmallString.h" |
19 | | #include "llvm/ADT/StringMap.h" |
20 | | #include "llvm/ADT/Triple.h" |
21 | | #include "llvm/Analysis/ValueTracking.h" |
22 | | #include "llvm/IR/DataLayout.h" |
23 | | #include "llvm/IR/DiagnosticInfo.h" |
24 | | #include "llvm/IR/Function.h" |
25 | | #include "llvm/IR/IRBuilder.h" |
26 | | #include "llvm/IR/IntrinsicInst.h" |
27 | | #include "llvm/IR/Intrinsics.h" |
28 | | #include "llvm/IR/LLVMContext.h" |
29 | | #include "llvm/IR/Module.h" |
30 | | #include "llvm/IR/PatternMatch.h" |
31 | | #include "llvm/Support/Allocator.h" |
32 | | #include "llvm/Support/CommandLine.h" |
33 | | #include "llvm/Analysis/TargetLibraryInfo.h" |
34 | | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
35 | | |
36 | | using namespace llvm; |
37 | | using namespace PatternMatch; |
38 | | |
39 | | #if 0 // HLSL Change Starts - option pending |
40 | | static cl::opt<bool> |
41 | | ColdErrorCalls("error-reporting-is-cold", cl::init(true), cl::Hidden, |
42 | | cl::desc("Treat error-reporting calls as cold")); |
43 | | |
44 | | static cl::opt<bool> |
45 | | EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden, |
46 | | cl::init(false), |
47 | | cl::desc("Enable unsafe double to float " |
48 | | "shrinking for math lib calls")); |
49 | | #else |
50 | | static const bool ColdErrorCalls = true; |
51 | | static const bool EnableUnsafeFPShrink = false; |
52 | | #endif // HLSL Change Ends |
53 | | |
54 | | //===----------------------------------------------------------------------===// |
55 | | // Helper Functions |
56 | | //===----------------------------------------------------------------------===// |
57 | | |
58 | 0 | static bool ignoreCallingConv(LibFunc::Func Func) { |
59 | 0 | switch (Func) { |
60 | 0 | case LibFunc::abs: |
61 | 0 | case LibFunc::labs: |
62 | 0 | case LibFunc::llabs: |
63 | 0 | case LibFunc::strlen: |
64 | 0 | return true; |
65 | 0 | default: |
66 | 0 | return false; |
67 | 0 | } |
68 | 0 | llvm_unreachable("All cases should be covered in the switch."); |
69 | 0 | } |
70 | | |
71 | | /// isOnlyUsedInZeroEqualityComparison - Return true if it only matters that the |
72 | | /// value is equal or not-equal to zero. |
73 | 0 | static bool isOnlyUsedInZeroEqualityComparison(Value *V) { |
74 | 0 | for (User *U : V->users()) { |
75 | 0 | if (ICmpInst *IC = dyn_cast<ICmpInst>(U)) |
76 | 0 | if (IC->isEquality()) |
77 | 0 | if (Constant *C = dyn_cast<Constant>(IC->getOperand(1))) |
78 | 0 | if (C->isNullValue()) |
79 | 0 | continue; |
80 | | // Unknown instruction. |
81 | 0 | return false; |
82 | 0 | } |
83 | 0 | return true; |
84 | 0 | } |
85 | | |
86 | | /// isOnlyUsedInEqualityComparison - Return true if it is only used in equality |
87 | | /// comparisons with With. |
88 | 0 | static bool isOnlyUsedInEqualityComparison(Value *V, Value *With) { |
89 | 0 | for (User *U : V->users()) { |
90 | 0 | if (ICmpInst *IC = dyn_cast<ICmpInst>(U)) |
91 | 0 | if (IC->isEquality() && IC->getOperand(1) == With) |
92 | 0 | continue; |
93 | | // Unknown instruction. |
94 | 0 | return false; |
95 | 0 | } |
96 | 0 | return true; |
97 | 0 | } |
98 | | |
99 | 0 | static bool callHasFloatingPointArgument(const CallInst *CI) { |
100 | 0 | for (CallInst::const_op_iterator it = CI->op_begin(), e = CI->op_end(); |
101 | 0 | it != e; ++it) { |
102 | 0 | if ((*it)->getType()->isFloatingPointTy()) |
103 | 0 | return true; |
104 | 0 | } |
105 | 0 | return false; |
106 | 0 | } |
107 | | |
108 | | /// \brief Check whether the overloaded unary floating point function |
109 | | /// corresponing to \a Ty is available. |
110 | | static bool hasUnaryFloatFn(const TargetLibraryInfo *TLI, Type *Ty, |
111 | | LibFunc::Func DoubleFn, LibFunc::Func FloatFn, |
112 | 0 | LibFunc::Func LongDoubleFn) { |
113 | 0 | switch (Ty->getTypeID()) { |
114 | 0 | case Type::FloatTyID: |
115 | 0 | return TLI->has(FloatFn); |
116 | 0 | case Type::DoubleTyID: |
117 | 0 | return TLI->has(DoubleFn); |
118 | 0 | default: |
119 | 0 | return TLI->has(LongDoubleFn); |
120 | 0 | } |
121 | 0 | } |
122 | | |
123 | | /// \brief Returns whether \p F matches the signature expected for the |
124 | | /// string/memory copying library function \p Func. |
125 | | /// Acceptable functions are st[rp][n]?cpy, memove, memcpy, and memset. |
126 | | /// Their fortified (_chk) counterparts are also accepted. |
127 | 0 | static bool checkStringCopyLibFuncSignature(Function *F, LibFunc::Func Func) { |
128 | 0 | const DataLayout &DL = F->getParent()->getDataLayout(); |
129 | 0 | FunctionType *FT = F->getFunctionType(); |
130 | 0 | LLVMContext &Context = F->getContext(); |
131 | 0 | Type *PCharTy = Type::getInt8PtrTy(Context); |
132 | 0 | Type *SizeTTy = DL.getIntPtrType(Context); |
133 | 0 | unsigned NumParams = FT->getNumParams(); |
134 | | |
135 | | // All string libfuncs return the same type as the first parameter. |
136 | 0 | if (FT->getReturnType() != FT->getParamType(0)) |
137 | 0 | return false; |
138 | | |
139 | 0 | switch (Func) { |
140 | 0 | default: |
141 | 0 | llvm_unreachable("Can't check signature for non-string-copy libfunc."); |
142 | 0 | case LibFunc::stpncpy_chk: |
143 | 0 | case LibFunc::strncpy_chk: |
144 | 0 | --NumParams; LLVM_FALLTHROUGH; // HLSL Change |
145 | 0 | case LibFunc::stpncpy: |
146 | 0 | case LibFunc::strncpy: { |
147 | 0 | if (NumParams != 3 || FT->getParamType(0) != FT->getParamType(1) || |
148 | 0 | FT->getParamType(0) != PCharTy || !FT->getParamType(2)->isIntegerTy()) |
149 | 0 | return false; |
150 | 0 | break; |
151 | 0 | } |
152 | 0 | case LibFunc::strcpy_chk: |
153 | 0 | case LibFunc::stpcpy_chk: |
154 | 0 | --NumParams; LLVM_FALLTHROUGH; // HLSL Change |
155 | 0 | case LibFunc::stpcpy: |
156 | 0 | case LibFunc::strcpy: { |
157 | 0 | if (NumParams != 2 || FT->getParamType(0) != FT->getParamType(1) || |
158 | 0 | FT->getParamType(0) != PCharTy) |
159 | 0 | return false; |
160 | 0 | break; |
161 | 0 | } |
162 | 0 | case LibFunc::memmove_chk: |
163 | 0 | case LibFunc::memcpy_chk: |
164 | 0 | --NumParams; LLVM_FALLTHROUGH; // HLSL Change |
165 | 0 | case LibFunc::memmove: |
166 | 0 | case LibFunc::memcpy: { |
167 | 0 | if (NumParams != 3 || !FT->getParamType(0)->isPointerTy() || |
168 | 0 | !FT->getParamType(1)->isPointerTy() || FT->getParamType(2) != SizeTTy) |
169 | 0 | return false; |
170 | 0 | break; |
171 | 0 | } |
172 | 0 | case LibFunc::memset_chk: |
173 | 0 | --NumParams; LLVM_FALLTHROUGH; // HLSL Change |
174 | 0 | case LibFunc::memset: { |
175 | 0 | if (NumParams != 3 || !FT->getParamType(0)->isPointerTy() || |
176 | 0 | !FT->getParamType(1)->isIntegerTy() || FT->getParamType(2) != SizeTTy) |
177 | 0 | return false; |
178 | 0 | break; |
179 | 0 | } |
180 | 0 | } |
181 | | // If this is a fortified libcall, the last parameter is a size_t. |
182 | 0 | if (NumParams == FT->getNumParams() - 1) |
183 | 0 | return FT->getParamType(FT->getNumParams() - 1) == SizeTTy; |
184 | 0 | return true; |
185 | 0 | } |
186 | | |
187 | | //===----------------------------------------------------------------------===// |
188 | | // String and Memory Library Call Optimizations |
189 | | //===----------------------------------------------------------------------===// |
190 | | |
191 | 0 | Value *LibCallSimplifier::optimizeStrCat(CallInst *CI, IRBuilder<> &B) { |
192 | 0 | Function *Callee = CI->getCalledFunction(); |
193 | | // Verify the "strcat" function prototype. |
194 | 0 | FunctionType *FT = Callee->getFunctionType(); |
195 | 0 | if (FT->getNumParams() != 2|| |
196 | 0 | FT->getReturnType() != B.getInt8PtrTy() || |
197 | 0 | FT->getParamType(0) != FT->getReturnType() || |
198 | 0 | FT->getParamType(1) != FT->getReturnType()) |
199 | 0 | return nullptr; |
200 | | |
201 | | // Extract some information from the instruction |
202 | 0 | Value *Dst = CI->getArgOperand(0); |
203 | 0 | Value *Src = CI->getArgOperand(1); |
204 | | |
205 | | // See if we can get the length of the input string. |
206 | 0 | uint64_t Len = GetStringLength(Src); |
207 | 0 | if (Len == 0) |
208 | 0 | return nullptr; |
209 | 0 | --Len; // Unbias length. |
210 | | |
211 | | // Handle the simple, do-nothing case: strcat(x, "") -> x |
212 | 0 | if (Len == 0) |
213 | 0 | return Dst; |
214 | | |
215 | 0 | return emitStrLenMemCpy(Src, Dst, Len, B); |
216 | 0 | } |
217 | | |
218 | | Value *LibCallSimplifier::emitStrLenMemCpy(Value *Src, Value *Dst, uint64_t Len, |
219 | 0 | IRBuilder<> &B) { |
220 | | // We need to find the end of the destination string. That's where the |
221 | | // memory is to be moved to. We just generate a call to strlen. |
222 | 0 | Value *DstLen = EmitStrLen(Dst, B, DL, TLI); |
223 | 0 | if (!DstLen) |
224 | 0 | return nullptr; |
225 | | |
226 | | // Now that we have the destination's length, we must index into the |
227 | | // destination's pointer to get the actual memcpy destination (end of |
228 | | // the string .. we're concatenating). |
229 | 0 | Value *CpyDst = B.CreateGEP(B.getInt8Ty(), Dst, DstLen, "endptr"); |
230 | | |
231 | | // We have enough information to now generate the memcpy call to do the |
232 | | // concatenation for us. Make a memcpy to copy the nul byte with align = 1. |
233 | 0 | B.CreateMemCpy(CpyDst, Src, |
234 | 0 | ConstantInt::get(DL.getIntPtrType(Src->getContext()), Len + 1), |
235 | 0 | 1); |
236 | 0 | return Dst; |
237 | 0 | } |
238 | | |
239 | 0 | Value *LibCallSimplifier::optimizeStrNCat(CallInst *CI, IRBuilder<> &B) { |
240 | 0 | Function *Callee = CI->getCalledFunction(); |
241 | | // Verify the "strncat" function prototype. |
242 | 0 | FunctionType *FT = Callee->getFunctionType(); |
243 | 0 | if (FT->getNumParams() != 3 || FT->getReturnType() != B.getInt8PtrTy() || |
244 | 0 | FT->getParamType(0) != FT->getReturnType() || |
245 | 0 | FT->getParamType(1) != FT->getReturnType() || |
246 | 0 | !FT->getParamType(2)->isIntegerTy()) |
247 | 0 | return nullptr; |
248 | | |
249 | | // Extract some information from the instruction |
250 | 0 | Value *Dst = CI->getArgOperand(0); |
251 | 0 | Value *Src = CI->getArgOperand(1); |
252 | 0 | uint64_t Len; |
253 | | |
254 | | // We don't do anything if length is not constant |
255 | 0 | if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2))) |
256 | 0 | Len = LengthArg->getZExtValue(); |
257 | 0 | else |
258 | 0 | return nullptr; |
259 | | |
260 | | // See if we can get the length of the input string. |
261 | 0 | uint64_t SrcLen = GetStringLength(Src); |
262 | 0 | if (SrcLen == 0) |
263 | 0 | return nullptr; |
264 | 0 | --SrcLen; // Unbias length. |
265 | | |
266 | | // Handle the simple, do-nothing cases: |
267 | | // strncat(x, "", c) -> x |
268 | | // strncat(x, c, 0) -> x |
269 | 0 | if (SrcLen == 0 || Len == 0) |
270 | 0 | return Dst; |
271 | | |
272 | | // We don't optimize this case |
273 | 0 | if (Len < SrcLen) |
274 | 0 | return nullptr; |
275 | | |
276 | | // strncat(x, s, c) -> strcat(x, s) |
277 | | // s is constant so the strcat can be optimized further |
278 | 0 | return emitStrLenMemCpy(Src, Dst, SrcLen, B); |
279 | 0 | } |
280 | | |
281 | 0 | Value *LibCallSimplifier::optimizeStrChr(CallInst *CI, IRBuilder<> &B) { |
282 | 0 | Function *Callee = CI->getCalledFunction(); |
283 | | // Verify the "strchr" function prototype. |
284 | 0 | FunctionType *FT = Callee->getFunctionType(); |
285 | 0 | if (FT->getNumParams() != 2 || FT->getReturnType() != B.getInt8PtrTy() || |
286 | 0 | FT->getParamType(0) != FT->getReturnType() || |
287 | 0 | !FT->getParamType(1)->isIntegerTy(32)) |
288 | 0 | return nullptr; |
289 | | |
290 | 0 | Value *SrcStr = CI->getArgOperand(0); |
291 | | |
292 | | // If the second operand is non-constant, see if we can compute the length |
293 | | // of the input string and turn this into memchr. |
294 | 0 | ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); |
295 | 0 | if (!CharC) { |
296 | 0 | uint64_t Len = GetStringLength(SrcStr); |
297 | 0 | if (Len == 0 || !FT->getParamType(1)->isIntegerTy(32)) // memchr needs i32. |
298 | 0 | return nullptr; |
299 | | |
300 | 0 | return EmitMemChr(SrcStr, CI->getArgOperand(1), // include nul. |
301 | 0 | ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), |
302 | 0 | B, DL, TLI); |
303 | 0 | } |
304 | | |
305 | | // Otherwise, the character is a constant, see if the first argument is |
306 | | // a string literal. If so, we can constant fold. |
307 | 0 | StringRef Str; |
308 | 0 | if (!getConstantStringInfo(SrcStr, Str)) { |
309 | 0 | if (CharC->isZero()) // strchr(p, 0) -> p + strlen(p) |
310 | 0 | return B.CreateGEP(B.getInt8Ty(), SrcStr, EmitStrLen(SrcStr, B, DL, TLI), "strchr"); |
311 | 0 | return nullptr; |
312 | 0 | } |
313 | | |
314 | | // Compute the offset, make sure to handle the case when we're searching for |
315 | | // zero (a weird way to spell strlen). |
316 | 0 | size_t I = (0xFF & CharC->getSExtValue()) == 0 |
317 | 0 | ? Str.size() |
318 | 0 | : Str.find(CharC->getSExtValue()); |
319 | 0 | if (I == StringRef::npos) // Didn't find the char. strchr returns null. |
320 | 0 | return Constant::getNullValue(CI->getType()); |
321 | | |
322 | | // strchr(s+n,c) -> gep(s+n+i,c) |
323 | 0 | return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strchr"); |
324 | 0 | } |
325 | | |
326 | 0 | Value *LibCallSimplifier::optimizeStrRChr(CallInst *CI, IRBuilder<> &B) { |
327 | 0 | Function *Callee = CI->getCalledFunction(); |
328 | | // Verify the "strrchr" function prototype. |
329 | 0 | FunctionType *FT = Callee->getFunctionType(); |
330 | 0 | if (FT->getNumParams() != 2 || FT->getReturnType() != B.getInt8PtrTy() || |
331 | 0 | FT->getParamType(0) != FT->getReturnType() || |
332 | 0 | !FT->getParamType(1)->isIntegerTy(32)) |
333 | 0 | return nullptr; |
334 | | |
335 | 0 | Value *SrcStr = CI->getArgOperand(0); |
336 | 0 | ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); |
337 | | |
338 | | // Cannot fold anything if we're not looking for a constant. |
339 | 0 | if (!CharC) |
340 | 0 | return nullptr; |
341 | | |
342 | 0 | StringRef Str; |
343 | 0 | if (!getConstantStringInfo(SrcStr, Str)) { |
344 | | // strrchr(s, 0) -> strchr(s, 0) |
345 | 0 | if (CharC->isZero()) |
346 | 0 | return EmitStrChr(SrcStr, '\0', B, TLI); |
347 | 0 | return nullptr; |
348 | 0 | } |
349 | | |
350 | | // Compute the offset. |
351 | 0 | size_t I = (0xFF & CharC->getSExtValue()) == 0 |
352 | 0 | ? Str.size() |
353 | 0 | : Str.rfind(CharC->getSExtValue()); |
354 | 0 | if (I == StringRef::npos) // Didn't find the char. Return null. |
355 | 0 | return Constant::getNullValue(CI->getType()); |
356 | | |
357 | | // strrchr(s+n,c) -> gep(s+n+i,c) |
358 | 0 | return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "strrchr"); |
359 | 0 | } |
360 | | |
361 | 0 | Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilder<> &B) { |
362 | 0 | Function *Callee = CI->getCalledFunction(); |
363 | | // Verify the "strcmp" function prototype. |
364 | 0 | FunctionType *FT = Callee->getFunctionType(); |
365 | 0 | if (FT->getNumParams() != 2 || !FT->getReturnType()->isIntegerTy(32) || |
366 | 0 | FT->getParamType(0) != FT->getParamType(1) || |
367 | 0 | FT->getParamType(0) != B.getInt8PtrTy()) |
368 | 0 | return nullptr; |
369 | | |
370 | 0 | Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1); |
371 | 0 | if (Str1P == Str2P) // strcmp(x,x) -> 0 |
372 | 0 | return ConstantInt::get(CI->getType(), 0); |
373 | | |
374 | 0 | StringRef Str1, Str2; |
375 | 0 | bool HasStr1 = getConstantStringInfo(Str1P, Str1); |
376 | 0 | bool HasStr2 = getConstantStringInfo(Str2P, Str2); |
377 | | |
378 | | // strcmp(x, y) -> cnst (if both x and y are constant strings) |
379 | 0 | if (HasStr1 && HasStr2) |
380 | 0 | return ConstantInt::get(CI->getType(), Str1.compare(Str2)); |
381 | | |
382 | 0 | if (HasStr1 && Str1.empty()) // strcmp("", x) -> -*x |
383 | 0 | return B.CreateNeg( |
384 | 0 | B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType())); |
385 | | |
386 | 0 | if (HasStr2 && Str2.empty()) // strcmp(x,"") -> *x |
387 | 0 | return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType()); |
388 | | |
389 | | // strcmp(P, "x") -> memcmp(P, "x", 2) |
390 | 0 | uint64_t Len1 = GetStringLength(Str1P); |
391 | 0 | uint64_t Len2 = GetStringLength(Str2P); |
392 | 0 | if (Len1 && Len2) { |
393 | 0 | return EmitMemCmp(Str1P, Str2P, |
394 | 0 | ConstantInt::get(DL.getIntPtrType(CI->getContext()), |
395 | 0 | std::min(Len1, Len2)), |
396 | 0 | B, DL, TLI); |
397 | 0 | } |
398 | | |
399 | 0 | return nullptr; |
400 | 0 | } |
401 | | |
402 | 0 | Value *LibCallSimplifier::optimizeStrNCmp(CallInst *CI, IRBuilder<> &B) { |
403 | 0 | Function *Callee = CI->getCalledFunction(); |
404 | | // Verify the "strncmp" function prototype. |
405 | 0 | FunctionType *FT = Callee->getFunctionType(); |
406 | 0 | if (FT->getNumParams() != 3 || !FT->getReturnType()->isIntegerTy(32) || |
407 | 0 | FT->getParamType(0) != FT->getParamType(1) || |
408 | 0 | FT->getParamType(0) != B.getInt8PtrTy() || |
409 | 0 | !FT->getParamType(2)->isIntegerTy()) |
410 | 0 | return nullptr; |
411 | | |
412 | 0 | Value *Str1P = CI->getArgOperand(0), *Str2P = CI->getArgOperand(1); |
413 | 0 | if (Str1P == Str2P) // strncmp(x,x,n) -> 0 |
414 | 0 | return ConstantInt::get(CI->getType(), 0); |
415 | | |
416 | | // Get the length argument if it is constant. |
417 | 0 | uint64_t Length; |
418 | 0 | if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(CI->getArgOperand(2))) |
419 | 0 | Length = LengthArg->getZExtValue(); |
420 | 0 | else |
421 | 0 | return nullptr; |
422 | | |
423 | 0 | if (Length == 0) // strncmp(x,y,0) -> 0 |
424 | 0 | return ConstantInt::get(CI->getType(), 0); |
425 | | |
426 | 0 | if (Length == 1) // strncmp(x,y,1) -> memcmp(x,y,1) |
427 | 0 | return EmitMemCmp(Str1P, Str2P, CI->getArgOperand(2), B, DL, TLI); |
428 | | |
429 | 0 | StringRef Str1, Str2; |
430 | 0 | bool HasStr1 = getConstantStringInfo(Str1P, Str1); |
431 | 0 | bool HasStr2 = getConstantStringInfo(Str2P, Str2); |
432 | | |
433 | | // strncmp(x, y) -> cnst (if both x and y are constant strings) |
434 | 0 | if (HasStr1 && HasStr2) { |
435 | 0 | StringRef SubStr1 = Str1.substr(0, Length); |
436 | 0 | StringRef SubStr2 = Str2.substr(0, Length); |
437 | 0 | return ConstantInt::get(CI->getType(), SubStr1.compare(SubStr2)); |
438 | 0 | } |
439 | | |
440 | 0 | if (HasStr1 && Str1.empty()) // strncmp("", x, n) -> -*x |
441 | 0 | return B.CreateNeg( |
442 | 0 | B.CreateZExt(B.CreateLoad(Str2P, "strcmpload"), CI->getType())); |
443 | | |
444 | 0 | if (HasStr2 && Str2.empty()) // strncmp(x, "", n) -> *x |
445 | 0 | return B.CreateZExt(B.CreateLoad(Str1P, "strcmpload"), CI->getType()); |
446 | | |
447 | 0 | return nullptr; |
448 | 0 | } |
449 | | |
450 | 0 | Value *LibCallSimplifier::optimizeStrCpy(CallInst *CI, IRBuilder<> &B) { |
451 | 0 | Function *Callee = CI->getCalledFunction(); |
452 | |
|
453 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::strcpy)) |
454 | 0 | return nullptr; |
455 | | |
456 | 0 | Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); |
457 | 0 | if (Dst == Src) // strcpy(x,x) -> x |
458 | 0 | return Src; |
459 | | |
460 | | // See if we can get the length of the input string. |
461 | 0 | uint64_t Len = GetStringLength(Src); |
462 | 0 | if (Len == 0) |
463 | 0 | return nullptr; |
464 | | |
465 | | // We have enough information to now generate the memcpy call to do the |
466 | | // copy for us. Make a memcpy to copy the nul byte with align = 1. |
467 | 0 | B.CreateMemCpy(Dst, Src, |
468 | 0 | ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len), 1); |
469 | 0 | return Dst; |
470 | 0 | } |
471 | | |
472 | 0 | Value *LibCallSimplifier::optimizeStpCpy(CallInst *CI, IRBuilder<> &B) { |
473 | 0 | Function *Callee = CI->getCalledFunction(); |
474 | | // Verify the "stpcpy" function prototype. |
475 | 0 | FunctionType *FT = Callee->getFunctionType(); |
476 | |
|
477 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::stpcpy)) |
478 | 0 | return nullptr; |
479 | | |
480 | 0 | Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1); |
481 | 0 | if (Dst == Src) { // stpcpy(x,x) -> x+strlen(x) |
482 | 0 | Value *StrLen = EmitStrLen(Src, B, DL, TLI); |
483 | 0 | return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr; |
484 | 0 | } |
485 | | |
486 | | // See if we can get the length of the input string. |
487 | 0 | uint64_t Len = GetStringLength(Src); |
488 | 0 | if (Len == 0) |
489 | 0 | return nullptr; |
490 | | |
491 | 0 | Type *PT = FT->getParamType(0); |
492 | 0 | Value *LenV = ConstantInt::get(DL.getIntPtrType(PT), Len); |
493 | 0 | Value *DstEnd = |
494 | 0 | B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(DL.getIntPtrType(PT), Len - 1)); |
495 | | |
496 | | // We have enough information to now generate the memcpy call to do the |
497 | | // copy for us. Make a memcpy to copy the nul byte with align = 1. |
498 | 0 | B.CreateMemCpy(Dst, Src, LenV, 1); |
499 | 0 | return DstEnd; |
500 | 0 | } |
501 | | |
502 | 0 | Value *LibCallSimplifier::optimizeStrNCpy(CallInst *CI, IRBuilder<> &B) { |
503 | 0 | Function *Callee = CI->getCalledFunction(); |
504 | 0 | FunctionType *FT = Callee->getFunctionType(); |
505 | |
|
506 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::strncpy)) |
507 | 0 | return nullptr; |
508 | | |
509 | 0 | Value *Dst = CI->getArgOperand(0); |
510 | 0 | Value *Src = CI->getArgOperand(1); |
511 | 0 | Value *LenOp = CI->getArgOperand(2); |
512 | | |
513 | | // See if we can get the length of the input string. |
514 | 0 | uint64_t SrcLen = GetStringLength(Src); |
515 | 0 | if (SrcLen == 0) |
516 | 0 | return nullptr; |
517 | 0 | --SrcLen; |
518 | |
|
519 | 0 | if (SrcLen == 0) { |
520 | | // strncpy(x, "", y) -> memset(x, '\0', y, 1) |
521 | 0 | B.CreateMemSet(Dst, B.getInt8('\0'), LenOp, 1); |
522 | 0 | return Dst; |
523 | 0 | } |
524 | | |
525 | 0 | uint64_t Len; |
526 | 0 | if (ConstantInt *LengthArg = dyn_cast<ConstantInt>(LenOp)) |
527 | 0 | Len = LengthArg->getZExtValue(); |
528 | 0 | else |
529 | 0 | return nullptr; |
530 | | |
531 | 0 | if (Len == 0) |
532 | 0 | return Dst; // strncpy(x, y, 0) -> x |
533 | | |
534 | | // Let strncpy handle the zero padding |
535 | 0 | if (Len > SrcLen + 1) |
536 | 0 | return nullptr; |
537 | | |
538 | 0 | Type *PT = FT->getParamType(0); |
539 | | // strncpy(x, s, c) -> memcpy(x, s, c, 1) [s and c are constant] |
540 | 0 | B.CreateMemCpy(Dst, Src, ConstantInt::get(DL.getIntPtrType(PT), Len), 1); |
541 | |
|
542 | 0 | return Dst; |
543 | 0 | } |
544 | | |
545 | 0 | Value *LibCallSimplifier::optimizeStrLen(CallInst *CI, IRBuilder<> &B) { |
546 | 0 | Function *Callee = CI->getCalledFunction(); |
547 | 0 | FunctionType *FT = Callee->getFunctionType(); |
548 | 0 | if (FT->getNumParams() != 1 || FT->getParamType(0) != B.getInt8PtrTy() || |
549 | 0 | !FT->getReturnType()->isIntegerTy()) |
550 | 0 | return nullptr; |
551 | | |
552 | 0 | Value *Src = CI->getArgOperand(0); |
553 | | |
554 | | // Constant folding: strlen("xyz") -> 3 |
555 | 0 | if (uint64_t Len = GetStringLength(Src)) |
556 | 0 | return ConstantInt::get(CI->getType(), Len - 1); |
557 | | |
558 | | // strlen(x?"foo":"bars") --> x ? 3 : 4 |
559 | 0 | if (SelectInst *SI = dyn_cast<SelectInst>(Src)) { |
560 | 0 | uint64_t LenTrue = GetStringLength(SI->getTrueValue()); |
561 | 0 | uint64_t LenFalse = GetStringLength(SI->getFalseValue()); |
562 | 0 | if (LenTrue && LenFalse) { |
563 | 0 | Function *Caller = CI->getParent()->getParent(); |
564 | 0 | emitOptimizationRemark(CI->getContext(), "simplify-libcalls", *Caller, |
565 | 0 | SI->getDebugLoc(), |
566 | 0 | "folded strlen(select) to select of constants"); |
567 | 0 | return B.CreateSelect(SI->getCondition(), |
568 | 0 | ConstantInt::get(CI->getType(), LenTrue - 1), |
569 | 0 | ConstantInt::get(CI->getType(), LenFalse - 1)); |
570 | 0 | } |
571 | 0 | } |
572 | | |
573 | | // strlen(x) != 0 --> *x != 0 |
574 | | // strlen(x) == 0 --> *x == 0 |
575 | 0 | if (isOnlyUsedInZeroEqualityComparison(CI)) |
576 | 0 | return B.CreateZExt(B.CreateLoad(Src, "strlenfirst"), CI->getType()); |
577 | | |
578 | 0 | return nullptr; |
579 | 0 | } |
580 | | |
581 | 0 | Value *LibCallSimplifier::optimizeStrPBrk(CallInst *CI, IRBuilder<> &B) { |
582 | 0 | Function *Callee = CI->getCalledFunction(); |
583 | 0 | FunctionType *FT = Callee->getFunctionType(); |
584 | 0 | if (FT->getNumParams() != 2 || FT->getParamType(0) != B.getInt8PtrTy() || |
585 | 0 | FT->getParamType(1) != FT->getParamType(0) || |
586 | 0 | FT->getReturnType() != FT->getParamType(0)) |
587 | 0 | return nullptr; |
588 | | |
589 | 0 | StringRef S1, S2; |
590 | 0 | bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); |
591 | 0 | bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); |
592 | | |
593 | | // strpbrk(s, "") -> nullptr |
594 | | // strpbrk("", s) -> nullptr |
595 | 0 | if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) |
596 | 0 | return Constant::getNullValue(CI->getType()); |
597 | | |
598 | | // Constant folding. |
599 | 0 | if (HasS1 && HasS2) { |
600 | 0 | size_t I = S1.find_first_of(S2); |
601 | 0 | if (I == StringRef::npos) // No match. |
602 | 0 | return Constant::getNullValue(CI->getType()); |
603 | | |
604 | 0 | return B.CreateGEP(B.getInt8Ty(), CI->getArgOperand(0), B.getInt64(I), "strpbrk"); |
605 | 0 | } |
606 | | |
607 | | // strpbrk(s, "a") -> strchr(s, 'a') |
608 | 0 | if (HasS2 && S2.size() == 1) |
609 | 0 | return EmitStrChr(CI->getArgOperand(0), S2[0], B, TLI); |
610 | | |
611 | 0 | return nullptr; |
612 | 0 | } |
613 | | |
614 | 0 | Value *LibCallSimplifier::optimizeStrTo(CallInst *CI, IRBuilder<> &B) { |
615 | 0 | Function *Callee = CI->getCalledFunction(); |
616 | 0 | FunctionType *FT = Callee->getFunctionType(); |
617 | 0 | if ((FT->getNumParams() != 2 && FT->getNumParams() != 3) || |
618 | 0 | !FT->getParamType(0)->isPointerTy() || |
619 | 0 | !FT->getParamType(1)->isPointerTy()) |
620 | 0 | return nullptr; |
621 | | |
622 | 0 | Value *EndPtr = CI->getArgOperand(1); |
623 | 0 | if (isa<ConstantPointerNull>(EndPtr)) { |
624 | | // With a null EndPtr, this function won't capture the main argument. |
625 | | // It would be readonly too, except that it still may write to errno. |
626 | 0 | CI->addAttribute(1, Attribute::NoCapture); |
627 | 0 | } |
628 | |
|
629 | 0 | return nullptr; |
630 | 0 | } |
631 | | |
632 | 0 | Value *LibCallSimplifier::optimizeStrSpn(CallInst *CI, IRBuilder<> &B) { |
633 | 0 | Function *Callee = CI->getCalledFunction(); |
634 | 0 | FunctionType *FT = Callee->getFunctionType(); |
635 | 0 | if (FT->getNumParams() != 2 || FT->getParamType(0) != B.getInt8PtrTy() || |
636 | 0 | FT->getParamType(1) != FT->getParamType(0) || |
637 | 0 | !FT->getReturnType()->isIntegerTy()) |
638 | 0 | return nullptr; |
639 | | |
640 | 0 | StringRef S1, S2; |
641 | 0 | bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); |
642 | 0 | bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); |
643 | | |
644 | | // strspn(s, "") -> 0 |
645 | | // strspn("", s) -> 0 |
646 | 0 | if ((HasS1 && S1.empty()) || (HasS2 && S2.empty())) |
647 | 0 | return Constant::getNullValue(CI->getType()); |
648 | | |
649 | | // Constant folding. |
650 | 0 | if (HasS1 && HasS2) { |
651 | 0 | size_t Pos = S1.find_first_not_of(S2); |
652 | 0 | if (Pos == StringRef::npos) |
653 | 0 | Pos = S1.size(); |
654 | 0 | return ConstantInt::get(CI->getType(), Pos); |
655 | 0 | } |
656 | | |
657 | 0 | return nullptr; |
658 | 0 | } |
659 | | |
660 | 0 | Value *LibCallSimplifier::optimizeStrCSpn(CallInst *CI, IRBuilder<> &B) { |
661 | 0 | Function *Callee = CI->getCalledFunction(); |
662 | 0 | FunctionType *FT = Callee->getFunctionType(); |
663 | 0 | if (FT->getNumParams() != 2 || FT->getParamType(0) != B.getInt8PtrTy() || |
664 | 0 | FT->getParamType(1) != FT->getParamType(0) || |
665 | 0 | !FT->getReturnType()->isIntegerTy()) |
666 | 0 | return nullptr; |
667 | | |
668 | 0 | StringRef S1, S2; |
669 | 0 | bool HasS1 = getConstantStringInfo(CI->getArgOperand(0), S1); |
670 | 0 | bool HasS2 = getConstantStringInfo(CI->getArgOperand(1), S2); |
671 | | |
672 | | // strcspn("", s) -> 0 |
673 | 0 | if (HasS1 && S1.empty()) |
674 | 0 | return Constant::getNullValue(CI->getType()); |
675 | | |
676 | | // Constant folding. |
677 | 0 | if (HasS1 && HasS2) { |
678 | 0 | size_t Pos = S1.find_first_of(S2); |
679 | 0 | if (Pos == StringRef::npos) |
680 | 0 | Pos = S1.size(); |
681 | 0 | return ConstantInt::get(CI->getType(), Pos); |
682 | 0 | } |
683 | | |
684 | | // strcspn(s, "") -> strlen(s) |
685 | 0 | if (HasS2 && S2.empty()) |
686 | 0 | return EmitStrLen(CI->getArgOperand(0), B, DL, TLI); |
687 | | |
688 | 0 | return nullptr; |
689 | 0 | } |
690 | | |
691 | 0 | Value *LibCallSimplifier::optimizeStrStr(CallInst *CI, IRBuilder<> &B) { |
692 | 0 | Function *Callee = CI->getCalledFunction(); |
693 | 0 | FunctionType *FT = Callee->getFunctionType(); |
694 | 0 | if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() || |
695 | 0 | !FT->getParamType(1)->isPointerTy() || |
696 | 0 | !FT->getReturnType()->isPointerTy()) |
697 | 0 | return nullptr; |
698 | | |
699 | | // fold strstr(x, x) -> x. |
700 | 0 | if (CI->getArgOperand(0) == CI->getArgOperand(1)) |
701 | 0 | return B.CreateBitCast(CI->getArgOperand(0), CI->getType()); |
702 | | |
703 | | // fold strstr(a, b) == a -> strncmp(a, b, strlen(b)) == 0 |
704 | 0 | if (isOnlyUsedInEqualityComparison(CI, CI->getArgOperand(0))) { |
705 | 0 | Value *StrLen = EmitStrLen(CI->getArgOperand(1), B, DL, TLI); |
706 | 0 | if (!StrLen) |
707 | 0 | return nullptr; |
708 | 0 | Value *StrNCmp = EmitStrNCmp(CI->getArgOperand(0), CI->getArgOperand(1), |
709 | 0 | StrLen, B, DL, TLI); |
710 | 0 | if (!StrNCmp) |
711 | 0 | return nullptr; |
712 | 0 | for (auto UI = CI->user_begin(), UE = CI->user_end(); UI != UE;) { |
713 | 0 | ICmpInst *Old = cast<ICmpInst>(*UI++); |
714 | 0 | Value *Cmp = |
715 | 0 | B.CreateICmp(Old->getPredicate(), StrNCmp, |
716 | 0 | ConstantInt::getNullValue(StrNCmp->getType()), "cmp"); |
717 | 0 | replaceAllUsesWith(Old, Cmp); |
718 | 0 | } |
719 | 0 | return CI; |
720 | 0 | } |
721 | | |
722 | | // See if either input string is a constant string. |
723 | 0 | StringRef SearchStr, ToFindStr; |
724 | 0 | bool HasStr1 = getConstantStringInfo(CI->getArgOperand(0), SearchStr); |
725 | 0 | bool HasStr2 = getConstantStringInfo(CI->getArgOperand(1), ToFindStr); |
726 | | |
727 | | // fold strstr(x, "") -> x. |
728 | 0 | if (HasStr2 && ToFindStr.empty()) |
729 | 0 | return B.CreateBitCast(CI->getArgOperand(0), CI->getType()); |
730 | | |
731 | | // If both strings are known, constant fold it. |
732 | 0 | if (HasStr1 && HasStr2) { |
733 | 0 | size_t Offset = SearchStr.find(ToFindStr); |
734 | |
|
735 | 0 | if (Offset == StringRef::npos) // strstr("foo", "bar") -> null |
736 | 0 | return Constant::getNullValue(CI->getType()); |
737 | | |
738 | | // strstr("abcd", "bc") -> gep((char*)"abcd", 1) |
739 | 0 | Value *Result = CastToCStr(CI->getArgOperand(0), B); |
740 | 0 | Result = B.CreateConstInBoundsGEP1_64(Result, Offset, "strstr"); |
741 | 0 | return B.CreateBitCast(Result, CI->getType()); |
742 | 0 | } |
743 | | |
744 | | // fold strstr(x, "y") -> strchr(x, 'y'). |
745 | 0 | if (HasStr2 && ToFindStr.size() == 1) { |
746 | 0 | Value *StrChr = EmitStrChr(CI->getArgOperand(0), ToFindStr[0], B, TLI); |
747 | 0 | return StrChr ? B.CreateBitCast(StrChr, CI->getType()) : nullptr; |
748 | 0 | } |
749 | 0 | return nullptr; |
750 | 0 | } |
751 | | |
752 | 0 | Value *LibCallSimplifier::optimizeMemChr(CallInst *CI, IRBuilder<> &B) { |
753 | 0 | Function *Callee = CI->getCalledFunction(); |
754 | 0 | FunctionType *FT = Callee->getFunctionType(); |
755 | 0 | if (FT->getNumParams() != 3 || !FT->getParamType(0)->isPointerTy() || |
756 | 0 | !FT->getParamType(1)->isIntegerTy(32) || |
757 | 0 | !FT->getParamType(2)->isIntegerTy() || |
758 | 0 | !FT->getReturnType()->isPointerTy()) |
759 | 0 | return nullptr; |
760 | | |
761 | 0 | Value *SrcStr = CI->getArgOperand(0); |
762 | 0 | ConstantInt *CharC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); |
763 | 0 | ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2)); |
764 | | |
765 | | // memchr(x, y, 0) -> null |
766 | 0 | if (LenC && LenC->isNullValue()) |
767 | 0 | return Constant::getNullValue(CI->getType()); |
768 | | |
769 | | // From now on we need at least constant length and string. |
770 | 0 | StringRef Str; |
771 | 0 | if (!LenC || !getConstantStringInfo(SrcStr, Str, 0, /*TrimAtNul=*/false)) |
772 | 0 | return nullptr; |
773 | | |
774 | | // Truncate the string to LenC. If Str is smaller than LenC we will still only |
775 | | // scan the string, as reading past the end of it is undefined and we can just |
776 | | // return null if we don't find the char. |
777 | 0 | Str = Str.substr(0, LenC->getZExtValue()); |
778 | | |
779 | | // If the char is variable but the input str and length are not we can turn |
780 | | // this memchr call into a simple bit field test. Of course this only works |
781 | | // when the return value is only checked against null. |
782 | | // |
783 | | // It would be really nice to reuse switch lowering here but we can't change |
784 | | // the CFG at this point. |
785 | | // |
786 | | // memchr("\r\n", C, 2) != nullptr -> (C & ((1 << '\r') | (1 << '\n'))) != 0 |
787 | | // after bounds check. |
788 | 0 | if (!CharC && !Str.empty() && isOnlyUsedInZeroEqualityComparison(CI)) { |
789 | 0 | unsigned char Max = |
790 | 0 | *std::max_element(reinterpret_cast<const unsigned char *>(Str.begin()), |
791 | 0 | reinterpret_cast<const unsigned char *>(Str.end())); |
792 | | |
793 | | // Make sure the bit field we're about to create fits in a register on the |
794 | | // target. |
795 | | // FIXME: On a 64 bit architecture this prevents us from using the |
796 | | // interesting range of alpha ascii chars. We could do better by emitting |
797 | | // two bitfields or shifting the range by 64 if no lower chars are used. |
798 | 0 | if (!DL.fitsInLegalInteger(Max + 1)) |
799 | 0 | return nullptr; |
800 | | |
801 | | // For the bit field use a power-of-2 type with at least 8 bits to avoid |
802 | | // creating unnecessary illegal types. |
803 | 0 | unsigned char Width = NextPowerOf2(std::max((unsigned char)7, Max)); |
804 | | |
805 | | // Now build the bit field. |
806 | 0 | APInt Bitfield(Width, 0); |
807 | 0 | for (char C : Str) |
808 | 0 | Bitfield.setBit((unsigned char)C); |
809 | 0 | Value *BitfieldC = B.getInt(Bitfield); |
810 | | |
811 | | // First check that the bit field access is within bounds. |
812 | 0 | Value *C = B.CreateZExtOrTrunc(CI->getArgOperand(1), BitfieldC->getType()); |
813 | 0 | Value *Bounds = B.CreateICmp(ICmpInst::ICMP_ULT, C, B.getIntN(Width, Width), |
814 | 0 | "memchr.bounds"); |
815 | | |
816 | | // Create code that checks if the given bit is set in the field. |
817 | 0 | Value *Shl = B.CreateShl(B.getIntN(Width, 1ULL), C); |
818 | 0 | Value *Bits = B.CreateIsNotNull(B.CreateAnd(Shl, BitfieldC), "memchr.bits"); |
819 | | |
820 | | // Finally merge both checks and cast to pointer type. The inttoptr |
821 | | // implicitly zexts the i1 to intptr type. |
822 | 0 | return B.CreateIntToPtr(B.CreateAnd(Bounds, Bits, "memchr"), CI->getType()); |
823 | 0 | } |
824 | | |
825 | | // Check if all arguments are constants. If so, we can constant fold. |
826 | 0 | if (!CharC) |
827 | 0 | return nullptr; |
828 | | |
829 | | // Compute the offset. |
830 | 0 | size_t I = Str.find(CharC->getSExtValue() & 0xFF); |
831 | 0 | if (I == StringRef::npos) // Didn't find the char. memchr returns null. |
832 | 0 | return Constant::getNullValue(CI->getType()); |
833 | | |
834 | | // memchr(s+n,c,l) -> gep(s+n+i,c) |
835 | 0 | return B.CreateGEP(B.getInt8Ty(), SrcStr, B.getInt64(I), "memchr"); |
836 | 0 | } |
837 | | |
838 | 0 | Value *LibCallSimplifier::optimizeMemCmp(CallInst *CI, IRBuilder<> &B) { |
839 | 0 | Function *Callee = CI->getCalledFunction(); |
840 | 0 | FunctionType *FT = Callee->getFunctionType(); |
841 | 0 | if (FT->getNumParams() != 3 || !FT->getParamType(0)->isPointerTy() || |
842 | 0 | !FT->getParamType(1)->isPointerTy() || |
843 | 0 | !FT->getReturnType()->isIntegerTy(32)) |
844 | 0 | return nullptr; |
845 | | |
846 | 0 | Value *LHS = CI->getArgOperand(0), *RHS = CI->getArgOperand(1); |
847 | |
|
848 | 0 | if (LHS == RHS) // memcmp(s,s,x) -> 0 |
849 | 0 | return Constant::getNullValue(CI->getType()); |
850 | | |
851 | | // Make sure we have a constant length. |
852 | 0 | ConstantInt *LenC = dyn_cast<ConstantInt>(CI->getArgOperand(2)); |
853 | 0 | if (!LenC) |
854 | 0 | return nullptr; |
855 | 0 | uint64_t Len = LenC->getZExtValue(); |
856 | |
|
857 | 0 | if (Len == 0) // memcmp(s1,s2,0) -> 0 |
858 | 0 | return Constant::getNullValue(CI->getType()); |
859 | | |
860 | | // memcmp(S1,S2,1) -> *(unsigned char*)LHS - *(unsigned char*)RHS |
861 | 0 | if (Len == 1) { |
862 | 0 | Value *LHSV = B.CreateZExt(B.CreateLoad(CastToCStr(LHS, B), "lhsc"), |
863 | 0 | CI->getType(), "lhsv"); |
864 | 0 | Value *RHSV = B.CreateZExt(B.CreateLoad(CastToCStr(RHS, B), "rhsc"), |
865 | 0 | CI->getType(), "rhsv"); |
866 | 0 | return B.CreateSub(LHSV, RHSV, "chardiff"); |
867 | 0 | } |
868 | | |
869 | | // Constant folding: memcmp(x, y, l) -> cnst (all arguments are constant) |
870 | 0 | StringRef LHSStr, RHSStr; |
871 | 0 | if (getConstantStringInfo(LHS, LHSStr) && |
872 | 0 | getConstantStringInfo(RHS, RHSStr)) { |
873 | | // Make sure we're not reading out-of-bounds memory. |
874 | 0 | if (Len > LHSStr.size() || Len > RHSStr.size()) |
875 | 0 | return nullptr; |
876 | | // Fold the memcmp and normalize the result. This way we get consistent |
877 | | // results across multiple platforms. |
878 | 0 | uint64_t Ret = 0; |
879 | 0 | int Cmp = memcmp(LHSStr.data(), RHSStr.data(), Len); |
880 | 0 | if (Cmp < 0) |
881 | 0 | Ret = -1; |
882 | 0 | else if (Cmp > 0) |
883 | 0 | Ret = 1; |
884 | 0 | return ConstantInt::get(CI->getType(), Ret); |
885 | 0 | } |
886 | | |
887 | 0 | return nullptr; |
888 | 0 | } |
889 | | |
890 | 0 | Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B) { |
891 | 0 | Function *Callee = CI->getCalledFunction(); |
892 | |
|
893 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::memcpy)) |
894 | 0 | return nullptr; |
895 | | |
896 | | // memcpy(x, y, n) -> llvm.memcpy(x, y, n, 1) |
897 | 0 | B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1), |
898 | 0 | CI->getArgOperand(2), 1); |
899 | 0 | return CI->getArgOperand(0); |
900 | 0 | } |
901 | | |
902 | 0 | Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B) { |
903 | 0 | Function *Callee = CI->getCalledFunction(); |
904 | |
|
905 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::memmove)) |
906 | 0 | return nullptr; |
907 | | |
908 | | // memmove(x, y, n) -> llvm.memmove(x, y, n, 1) |
909 | 0 | B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1), |
910 | 0 | CI->getArgOperand(2), 1); |
911 | 0 | return CI->getArgOperand(0); |
912 | 0 | } |
913 | | |
914 | 0 | Value *LibCallSimplifier::optimizeMemSet(CallInst *CI, IRBuilder<> &B) { |
915 | 0 | Function *Callee = CI->getCalledFunction(); |
916 | |
|
917 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::memset)) |
918 | 0 | return nullptr; |
919 | | |
920 | | // memset(p, v, n) -> llvm.memset(p, v, n, 1) |
921 | 0 | Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false); |
922 | 0 | B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1); |
923 | 0 | return CI->getArgOperand(0); |
924 | 0 | } |
925 | | |
926 | | //===----------------------------------------------------------------------===// |
927 | | // Math Library Optimizations |
928 | | //===----------------------------------------------------------------------===// |
929 | | |
930 | | /// Return a variant of Val with float type. |
931 | | /// Currently this works in two cases: If Val is an FPExtension of a float |
932 | | /// value to something bigger, simply return the operand. |
933 | | /// If Val is a ConstantFP but can be converted to a float ConstantFP without |
934 | | /// loss of precision do so. |
935 | 0 | static Value *valueHasFloatPrecision(Value *Val) { |
936 | 0 | if (FPExtInst *Cast = dyn_cast<FPExtInst>(Val)) { |
937 | 0 | Value *Op = Cast->getOperand(0); |
938 | 0 | if (Op->getType()->isFloatTy()) |
939 | 0 | return Op; |
940 | 0 | } |
941 | 0 | if (ConstantFP *Const = dyn_cast<ConstantFP>(Val)) { |
942 | 0 | APFloat F = Const->getValueAPF(); |
943 | 0 | bool losesInfo; |
944 | 0 | (void)F.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, |
945 | 0 | &losesInfo); |
946 | 0 | if (!losesInfo) |
947 | 0 | return ConstantFP::get(Const->getContext(), F); |
948 | 0 | } |
949 | 0 | return nullptr; |
950 | 0 | } |
951 | | |
952 | | //===----------------------------------------------------------------------===// |
953 | | // Double -> Float Shrinking Optimizations for Unary Functions like 'floor' |
954 | | |
955 | | Value *LibCallSimplifier::optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B, |
956 | 0 | bool CheckRetType) { |
957 | 0 | Function *Callee = CI->getCalledFunction(); |
958 | 0 | FunctionType *FT = Callee->getFunctionType(); |
959 | 0 | if (FT->getNumParams() != 1 || !FT->getReturnType()->isDoubleTy() || |
960 | 0 | !FT->getParamType(0)->isDoubleTy()) |
961 | 0 | return nullptr; |
962 | | |
963 | 0 | if (CheckRetType) { |
964 | | // Check if all the uses for function like 'sin' are converted to float. |
965 | 0 | for (User *U : CI->users()) { |
966 | 0 | FPTruncInst *Cast = dyn_cast<FPTruncInst>(U); |
967 | 0 | if (!Cast || !Cast->getType()->isFloatTy()) |
968 | 0 | return nullptr; |
969 | 0 | } |
970 | 0 | } |
971 | | |
972 | | // If this is something like 'floor((double)floatval)', convert to floorf. |
973 | 0 | Value *V = valueHasFloatPrecision(CI->getArgOperand(0)); |
974 | 0 | if (V == nullptr) |
975 | 0 | return nullptr; |
976 | | |
977 | | // floor((double)floatval) -> (double)floorf(floatval) |
978 | 0 | if (Callee->isIntrinsic()) { |
979 | 0 | Module *M = CI->getParent()->getParent()->getParent(); |
980 | 0 | Intrinsic::ID IID = Callee->getIntrinsicID(); |
981 | 0 | Function *F = Intrinsic::getDeclaration(M, IID, B.getFloatTy()); |
982 | 0 | V = B.CreateCall(F, V); |
983 | 0 | } else { |
984 | | // The call is a library call rather than an intrinsic. |
985 | 0 | V = EmitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes()); |
986 | 0 | } |
987 | |
|
988 | 0 | return B.CreateFPExt(V, B.getDoubleTy()); |
989 | 0 | } |
990 | | |
991 | | // Double -> Float Shrinking Optimizations for Binary Functions like 'fmin/fmax' |
992 | 0 | Value *LibCallSimplifier::optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) { |
993 | 0 | Function *Callee = CI->getCalledFunction(); |
994 | 0 | FunctionType *FT = Callee->getFunctionType(); |
995 | | // Just make sure this has 2 arguments of the same FP type, which match the |
996 | | // result type. |
997 | 0 | if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) || |
998 | 0 | FT->getParamType(0) != FT->getParamType(1) || |
999 | 0 | !FT->getParamType(0)->isFloatingPointTy()) |
1000 | 0 | return nullptr; |
1001 | | |
1002 | | // If this is something like 'fmin((double)floatval1, (double)floatval2)', |
1003 | | // or fmin(1.0, (double)floatval), then we convert it to fminf. |
1004 | 0 | Value *V1 = valueHasFloatPrecision(CI->getArgOperand(0)); |
1005 | 0 | if (V1 == nullptr) |
1006 | 0 | return nullptr; |
1007 | 0 | Value *V2 = valueHasFloatPrecision(CI->getArgOperand(1)); |
1008 | 0 | if (V2 == nullptr) |
1009 | 0 | return nullptr; |
1010 | | |
1011 | | // fmin((double)floatval1, (double)floatval2) |
1012 | | // -> (double)fminf(floatval1, floatval2) |
1013 | | // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP(). |
1014 | 0 | Value *V = EmitBinaryFloatFnCall(V1, V2, Callee->getName(), B, |
1015 | 0 | Callee->getAttributes()); |
1016 | 0 | return B.CreateFPExt(V, B.getDoubleTy()); |
1017 | 0 | } |
1018 | | |
1019 | 0 | Value *LibCallSimplifier::optimizeCos(CallInst *CI, IRBuilder<> &B) { |
1020 | 0 | Function *Callee = CI->getCalledFunction(); |
1021 | 0 | Value *Ret = nullptr; |
1022 | 0 | if (UnsafeFPShrink && Callee->getName() == "cos" && TLI->has(LibFunc::cosf)) { |
1023 | 0 | Ret = optimizeUnaryDoubleFP(CI, B, true); |
1024 | 0 | } |
1025 | |
|
1026 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1027 | | // Just make sure this has 1 argument of FP type, which matches the |
1028 | | // result type. |
1029 | 0 | if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || |
1030 | 0 | !FT->getParamType(0)->isFloatingPointTy()) |
1031 | 0 | return Ret; |
1032 | | |
1033 | | // cos(-x) -> cos(x) |
1034 | 0 | Value *Op1 = CI->getArgOperand(0); |
1035 | 0 | if (BinaryOperator::isFNeg(Op1)) { |
1036 | 0 | BinaryOperator *BinExpr = cast<BinaryOperator>(Op1); |
1037 | 0 | return B.CreateCall(Callee, BinExpr->getOperand(1), "cos"); |
1038 | 0 | } |
1039 | 0 | return Ret; |
1040 | 0 | } |
1041 | | |
1042 | 0 | Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) { |
1043 | 0 | Function *Callee = CI->getCalledFunction(); |
1044 | |
|
1045 | 0 | Value *Ret = nullptr; |
1046 | 0 | if (UnsafeFPShrink && Callee->getName() == "pow" && TLI->has(LibFunc::powf)) { |
1047 | 0 | Ret = optimizeUnaryDoubleFP(CI, B, true); |
1048 | 0 | } |
1049 | |
|
1050 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1051 | | // Just make sure this has 2 arguments of the same FP type, which match the |
1052 | | // result type. |
1053 | 0 | if (FT->getNumParams() != 2 || FT->getReturnType() != FT->getParamType(0) || |
1054 | 0 | FT->getParamType(0) != FT->getParamType(1) || |
1055 | 0 | !FT->getParamType(0)->isFloatingPointTy()) |
1056 | 0 | return Ret; |
1057 | | |
1058 | 0 | Value *Op1 = CI->getArgOperand(0), *Op2 = CI->getArgOperand(1); |
1059 | 0 | if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) { |
1060 | | // pow(1.0, x) -> 1.0 |
1061 | 0 | if (Op1C->isExactlyValue(1.0)) |
1062 | 0 | return Op1C; |
1063 | | // pow(2.0, x) -> exp2(x) |
1064 | 0 | if (Op1C->isExactlyValue(2.0) && |
1065 | 0 | hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp2, LibFunc::exp2f, |
1066 | 0 | LibFunc::exp2l)) |
1067 | 0 | return EmitUnaryFloatFnCall(Op2, "exp2", B, Callee->getAttributes()); |
1068 | | // pow(10.0, x) -> exp10(x) |
1069 | 0 | if (Op1C->isExactlyValue(10.0) && |
1070 | 0 | hasUnaryFloatFn(TLI, Op1->getType(), LibFunc::exp10, LibFunc::exp10f, |
1071 | 0 | LibFunc::exp10l)) |
1072 | 0 | return EmitUnaryFloatFnCall(Op2, TLI->getName(LibFunc::exp10), B, |
1073 | 0 | Callee->getAttributes()); |
1074 | 0 | } |
1075 | | |
1076 | 0 | ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2); |
1077 | 0 | if (!Op2C) |
1078 | 0 | return Ret; |
1079 | | |
1080 | 0 | if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0 |
1081 | 0 | return ConstantFP::get(CI->getType(), 1.0); |
1082 | | |
1083 | 0 | if (Op2C->isExactlyValue(0.5) && |
1084 | 0 | hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::sqrt, LibFunc::sqrtf, |
1085 | 0 | LibFunc::sqrtl) && |
1086 | 0 | hasUnaryFloatFn(TLI, Op2->getType(), LibFunc::fabs, LibFunc::fabsf, |
1087 | 0 | LibFunc::fabsl)) { |
1088 | | // Expand pow(x, 0.5) to (x == -infinity ? +infinity : fabs(sqrt(x))). |
1089 | | // This is faster than calling pow, and still handles negative zero |
1090 | | // and negative infinity correctly. |
1091 | | // TODO: In fast-math mode, this could be just sqrt(x). |
1092 | | // TODO: In finite-only mode, this could be just fabs(sqrt(x)). |
1093 | 0 | Value *Inf = ConstantFP::getInfinity(CI->getType()); |
1094 | 0 | Value *NegInf = ConstantFP::getInfinity(CI->getType(), true); |
1095 | 0 | Value *Sqrt = EmitUnaryFloatFnCall(Op1, "sqrt", B, Callee->getAttributes()); |
1096 | 0 | Value *FAbs = |
1097 | 0 | EmitUnaryFloatFnCall(Sqrt, "fabs", B, Callee->getAttributes()); |
1098 | 0 | Value *FCmp = B.CreateFCmpOEQ(Op1, NegInf); |
1099 | 0 | Value *Sel = B.CreateSelect(FCmp, Inf, FAbs); |
1100 | 0 | return Sel; |
1101 | 0 | } |
1102 | | |
1103 | 0 | if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x |
1104 | 0 | return Op1; |
1105 | 0 | if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x |
1106 | 0 | return B.CreateFMul(Op1, Op1, "pow2"); |
1107 | 0 | if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x |
1108 | 0 | return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip"); |
1109 | 0 | return nullptr; |
1110 | 0 | } |
1111 | | |
1112 | 0 | Value *LibCallSimplifier::optimizeExp2(CallInst *CI, IRBuilder<> &B) { |
1113 | 0 | Function *Callee = CI->getCalledFunction(); |
1114 | 0 | Function *Caller = CI->getParent()->getParent(); |
1115 | |
|
1116 | 0 | Value *Ret = nullptr; |
1117 | 0 | if (UnsafeFPShrink && Callee->getName() == "exp2" && |
1118 | 0 | TLI->has(LibFunc::exp2f)) { |
1119 | 0 | Ret = optimizeUnaryDoubleFP(CI, B, true); |
1120 | 0 | } |
1121 | |
|
1122 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1123 | | // Just make sure this has 1 argument of FP type, which matches the |
1124 | | // result type. |
1125 | 0 | if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || |
1126 | 0 | !FT->getParamType(0)->isFloatingPointTy()) |
1127 | 0 | return Ret; |
1128 | | |
1129 | 0 | Value *Op = CI->getArgOperand(0); |
1130 | | // Turn exp2(sitofp(x)) -> ldexp(1.0, sext(x)) if sizeof(x) <= 32 |
1131 | | // Turn exp2(uitofp(x)) -> ldexp(1.0, zext(x)) if sizeof(x) < 32 |
1132 | 0 | LibFunc::Func LdExp = LibFunc::ldexpl; |
1133 | 0 | if (Op->getType()->isFloatTy()) |
1134 | 0 | LdExp = LibFunc::ldexpf; |
1135 | 0 | else if (Op->getType()->isDoubleTy()) |
1136 | 0 | LdExp = LibFunc::ldexp; |
1137 | |
|
1138 | 0 | if (TLI->has(LdExp)) { |
1139 | 0 | Value *LdExpArg = nullptr; |
1140 | 0 | if (SIToFPInst *OpC = dyn_cast<SIToFPInst>(Op)) { |
1141 | 0 | if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() <= 32) |
1142 | 0 | LdExpArg = B.CreateSExt(OpC->getOperand(0), B.getInt32Ty()); |
1143 | 0 | } else if (UIToFPInst *OpC = dyn_cast<UIToFPInst>(Op)) { |
1144 | 0 | if (OpC->getOperand(0)->getType()->getPrimitiveSizeInBits() < 32) |
1145 | 0 | LdExpArg = B.CreateZExt(OpC->getOperand(0), B.getInt32Ty()); |
1146 | 0 | } |
1147 | |
|
1148 | 0 | if (LdExpArg) { |
1149 | 0 | Constant *One = ConstantFP::get(CI->getContext(), APFloat(1.0f)); |
1150 | 0 | if (!Op->getType()->isFloatTy()) |
1151 | 0 | One = ConstantExpr::getFPExtend(One, Op->getType()); |
1152 | |
|
1153 | 0 | Module *M = Caller->getParent(); |
1154 | 0 | Value *Callee = |
1155 | 0 | M->getOrInsertFunction(TLI->getName(LdExp), Op->getType(), |
1156 | 0 | Op->getType(), B.getInt32Ty(), nullptr); |
1157 | 0 | CallInst *CI = B.CreateCall(Callee, {One, LdExpArg}); |
1158 | 0 | if (const Function *F = dyn_cast<Function>(Callee->stripPointerCasts())) |
1159 | 0 | CI->setCallingConv(F->getCallingConv()); |
1160 | |
|
1161 | 0 | return CI; |
1162 | 0 | } |
1163 | 0 | } |
1164 | 0 | return Ret; |
1165 | 0 | } |
1166 | | |
1167 | 0 | Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) { |
1168 | 0 | Function *Callee = CI->getCalledFunction(); |
1169 | |
|
1170 | 0 | Value *Ret = nullptr; |
1171 | 0 | if (Callee->getName() == "fabs" && TLI->has(LibFunc::fabsf)) { |
1172 | 0 | Ret = optimizeUnaryDoubleFP(CI, B, false); |
1173 | 0 | } |
1174 | |
|
1175 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1176 | | // Make sure this has 1 argument of FP type which matches the result type. |
1177 | 0 | if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || |
1178 | 0 | !FT->getParamType(0)->isFloatingPointTy()) |
1179 | 0 | return Ret; |
1180 | | |
1181 | 0 | Value *Op = CI->getArgOperand(0); |
1182 | 0 | if (Instruction *I = dyn_cast<Instruction>(Op)) { |
1183 | | // Fold fabs(x * x) -> x * x; any squared FP value must already be positive. |
1184 | 0 | if (I->getOpcode() == Instruction::FMul) |
1185 | 0 | if (I->getOperand(0) == I->getOperand(1)) |
1186 | 0 | return Op; |
1187 | 0 | } |
1188 | 0 | return Ret; |
1189 | 0 | } |
1190 | | |
1191 | 0 | Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) { |
1192 | 0 | Function *Callee = CI->getCalledFunction(); |
1193 | | |
1194 | 0 | Value *Ret = nullptr; |
1195 | 0 | if (TLI->has(LibFunc::sqrtf) && (Callee->getName() == "sqrt" || |
1196 | 0 | Callee->getIntrinsicID() == Intrinsic::sqrt)) |
1197 | 0 | Ret = optimizeUnaryDoubleFP(CI, B, true); |
1198 | | |
1199 | | // FIXME: For finer-grain optimization, we need intrinsics to have the same |
1200 | | // fast-math flag decorations that are applied to FP instructions. For now, |
1201 | | // we have to rely on the function-level unsafe-fp-math attribute to do this |
1202 | | // optimization because there's no other way to express that the sqrt can be |
1203 | | // reassociated. |
1204 | 0 | Function *F = CI->getParent()->getParent(); |
1205 | 0 | if (F->hasFnAttribute("unsafe-fp-math")) { |
1206 | | // Check for unsafe-fp-math = true. |
1207 | 0 | Attribute Attr = F->getFnAttribute("unsafe-fp-math"); |
1208 | 0 | if (Attr.getValueAsString() != "true") |
1209 | 0 | return Ret; |
1210 | 0 | } |
1211 | 0 | Value *Op = CI->getArgOperand(0); |
1212 | 0 | if (Instruction *I = dyn_cast<Instruction>(Op)) { |
1213 | 0 | if (I->getOpcode() == Instruction::FMul && I->hasUnsafeAlgebra()) { |
1214 | | // We're looking for a repeated factor in a multiplication tree, |
1215 | | // so we can do this fold: sqrt(x * x) -> fabs(x); |
1216 | | // or this fold: sqrt(x * x * y) -> fabs(x) * sqrt(y). |
1217 | 0 | Value *Op0 = I->getOperand(0); |
1218 | 0 | Value *Op1 = I->getOperand(1); |
1219 | 0 | Value *RepeatOp = nullptr; |
1220 | 0 | Value *OtherOp = nullptr; |
1221 | 0 | if (Op0 == Op1) { |
1222 | | // Simple match: the operands of the multiply are identical. |
1223 | 0 | RepeatOp = Op0; |
1224 | 0 | } else { |
1225 | | // Look for a more complicated pattern: one of the operands is itself |
1226 | | // a multiply, so search for a common factor in that multiply. |
1227 | | // Note: We don't bother looking any deeper than this first level or for |
1228 | | // variations of this pattern because instcombine's visitFMUL and/or the |
1229 | | // reassociation pass should give us this form. |
1230 | 0 | Value *OtherMul0, *OtherMul1; |
1231 | 0 | if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) { |
1232 | | // Pattern: sqrt((x * y) * z) |
1233 | 0 | if (OtherMul0 == OtherMul1) { |
1234 | | // Matched: sqrt((x * x) * z) |
1235 | 0 | RepeatOp = OtherMul0; |
1236 | 0 | OtherOp = Op1; |
1237 | 0 | } |
1238 | 0 | } |
1239 | 0 | } |
1240 | 0 | if (RepeatOp) { |
1241 | | // Fast math flags for any created instructions should match the sqrt |
1242 | | // and multiply. |
1243 | | // FIXME: We're not checking the sqrt because it doesn't have |
1244 | | // fast-math-flags (see earlier comment). |
1245 | 0 | IRBuilder<true, ConstantFolder, |
1246 | 0 | IRBuilderDefaultInserter<true> >::FastMathFlagGuard Guard(B); |
1247 | 0 | B.SetFastMathFlags(I->getFastMathFlags()); |
1248 | | // If we found a repeated factor, hoist it out of the square root and |
1249 | | // replace it with the fabs of that factor. |
1250 | 0 | Module *M = Callee->getParent(); |
1251 | 0 | Type *ArgType = Op->getType(); |
1252 | 0 | Value *Fabs = Intrinsic::getDeclaration(M, Intrinsic::fabs, ArgType); |
1253 | 0 | Value *FabsCall = B.CreateCall(Fabs, RepeatOp, "fabs"); |
1254 | 0 | if (OtherOp) { |
1255 | | // If we found a non-repeated factor, we still need to get its square |
1256 | | // root. We then multiply that by the value that was simplified out |
1257 | | // of the square root calculation. |
1258 | 0 | Value *Sqrt = Intrinsic::getDeclaration(M, Intrinsic::sqrt, ArgType); |
1259 | 0 | Value *SqrtCall = B.CreateCall(Sqrt, OtherOp, "sqrt"); |
1260 | 0 | return B.CreateFMul(FabsCall, SqrtCall); |
1261 | 0 | } |
1262 | 0 | return FabsCall; |
1263 | 0 | } |
1264 | 0 | } |
1265 | 0 | } |
1266 | 0 | return Ret; |
1267 | 0 | } |
1268 | | |
1269 | | static bool isTrigLibCall(CallInst *CI); |
1270 | | static void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg, |
1271 | | bool UseFloat, Value *&Sin, Value *&Cos, |
1272 | | Value *&SinCos); |
1273 | | |
1274 | 0 | Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, IRBuilder<> &B) { |
1275 | | |
1276 | | // Make sure the prototype is as expected, otherwise the rest of the |
1277 | | // function is probably invalid and likely to abort. |
1278 | 0 | if (!isTrigLibCall(CI)) |
1279 | 0 | return nullptr; |
1280 | | |
1281 | 0 | Value *Arg = CI->getArgOperand(0); |
1282 | 0 | SmallVector<CallInst *, 1> SinCalls; |
1283 | 0 | SmallVector<CallInst *, 1> CosCalls; |
1284 | 0 | SmallVector<CallInst *, 1> SinCosCalls; |
1285 | |
|
1286 | 0 | bool IsFloat = Arg->getType()->isFloatTy(); |
1287 | | |
1288 | | // Look for all compatible sinpi, cospi and sincospi calls with the same |
1289 | | // argument. If there are enough (in some sense) we can make the |
1290 | | // substitution. |
1291 | 0 | for (User *U : Arg->users()) |
1292 | 0 | classifyArgUse(U, CI->getParent(), IsFloat, SinCalls, CosCalls, |
1293 | 0 | SinCosCalls); |
1294 | | |
1295 | | // It's only worthwhile if both sinpi and cospi are actually used. |
1296 | 0 | if (SinCosCalls.empty() && (SinCalls.empty() || CosCalls.empty())) |
1297 | 0 | return nullptr; |
1298 | | |
1299 | 0 | Value *Sin, *Cos, *SinCos; |
1300 | 0 | insertSinCosCall(B, CI->getCalledFunction(), Arg, IsFloat, Sin, Cos, SinCos); |
1301 | |
|
1302 | 0 | replaceTrigInsts(SinCalls, Sin); |
1303 | 0 | replaceTrigInsts(CosCalls, Cos); |
1304 | 0 | replaceTrigInsts(SinCosCalls, SinCos); |
1305 | |
|
1306 | 0 | return nullptr; |
1307 | 0 | } |
1308 | | |
1309 | 0 | static bool isTrigLibCall(CallInst *CI) { |
1310 | 0 | Function *Callee = CI->getCalledFunction(); |
1311 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1312 | | |
1313 | | // We can only hope to do anything useful if we can ignore things like errno |
1314 | | // and floating-point exceptions. |
1315 | 0 | bool AttributesSafe = |
1316 | 0 | CI->hasFnAttr(Attribute::NoUnwind) && CI->hasFnAttr(Attribute::ReadNone); |
1317 | | |
1318 | | // Other than that we need float(float) or double(double) |
1319 | 0 | return AttributesSafe && FT->getNumParams() == 1 && |
1320 | 0 | FT->getReturnType() == FT->getParamType(0) && |
1321 | 0 | (FT->getParamType(0)->isFloatTy() || |
1322 | 0 | FT->getParamType(0)->isDoubleTy()); |
1323 | 0 | } |
1324 | | |
1325 | | void |
1326 | | LibCallSimplifier::classifyArgUse(Value *Val, BasicBlock *BB, bool IsFloat, |
1327 | | SmallVectorImpl<CallInst *> &SinCalls, |
1328 | | SmallVectorImpl<CallInst *> &CosCalls, |
1329 | 0 | SmallVectorImpl<CallInst *> &SinCosCalls) { |
1330 | 0 | CallInst *CI = dyn_cast<CallInst>(Val); |
1331 | |
|
1332 | 0 | if (!CI) |
1333 | 0 | return; |
1334 | | |
1335 | 0 | Function *Callee = CI->getCalledFunction(); |
1336 | 0 | StringRef FuncName = Callee->getName(); |
1337 | 0 | LibFunc::Func Func; |
1338 | 0 | if (!TLI->getLibFunc(FuncName, Func) || !TLI->has(Func) || !isTrigLibCall(CI)) |
1339 | 0 | return; |
1340 | | |
1341 | 0 | if (IsFloat) { |
1342 | 0 | if (Func == LibFunc::sinpif) |
1343 | 0 | SinCalls.push_back(CI); |
1344 | 0 | else if (Func == LibFunc::cospif) |
1345 | 0 | CosCalls.push_back(CI); |
1346 | 0 | else if (Func == LibFunc::sincospif_stret) |
1347 | 0 | SinCosCalls.push_back(CI); |
1348 | 0 | } else { |
1349 | 0 | if (Func == LibFunc::sinpi) |
1350 | 0 | SinCalls.push_back(CI); |
1351 | 0 | else if (Func == LibFunc::cospi) |
1352 | 0 | CosCalls.push_back(CI); |
1353 | 0 | else if (Func == LibFunc::sincospi_stret) |
1354 | 0 | SinCosCalls.push_back(CI); |
1355 | 0 | } |
1356 | 0 | } |
1357 | | |
1358 | | void LibCallSimplifier::replaceTrigInsts(SmallVectorImpl<CallInst *> &Calls, |
1359 | 0 | Value *Res) { |
1360 | 0 | for (SmallVectorImpl<CallInst *>::iterator I = Calls.begin(), E = Calls.end(); |
1361 | 0 | I != E; ++I) { |
1362 | 0 | replaceAllUsesWith(*I, Res); |
1363 | 0 | } |
1364 | 0 | } |
1365 | | |
1366 | | void insertSinCosCall(IRBuilder<> &B, Function *OrigCallee, Value *Arg, |
1367 | 0 | bool UseFloat, Value *&Sin, Value *&Cos, Value *&SinCos) { |
1368 | 0 | Type *ArgTy = Arg->getType(); |
1369 | 0 | Type *ResTy; |
1370 | 0 | StringRef Name; |
1371 | |
|
1372 | 0 | Triple T(OrigCallee->getParent()->getTargetTriple()); |
1373 | 0 | if (UseFloat) { |
1374 | 0 | Name = "__sincospif_stret"; |
1375 | |
|
1376 | 0 | assert(T.getArch() != Triple::x86 && "x86 messy and unsupported for now"); |
1377 | | // x86_64 can't use {float, float} since that would be returned in both |
1378 | | // xmm0 and xmm1, which isn't what a real struct would do. |
1379 | 0 | ResTy = T.getArch() == Triple::x86_64 |
1380 | 0 | ? static_cast<Type *>(VectorType::get(ArgTy, 2)) |
1381 | 0 | : static_cast<Type *>(StructType::get(ArgTy, ArgTy, nullptr)); |
1382 | 0 | } else { |
1383 | 0 | Name = "__sincospi_stret"; |
1384 | 0 | ResTy = StructType::get(ArgTy, ArgTy, nullptr); |
1385 | 0 | } |
1386 | |
|
1387 | 0 | Module *M = OrigCallee->getParent(); |
1388 | 0 | Value *Callee = M->getOrInsertFunction(Name, OrigCallee->getAttributes(), |
1389 | 0 | ResTy, ArgTy, nullptr); |
1390 | |
|
1391 | 0 | if (Instruction *ArgInst = dyn_cast<Instruction>(Arg)) { |
1392 | | // If the argument is an instruction, it must dominate all uses so put our |
1393 | | // sincos call there. |
1394 | 0 | BasicBlock::iterator Loc = ArgInst; |
1395 | 0 | B.SetInsertPoint(ArgInst->getParent(), ++Loc); |
1396 | 0 | } else { |
1397 | | // Otherwise (e.g. for a constant) the beginning of the function is as |
1398 | | // good a place as any. |
1399 | 0 | BasicBlock &EntryBB = B.GetInsertBlock()->getParent()->getEntryBlock(); |
1400 | 0 | B.SetInsertPoint(&EntryBB, EntryBB.begin()); |
1401 | 0 | } |
1402 | |
|
1403 | 0 | SinCos = B.CreateCall(Callee, Arg, "sincospi"); |
1404 | |
|
1405 | 0 | if (SinCos->getType()->isStructTy()) { |
1406 | 0 | Sin = B.CreateExtractValue(SinCos, 0, "sinpi"); |
1407 | 0 | Cos = B.CreateExtractValue(SinCos, 1, "cospi"); |
1408 | 0 | } else { |
1409 | 0 | Sin = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 0), |
1410 | 0 | "sinpi"); |
1411 | 0 | Cos = B.CreateExtractElement(SinCos, ConstantInt::get(B.getInt32Ty(), 1), |
1412 | 0 | "cospi"); |
1413 | 0 | } |
1414 | 0 | } |
1415 | | |
1416 | | //===----------------------------------------------------------------------===// |
1417 | | // Integer Library Call Optimizations |
1418 | | //===----------------------------------------------------------------------===// |
1419 | | |
1420 | 0 | Value *LibCallSimplifier::optimizeFFS(CallInst *CI, IRBuilder<> &B) { |
1421 | 0 | Function *Callee = CI->getCalledFunction(); |
1422 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1423 | | // Just make sure this has 2 arguments of the same FP type, which match the |
1424 | | // result type. |
1425 | 0 | if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy(32) || |
1426 | 0 | !FT->getParamType(0)->isIntegerTy()) |
1427 | 0 | return nullptr; |
1428 | | |
1429 | 0 | Value *Op = CI->getArgOperand(0); |
1430 | | |
1431 | | // Constant fold. |
1432 | 0 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { |
1433 | 0 | if (CI->isZero()) // ffs(0) -> 0. |
1434 | 0 | return B.getInt32(0); |
1435 | | // ffs(c) -> cttz(c)+1 |
1436 | 0 | return B.getInt32(CI->getValue().countTrailingZeros() + 1); |
1437 | 0 | } |
1438 | | |
1439 | | // ffs(x) -> x != 0 ? (i32)llvm.cttz(x)+1 : 0 |
1440 | 0 | Type *ArgType = Op->getType(); |
1441 | 0 | Value *F = |
1442 | 0 | Intrinsic::getDeclaration(Callee->getParent(), Intrinsic::cttz, ArgType); |
1443 | 0 | Value *V = B.CreateCall(F, {Op, B.getFalse()}, "cttz"); |
1444 | 0 | V = B.CreateAdd(V, ConstantInt::get(V->getType(), 1)); |
1445 | 0 | V = B.CreateIntCast(V, B.getInt32Ty(), false); |
1446 | |
|
1447 | 0 | Value *Cond = B.CreateICmpNE(Op, Constant::getNullValue(ArgType)); |
1448 | 0 | return B.CreateSelect(Cond, V, B.getInt32(0)); |
1449 | 0 | } |
1450 | | |
1451 | 0 | Value *LibCallSimplifier::optimizeAbs(CallInst *CI, IRBuilder<> &B) { |
1452 | 0 | Function *Callee = CI->getCalledFunction(); |
1453 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1454 | | // We require integer(integer) where the types agree. |
1455 | 0 | if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy() || |
1456 | 0 | FT->getParamType(0) != FT->getReturnType()) |
1457 | 0 | return nullptr; |
1458 | | |
1459 | | // abs(x) -> x >s -1 ? x : -x |
1460 | 0 | Value *Op = CI->getArgOperand(0); |
1461 | 0 | Value *Pos = |
1462 | 0 | B.CreateICmpSGT(Op, Constant::getAllOnesValue(Op->getType()), "ispos"); |
1463 | 0 | Value *Neg = B.CreateNeg(Op, "neg"); |
1464 | 0 | return B.CreateSelect(Pos, Op, Neg); |
1465 | 0 | } |
1466 | | |
1467 | 0 | Value *LibCallSimplifier::optimizeIsDigit(CallInst *CI, IRBuilder<> &B) { |
1468 | 0 | Function *Callee = CI->getCalledFunction(); |
1469 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1470 | | // We require integer(i32) |
1471 | 0 | if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy() || |
1472 | 0 | !FT->getParamType(0)->isIntegerTy(32)) |
1473 | 0 | return nullptr; |
1474 | | |
1475 | | // isdigit(c) -> (c-'0') <u 10 |
1476 | 0 | Value *Op = CI->getArgOperand(0); |
1477 | 0 | Op = B.CreateSub(Op, B.getInt32('0'), "isdigittmp"); |
1478 | 0 | Op = B.CreateICmpULT(Op, B.getInt32(10), "isdigit"); |
1479 | 0 | return B.CreateZExt(Op, CI->getType()); |
1480 | 0 | } |
1481 | | |
1482 | 0 | Value *LibCallSimplifier::optimizeIsAscii(CallInst *CI, IRBuilder<> &B) { |
1483 | 0 | Function *Callee = CI->getCalledFunction(); |
1484 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1485 | | // We require integer(i32) |
1486 | 0 | if (FT->getNumParams() != 1 || !FT->getReturnType()->isIntegerTy() || |
1487 | 0 | !FT->getParamType(0)->isIntegerTy(32)) |
1488 | 0 | return nullptr; |
1489 | | |
1490 | | // isascii(c) -> c <u 128 |
1491 | 0 | Value *Op = CI->getArgOperand(0); |
1492 | 0 | Op = B.CreateICmpULT(Op, B.getInt32(128), "isascii"); |
1493 | 0 | return B.CreateZExt(Op, CI->getType()); |
1494 | 0 | } |
1495 | | |
1496 | 0 | Value *LibCallSimplifier::optimizeToAscii(CallInst *CI, IRBuilder<> &B) { |
1497 | 0 | Function *Callee = CI->getCalledFunction(); |
1498 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1499 | | // We require i32(i32) |
1500 | 0 | if (FT->getNumParams() != 1 || FT->getReturnType() != FT->getParamType(0) || |
1501 | 0 | !FT->getParamType(0)->isIntegerTy(32)) |
1502 | 0 | return nullptr; |
1503 | | |
1504 | | // toascii(c) -> c & 0x7f |
1505 | 0 | return B.CreateAnd(CI->getArgOperand(0), |
1506 | 0 | ConstantInt::get(CI->getType(), 0x7F)); |
1507 | 0 | } |
1508 | | |
1509 | | //===----------------------------------------------------------------------===// |
1510 | | // Formatting and IO Library Call Optimizations |
1511 | | //===----------------------------------------------------------------------===// |
1512 | | |
1513 | | static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg); |
1514 | | |
1515 | | Value *LibCallSimplifier::optimizeErrorReporting(CallInst *CI, IRBuilder<> &B, |
1516 | 0 | int StreamArg) { |
1517 | | // Error reporting calls should be cold, mark them as such. |
1518 | | // This applies even to non-builtin calls: it is only a hint and applies to |
1519 | | // functions that the frontend might not understand as builtins. |
1520 | | |
1521 | | // This heuristic was suggested in: |
1522 | | // Improving Static Branch Prediction in a Compiler |
1523 | | // Brian L. Deitrich, Ben-Chung Cheng, Wen-mei W. Hwu |
1524 | | // Proceedings of PACT'98, Oct. 1998, IEEE |
1525 | 0 | Function *Callee = CI->getCalledFunction(); |
1526 | |
|
1527 | 0 | if (!CI->hasFnAttr(Attribute::Cold) && |
1528 | 0 | isReportingError(Callee, CI, StreamArg)) { |
1529 | 0 | CI->addAttribute(AttributeSet::FunctionIndex, Attribute::Cold); |
1530 | 0 | } |
1531 | |
|
1532 | 0 | return nullptr; |
1533 | 0 | } |
1534 | | |
1535 | 0 | static bool isReportingError(Function *Callee, CallInst *CI, int StreamArg) { |
1536 | 0 | if (!ColdErrorCalls) |
1537 | 0 | return false; |
1538 | | |
1539 | 0 | if (!Callee || !Callee->isDeclaration()) |
1540 | 0 | return false; |
1541 | | |
1542 | 0 | if (StreamArg < 0) |
1543 | 0 | return true; |
1544 | | |
1545 | | // These functions might be considered cold, but only if their stream |
1546 | | // argument is stderr. |
1547 | | |
1548 | 0 | if (StreamArg >= (int)CI->getNumArgOperands()) |
1549 | 0 | return false; |
1550 | 0 | LoadInst *LI = dyn_cast<LoadInst>(CI->getArgOperand(StreamArg)); |
1551 | 0 | if (!LI) |
1552 | 0 | return false; |
1553 | 0 | GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand()); |
1554 | 0 | if (!GV || !GV->isDeclaration()) |
1555 | 0 | return false; |
1556 | 0 | return GV->getName() == "stderr"; |
1557 | 0 | } |
1558 | | |
1559 | 0 | Value *LibCallSimplifier::optimizePrintFString(CallInst *CI, IRBuilder<> &B) { |
1560 | | // Check for a fixed format string. |
1561 | 0 | StringRef FormatStr; |
1562 | 0 | if (!getConstantStringInfo(CI->getArgOperand(0), FormatStr)) |
1563 | 0 | return nullptr; |
1564 | | |
1565 | | // Empty format string -> noop. |
1566 | 0 | if (FormatStr.empty()) // Tolerate printf's declared void. |
1567 | 0 | return CI->use_empty() ? (Value *)CI : ConstantInt::get(CI->getType(), 0); |
1568 | | |
1569 | | // Do not do any of the following transformations if the printf return value |
1570 | | // is used, in general the printf return value is not compatible with either |
1571 | | // putchar() or puts(). |
1572 | 0 | if (!CI->use_empty()) |
1573 | 0 | return nullptr; |
1574 | | |
1575 | | // printf("x") -> putchar('x'), even for '%'. |
1576 | 0 | if (FormatStr.size() == 1) { |
1577 | 0 | Value *Res = EmitPutChar(B.getInt32(FormatStr[0]), B, TLI); |
1578 | 0 | if (CI->use_empty() || !Res) |
1579 | 0 | return Res; |
1580 | 0 | return B.CreateIntCast(Res, CI->getType(), true); |
1581 | 0 | } |
1582 | | |
1583 | | // printf("foo\n") --> puts("foo") |
1584 | 0 | if (FormatStr[FormatStr.size() - 1] == '\n' && |
1585 | 0 | FormatStr.find('%') == StringRef::npos) { // No format characters. |
1586 | | // Create a string literal with no \n on it. We expect the constant merge |
1587 | | // pass to be run after this pass, to merge duplicate strings. |
1588 | 0 | FormatStr = FormatStr.drop_back(); |
1589 | 0 | Value *GV = B.CreateGlobalString(FormatStr, "str"); |
1590 | 0 | Value *NewCI = EmitPutS(GV, B, TLI); |
1591 | 0 | return (CI->use_empty() || !NewCI) |
1592 | 0 | ? NewCI |
1593 | 0 | : ConstantInt::get(CI->getType(), FormatStr.size() + 1); |
1594 | 0 | } |
1595 | | |
1596 | | // Optimize specific format strings. |
1597 | | // printf("%c", chr) --> putchar(chr) |
1598 | 0 | if (FormatStr == "%c" && CI->getNumArgOperands() > 1 && |
1599 | 0 | CI->getArgOperand(1)->getType()->isIntegerTy()) { |
1600 | 0 | Value *Res = EmitPutChar(CI->getArgOperand(1), B, TLI); |
1601 | |
|
1602 | 0 | if (CI->use_empty() || !Res) |
1603 | 0 | return Res; |
1604 | 0 | return B.CreateIntCast(Res, CI->getType(), true); |
1605 | 0 | } |
1606 | | |
1607 | | // printf("%s\n", str) --> puts(str) |
1608 | 0 | if (FormatStr == "%s\n" && CI->getNumArgOperands() > 1 && |
1609 | 0 | CI->getArgOperand(1)->getType()->isPointerTy()) { |
1610 | 0 | return EmitPutS(CI->getArgOperand(1), B, TLI); |
1611 | 0 | } |
1612 | 0 | return nullptr; |
1613 | 0 | } |
1614 | | |
1615 | 0 | Value *LibCallSimplifier::optimizePrintF(CallInst *CI, IRBuilder<> &B) { |
1616 | |
|
1617 | 0 | Function *Callee = CI->getCalledFunction(); |
1618 | | // Require one fixed pointer argument and an integer/void result. |
1619 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1620 | 0 | if (FT->getNumParams() < 1 || !FT->getParamType(0)->isPointerTy() || |
1621 | 0 | !(FT->getReturnType()->isIntegerTy() || FT->getReturnType()->isVoidTy())) |
1622 | 0 | return nullptr; |
1623 | | |
1624 | 0 | if (Value *V = optimizePrintFString(CI, B)) { |
1625 | 0 | return V; |
1626 | 0 | } |
1627 | | |
1628 | | // printf(format, ...) -> iprintf(format, ...) if no floating point |
1629 | | // arguments. |
1630 | 0 | if (TLI->has(LibFunc::iprintf) && !callHasFloatingPointArgument(CI)) { |
1631 | 0 | Module *M = B.GetInsertBlock()->getParent()->getParent(); |
1632 | 0 | Constant *IPrintFFn = |
1633 | 0 | M->getOrInsertFunction("iprintf", FT, Callee->getAttributes()); |
1634 | 0 | CallInst *New = cast<CallInst>(CI->clone()); |
1635 | 0 | New->setCalledFunction(IPrintFFn); |
1636 | 0 | B.Insert(New); |
1637 | 0 | return New; |
1638 | 0 | } |
1639 | 0 | return nullptr; |
1640 | 0 | } |
1641 | | |
1642 | 0 | Value *LibCallSimplifier::optimizeSPrintFString(CallInst *CI, IRBuilder<> &B) { |
1643 | | // Check for a fixed format string. |
1644 | 0 | StringRef FormatStr; |
1645 | 0 | if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr)) |
1646 | 0 | return nullptr; |
1647 | | |
1648 | | // If we just have a format string (nothing else crazy) transform it. |
1649 | 0 | if (CI->getNumArgOperands() == 2) { |
1650 | | // Make sure there's no % in the constant array. We could try to handle |
1651 | | // %% -> % in the future if we cared. |
1652 | 0 | for (unsigned i = 0, e = FormatStr.size(); i != e; ++i) |
1653 | 0 | if (FormatStr[i] == '%') |
1654 | 0 | return nullptr; // we found a format specifier, bail out. |
1655 | | |
1656 | | // sprintf(str, fmt) -> llvm.memcpy(str, fmt, strlen(fmt)+1, 1) |
1657 | 0 | B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1), |
1658 | 0 | ConstantInt::get(DL.getIntPtrType(CI->getContext()), |
1659 | 0 | FormatStr.size() + 1), |
1660 | 0 | 1); // Copy the null byte. |
1661 | 0 | return ConstantInt::get(CI->getType(), FormatStr.size()); |
1662 | 0 | } |
1663 | | |
1664 | | // The remaining optimizations require the format string to be "%s" or "%c" |
1665 | | // and have an extra operand. |
1666 | 0 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || |
1667 | 0 | CI->getNumArgOperands() < 3) |
1668 | 0 | return nullptr; |
1669 | | |
1670 | | // Decode the second character of the format string. |
1671 | 0 | if (FormatStr[1] == 'c') { |
1672 | | // sprintf(dst, "%c", chr) --> *(i8*)dst = chr; *((i8*)dst+1) = 0 |
1673 | 0 | if (!CI->getArgOperand(2)->getType()->isIntegerTy()) |
1674 | 0 | return nullptr; |
1675 | 0 | Value *V = B.CreateTrunc(CI->getArgOperand(2), B.getInt8Ty(), "char"); |
1676 | 0 | Value *Ptr = CastToCStr(CI->getArgOperand(0), B); |
1677 | 0 | B.CreateStore(V, Ptr); |
1678 | 0 | Ptr = B.CreateGEP(B.getInt8Ty(), Ptr, B.getInt32(1), "nul"); |
1679 | 0 | B.CreateStore(B.getInt8(0), Ptr); |
1680 | |
|
1681 | 0 | return ConstantInt::get(CI->getType(), 1); |
1682 | 0 | } |
1683 | | |
1684 | 0 | if (FormatStr[1] == 's') { |
1685 | | // sprintf(dest, "%s", str) -> llvm.memcpy(dest, str, strlen(str)+1, 1) |
1686 | 0 | if (!CI->getArgOperand(2)->getType()->isPointerTy()) |
1687 | 0 | return nullptr; |
1688 | | |
1689 | 0 | Value *Len = EmitStrLen(CI->getArgOperand(2), B, DL, TLI); |
1690 | 0 | if (!Len) |
1691 | 0 | return nullptr; |
1692 | 0 | Value *IncLen = |
1693 | 0 | B.CreateAdd(Len, ConstantInt::get(Len->getType(), 1), "leninc"); |
1694 | 0 | B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(2), IncLen, 1); |
1695 | | |
1696 | | // The sprintf result is the unincremented number of bytes in the string. |
1697 | 0 | return B.CreateIntCast(Len, CI->getType(), false); |
1698 | 0 | } |
1699 | 0 | return nullptr; |
1700 | 0 | } |
1701 | | |
1702 | 0 | Value *LibCallSimplifier::optimizeSPrintF(CallInst *CI, IRBuilder<> &B) { |
1703 | 0 | Function *Callee = CI->getCalledFunction(); |
1704 | | // Require two fixed pointer arguments and an integer result. |
1705 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1706 | 0 | if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() || |
1707 | 0 | !FT->getParamType(1)->isPointerTy() || |
1708 | 0 | !FT->getReturnType()->isIntegerTy()) |
1709 | 0 | return nullptr; |
1710 | | |
1711 | 0 | if (Value *V = optimizeSPrintFString(CI, B)) { |
1712 | 0 | return V; |
1713 | 0 | } |
1714 | | |
1715 | | // sprintf(str, format, ...) -> siprintf(str, format, ...) if no floating |
1716 | | // point arguments. |
1717 | 0 | if (TLI->has(LibFunc::siprintf) && !callHasFloatingPointArgument(CI)) { |
1718 | 0 | Module *M = B.GetInsertBlock()->getParent()->getParent(); |
1719 | 0 | Constant *SIPrintFFn = |
1720 | 0 | M->getOrInsertFunction("siprintf", FT, Callee->getAttributes()); |
1721 | 0 | CallInst *New = cast<CallInst>(CI->clone()); |
1722 | 0 | New->setCalledFunction(SIPrintFFn); |
1723 | 0 | B.Insert(New); |
1724 | 0 | return New; |
1725 | 0 | } |
1726 | 0 | return nullptr; |
1727 | 0 | } |
1728 | | |
1729 | 0 | Value *LibCallSimplifier::optimizeFPrintFString(CallInst *CI, IRBuilder<> &B) { |
1730 | 0 | optimizeErrorReporting(CI, B, 0); |
1731 | | |
1732 | | // All the optimizations depend on the format string. |
1733 | 0 | StringRef FormatStr; |
1734 | 0 | if (!getConstantStringInfo(CI->getArgOperand(1), FormatStr)) |
1735 | 0 | return nullptr; |
1736 | | |
1737 | | // Do not do any of the following transformations if the fprintf return |
1738 | | // value is used, in general the fprintf return value is not compatible |
1739 | | // with fwrite(), fputc() or fputs(). |
1740 | 0 | if (!CI->use_empty()) |
1741 | 0 | return nullptr; |
1742 | | |
1743 | | // fprintf(F, "foo") --> fwrite("foo", 3, 1, F) |
1744 | 0 | if (CI->getNumArgOperands() == 2) { |
1745 | 0 | for (unsigned i = 0, e = FormatStr.size(); i != e; ++i) |
1746 | 0 | if (FormatStr[i] == '%') // Could handle %% -> % if we cared. |
1747 | 0 | return nullptr; // We found a format specifier. |
1748 | | |
1749 | 0 | return EmitFWrite( |
1750 | 0 | CI->getArgOperand(1), |
1751 | 0 | ConstantInt::get(DL.getIntPtrType(CI->getContext()), FormatStr.size()), |
1752 | 0 | CI->getArgOperand(0), B, DL, TLI); |
1753 | 0 | } |
1754 | | |
1755 | | // The remaining optimizations require the format string to be "%s" or "%c" |
1756 | | // and have an extra operand. |
1757 | 0 | if (FormatStr.size() != 2 || FormatStr[0] != '%' || |
1758 | 0 | CI->getNumArgOperands() < 3) |
1759 | 0 | return nullptr; |
1760 | | |
1761 | | // Decode the second character of the format string. |
1762 | 0 | if (FormatStr[1] == 'c') { |
1763 | | // fprintf(F, "%c", chr) --> fputc(chr, F) |
1764 | 0 | if (!CI->getArgOperand(2)->getType()->isIntegerTy()) |
1765 | 0 | return nullptr; |
1766 | 0 | return EmitFPutC(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI); |
1767 | 0 | } |
1768 | | |
1769 | 0 | if (FormatStr[1] == 's') { |
1770 | | // fprintf(F, "%s", str) --> fputs(str, F) |
1771 | 0 | if (!CI->getArgOperand(2)->getType()->isPointerTy()) |
1772 | 0 | return nullptr; |
1773 | 0 | return EmitFPutS(CI->getArgOperand(2), CI->getArgOperand(0), B, TLI); |
1774 | 0 | } |
1775 | 0 | return nullptr; |
1776 | 0 | } |
1777 | | |
1778 | 0 | Value *LibCallSimplifier::optimizeFPrintF(CallInst *CI, IRBuilder<> &B) { |
1779 | 0 | Function *Callee = CI->getCalledFunction(); |
1780 | | // Require two fixed paramters as pointers and integer result. |
1781 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1782 | 0 | if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() || |
1783 | 0 | !FT->getParamType(1)->isPointerTy() || |
1784 | 0 | !FT->getReturnType()->isIntegerTy()) |
1785 | 0 | return nullptr; |
1786 | | |
1787 | 0 | if (Value *V = optimizeFPrintFString(CI, B)) { |
1788 | 0 | return V; |
1789 | 0 | } |
1790 | | |
1791 | | // fprintf(stream, format, ...) -> fiprintf(stream, format, ...) if no |
1792 | | // floating point arguments. |
1793 | 0 | if (TLI->has(LibFunc::fiprintf) && !callHasFloatingPointArgument(CI)) { |
1794 | 0 | Module *M = B.GetInsertBlock()->getParent()->getParent(); |
1795 | 0 | Constant *FIPrintFFn = |
1796 | 0 | M->getOrInsertFunction("fiprintf", FT, Callee->getAttributes()); |
1797 | 0 | CallInst *New = cast<CallInst>(CI->clone()); |
1798 | 0 | New->setCalledFunction(FIPrintFFn); |
1799 | 0 | B.Insert(New); |
1800 | 0 | return New; |
1801 | 0 | } |
1802 | 0 | return nullptr; |
1803 | 0 | } |
1804 | | |
1805 | 0 | Value *LibCallSimplifier::optimizeFWrite(CallInst *CI, IRBuilder<> &B) { |
1806 | 0 | optimizeErrorReporting(CI, B, 3); |
1807 | |
|
1808 | 0 | Function *Callee = CI->getCalledFunction(); |
1809 | | // Require a pointer, an integer, an integer, a pointer, returning integer. |
1810 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1811 | 0 | if (FT->getNumParams() != 4 || !FT->getParamType(0)->isPointerTy() || |
1812 | 0 | !FT->getParamType(1)->isIntegerTy() || |
1813 | 0 | !FT->getParamType(2)->isIntegerTy() || |
1814 | 0 | !FT->getParamType(3)->isPointerTy() || |
1815 | 0 | !FT->getReturnType()->isIntegerTy()) |
1816 | 0 | return nullptr; |
1817 | | |
1818 | | // Get the element size and count. |
1819 | 0 | ConstantInt *SizeC = dyn_cast<ConstantInt>(CI->getArgOperand(1)); |
1820 | 0 | ConstantInt *CountC = dyn_cast<ConstantInt>(CI->getArgOperand(2)); |
1821 | 0 | if (!SizeC || !CountC) |
1822 | 0 | return nullptr; |
1823 | 0 | uint64_t Bytes = SizeC->getZExtValue() * CountC->getZExtValue(); |
1824 | | |
1825 | | // If this is writing zero records, remove the call (it's a noop). |
1826 | 0 | if (Bytes == 0) |
1827 | 0 | return ConstantInt::get(CI->getType(), 0); |
1828 | | |
1829 | | // If this is writing one byte, turn it into fputc. |
1830 | | // This optimisation is only valid, if the return value is unused. |
1831 | 0 | if (Bytes == 1 && CI->use_empty()) { // fwrite(S,1,1,F) -> fputc(S[0],F) |
1832 | 0 | Value *Char = B.CreateLoad(CastToCStr(CI->getArgOperand(0), B), "char"); |
1833 | 0 | Value *NewCI = EmitFPutC(Char, CI->getArgOperand(3), B, TLI); |
1834 | 0 | return NewCI ? ConstantInt::get(CI->getType(), 1) : nullptr; |
1835 | 0 | } |
1836 | | |
1837 | 0 | return nullptr; |
1838 | 0 | } |
1839 | | |
1840 | 0 | Value *LibCallSimplifier::optimizeFPuts(CallInst *CI, IRBuilder<> &B) { |
1841 | 0 | optimizeErrorReporting(CI, B, 1); |
1842 | |
|
1843 | 0 | Function *Callee = CI->getCalledFunction(); |
1844 | | |
1845 | | // Require two pointers. Also, we can't optimize if return value is used. |
1846 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1847 | 0 | if (FT->getNumParams() != 2 || !FT->getParamType(0)->isPointerTy() || |
1848 | 0 | !FT->getParamType(1)->isPointerTy() || !CI->use_empty()) |
1849 | 0 | return nullptr; |
1850 | | |
1851 | | // fputs(s,F) --> fwrite(s,1,strlen(s),F) |
1852 | 0 | uint64_t Len = GetStringLength(CI->getArgOperand(0)); |
1853 | 0 | if (!Len) |
1854 | 0 | return nullptr; |
1855 | | |
1856 | | // Known to have no uses (see above). |
1857 | 0 | return EmitFWrite( |
1858 | 0 | CI->getArgOperand(0), |
1859 | 0 | ConstantInt::get(DL.getIntPtrType(CI->getContext()), Len - 1), |
1860 | 0 | CI->getArgOperand(1), B, DL, TLI); |
1861 | 0 | } |
1862 | | |
1863 | 0 | Value *LibCallSimplifier::optimizePuts(CallInst *CI, IRBuilder<> &B) { |
1864 | 0 | Function *Callee = CI->getCalledFunction(); |
1865 | | // Require one fixed pointer argument and an integer/void result. |
1866 | 0 | FunctionType *FT = Callee->getFunctionType(); |
1867 | 0 | if (FT->getNumParams() < 1 || !FT->getParamType(0)->isPointerTy() || |
1868 | 0 | !(FT->getReturnType()->isIntegerTy() || FT->getReturnType()->isVoidTy())) |
1869 | 0 | return nullptr; |
1870 | | |
1871 | | // Check for a constant string. |
1872 | 0 | StringRef Str; |
1873 | 0 | if (!getConstantStringInfo(CI->getArgOperand(0), Str)) |
1874 | 0 | return nullptr; |
1875 | | |
1876 | 0 | if (Str.empty() && CI->use_empty()) { |
1877 | | // puts("") -> putchar('\n') |
1878 | 0 | Value *Res = EmitPutChar(B.getInt32('\n'), B, TLI); |
1879 | 0 | if (CI->use_empty() || !Res) |
1880 | 0 | return Res; |
1881 | 0 | return B.CreateIntCast(Res, CI->getType(), true); |
1882 | 0 | } |
1883 | | |
1884 | 0 | return nullptr; |
1885 | 0 | } |
1886 | | |
1887 | 0 | bool LibCallSimplifier::hasFloatVersion(StringRef FuncName) { |
1888 | 0 | LibFunc::Func Func; |
1889 | 0 | SmallString<20> FloatFuncName = FuncName; |
1890 | 0 | FloatFuncName += 'f'; |
1891 | 0 | if (TLI->getLibFunc(FloatFuncName, Func)) |
1892 | 0 | return TLI->has(Func); |
1893 | 0 | return false; |
1894 | 0 | } |
1895 | | |
1896 | | Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI, |
1897 | 0 | IRBuilder<> &Builder) { |
1898 | 0 | LibFunc::Func Func; |
1899 | 0 | Function *Callee = CI->getCalledFunction(); |
1900 | 0 | StringRef FuncName = Callee->getName(); |
1901 | | |
1902 | | // Check for string/memory library functions. |
1903 | 0 | if (TLI->getLibFunc(FuncName, Func) && TLI->has(Func)) { |
1904 | | // Make sure we never change the calling convention. |
1905 | 0 | assert((ignoreCallingConv(Func) || |
1906 | 0 | CI->getCallingConv() == llvm::CallingConv::C) && |
1907 | 0 | "Optimizing string/memory libcall would change the calling convention"); |
1908 | 0 | switch (Func) { |
1909 | 0 | case LibFunc::strcat: |
1910 | 0 | return optimizeStrCat(CI, Builder); |
1911 | 0 | case LibFunc::strncat: |
1912 | 0 | return optimizeStrNCat(CI, Builder); |
1913 | 0 | case LibFunc::strchr: |
1914 | 0 | return optimizeStrChr(CI, Builder); |
1915 | 0 | case LibFunc::strrchr: |
1916 | 0 | return optimizeStrRChr(CI, Builder); |
1917 | 0 | case LibFunc::strcmp: |
1918 | 0 | return optimizeStrCmp(CI, Builder); |
1919 | 0 | case LibFunc::strncmp: |
1920 | 0 | return optimizeStrNCmp(CI, Builder); |
1921 | 0 | case LibFunc::strcpy: |
1922 | 0 | return optimizeStrCpy(CI, Builder); |
1923 | 0 | case LibFunc::stpcpy: |
1924 | 0 | return optimizeStpCpy(CI, Builder); |
1925 | 0 | case LibFunc::strncpy: |
1926 | 0 | return optimizeStrNCpy(CI, Builder); |
1927 | 0 | case LibFunc::strlen: |
1928 | 0 | return optimizeStrLen(CI, Builder); |
1929 | 0 | case LibFunc::strpbrk: |
1930 | 0 | return optimizeStrPBrk(CI, Builder); |
1931 | 0 | case LibFunc::strtol: |
1932 | 0 | case LibFunc::strtod: |
1933 | 0 | case LibFunc::strtof: |
1934 | 0 | case LibFunc::strtoul: |
1935 | 0 | case LibFunc::strtoll: |
1936 | 0 | case LibFunc::strtold: |
1937 | 0 | case LibFunc::strtoull: |
1938 | 0 | return optimizeStrTo(CI, Builder); |
1939 | 0 | case LibFunc::strspn: |
1940 | 0 | return optimizeStrSpn(CI, Builder); |
1941 | 0 | case LibFunc::strcspn: |
1942 | 0 | return optimizeStrCSpn(CI, Builder); |
1943 | 0 | case LibFunc::strstr: |
1944 | 0 | return optimizeStrStr(CI, Builder); |
1945 | 0 | case LibFunc::memchr: |
1946 | 0 | return optimizeMemChr(CI, Builder); |
1947 | 0 | case LibFunc::memcmp: |
1948 | 0 | return optimizeMemCmp(CI, Builder); |
1949 | 0 | case LibFunc::memcpy: |
1950 | 0 | return optimizeMemCpy(CI, Builder); |
1951 | 0 | case LibFunc::memmove: |
1952 | 0 | return optimizeMemMove(CI, Builder); |
1953 | 0 | case LibFunc::memset: |
1954 | 0 | return optimizeMemSet(CI, Builder); |
1955 | 0 | default: |
1956 | 0 | break; |
1957 | 0 | } |
1958 | 0 | } |
1959 | 0 | return nullptr; |
1960 | 0 | } |
1961 | | |
1962 | 3.61M | Value *LibCallSimplifier::optimizeCall(CallInst *CI) { |
1963 | 3.61M | if (CI->isNoBuiltin()) |
1964 | 0 | return nullptr; |
1965 | | |
1966 | 3.61M | LibFunc::Func Func; |
1967 | 3.61M | Function *Callee = CI->getCalledFunction(); |
1968 | 3.61M | StringRef FuncName = Callee->getName(); |
1969 | 3.61M | IRBuilder<> Builder(CI); |
1970 | 3.61M | bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C; |
1971 | | |
1972 | | // Command-line parameter overrides function attribute. |
1973 | 3.61M | if (false) // HLSL Change - EnableUnsafeFPShrink.getNumOccurrences() > 0) |
1974 | 0 | UnsafeFPShrink = EnableUnsafeFPShrink; |
1975 | 3.61M | else if (Callee->hasFnAttribute("unsafe-fp-math")) { |
1976 | | // FIXME: This is the same problem as described in optimizeSqrt(). |
1977 | | // If calls gain access to IR-level FMF, then use that instead of a |
1978 | | // function attribute. |
1979 | | |
1980 | | // Check for unsafe-fp-math = true. |
1981 | 0 | Attribute Attr = Callee->getFnAttribute("unsafe-fp-math"); |
1982 | 0 | if (Attr.getValueAsString() == "true") |
1983 | 0 | UnsafeFPShrink = true; |
1984 | 0 | } |
1985 | | |
1986 | | // First, check for intrinsics. |
1987 | 3.61M | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { |
1988 | 1.00M | if (!isCallingConvC) |
1989 | 0 | return nullptr; |
1990 | 1.00M | switch (II->getIntrinsicID()) { |
1991 | 0 | case Intrinsic::pow: |
1992 | 0 | return optimizePow(CI, Builder); |
1993 | 0 | case Intrinsic::exp2: |
1994 | 0 | return optimizeExp2(CI, Builder); |
1995 | 0 | case Intrinsic::fabs: |
1996 | 0 | return optimizeFabs(CI, Builder); |
1997 | 0 | case Intrinsic::sqrt: |
1998 | 0 | return optimizeSqrt(CI, Builder); |
1999 | 1.00M | default: |
2000 | 1.00M | return nullptr; |
2001 | 1.00M | } |
2002 | 1.00M | } |
2003 | | |
2004 | | // Also try to simplify calls to fortified library functions. |
2005 | 2.60M | if (Value *SimplifiedFortifiedCI = FortifiedSimplifier.optimizeCall(CI)) { |
2006 | | // Try to further simplify the result. |
2007 | 0 | CallInst *SimplifiedCI = dyn_cast<CallInst>(SimplifiedFortifiedCI); |
2008 | 0 | if (SimplifiedCI && SimplifiedCI->getCalledFunction()) |
2009 | 0 | if (Value *V = optimizeStringMemoryLibCall(SimplifiedCI, Builder)) { |
2010 | | // If we were able to further simplify, remove the now redundant call. |
2011 | 0 | SimplifiedCI->replaceAllUsesWith(V); |
2012 | 0 | SimplifiedCI->eraseFromParent(); |
2013 | 0 | return V; |
2014 | 0 | } |
2015 | 0 | return SimplifiedFortifiedCI; |
2016 | 0 | } |
2017 | | |
2018 | | // Then check for known library functions. |
2019 | 2.60M | if (TLI->getLibFunc(FuncName, Func) && TLI->has(Func)0 ) { |
2020 | | // We never change the calling convention. |
2021 | 0 | if (!ignoreCallingConv(Func) && !isCallingConvC) |
2022 | 0 | return nullptr; |
2023 | 0 | if (Value *V = optimizeStringMemoryLibCall(CI, Builder)) |
2024 | 0 | return V; |
2025 | 0 | switch (Func) { |
2026 | 0 | case LibFunc::cosf: |
2027 | 0 | case LibFunc::cos: |
2028 | 0 | case LibFunc::cosl: |
2029 | 0 | return optimizeCos(CI, Builder); |
2030 | 0 | case LibFunc::sinpif: |
2031 | 0 | case LibFunc::sinpi: |
2032 | 0 | case LibFunc::cospif: |
2033 | 0 | case LibFunc::cospi: |
2034 | 0 | return optimizeSinCosPi(CI, Builder); |
2035 | 0 | case LibFunc::powf: |
2036 | 0 | case LibFunc::pow: |
2037 | 0 | case LibFunc::powl: |
2038 | 0 | return optimizePow(CI, Builder); |
2039 | 0 | case LibFunc::exp2l: |
2040 | 0 | case LibFunc::exp2: |
2041 | 0 | case LibFunc::exp2f: |
2042 | 0 | return optimizeExp2(CI, Builder); |
2043 | 0 | case LibFunc::fabsf: |
2044 | 0 | case LibFunc::fabs: |
2045 | 0 | case LibFunc::fabsl: |
2046 | 0 | return optimizeFabs(CI, Builder); |
2047 | 0 | case LibFunc::sqrtf: |
2048 | 0 | case LibFunc::sqrt: |
2049 | 0 | case LibFunc::sqrtl: |
2050 | 0 | return optimizeSqrt(CI, Builder); |
2051 | 0 | case LibFunc::ffs: |
2052 | 0 | case LibFunc::ffsl: |
2053 | 0 | case LibFunc::ffsll: |
2054 | 0 | return optimizeFFS(CI, Builder); |
2055 | 0 | case LibFunc::abs: |
2056 | 0 | case LibFunc::labs: |
2057 | 0 | case LibFunc::llabs: |
2058 | 0 | return optimizeAbs(CI, Builder); |
2059 | 0 | case LibFunc::isdigit: |
2060 | 0 | return optimizeIsDigit(CI, Builder); |
2061 | 0 | case LibFunc::isascii: |
2062 | 0 | return optimizeIsAscii(CI, Builder); |
2063 | 0 | case LibFunc::toascii: |
2064 | 0 | return optimizeToAscii(CI, Builder); |
2065 | 0 | case LibFunc::printf: |
2066 | 0 | return optimizePrintF(CI, Builder); |
2067 | 0 | case LibFunc::sprintf: |
2068 | 0 | return optimizeSPrintF(CI, Builder); |
2069 | 0 | case LibFunc::fprintf: |
2070 | 0 | return optimizeFPrintF(CI, Builder); |
2071 | 0 | case LibFunc::fwrite: |
2072 | 0 | return optimizeFWrite(CI, Builder); |
2073 | 0 | case LibFunc::fputs: |
2074 | 0 | return optimizeFPuts(CI, Builder); |
2075 | 0 | case LibFunc::puts: |
2076 | 0 | return optimizePuts(CI, Builder); |
2077 | 0 | case LibFunc::perror: |
2078 | 0 | return optimizeErrorReporting(CI, Builder); |
2079 | 0 | case LibFunc::vfprintf: |
2080 | 0 | case LibFunc::fiprintf: |
2081 | 0 | return optimizeErrorReporting(CI, Builder, 0); |
2082 | 0 | case LibFunc::fputc: |
2083 | 0 | return optimizeErrorReporting(CI, Builder, 1); |
2084 | 0 | case LibFunc::ceil: |
2085 | 0 | case LibFunc::floor: |
2086 | 0 | case LibFunc::rint: |
2087 | 0 | case LibFunc::round: |
2088 | 0 | case LibFunc::nearbyint: |
2089 | 0 | case LibFunc::trunc: |
2090 | 0 | if (hasFloatVersion(FuncName)) |
2091 | 0 | return optimizeUnaryDoubleFP(CI, Builder, false); |
2092 | 0 | return nullptr; |
2093 | 0 | case LibFunc::acos: |
2094 | 0 | case LibFunc::acosh: |
2095 | 0 | case LibFunc::asin: |
2096 | 0 | case LibFunc::asinh: |
2097 | 0 | case LibFunc::atan: |
2098 | 0 | case LibFunc::atanh: |
2099 | 0 | case LibFunc::cbrt: |
2100 | 0 | case LibFunc::cosh: |
2101 | 0 | case LibFunc::exp: |
2102 | 0 | case LibFunc::exp10: |
2103 | 0 | case LibFunc::expm1: |
2104 | 0 | case LibFunc::log: |
2105 | 0 | case LibFunc::log10: |
2106 | 0 | case LibFunc::log1p: |
2107 | 0 | case LibFunc::log2: |
2108 | 0 | case LibFunc::logb: |
2109 | 0 | case LibFunc::sin: |
2110 | 0 | case LibFunc::sinh: |
2111 | 0 | case LibFunc::tan: |
2112 | 0 | case LibFunc::tanh: |
2113 | 0 | if (UnsafeFPShrink && hasFloatVersion(FuncName)) |
2114 | 0 | return optimizeUnaryDoubleFP(CI, Builder, true); |
2115 | 0 | return nullptr; |
2116 | 0 | case LibFunc::copysign: |
2117 | 0 | case LibFunc::fmin: |
2118 | 0 | case LibFunc::fmax: |
2119 | 0 | if (hasFloatVersion(FuncName)) |
2120 | 0 | return optimizeBinaryDoubleFP(CI, Builder); |
2121 | 0 | return nullptr; |
2122 | 0 | default: |
2123 | 0 | return nullptr; |
2124 | 0 | } |
2125 | 0 | } |
2126 | 2.60M | return nullptr; |
2127 | 2.60M | } |
2128 | | |
2129 | | LibCallSimplifier::LibCallSimplifier( |
2130 | | const DataLayout &DL, const TargetLibraryInfo *TLI, |
2131 | | function_ref<void(Instruction *, Value *)> Replacer) |
2132 | | : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), UnsafeFPShrink(false), |
2133 | 3.61M | Replacer(Replacer) {} |
2134 | | |
2135 | 0 | void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) { |
2136 | | // Indirect through the replacer used in this instance. |
2137 | 0 | Replacer(I, With); |
2138 | 0 | } |
2139 | | |
2140 | | /*static*/ void LibCallSimplifier::replaceAllUsesWithDefault(Instruction *I, |
2141 | 0 | Value *With) { |
2142 | 0 | I->replaceAllUsesWith(With); |
2143 | 0 | I->eraseFromParent(); |
2144 | 0 | } |
2145 | | |
2146 | | // TODO: |
2147 | | // Additional cases that we need to add to this file: |
2148 | | // |
2149 | | // cbrt: |
2150 | | // * cbrt(expN(X)) -> expN(x/3) |
2151 | | // * cbrt(sqrt(x)) -> pow(x,1/6) |
2152 | | // * cbrt(sqrt(x)) -> pow(x,1/9) |
2153 | | // |
2154 | | // exp, expf, expl: |
2155 | | // * exp(log(x)) -> x |
2156 | | // |
2157 | | // log, logf, logl: |
2158 | | // * log(exp(x)) -> x |
2159 | | // * log(x**y) -> y*log(x) |
2160 | | // * log(exp(y)) -> y*log(e) |
2161 | | // * log(exp2(y)) -> y*log(2) |
2162 | | // * log(exp10(y)) -> y*log(10) |
2163 | | // * log(sqrt(x)) -> 0.5*log(x) |
2164 | | // * log(pow(x,y)) -> y*log(x) |
2165 | | // |
2166 | | // lround, lroundf, lroundl: |
2167 | | // * lround(cnst) -> cnst' |
2168 | | // |
2169 | | // pow, powf, powl: |
2170 | | // * pow(exp(x),y) -> exp(x*y) |
2171 | | // * pow(sqrt(x),y) -> pow(x,y*0.5) |
2172 | | // * pow(pow(x,y),z)-> pow(x,y*z) |
2173 | | // |
2174 | | // round, roundf, roundl: |
2175 | | // * round(cnst) -> cnst' |
2176 | | // |
2177 | | // signbit: |
2178 | | // * signbit(cnst) -> cnst' |
2179 | | // * signbit(nncst) -> 0 (if pstv is a non-negative constant) |
2180 | | // |
2181 | | // sqrt, sqrtf, sqrtl: |
2182 | | // * sqrt(expN(x)) -> expN(x*0.5) |
2183 | | // * sqrt(Nroot(x)) -> pow(x,1/(2*N)) |
2184 | | // * sqrt(pow(x,y)) -> pow(|x|,y*0.5) |
2185 | | // |
2186 | | // tan, tanf, tanl: |
2187 | | // * tan(atan(x)) -> x |
2188 | | // |
2189 | | // trunc, truncf, truncl: |
2190 | | // * trunc(cnst) -> cnst' |
2191 | | // |
2192 | | // |
2193 | | |
2194 | | //===----------------------------------------------------------------------===// |
2195 | | // Fortified Library Call Optimizations |
2196 | | //===----------------------------------------------------------------------===// |
2197 | | |
2198 | | bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI, |
2199 | | unsigned ObjSizeOp, |
2200 | | unsigned SizeOp, |
2201 | 0 | bool isString) { |
2202 | 0 | if (CI->getArgOperand(ObjSizeOp) == CI->getArgOperand(SizeOp)) |
2203 | 0 | return true; |
2204 | 0 | if (ConstantInt *ObjSizeCI = |
2205 | 0 | dyn_cast<ConstantInt>(CI->getArgOperand(ObjSizeOp))) { |
2206 | 0 | if (ObjSizeCI->isAllOnesValue()) |
2207 | 0 | return true; |
2208 | | // If the object size wasn't -1 (unknown), bail out if we were asked to. |
2209 | 0 | if (OnlyLowerUnknownSize) |
2210 | 0 | return false; |
2211 | 0 | if (isString) { |
2212 | 0 | uint64_t Len = GetStringLength(CI->getArgOperand(SizeOp)); |
2213 | | // If the length is 0 we don't know how long it is and so we can't |
2214 | | // remove the check. |
2215 | 0 | if (Len == 0) |
2216 | 0 | return false; |
2217 | 0 | return ObjSizeCI->getZExtValue() >= Len; |
2218 | 0 | } |
2219 | 0 | if (ConstantInt *SizeCI = dyn_cast<ConstantInt>(CI->getArgOperand(SizeOp))) |
2220 | 0 | return ObjSizeCI->getZExtValue() >= SizeCI->getZExtValue(); |
2221 | 0 | } |
2222 | 0 | return false; |
2223 | 0 | } |
2224 | | |
2225 | 0 | Value *FortifiedLibCallSimplifier::optimizeMemCpyChk(CallInst *CI, IRBuilder<> &B) { |
2226 | 0 | Function *Callee = CI->getCalledFunction(); |
2227 | |
|
2228 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::memcpy_chk)) |
2229 | 0 | return nullptr; |
2230 | | |
2231 | 0 | if (isFortifiedCallFoldable(CI, 3, 2, false)) { |
2232 | 0 | B.CreateMemCpy(CI->getArgOperand(0), CI->getArgOperand(1), |
2233 | 0 | CI->getArgOperand(2), 1); |
2234 | 0 | return CI->getArgOperand(0); |
2235 | 0 | } |
2236 | 0 | return nullptr; |
2237 | 0 | } |
2238 | | |
2239 | 0 | Value *FortifiedLibCallSimplifier::optimizeMemMoveChk(CallInst *CI, IRBuilder<> &B) { |
2240 | 0 | Function *Callee = CI->getCalledFunction(); |
2241 | |
|
2242 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::memmove_chk)) |
2243 | 0 | return nullptr; |
2244 | | |
2245 | 0 | if (isFortifiedCallFoldable(CI, 3, 2, false)) { |
2246 | 0 | B.CreateMemMove(CI->getArgOperand(0), CI->getArgOperand(1), |
2247 | 0 | CI->getArgOperand(2), 1); |
2248 | 0 | return CI->getArgOperand(0); |
2249 | 0 | } |
2250 | 0 | return nullptr; |
2251 | 0 | } |
2252 | | |
2253 | 0 | Value *FortifiedLibCallSimplifier::optimizeMemSetChk(CallInst *CI, IRBuilder<> &B) { |
2254 | 0 | Function *Callee = CI->getCalledFunction(); |
2255 | |
|
2256 | 0 | if (!checkStringCopyLibFuncSignature(Callee, LibFunc::memset_chk)) |
2257 | 0 | return nullptr; |
2258 | | |
2259 | 0 | if (isFortifiedCallFoldable(CI, 3, 2, false)) { |
2260 | 0 | Value *Val = B.CreateIntCast(CI->getArgOperand(1), B.getInt8Ty(), false); |
2261 | 0 | B.CreateMemSet(CI->getArgOperand(0), Val, CI->getArgOperand(2), 1); |
2262 | 0 | return CI->getArgOperand(0); |
2263 | 0 | } |
2264 | 0 | return nullptr; |
2265 | 0 | } |
2266 | | |
2267 | | Value *FortifiedLibCallSimplifier::optimizeStrpCpyChk(CallInst *CI, |
2268 | | IRBuilder<> &B, |
2269 | 0 | LibFunc::Func Func) { |
2270 | 0 | Function *Callee = CI->getCalledFunction(); |
2271 | 0 | StringRef Name = Callee->getName(); |
2272 | 0 | const DataLayout &DL = CI->getModule()->getDataLayout(); |
2273 | |
|
2274 | 0 | if (!checkStringCopyLibFuncSignature(Callee, Func)) |
2275 | 0 | return nullptr; |
2276 | | |
2277 | 0 | Value *Dst = CI->getArgOperand(0), *Src = CI->getArgOperand(1), |
2278 | 0 | *ObjSize = CI->getArgOperand(2); |
2279 | | |
2280 | | // __stpcpy_chk(x,x,...) -> x+strlen(x) |
2281 | 0 | if (Func == LibFunc::stpcpy_chk && !OnlyLowerUnknownSize && Dst == Src) { |
2282 | 0 | Value *StrLen = EmitStrLen(Src, B, DL, TLI); |
2283 | 0 | return StrLen ? B.CreateInBoundsGEP(B.getInt8Ty(), Dst, StrLen) : nullptr; |
2284 | 0 | } |
2285 | | |
2286 | | // If a) we don't have any length information, or b) we know this will |
2287 | | // fit then just lower to a plain st[rp]cpy. Otherwise we'll keep our |
2288 | | // st[rp]cpy_chk call which may fail at runtime if the size is too long. |
2289 | | // TODO: It might be nice to get a maximum length out of the possible |
2290 | | // string lengths for varying. |
2291 | 0 | if (isFortifiedCallFoldable(CI, 2, 1, true)) |
2292 | 0 | return EmitStrCpy(Dst, Src, B, TLI, Name.substr(2, 6)); |
2293 | | |
2294 | 0 | if (OnlyLowerUnknownSize) |
2295 | 0 | return nullptr; |
2296 | | |
2297 | | // Maybe we can stil fold __st[rp]cpy_chk to __memcpy_chk. |
2298 | 0 | uint64_t Len = GetStringLength(Src); |
2299 | 0 | if (Len == 0) |
2300 | 0 | return nullptr; |
2301 | | |
2302 | 0 | Type *SizeTTy = DL.getIntPtrType(CI->getContext()); |
2303 | 0 | Value *LenV = ConstantInt::get(SizeTTy, Len); |
2304 | 0 | Value *Ret = EmitMemCpyChk(Dst, Src, LenV, ObjSize, B, DL, TLI); |
2305 | | // If the function was an __stpcpy_chk, and we were able to fold it into |
2306 | | // a __memcpy_chk, we still need to return the correct end pointer. |
2307 | 0 | if (Ret && Func == LibFunc::stpcpy_chk) |
2308 | 0 | return B.CreateGEP(B.getInt8Ty(), Dst, ConstantInt::get(SizeTTy, Len - 1)); |
2309 | 0 | return Ret; |
2310 | 0 | } |
2311 | | |
2312 | | Value *FortifiedLibCallSimplifier::optimizeStrpNCpyChk(CallInst *CI, |
2313 | | IRBuilder<> &B, |
2314 | 0 | LibFunc::Func Func) { |
2315 | 0 | Function *Callee = CI->getCalledFunction(); |
2316 | 0 | StringRef Name = Callee->getName(); |
2317 | |
|
2318 | 0 | if (!checkStringCopyLibFuncSignature(Callee, Func)) |
2319 | 0 | return nullptr; |
2320 | 0 | if (isFortifiedCallFoldable(CI, 3, 2, false)) { |
2321 | 0 | Value *Ret = EmitStrNCpy(CI->getArgOperand(0), CI->getArgOperand(1), |
2322 | 0 | CI->getArgOperand(2), B, TLI, Name.substr(2, 7)); |
2323 | 0 | return Ret; |
2324 | 0 | } |
2325 | 0 | return nullptr; |
2326 | 0 | } |
2327 | | |
2328 | 2.60M | Value *FortifiedLibCallSimplifier::optimizeCall(CallInst *CI) { |
2329 | | // FIXME: We shouldn't be changing "nobuiltin" or TLI unavailable calls here. |
2330 | | // Some clang users checked for _chk libcall availability using: |
2331 | | // __has_builtin(__builtin___memcpy_chk) |
2332 | | // When compiling with -fno-builtin, this is always true. |
2333 | | // When passing -ffreestanding/-mkernel, which both imply -fno-builtin, we |
2334 | | // end up with fortified libcalls, which isn't acceptable in a freestanding |
2335 | | // environment which only provides their non-fortified counterparts. |
2336 | | // |
2337 | | // Until we change clang and/or teach external users to check for availability |
2338 | | // differently, disregard the "nobuiltin" attribute and TLI::has. |
2339 | | // |
2340 | | // PR23093. |
2341 | | |
2342 | 2.60M | LibFunc::Func Func; |
2343 | 2.60M | Function *Callee = CI->getCalledFunction(); |
2344 | 2.60M | StringRef FuncName = Callee->getName(); |
2345 | 2.60M | IRBuilder<> Builder(CI); |
2346 | 2.60M | bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C; |
2347 | | |
2348 | | // First, check that this is a known library functions. |
2349 | 2.60M | if (!TLI->getLibFunc(FuncName, Func)) |
2350 | 2.60M | return nullptr; |
2351 | | |
2352 | | // We never change the calling convention. |
2353 | 0 | if (!ignoreCallingConv(Func) && !isCallingConvC) |
2354 | 0 | return nullptr; |
2355 | | |
2356 | 0 | switch (Func) { |
2357 | 0 | case LibFunc::memcpy_chk: |
2358 | 0 | return optimizeMemCpyChk(CI, Builder); |
2359 | 0 | case LibFunc::memmove_chk: |
2360 | 0 | return optimizeMemMoveChk(CI, Builder); |
2361 | 0 | case LibFunc::memset_chk: |
2362 | 0 | return optimizeMemSetChk(CI, Builder); |
2363 | 0 | case LibFunc::stpcpy_chk: |
2364 | 0 | case LibFunc::strcpy_chk: |
2365 | 0 | return optimizeStrpCpyChk(CI, Builder, Func); |
2366 | 0 | case LibFunc::stpncpy_chk: |
2367 | 0 | case LibFunc::strncpy_chk: |
2368 | 0 | return optimizeStrpNCpyChk(CI, Builder, Func); |
2369 | 0 | default: |
2370 | 0 | break; |
2371 | 0 | } |
2372 | 0 | return nullptr; |
2373 | 0 | } |
2374 | | |
2375 | | FortifiedLibCallSimplifier::FortifiedLibCallSimplifier( |
2376 | | const TargetLibraryInfo *TLI, bool OnlyLowerUnknownSize) |
2377 | 3.61M | : TLI(TLI), OnlyLowerUnknownSize(OnlyLowerUnknownSize) {} |