1 | //===-- ConstantFolding.cpp - Fold instructions into constants ------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines routines for folding instructions into constants. |
10 | // |
11 | // Also, to supplement the basic IR ConstantExpr simplifications, |
12 | // this file defines some additional folding routines that can make use of |
13 | // DataLayout information. These functions cannot go in IR due to library |
14 | // dependency issues. |
15 | // |
16 | //===----------------------------------------------------------------------===// |
17 | |
18 | #include "llvm/Analysis/ConstantFolding.h" |
19 | #include "llvm/ADT/APFloat.h" |
20 | #include "llvm/ADT/APInt.h" |
21 | #include "llvm/ADT/APSInt.h" |
22 | #include "llvm/ADT/ArrayRef.h" |
23 | #include "llvm/ADT/DenseMap.h" |
24 | #include "llvm/ADT/STLExtras.h" |
25 | #include "llvm/ADT/SmallVector.h" |
26 | #include "llvm/ADT/StringRef.h" |
27 | #include "llvm/Analysis/TargetFolder.h" |
28 | #include "llvm/Analysis/TargetLibraryInfo.h" |
29 | #include "llvm/Analysis/ValueTracking.h" |
30 | #include "llvm/Analysis/VectorUtils.h" |
31 | #include "llvm/Config/config.h" |
32 | #include "llvm/IR/Constant.h" |
33 | #include "llvm/IR/ConstantFold.h" |
34 | #include "llvm/IR/Constants.h" |
35 | #include "llvm/IR/DataLayout.h" |
36 | #include "llvm/IR/DerivedTypes.h" |
37 | #include "llvm/IR/Function.h" |
38 | #include "llvm/IR/GlobalValue.h" |
39 | #include "llvm/IR/GlobalVariable.h" |
40 | #include "llvm/IR/InstrTypes.h" |
41 | #include "llvm/IR/Instruction.h" |
42 | #include "llvm/IR/Instructions.h" |
43 | #include "llvm/IR/IntrinsicInst.h" |
44 | #include "llvm/IR/Intrinsics.h" |
45 | #include "llvm/IR/IntrinsicsAArch64.h" |
46 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
47 | #include "llvm/IR/IntrinsicsARM.h" |
48 | #include "llvm/IR/IntrinsicsNVPTX.h" |
49 | #include "llvm/IR/IntrinsicsWebAssembly.h" |
50 | #include "llvm/IR/IntrinsicsX86.h" |
51 | #include "llvm/IR/NVVMIntrinsicUtils.h" |
52 | #include "llvm/IR/Operator.h" |
53 | #include "llvm/IR/Type.h" |
54 | #include "llvm/IR/Value.h" |
55 | #include "llvm/Support/Casting.h" |
56 | #include "llvm/Support/ErrorHandling.h" |
57 | #include "llvm/Support/KnownBits.h" |
58 | #include "llvm/Support/MathExtras.h" |
59 | #include <cassert> |
60 | #include <cerrno> |
61 | #include <cfenv> |
62 | #include <cmath> |
63 | #include <cstdint> |
64 | |
65 | using namespace llvm; |
66 | |
67 | static cl::opt<bool> DisableFPCallFolding( |
68 | "disable-fp-call-folding" , |
69 | cl::desc("Disable constant-folding of FP intrinsics and libcalls." ), |
70 | cl::init(Val: false), cl::Hidden); |
71 | |
72 | namespace { |
73 | |
74 | //===----------------------------------------------------------------------===// |
75 | // Constant Folding internal helper functions |
76 | //===----------------------------------------------------------------------===// |
77 | |
78 | static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy, |
79 | Constant *C, Type *SrcEltTy, |
80 | unsigned NumSrcElts, |
81 | const DataLayout &DL) { |
82 | // Now that we know that the input value is a vector of integers, just shift |
83 | // and insert them into our result. |
84 | unsigned BitShift = DL.getTypeSizeInBits(Ty: SrcEltTy); |
85 | for (unsigned i = 0; i != NumSrcElts; ++i) { |
86 | Constant *Element; |
87 | if (DL.isLittleEndian()) |
88 | Element = C->getAggregateElement(Elt: NumSrcElts - i - 1); |
89 | else |
90 | Element = C->getAggregateElement(Elt: i); |
91 | |
92 | if (isa_and_nonnull<UndefValue>(Val: Element)) { |
93 | Result <<= BitShift; |
94 | continue; |
95 | } |
96 | |
97 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Val: Element); |
98 | if (!ElementCI) |
99 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
100 | |
101 | Result <<= BitShift; |
102 | Result |= ElementCI->getValue().zext(width: Result.getBitWidth()); |
103 | } |
104 | |
105 | return nullptr; |
106 | } |
107 | |
108 | /// Constant fold bitcast, symbolically evaluating it with DataLayout. |
109 | /// This always returns a non-null constant, but it may be a |
110 | /// ConstantExpr if unfoldable. |
111 | Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) { |
112 | assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) && |
113 | "Invalid constantexpr bitcast!" ); |
114 | |
115 | // Catch the obvious splat cases. |
116 | if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL)) |
117 | return Res; |
118 | |
119 | if (auto *VTy = dyn_cast<VectorType>(Val: C->getType())) { |
120 | // Handle a vector->scalar integer/fp cast. |
121 | if (isa<IntegerType>(Val: DestTy) || DestTy->isFloatingPointTy()) { |
122 | unsigned NumSrcElts = cast<FixedVectorType>(Val: VTy)->getNumElements(); |
123 | Type *SrcEltTy = VTy->getElementType(); |
124 | |
125 | // If the vector is a vector of floating point, convert it to vector of int |
126 | // to simplify things. |
127 | if (SrcEltTy->isFloatingPointTy()) { |
128 | unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); |
129 | auto *SrcIVTy = FixedVectorType::get( |
130 | ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElts); |
131 | // Ask IR to do the conversion now that #elts line up. |
132 | C = ConstantExpr::getBitCast(C, Ty: SrcIVTy); |
133 | } |
134 | |
135 | APInt Result(DL.getTypeSizeInBits(Ty: DestTy), 0); |
136 | if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C, |
137 | SrcEltTy, NumSrcElts, DL)) |
138 | return CE; |
139 | |
140 | if (isa<IntegerType>(Val: DestTy)) |
141 | return ConstantInt::get(Ty: DestTy, V: Result); |
142 | |
143 | APFloat FP(DestTy->getFltSemantics(), Result); |
144 | return ConstantFP::get(Context&: DestTy->getContext(), V: FP); |
145 | } |
146 | } |
147 | |
148 | // The code below only handles casts to vectors currently. |
149 | auto *DestVTy = dyn_cast<VectorType>(Val: DestTy); |
150 | if (!DestVTy) |
151 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
152 | |
153 | // If this is a scalar -> vector cast, convert the input into a <1 x scalar> |
154 | // vector so the code below can handle it uniformly. |
155 | if (!isa<VectorType>(Val: C->getType()) && |
156 | (isa<ConstantFP>(Val: C) || isa<ConstantInt>(Val: C))) { |
157 | Constant *Ops = C; // don't take the address of C! |
158 | return FoldBitCast(C: ConstantVector::get(V: Ops), DestTy, DL); |
159 | } |
160 | |
161 | // Some of what follows may extend to cover scalable vectors but the current |
162 | // implementation is fixed length specific. |
163 | if (!isa<FixedVectorType>(Val: C->getType())) |
164 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
165 | |
166 | // If this is a bitcast from constant vector -> vector, fold it. |
167 | if (!isa<ConstantDataVector>(Val: C) && !isa<ConstantVector>(Val: C) && |
168 | !isa<ConstantInt>(Val: C) && !isa<ConstantFP>(Val: C)) |
169 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
170 | |
171 | // If the element types match, IR can fold it. |
172 | unsigned NumDstElt = cast<FixedVectorType>(Val: DestVTy)->getNumElements(); |
173 | unsigned NumSrcElt = cast<FixedVectorType>(Val: C->getType())->getNumElements(); |
174 | if (NumDstElt == NumSrcElt) |
175 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
176 | |
177 | Type *SrcEltTy = cast<VectorType>(Val: C->getType())->getElementType(); |
178 | Type *DstEltTy = DestVTy->getElementType(); |
179 | |
180 | // Otherwise, we're changing the number of elements in a vector, which |
181 | // requires endianness information to do the right thing. For example, |
182 | // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) |
183 | // folds to (little endian): |
184 | // <4 x i32> <i32 0, i32 0, i32 1, i32 0> |
185 | // and to (big endian): |
186 | // <4 x i32> <i32 0, i32 0, i32 0, i32 1> |
187 | |
188 | // First thing is first. We only want to think about integer here, so if |
189 | // we have something in FP form, recast it as integer. |
190 | if (DstEltTy->isFloatingPointTy()) { |
191 | // Fold to an vector of integers with same size as our FP type. |
192 | unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); |
193 | auto *DestIVTy = FixedVectorType::get( |
194 | ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumDstElt); |
195 | // Recursively handle this integer conversion, if possible. |
196 | C = FoldBitCast(C, DestTy: DestIVTy, DL); |
197 | |
198 | // Finally, IR can handle this now that #elts line up. |
199 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
200 | } |
201 | |
202 | // Okay, we know the destination is integer, if the input is FP, convert |
203 | // it to integer first. |
204 | if (SrcEltTy->isFloatingPointTy()) { |
205 | unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); |
206 | auto *SrcIVTy = FixedVectorType::get( |
207 | ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElt); |
208 | // Ask IR to do the conversion now that #elts line up. |
209 | C = ConstantExpr::getBitCast(C, Ty: SrcIVTy); |
210 | assert((isa<ConstantVector>(C) || // FIXME: Remove ConstantVector. |
211 | isa<ConstantDataVector>(C) || isa<ConstantInt>(C)) && |
212 | "Constant folding cannot fail for plain fp->int bitcast!" ); |
213 | } |
214 | |
215 | // Now we know that the input and output vectors are both integer vectors |
216 | // of the same size, and that their #elements is not the same. Do the |
217 | // conversion here, which depends on whether the input or output has |
218 | // more elements. |
219 | bool isLittleEndian = DL.isLittleEndian(); |
220 | |
221 | SmallVector<Constant*, 32> Result; |
222 | if (NumDstElt < NumSrcElt) { |
223 | // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) |
224 | Constant *Zero = Constant::getNullValue(Ty: DstEltTy); |
225 | unsigned Ratio = NumSrcElt/NumDstElt; |
226 | unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); |
227 | unsigned SrcElt = 0; |
228 | for (unsigned i = 0; i != NumDstElt; ++i) { |
229 | // Build each element of the result. |
230 | Constant *Elt = Zero; |
231 | unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); |
232 | for (unsigned j = 0; j != Ratio; ++j) { |
233 | Constant *Src = C->getAggregateElement(Elt: SrcElt++); |
234 | if (isa_and_nonnull<UndefValue>(Val: Src)) |
235 | Src = Constant::getNullValue( |
236 | Ty: cast<VectorType>(Val: C->getType())->getElementType()); |
237 | else |
238 | Src = dyn_cast_or_null<ConstantInt>(Val: Src); |
239 | if (!Src) // Reject constantexpr elements. |
240 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
241 | |
242 | // Zero extend the element to the right size. |
243 | Src = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: Src, DestTy: Elt->getType(), |
244 | DL); |
245 | assert(Src && "Constant folding cannot fail on plain integers" ); |
246 | |
247 | // Shift it to the right place, depending on endianness. |
248 | Src = ConstantFoldBinaryOpOperands( |
249 | Opcode: Instruction::Shl, LHS: Src, RHS: ConstantInt::get(Ty: Src->getType(), V: ShiftAmt), |
250 | DL); |
251 | assert(Src && "Constant folding cannot fail on plain integers" ); |
252 | |
253 | ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; |
254 | |
255 | // Mix it in. |
256 | Elt = ConstantFoldBinaryOpOperands(Opcode: Instruction::Or, LHS: Elt, RHS: Src, DL); |
257 | assert(Elt && "Constant folding cannot fail on plain integers" ); |
258 | } |
259 | Result.push_back(Elt); |
260 | } |
261 | return ConstantVector::get(V: Result); |
262 | } |
263 | |
264 | // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) |
265 | unsigned Ratio = NumDstElt/NumSrcElt; |
266 | unsigned DstBitSize = DL.getTypeSizeInBits(Ty: DstEltTy); |
267 | |
268 | // Loop over each source value, expanding into multiple results. |
269 | for (unsigned i = 0; i != NumSrcElt; ++i) { |
270 | auto *Element = C->getAggregateElement(Elt: i); |
271 | |
272 | if (!Element) // Reject constantexpr elements. |
273 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
274 | |
275 | if (isa<UndefValue>(Val: Element)) { |
276 | // Correctly Propagate undef values. |
277 | Result.append(NumInputs: Ratio, Elt: UndefValue::get(T: DstEltTy)); |
278 | continue; |
279 | } |
280 | |
281 | auto *Src = dyn_cast<ConstantInt>(Val: Element); |
282 | if (!Src) |
283 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
284 | |
285 | unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); |
286 | for (unsigned j = 0; j != Ratio; ++j) { |
287 | // Shift the piece of the value into the right place, depending on |
288 | // endianness. |
289 | APInt Elt = Src->getValue().lshr(shiftAmt: ShiftAmt); |
290 | ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; |
291 | |
292 | // Truncate and remember this piece. |
293 | Result.push_back(Elt: ConstantInt::get(Ty: DstEltTy, V: Elt.trunc(width: DstBitSize))); |
294 | } |
295 | } |
296 | |
297 | return ConstantVector::get(V: Result); |
298 | } |
299 | |
300 | } // end anonymous namespace |
301 | |
302 | /// If this constant is a constant offset from a global, return the global and |
303 | /// the constant. Because of constantexprs, this function is recursive. |
304 | bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, |
305 | APInt &Offset, const DataLayout &DL, |
306 | DSOLocalEquivalent **DSOEquiv) { |
307 | if (DSOEquiv) |
308 | *DSOEquiv = nullptr; |
309 | |
310 | // Trivial case, constant is the global. |
311 | if ((GV = dyn_cast<GlobalValue>(Val: C))) { |
312 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType()); |
313 | Offset = APInt(BitWidth, 0); |
314 | return true; |
315 | } |
316 | |
317 | if (auto *FoundDSOEquiv = dyn_cast<DSOLocalEquivalent>(Val: C)) { |
318 | if (DSOEquiv) |
319 | *DSOEquiv = FoundDSOEquiv; |
320 | GV = FoundDSOEquiv->getGlobalValue(); |
321 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType()); |
322 | Offset = APInt(BitWidth, 0); |
323 | return true; |
324 | } |
325 | |
326 | // Otherwise, if this isn't a constant expr, bail out. |
327 | auto *CE = dyn_cast<ConstantExpr>(Val: C); |
328 | if (!CE) return false; |
329 | |
330 | // Look through ptr->int and ptr->ptr casts. |
331 | if (CE->getOpcode() == Instruction::PtrToInt || |
332 | CE->getOpcode() == Instruction::BitCast) |
333 | return IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset, DL, |
334 | DSOEquiv); |
335 | |
336 | // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) |
337 | auto *GEP = dyn_cast<GEPOperator>(Val: CE); |
338 | if (!GEP) |
339 | return false; |
340 | |
341 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType()); |
342 | APInt TmpOffset(BitWidth, 0); |
343 | |
344 | // If the base isn't a global+constant, we aren't either. |
345 | if (!IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset&: TmpOffset, DL, |
346 | DSOEquiv)) |
347 | return false; |
348 | |
349 | // Otherwise, add any offset that our operands provide. |
350 | if (!GEP->accumulateConstantOffset(DL, Offset&: TmpOffset)) |
351 | return false; |
352 | |
353 | Offset = TmpOffset; |
354 | return true; |
355 | } |
356 | |
357 | Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy, |
358 | const DataLayout &DL) { |
359 | do { |
360 | Type *SrcTy = C->getType(); |
361 | if (SrcTy == DestTy) |
362 | return C; |
363 | |
364 | TypeSize DestSize = DL.getTypeSizeInBits(Ty: DestTy); |
365 | TypeSize SrcSize = DL.getTypeSizeInBits(Ty: SrcTy); |
366 | if (!TypeSize::isKnownGE(LHS: SrcSize, RHS: DestSize)) |
367 | return nullptr; |
368 | |
369 | // Catch the obvious splat cases (since all-zeros can coerce non-integral |
370 | // pointers legally). |
371 | if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL)) |
372 | return Res; |
373 | |
374 | // If the type sizes are the same and a cast is legal, just directly |
375 | // cast the constant. |
376 | // But be careful not to coerce non-integral pointers illegally. |
377 | if (SrcSize == DestSize && |
378 | DL.isNonIntegralPointerType(Ty: SrcTy->getScalarType()) == |
379 | DL.isNonIntegralPointerType(Ty: DestTy->getScalarType())) { |
380 | Instruction::CastOps Cast = Instruction::BitCast; |
381 | // If we are going from a pointer to int or vice versa, we spell the cast |
382 | // differently. |
383 | if (SrcTy->isIntegerTy() && DestTy->isPointerTy()) |
384 | Cast = Instruction::IntToPtr; |
385 | else if (SrcTy->isPointerTy() && DestTy->isIntegerTy()) |
386 | Cast = Instruction::PtrToInt; |
387 | |
388 | if (CastInst::castIsValid(op: Cast, S: C, DstTy: DestTy)) |
389 | return ConstantFoldCastOperand(Opcode: Cast, C, DestTy, DL); |
390 | } |
391 | |
392 | // If this isn't an aggregate type, there is nothing we can do to drill down |
393 | // and find a bitcastable constant. |
394 | if (!SrcTy->isAggregateType() && !SrcTy->isVectorTy()) |
395 | return nullptr; |
396 | |
397 | // We're simulating a load through a pointer that was bitcast to point to |
398 | // a different type, so we can try to walk down through the initial |
399 | // elements of an aggregate to see if some part of the aggregate is |
400 | // castable to implement the "load" semantic model. |
401 | if (SrcTy->isStructTy()) { |
402 | // Struct types might have leading zero-length elements like [0 x i32], |
403 | // which are certainly not what we are looking for, so skip them. |
404 | unsigned Elem = 0; |
405 | Constant *ElemC; |
406 | do { |
407 | ElemC = C->getAggregateElement(Elt: Elem++); |
408 | } while (ElemC && DL.getTypeSizeInBits(Ty: ElemC->getType()).isZero()); |
409 | C = ElemC; |
410 | } else { |
411 | // For non-byte-sized vector elements, the first element is not |
412 | // necessarily located at the vector base address. |
413 | if (auto *VT = dyn_cast<VectorType>(Val: SrcTy)) |
414 | if (!DL.typeSizeEqualsStoreSize(Ty: VT->getElementType())) |
415 | return nullptr; |
416 | |
417 | C = C->getAggregateElement(Elt: 0u); |
418 | } |
419 | } while (C); |
420 | |
421 | return nullptr; |
422 | } |
423 | |
424 | namespace { |
425 | |
426 | /// Recursive helper to read bits out of global. C is the constant being copied |
427 | /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy |
428 | /// results into and BytesLeft is the number of bytes left in |
429 | /// the CurPtr buffer. DL is the DataLayout. |
430 | bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr, |
431 | unsigned BytesLeft, const DataLayout &DL) { |
432 | assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) && |
433 | "Out of range access" ); |
434 | |
435 | // Reading type padding, return zero. |
436 | if (ByteOffset >= DL.getTypeStoreSize(Ty: C->getType())) |
437 | return true; |
438 | |
439 | // If this element is zero or undefined, we can just return since *CurPtr is |
440 | // zero initialized. |
441 | if (isa<ConstantAggregateZero>(Val: C) || isa<UndefValue>(Val: C)) |
442 | return true; |
443 | |
444 | if (auto *CI = dyn_cast<ConstantInt>(Val: C)) { |
445 | if ((CI->getBitWidth() & 7) != 0) |
446 | return false; |
447 | const APInt &Val = CI->getValue(); |
448 | unsigned IntBytes = unsigned(CI->getBitWidth()/8); |
449 | |
450 | for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { |
451 | unsigned n = ByteOffset; |
452 | if (!DL.isLittleEndian()) |
453 | n = IntBytes - n - 1; |
454 | CurPtr[i] = Val.extractBits(numBits: 8, bitPosition: n * 8).getZExtValue(); |
455 | ++ByteOffset; |
456 | } |
457 | return true; |
458 | } |
459 | |
460 | if (auto *CFP = dyn_cast<ConstantFP>(Val: C)) { |
461 | if (CFP->getType()->isDoubleTy()) { |
462 | C = FoldBitCast(C, DestTy: Type::getInt64Ty(C&: C->getContext()), DL); |
463 | return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); |
464 | } |
465 | if (CFP->getType()->isFloatTy()){ |
466 | C = FoldBitCast(C, DestTy: Type::getInt32Ty(C&: C->getContext()), DL); |
467 | return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); |
468 | } |
469 | if (CFP->getType()->isHalfTy()){ |
470 | C = FoldBitCast(C, DestTy: Type::getInt16Ty(C&: C->getContext()), DL); |
471 | return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); |
472 | } |
473 | return false; |
474 | } |
475 | |
476 | if (auto *CS = dyn_cast<ConstantStruct>(Val: C)) { |
477 | const StructLayout *SL = DL.getStructLayout(Ty: CS->getType()); |
478 | unsigned Index = SL->getElementContainingOffset(FixedOffset: ByteOffset); |
479 | uint64_t CurEltOffset = SL->getElementOffset(Idx: Index); |
480 | ByteOffset -= CurEltOffset; |
481 | |
482 | while (true) { |
483 | // If the element access is to the element itself and not to tail padding, |
484 | // read the bytes from the element. |
485 | uint64_t EltSize = DL.getTypeAllocSize(Ty: CS->getOperand(i_nocapture: Index)->getType()); |
486 | |
487 | if (ByteOffset < EltSize && |
488 | !ReadDataFromGlobal(C: CS->getOperand(i_nocapture: Index), ByteOffset, CurPtr, |
489 | BytesLeft, DL)) |
490 | return false; |
491 | |
492 | ++Index; |
493 | |
494 | // Check to see if we read from the last struct element, if so we're done. |
495 | if (Index == CS->getType()->getNumElements()) |
496 | return true; |
497 | |
498 | // If we read all of the bytes we needed from this element we're done. |
499 | uint64_t NextEltOffset = SL->getElementOffset(Idx: Index); |
500 | |
501 | if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset) |
502 | return true; |
503 | |
504 | // Move to the next element of the struct. |
505 | CurPtr += NextEltOffset - CurEltOffset - ByteOffset; |
506 | BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset; |
507 | ByteOffset = 0; |
508 | CurEltOffset = NextEltOffset; |
509 | } |
510 | // not reached. |
511 | } |
512 | |
513 | if (isa<ConstantArray>(Val: C) || isa<ConstantVector>(Val: C) || |
514 | isa<ConstantDataSequential>(Val: C)) { |
515 | uint64_t NumElts, EltSize; |
516 | Type *EltTy; |
517 | if (auto *AT = dyn_cast<ArrayType>(Val: C->getType())) { |
518 | NumElts = AT->getNumElements(); |
519 | EltTy = AT->getElementType(); |
520 | EltSize = DL.getTypeAllocSize(Ty: EltTy); |
521 | } else { |
522 | NumElts = cast<FixedVectorType>(Val: C->getType())->getNumElements(); |
523 | EltTy = cast<FixedVectorType>(Val: C->getType())->getElementType(); |
524 | // TODO: For non-byte-sized vectors, current implementation assumes there is |
525 | // padding to the next byte boundary between elements. |
526 | if (!DL.typeSizeEqualsStoreSize(Ty: EltTy)) |
527 | return false; |
528 | |
529 | EltSize = DL.getTypeStoreSize(Ty: EltTy); |
530 | } |
531 | uint64_t Index = ByteOffset / EltSize; |
532 | uint64_t Offset = ByteOffset - Index * EltSize; |
533 | |
534 | for (; Index != NumElts; ++Index) { |
535 | if (!ReadDataFromGlobal(C: C->getAggregateElement(Elt: Index), ByteOffset: Offset, CurPtr, |
536 | BytesLeft, DL)) |
537 | return false; |
538 | |
539 | uint64_t BytesWritten = EltSize - Offset; |
540 | assert(BytesWritten <= EltSize && "Not indexing into this element?" ); |
541 | if (BytesWritten >= BytesLeft) |
542 | return true; |
543 | |
544 | Offset = 0; |
545 | BytesLeft -= BytesWritten; |
546 | CurPtr += BytesWritten; |
547 | } |
548 | return true; |
549 | } |
550 | |
551 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
552 | if (CE->getOpcode() == Instruction::IntToPtr && |
553 | CE->getOperand(i_nocapture: 0)->getType() == DL.getIntPtrType(CE->getType())) { |
554 | return ReadDataFromGlobal(C: CE->getOperand(i_nocapture: 0), ByteOffset, CurPtr, |
555 | BytesLeft, DL); |
556 | } |
557 | } |
558 | |
559 | // Otherwise, unknown initializer type. |
560 | return false; |
561 | } |
562 | |
563 | Constant *FoldReinterpretLoadFromConst(Constant *C, Type *LoadTy, |
564 | int64_t Offset, const DataLayout &DL) { |
565 | // Bail out early. Not expect to load from scalable global variable. |
566 | if (isa<ScalableVectorType>(Val: LoadTy)) |
567 | return nullptr; |
568 | |
569 | auto *IntType = dyn_cast<IntegerType>(Val: LoadTy); |
570 | |
571 | // If this isn't an integer load we can't fold it directly. |
572 | if (!IntType) { |
573 | // If this is a non-integer load, we can try folding it as an int load and |
574 | // then bitcast the result. This can be useful for union cases. Note |
575 | // that address spaces don't matter here since we're not going to result in |
576 | // an actual new load. |
577 | if (!LoadTy->isFloatingPointTy() && !LoadTy->isPointerTy() && |
578 | !LoadTy->isVectorTy()) |
579 | return nullptr; |
580 | |
581 | Type *MapTy = Type::getIntNTy(C&: C->getContext(), |
582 | N: DL.getTypeSizeInBits(Ty: LoadTy).getFixedValue()); |
583 | if (Constant *Res = FoldReinterpretLoadFromConst(C, LoadTy: MapTy, Offset, DL)) { |
584 | if (Res->isNullValue() && !LoadTy->isX86_AMXTy()) |
585 | // Materializing a zero can be done trivially without a bitcast |
586 | return Constant::getNullValue(Ty: LoadTy); |
587 | Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy; |
588 | Res = FoldBitCast(C: Res, DestTy: CastTy, DL); |
589 | if (LoadTy->isPtrOrPtrVectorTy()) { |
590 | // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr |
591 | if (Res->isNullValue() && !LoadTy->isX86_AMXTy()) |
592 | return Constant::getNullValue(Ty: LoadTy); |
593 | if (DL.isNonIntegralPointerType(Ty: LoadTy->getScalarType())) |
594 | // Be careful not to replace a load of an addrspace value with an inttoptr here |
595 | return nullptr; |
596 | Res = ConstantExpr::getIntToPtr(C: Res, Ty: LoadTy); |
597 | } |
598 | return Res; |
599 | } |
600 | return nullptr; |
601 | } |
602 | |
603 | unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; |
604 | if (BytesLoaded > 32 || BytesLoaded == 0) |
605 | return nullptr; |
606 | |
607 | // If we're not accessing anything in this constant, the result is undefined. |
608 | if (Offset <= -1 * static_cast<int64_t>(BytesLoaded)) |
609 | return PoisonValue::get(T: IntType); |
610 | |
611 | // TODO: We should be able to support scalable types. |
612 | TypeSize InitializerSize = DL.getTypeAllocSize(Ty: C->getType()); |
613 | if (InitializerSize.isScalable()) |
614 | return nullptr; |
615 | |
616 | // If we're not accessing anything in this constant, the result is undefined. |
617 | if (Offset >= (int64_t)InitializerSize.getFixedValue()) |
618 | return PoisonValue::get(T: IntType); |
619 | |
620 | unsigned char RawBytes[32] = {0}; |
621 | unsigned char *CurPtr = RawBytes; |
622 | unsigned BytesLeft = BytesLoaded; |
623 | |
624 | // If we're loading off the beginning of the global, some bytes may be valid. |
625 | if (Offset < 0) { |
626 | CurPtr += -Offset; |
627 | BytesLeft += Offset; |
628 | Offset = 0; |
629 | } |
630 | |
631 | if (!ReadDataFromGlobal(C, ByteOffset: Offset, CurPtr, BytesLeft, DL)) |
632 | return nullptr; |
633 | |
634 | APInt ResultVal = APInt(IntType->getBitWidth(), 0); |
635 | if (DL.isLittleEndian()) { |
636 | ResultVal = RawBytes[BytesLoaded - 1]; |
637 | for (unsigned i = 1; i != BytesLoaded; ++i) { |
638 | ResultVal <<= 8; |
639 | ResultVal |= RawBytes[BytesLoaded - 1 - i]; |
640 | } |
641 | } else { |
642 | ResultVal = RawBytes[0]; |
643 | for (unsigned i = 1; i != BytesLoaded; ++i) { |
644 | ResultVal <<= 8; |
645 | ResultVal |= RawBytes[i]; |
646 | } |
647 | } |
648 | |
649 | return ConstantInt::get(Context&: IntType->getContext(), V: ResultVal); |
650 | } |
651 | |
652 | } // anonymous namespace |
653 | |
654 | // If GV is a constant with an initializer read its representation starting |
655 | // at Offset and return it as a constant array of unsigned char. Otherwise |
656 | // return null. |
657 | Constant *llvm::ReadByteArrayFromGlobal(const GlobalVariable *GV, |
658 | uint64_t Offset) { |
659 | if (!GV->isConstant() || !GV->hasDefinitiveInitializer()) |
660 | return nullptr; |
661 | |
662 | const DataLayout &DL = GV->getDataLayout(); |
663 | Constant *Init = const_cast<Constant *>(GV->getInitializer()); |
664 | TypeSize InitSize = DL.getTypeAllocSize(Ty: Init->getType()); |
665 | if (InitSize < Offset) |
666 | return nullptr; |
667 | |
668 | uint64_t NBytes = InitSize - Offset; |
669 | if (NBytes > UINT16_MAX) |
670 | // Bail for large initializers in excess of 64K to avoid allocating |
671 | // too much memory. |
672 | // Offset is assumed to be less than or equal than InitSize (this |
673 | // is enforced in ReadDataFromGlobal). |
674 | return nullptr; |
675 | |
676 | SmallVector<unsigned char, 256> RawBytes(static_cast<size_t>(NBytes)); |
677 | unsigned char *CurPtr = RawBytes.data(); |
678 | |
679 | if (!ReadDataFromGlobal(C: Init, ByteOffset: Offset, CurPtr, BytesLeft: NBytes, DL)) |
680 | return nullptr; |
681 | |
682 | return ConstantDataArray::get(Context&: GV->getContext(), Elts&: RawBytes); |
683 | } |
684 | |
685 | /// If this Offset points exactly to the start of an aggregate element, return |
686 | /// that element, otherwise return nullptr. |
687 | Constant *getConstantAtOffset(Constant *Base, APInt Offset, |
688 | const DataLayout &DL) { |
689 | if (Offset.isZero()) |
690 | return Base; |
691 | |
692 | if (!isa<ConstantAggregate>(Val: Base) && !isa<ConstantDataSequential>(Val: Base)) |
693 | return nullptr; |
694 | |
695 | Type *ElemTy = Base->getType(); |
696 | SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset); |
697 | if (!Offset.isZero() || !Indices[0].isZero()) |
698 | return nullptr; |
699 | |
700 | Constant *C = Base; |
701 | for (const APInt &Index : drop_begin(RangeOrContainer&: Indices)) { |
702 | if (Index.isNegative() || Index.getActiveBits() >= 32) |
703 | return nullptr; |
704 | |
705 | C = C->getAggregateElement(Elt: Index.getZExtValue()); |
706 | if (!C) |
707 | return nullptr; |
708 | } |
709 | |
710 | return C; |
711 | } |
712 | |
713 | Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty, |
714 | const APInt &Offset, |
715 | const DataLayout &DL) { |
716 | if (Constant *AtOffset = getConstantAtOffset(Base: C, Offset, DL)) |
717 | if (Constant *Result = ConstantFoldLoadThroughBitcast(C: AtOffset, DestTy: Ty, DL)) |
718 | return Result; |
719 | |
720 | // Explicitly check for out-of-bounds access, so we return poison even if the |
721 | // constant is a uniform value. |
722 | TypeSize Size = DL.getTypeAllocSize(Ty: C->getType()); |
723 | if (!Size.isScalable() && Offset.sge(RHS: Size.getFixedValue())) |
724 | return PoisonValue::get(T: Ty); |
725 | |
726 | // Try an offset-independent fold of a uniform value. |
727 | if (Constant *Result = ConstantFoldLoadFromUniformValue(C, Ty, DL)) |
728 | return Result; |
729 | |
730 | // Try hard to fold loads from bitcasted strange and non-type-safe things. |
731 | if (Offset.getSignificantBits() <= 64) |
732 | if (Constant *Result = |
733 | FoldReinterpretLoadFromConst(C, LoadTy: Ty, Offset: Offset.getSExtValue(), DL)) |
734 | return Result; |
735 | |
736 | return nullptr; |
737 | } |
738 | |
739 | Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty, |
740 | const DataLayout &DL) { |
741 | return ConstantFoldLoadFromConst(C, Ty, Offset: APInt(64, 0), DL); |
742 | } |
743 | |
744 | Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, |
745 | APInt Offset, |
746 | const DataLayout &DL) { |
747 | // We can only fold loads from constant globals with a definitive initializer. |
748 | // Check this upfront, to skip expensive offset calculations. |
749 | auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: C)); |
750 | if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) |
751 | return nullptr; |
752 | |
753 | C = cast<Constant>(Val: C->stripAndAccumulateConstantOffsets( |
754 | DL, Offset, /* AllowNonInbounds */ true)); |
755 | |
756 | if (C == GV) |
757 | if (Constant *Result = ConstantFoldLoadFromConst(C: GV->getInitializer(), Ty, |
758 | Offset, DL)) |
759 | return Result; |
760 | |
761 | // If this load comes from anywhere in a uniform constant global, the value |
762 | // is always the same, regardless of the loaded offset. |
763 | return ConstantFoldLoadFromUniformValue(C: GV->getInitializer(), Ty, DL); |
764 | } |
765 | |
766 | Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, |
767 | const DataLayout &DL) { |
768 | APInt Offset(DL.getIndexTypeSizeInBits(Ty: C->getType()), 0); |
769 | return ConstantFoldLoadFromConstPtr(C, Ty, Offset: std::move(Offset), DL); |
770 | } |
771 | |
772 | Constant *llvm::ConstantFoldLoadFromUniformValue(Constant *C, Type *Ty, |
773 | const DataLayout &DL) { |
774 | if (isa<PoisonValue>(Val: C)) |
775 | return PoisonValue::get(T: Ty); |
776 | if (isa<UndefValue>(Val: C)) |
777 | return UndefValue::get(T: Ty); |
778 | // If padding is needed when storing C to memory, then it isn't considered as |
779 | // uniform. |
780 | if (!DL.typeSizeEqualsStoreSize(Ty: C->getType())) |
781 | return nullptr; |
782 | if (C->isNullValue() && !Ty->isX86_AMXTy()) |
783 | return Constant::getNullValue(Ty); |
784 | if (C->isAllOnesValue() && |
785 | (Ty->isIntOrIntVectorTy() || Ty->isFPOrFPVectorTy())) |
786 | return Constant::getAllOnesValue(Ty); |
787 | return nullptr; |
788 | } |
789 | |
790 | namespace { |
791 | |
792 | /// One of Op0/Op1 is a constant expression. |
793 | /// Attempt to symbolically evaluate the result of a binary operator merging |
794 | /// these together. If target data info is available, it is provided as DL, |
795 | /// otherwise DL is null. |
796 | Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1, |
797 | const DataLayout &DL) { |
798 | // SROA |
799 | |
800 | // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. |
801 | // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute |
802 | // bits. |
803 | |
804 | if (Opc == Instruction::And) { |
805 | KnownBits Known0 = computeKnownBits(V: Op0, DL); |
806 | KnownBits Known1 = computeKnownBits(V: Op1, DL); |
807 | if ((Known1.One | Known0.Zero).isAllOnes()) { |
808 | // All the bits of Op0 that the 'and' could be masking are already zero. |
809 | return Op0; |
810 | } |
811 | if ((Known0.One | Known1.Zero).isAllOnes()) { |
812 | // All the bits of Op1 that the 'and' could be masking are already zero. |
813 | return Op1; |
814 | } |
815 | |
816 | Known0 &= Known1; |
817 | if (Known0.isConstant()) |
818 | return ConstantInt::get(Ty: Op0->getType(), V: Known0.getConstant()); |
819 | } |
820 | |
821 | // If the constant expr is something like &A[123] - &A[4].f, fold this into a |
822 | // constant. This happens frequently when iterating over a global array. |
823 | if (Opc == Instruction::Sub) { |
824 | GlobalValue *GV1, *GV2; |
825 | APInt Offs1, Offs2; |
826 | |
827 | if (IsConstantOffsetFromGlobal(C: Op0, GV&: GV1, Offset&: Offs1, DL)) |
828 | if (IsConstantOffsetFromGlobal(C: Op1, GV&: GV2, Offset&: Offs2, DL) && GV1 == GV2) { |
829 | unsigned OpSize = DL.getTypeSizeInBits(Ty: Op0->getType()); |
830 | |
831 | // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. |
832 | // PtrToInt may change the bitwidth so we have convert to the right size |
833 | // first. |
834 | return ConstantInt::get(Ty: Op0->getType(), V: Offs1.zextOrTrunc(width: OpSize) - |
835 | Offs2.zextOrTrunc(width: OpSize)); |
836 | } |
837 | } |
838 | |
839 | return nullptr; |
840 | } |
841 | |
842 | /// If array indices are not pointer-sized integers, explicitly cast them so |
843 | /// that they aren't implicitly casted by the getelementptr. |
844 | Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops, |
845 | Type *ResultTy, GEPNoWrapFlags NW, |
846 | std::optional<ConstantRange> InRange, |
847 | const DataLayout &DL, const TargetLibraryInfo *TLI) { |
848 | Type *IntIdxTy = DL.getIndexType(PtrTy: ResultTy); |
849 | Type *IntIdxScalarTy = IntIdxTy->getScalarType(); |
850 | |
851 | bool Any = false; |
852 | SmallVector<Constant*, 32> NewIdxs; |
853 | for (unsigned i = 1, e = Ops.size(); i != e; ++i) { |
854 | if ((i == 1 || |
855 | !isa<StructType>(Val: GetElementPtrInst::getIndexedType( |
856 | Ty: SrcElemTy, IdxList: Ops.slice(N: 1, M: i - 1)))) && |
857 | Ops[i]->getType()->getScalarType() != IntIdxScalarTy) { |
858 | Any = true; |
859 | Type *NewType = |
860 | Ops[i]->getType()->isVectorTy() ? IntIdxTy : IntIdxScalarTy; |
861 | Constant *NewIdx = ConstantFoldCastOperand( |
862 | Opcode: CastInst::getCastOpcode(Val: Ops[i], SrcIsSigned: true, Ty: NewType, DstIsSigned: true), C: Ops[i], DestTy: NewType, |
863 | DL); |
864 | if (!NewIdx) |
865 | return nullptr; |
866 | NewIdxs.push_back(Elt: NewIdx); |
867 | } else |
868 | NewIdxs.push_back(Elt: Ops[i]); |
869 | } |
870 | |
871 | if (!Any) |
872 | return nullptr; |
873 | |
874 | Constant *C = |
875 | ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: NewIdxs, NW, InRange); |
876 | return ConstantFoldConstant(C, DL, TLI); |
877 | } |
878 | |
879 | /// If we can symbolically evaluate the GEP constant expression, do so. |
880 | Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP, |
881 | ArrayRef<Constant *> Ops, |
882 | const DataLayout &DL, |
883 | const TargetLibraryInfo *TLI) { |
884 | Type *SrcElemTy = GEP->getSourceElementType(); |
885 | Type *ResTy = GEP->getType(); |
886 | if (!SrcElemTy->isSized() || isa<ScalableVectorType>(Val: SrcElemTy)) |
887 | return nullptr; |
888 | |
889 | if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResultTy: ResTy, NW: GEP->getNoWrapFlags(), |
890 | InRange: GEP->getInRange(), DL, TLI)) |
891 | return C; |
892 | |
893 | Constant *Ptr = Ops[0]; |
894 | if (!Ptr->getType()->isPointerTy()) |
895 | return nullptr; |
896 | |
897 | Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType()); |
898 | |
899 | for (unsigned i = 1, e = Ops.size(); i != e; ++i) |
900 | if (!isa<ConstantInt>(Val: Ops[i]) || !Ops[i]->getType()->isIntegerTy()) |
901 | return nullptr; |
902 | |
903 | unsigned BitWidth = DL.getTypeSizeInBits(Ty: IntIdxTy); |
904 | APInt Offset = APInt( |
905 | BitWidth, |
906 | DL.getIndexedOffsetInType( |
907 | ElemTy: SrcElemTy, Indices: ArrayRef((Value *const *)Ops.data() + 1, Ops.size() - 1)), |
908 | /*isSigned=*/true, /*implicitTrunc=*/true); |
909 | |
910 | std::optional<ConstantRange> InRange = GEP->getInRange(); |
911 | if (InRange) |
912 | InRange = InRange->sextOrTrunc(BitWidth); |
913 | |
914 | // If this is a GEP of a GEP, fold it all into a single GEP. |
915 | GEPNoWrapFlags NW = GEP->getNoWrapFlags(); |
916 | bool Overflow = false; |
917 | while (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr)) { |
918 | NW &= GEP->getNoWrapFlags(); |
919 | |
920 | SmallVector<Value *, 4> NestedOps(llvm::drop_begin(RangeOrContainer: GEP->operands())); |
921 | |
922 | // Do not try the incorporate the sub-GEP if some index is not a number. |
923 | bool AllConstantInt = true; |
924 | for (Value *NestedOp : NestedOps) |
925 | if (!isa<ConstantInt>(Val: NestedOp)) { |
926 | AllConstantInt = false; |
927 | break; |
928 | } |
929 | if (!AllConstantInt) |
930 | break; |
931 | |
932 | // TODO: Try to intersect two inrange attributes? |
933 | if (!InRange) { |
934 | InRange = GEP->getInRange(); |
935 | if (InRange) |
936 | // Adjust inrange by offset until now. |
937 | InRange = InRange->sextOrTrunc(BitWidth).subtract(CI: Offset); |
938 | } |
939 | |
940 | Ptr = cast<Constant>(Val: GEP->getOperand(i_nocapture: 0)); |
941 | SrcElemTy = GEP->getSourceElementType(); |
942 | Offset = Offset.sadd_ov( |
943 | RHS: APInt(BitWidth, DL.getIndexedOffsetInType(ElemTy: SrcElemTy, Indices: NestedOps), |
944 | /*isSigned=*/true, /*implicitTrunc=*/true), |
945 | Overflow); |
946 | } |
947 | |
948 | // Preserving nusw (without inbounds) also requires that the offset |
949 | // additions did not overflow. |
950 | if (NW.hasNoUnsignedSignedWrap() && !NW.isInBounds() && Overflow) |
951 | NW = NW.withoutNoUnsignedSignedWrap(); |
952 | |
953 | // If the base value for this address is a literal integer value, fold the |
954 | // getelementptr to the resulting integer value casted to the pointer type. |
955 | APInt BasePtr(DL.getPointerTypeSizeInBits(Ptr->getType()), 0); |
956 | if (auto *CE = dyn_cast<ConstantExpr>(Val: Ptr)) { |
957 | if (CE->getOpcode() == Instruction::IntToPtr) { |
958 | if (auto *Base = dyn_cast<ConstantInt>(Val: CE->getOperand(i_nocapture: 0))) |
959 | BasePtr = Base->getValue().zextOrTrunc(width: BasePtr.getBitWidth()); |
960 | } |
961 | } |
962 | |
963 | auto *PTy = cast<PointerType>(Val: Ptr->getType()); |
964 | if ((Ptr->isNullValue() || BasePtr != 0) && |
965 | !DL.isNonIntegralPointerType(PT: PTy)) { |
966 | // If the index size is smaller than the pointer size, add to the low |
967 | // bits only. |
968 | BasePtr.insertBits(SubBits: BasePtr.trunc(width: BitWidth) + Offset, bitPosition: 0); |
969 | Constant *C = ConstantInt::get(Context&: Ptr->getContext(), V: BasePtr); |
970 | return ConstantExpr::getIntToPtr(C, Ty: ResTy); |
971 | } |
972 | |
973 | // Try to infer inbounds for GEPs of globals. |
974 | if (!NW.isInBounds() && Offset.isNonNegative()) { |
975 | bool CanBeNull, CanBeFreed; |
976 | uint64_t DerefBytes = |
977 | Ptr->getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed); |
978 | if (DerefBytes != 0 && !CanBeNull && Offset.sle(RHS: DerefBytes)) |
979 | NW |= GEPNoWrapFlags::inBounds(); |
980 | } |
981 | |
982 | // nusw + nneg -> nuw |
983 | if (NW.hasNoUnsignedSignedWrap() && Offset.isNonNegative()) |
984 | NW |= GEPNoWrapFlags::noUnsignedWrap(); |
985 | |
986 | // Otherwise canonicalize this to a single ptradd. |
987 | LLVMContext &Ctx = Ptr->getContext(); |
988 | return ConstantExpr::getGetElementPtr(Ty: Type::getInt8Ty(C&: Ctx), C: Ptr, |
989 | Idx: ConstantInt::get(Context&: Ctx, V: Offset), NW, |
990 | InRange); |
991 | } |
992 | |
993 | /// Attempt to constant fold an instruction with the |
994 | /// specified opcode and operands. If successful, the constant result is |
995 | /// returned, if not, null is returned. Note that this function can fail when |
996 | /// attempting to fold instructions like loads and stores, which have no |
997 | /// constant expression form. |
998 | Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode, |
999 | ArrayRef<Constant *> Ops, |
1000 | const DataLayout &DL, |
1001 | const TargetLibraryInfo *TLI, |
1002 | bool AllowNonDeterministic) { |
1003 | Type *DestTy = InstOrCE->getType(); |
1004 | |
1005 | if (Instruction::isUnaryOp(Opcode)) |
1006 | return ConstantFoldUnaryOpOperand(Opcode, Op: Ops[0], DL); |
1007 | |
1008 | if (Instruction::isBinaryOp(Opcode)) { |
1009 | switch (Opcode) { |
1010 | default: |
1011 | break; |
1012 | case Instruction::FAdd: |
1013 | case Instruction::FSub: |
1014 | case Instruction::FMul: |
1015 | case Instruction::FDiv: |
1016 | case Instruction::FRem: |
1017 | // Handle floating point instructions separately to account for denormals |
1018 | // TODO: If a constant expression is being folded rather than an |
1019 | // instruction, denormals will not be flushed/treated as zero |
1020 | if (const auto *I = dyn_cast<Instruction>(Val: InstOrCE)) { |
1021 | return ConstantFoldFPInstOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL, I, |
1022 | AllowNonDeterministic); |
1023 | } |
1024 | } |
1025 | return ConstantFoldBinaryOpOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL); |
1026 | } |
1027 | |
1028 | if (Instruction::isCast(Opcode)) |
1029 | return ConstantFoldCastOperand(Opcode, C: Ops[0], DestTy, DL); |
1030 | |
1031 | if (auto *GEP = dyn_cast<GEPOperator>(Val: InstOrCE)) { |
1032 | Type *SrcElemTy = GEP->getSourceElementType(); |
1033 | if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy)) |
1034 | return nullptr; |
1035 | |
1036 | if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI)) |
1037 | return C; |
1038 | |
1039 | return ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: Ops.slice(N: 1), |
1040 | NW: GEP->getNoWrapFlags(), |
1041 | InRange: GEP->getInRange()); |
1042 | } |
1043 | |
1044 | if (auto *CE = dyn_cast<ConstantExpr>(Val: InstOrCE)) |
1045 | return CE->getWithOperands(Ops); |
1046 | |
1047 | switch (Opcode) { |
1048 | default: return nullptr; |
1049 | case Instruction::ICmp: |
1050 | case Instruction::FCmp: { |
1051 | auto *C = cast<CmpInst>(Val: InstOrCE); |
1052 | return ConstantFoldCompareInstOperands(Predicate: C->getPredicate(), LHS: Ops[0], RHS: Ops[1], |
1053 | DL, TLI, I: C); |
1054 | } |
1055 | case Instruction::Freeze: |
1056 | return isGuaranteedNotToBeUndefOrPoison(V: Ops[0]) ? Ops[0] : nullptr; |
1057 | case Instruction::Call: |
1058 | if (auto *F = dyn_cast<Function>(Val: Ops.back())) { |
1059 | const auto *Call = cast<CallBase>(Val: InstOrCE); |
1060 | if (canConstantFoldCallTo(Call, F)) |
1061 | return ConstantFoldCall(Call, F, Operands: Ops.slice(N: 0, M: Ops.size() - 1), TLI, |
1062 | AllowNonDeterministic); |
1063 | } |
1064 | return nullptr; |
1065 | case Instruction::Select: |
1066 | return ConstantFoldSelectInstruction(Cond: Ops[0], V1: Ops[1], V2: Ops[2]); |
1067 | case Instruction::ExtractElement: |
1068 | return ConstantExpr::getExtractElement(Vec: Ops[0], Idx: Ops[1]); |
1069 | case Instruction::ExtractValue: |
1070 | return ConstantFoldExtractValueInstruction( |
1071 | Agg: Ops[0], Idxs: cast<ExtractValueInst>(Val: InstOrCE)->getIndices()); |
1072 | case Instruction::InsertElement: |
1073 | return ConstantExpr::getInsertElement(Vec: Ops[0], Elt: Ops[1], Idx: Ops[2]); |
1074 | case Instruction::InsertValue: |
1075 | return ConstantFoldInsertValueInstruction( |
1076 | Agg: Ops[0], Val: Ops[1], Idxs: cast<InsertValueInst>(Val: InstOrCE)->getIndices()); |
1077 | case Instruction::ShuffleVector: |
1078 | return ConstantExpr::getShuffleVector( |
1079 | V1: Ops[0], V2: Ops[1], Mask: cast<ShuffleVectorInst>(Val: InstOrCE)->getShuffleMask()); |
1080 | case Instruction::Load: { |
1081 | const auto *LI = dyn_cast<LoadInst>(Val: InstOrCE); |
1082 | if (LI->isVolatile()) |
1083 | return nullptr; |
1084 | return ConstantFoldLoadFromConstPtr(C: Ops[0], Ty: LI->getType(), DL); |
1085 | } |
1086 | } |
1087 | } |
1088 | |
1089 | } // end anonymous namespace |
1090 | |
1091 | //===----------------------------------------------------------------------===// |
1092 | // Constant Folding public APIs |
1093 | //===----------------------------------------------------------------------===// |
1094 | |
1095 | namespace { |
1096 | |
1097 | Constant * |
1098 | ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL, |
1099 | const TargetLibraryInfo *TLI, |
1100 | SmallDenseMap<Constant *, Constant *> &FoldedOps) { |
1101 | if (!isa<ConstantVector>(Val: C) && !isa<ConstantExpr>(Val: C)) |
1102 | return const_cast<Constant *>(C); |
1103 | |
1104 | SmallVector<Constant *, 8> Ops; |
1105 | for (const Use &OldU : C->operands()) { |
1106 | Constant *OldC = cast<Constant>(Val: &OldU); |
1107 | Constant *NewC = OldC; |
1108 | // Recursively fold the ConstantExpr's operands. If we have already folded |
1109 | // a ConstantExpr, we don't have to process it again. |
1110 | if (isa<ConstantVector>(Val: OldC) || isa<ConstantExpr>(Val: OldC)) { |
1111 | auto It = FoldedOps.find(Val: OldC); |
1112 | if (It == FoldedOps.end()) { |
1113 | NewC = ConstantFoldConstantImpl(C: OldC, DL, TLI, FoldedOps); |
1114 | FoldedOps.insert(KV: {OldC, NewC}); |
1115 | } else { |
1116 | NewC = It->second; |
1117 | } |
1118 | } |
1119 | Ops.push_back(Elt: NewC); |
1120 | } |
1121 | |
1122 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
1123 | if (Constant *Res = ConstantFoldInstOperandsImpl( |
1124 | InstOrCE: CE, Opcode: CE->getOpcode(), Ops, DL, TLI, /*AllowNonDeterministic=*/true)) |
1125 | return Res; |
1126 | return const_cast<Constant *>(C); |
1127 | } |
1128 | |
1129 | assert(isa<ConstantVector>(C)); |
1130 | return ConstantVector::get(V: Ops); |
1131 | } |
1132 | |
1133 | } // end anonymous namespace |
1134 | |
1135 | Constant *llvm::ConstantFoldInstruction(const Instruction *I, |
1136 | const DataLayout &DL, |
1137 | const TargetLibraryInfo *TLI) { |
1138 | // Handle PHI nodes quickly here... |
1139 | if (auto *PN = dyn_cast<PHINode>(Val: I)) { |
1140 | Constant *CommonValue = nullptr; |
1141 | |
1142 | SmallDenseMap<Constant *, Constant *> FoldedOps; |
1143 | for (Value *Incoming : PN->incoming_values()) { |
1144 | // If the incoming value is undef then skip it. Note that while we could |
1145 | // skip the value if it is equal to the phi node itself we choose not to |
1146 | // because that would break the rule that constant folding only applies if |
1147 | // all operands are constants. |
1148 | if (isa<UndefValue>(Val: Incoming)) |
1149 | continue; |
1150 | // If the incoming value is not a constant, then give up. |
1151 | auto *C = dyn_cast<Constant>(Val: Incoming); |
1152 | if (!C) |
1153 | return nullptr; |
1154 | // Fold the PHI's operands. |
1155 | C = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps); |
1156 | // If the incoming value is a different constant to |
1157 | // the one we saw previously, then give up. |
1158 | if (CommonValue && C != CommonValue) |
1159 | return nullptr; |
1160 | CommonValue = C; |
1161 | } |
1162 | |
1163 | // If we reach here, all incoming values are the same constant or undef. |
1164 | return CommonValue ? CommonValue : UndefValue::get(T: PN->getType()); |
1165 | } |
1166 | |
1167 | // Scan the operand list, checking to see if they are all constants, if so, |
1168 | // hand off to ConstantFoldInstOperandsImpl. |
1169 | if (!all_of(Range: I->operands(), P: [](const Use &U) { return isa<Constant>(Val: U); })) |
1170 | return nullptr; |
1171 | |
1172 | SmallDenseMap<Constant *, Constant *> FoldedOps; |
1173 | SmallVector<Constant *, 8> Ops; |
1174 | for (const Use &OpU : I->operands()) { |
1175 | auto *Op = cast<Constant>(Val: &OpU); |
1176 | // Fold the Instruction's operands. |
1177 | Op = ConstantFoldConstantImpl(C: Op, DL, TLI, FoldedOps); |
1178 | Ops.push_back(Elt: Op); |
1179 | } |
1180 | |
1181 | return ConstantFoldInstOperands(I, Ops, DL, TLI); |
1182 | } |
1183 | |
1184 | Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL, |
1185 | const TargetLibraryInfo *TLI) { |
1186 | SmallDenseMap<Constant *, Constant *> FoldedOps; |
1187 | return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps); |
1188 | } |
1189 | |
1190 | Constant *llvm::ConstantFoldInstOperands(const Instruction *I, |
1191 | ArrayRef<Constant *> Ops, |
1192 | const DataLayout &DL, |
1193 | const TargetLibraryInfo *TLI, |
1194 | bool AllowNonDeterministic) { |
1195 | return ConstantFoldInstOperandsImpl(InstOrCE: I, Opcode: I->getOpcode(), Ops, DL, TLI, |
1196 | AllowNonDeterministic); |
1197 | } |
1198 | |
1199 | Constant *llvm::ConstantFoldCompareInstOperands( |
1200 | unsigned IntPredicate, Constant *Ops0, Constant *Ops1, const DataLayout &DL, |
1201 | const TargetLibraryInfo *TLI, const Instruction *I) { |
1202 | CmpInst::Predicate Predicate = (CmpInst::Predicate)IntPredicate; |
1203 | // fold: icmp (inttoptr x), null -> icmp x, 0 |
1204 | // fold: icmp null, (inttoptr x) -> icmp 0, x |
1205 | // fold: icmp (ptrtoint x), 0 -> icmp x, null |
1206 | // fold: icmp 0, (ptrtoint x) -> icmp null, x |
1207 | // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y |
1208 | // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y |
1209 | // |
1210 | // FIXME: The following comment is out of data and the DataLayout is here now. |
1211 | // ConstantExpr::getCompare cannot do this, because it doesn't have DL |
1212 | // around to know if bit truncation is happening. |
1213 | if (auto *CE0 = dyn_cast<ConstantExpr>(Val: Ops0)) { |
1214 | if (Ops1->isNullValue()) { |
1215 | if (CE0->getOpcode() == Instruction::IntToPtr) { |
1216 | Type *IntPtrTy = DL.getIntPtrType(CE0->getType()); |
1217 | // Convert the integer value to the right size to ensure we get the |
1218 | // proper extension or truncation. |
1219 | if (Constant *C = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy, |
1220 | /*IsSigned*/ false, DL)) { |
1221 | Constant *Null = Constant::getNullValue(Ty: C->getType()); |
1222 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI); |
1223 | } |
1224 | } |
1225 | |
1226 | // Only do this transformation if the int is intptrty in size, otherwise |
1227 | // there is a truncation or extension that we aren't modeling. |
1228 | if (CE0->getOpcode() == Instruction::PtrToInt) { |
1229 | Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(i_nocapture: 0)->getType()); |
1230 | if (CE0->getType() == IntPtrTy) { |
1231 | Constant *C = CE0->getOperand(i_nocapture: 0); |
1232 | Constant *Null = Constant::getNullValue(Ty: C->getType()); |
1233 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI); |
1234 | } |
1235 | } |
1236 | } |
1237 | |
1238 | if (auto *CE1 = dyn_cast<ConstantExpr>(Val: Ops1)) { |
1239 | if (CE0->getOpcode() == CE1->getOpcode()) { |
1240 | if (CE0->getOpcode() == Instruction::IntToPtr) { |
1241 | Type *IntPtrTy = DL.getIntPtrType(CE0->getType()); |
1242 | |
1243 | // Convert the integer value to the right size to ensure we get the |
1244 | // proper extension or truncation. |
1245 | Constant *C0 = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy, |
1246 | /*IsSigned*/ false, DL); |
1247 | Constant *C1 = ConstantFoldIntegerCast(C: CE1->getOperand(i_nocapture: 0), DestTy: IntPtrTy, |
1248 | /*IsSigned*/ false, DL); |
1249 | if (C0 && C1) |
1250 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C0, Ops1: C1, DL, TLI); |
1251 | } |
1252 | |
1253 | // Only do this transformation if the int is intptrty in size, otherwise |
1254 | // there is a truncation or extension that we aren't modeling. |
1255 | if (CE0->getOpcode() == Instruction::PtrToInt) { |
1256 | Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(i_nocapture: 0)->getType()); |
1257 | if (CE0->getType() == IntPtrTy && |
1258 | CE0->getOperand(i_nocapture: 0)->getType() == CE1->getOperand(i_nocapture: 0)->getType()) { |
1259 | return ConstantFoldCompareInstOperands( |
1260 | IntPredicate: Predicate, Ops0: CE0->getOperand(i_nocapture: 0), Ops1: CE1->getOperand(i_nocapture: 0), DL, TLI); |
1261 | } |
1262 | } |
1263 | } |
1264 | } |
1265 | |
1266 | // Convert pointer comparison (base+offset1) pred (base+offset2) into |
1267 | // offset1 pred offset2, for the case where the offset is inbounds. This |
1268 | // only works for equality and unsigned comparison, as inbounds permits |
1269 | // crossing the sign boundary. However, the offset comparison itself is |
1270 | // signed. |
1271 | if (Ops0->getType()->isPointerTy() && !ICmpInst::isSigned(predicate: Predicate)) { |
1272 | unsigned IndexWidth = DL.getIndexTypeSizeInBits(Ty: Ops0->getType()); |
1273 | APInt Offset0(IndexWidth, 0); |
1274 | bool IsEqPred = ICmpInst::isEquality(P: Predicate); |
1275 | Value *Stripped0 = Ops0->stripAndAccumulateConstantOffsets( |
1276 | DL, Offset&: Offset0, /*AllowNonInbounds=*/IsEqPred, |
1277 | /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr, |
1278 | /*LookThroughIntToPtr=*/IsEqPred); |
1279 | APInt Offset1(IndexWidth, 0); |
1280 | Value *Stripped1 = Ops1->stripAndAccumulateConstantOffsets( |
1281 | DL, Offset&: Offset1, /*AllowNonInbounds=*/IsEqPred, |
1282 | /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr, |
1283 | /*LookThroughIntToPtr=*/IsEqPred); |
1284 | if (Stripped0 == Stripped1) |
1285 | return ConstantInt::getBool( |
1286 | Context&: Ops0->getContext(), |
1287 | V: ICmpInst::compare(LHS: Offset0, RHS: Offset1, |
1288 | Pred: ICmpInst::getSignedPredicate(Pred: Predicate))); |
1289 | } |
1290 | } else if (isa<ConstantExpr>(Val: Ops1)) { |
1291 | // If RHS is a constant expression, but the left side isn't, swap the |
1292 | // operands and try again. |
1293 | Predicate = ICmpInst::getSwappedPredicate(pred: Predicate); |
1294 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: Ops1, Ops1: Ops0, DL, TLI); |
1295 | } |
1296 | |
1297 | if (CmpInst::isFPPredicate(P: Predicate)) { |
1298 | // Flush any denormal constant float input according to denormal handling |
1299 | // mode. |
1300 | Ops0 = FlushFPConstant(Operand: Ops0, I, /*IsOutput=*/false); |
1301 | if (!Ops0) |
1302 | return nullptr; |
1303 | Ops1 = FlushFPConstant(Operand: Ops1, I, /*IsOutput=*/false); |
1304 | if (!Ops1) |
1305 | return nullptr; |
1306 | } |
1307 | |
1308 | return ConstantFoldCompareInstruction(Predicate, C1: Ops0, C2: Ops1); |
1309 | } |
1310 | |
1311 | Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op, |
1312 | const DataLayout &DL) { |
1313 | assert(Instruction::isUnaryOp(Opcode)); |
1314 | |
1315 | return ConstantFoldUnaryInstruction(Opcode, V: Op); |
1316 | } |
1317 | |
1318 | Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, |
1319 | Constant *RHS, |
1320 | const DataLayout &DL) { |
1321 | assert(Instruction::isBinaryOp(Opcode)); |
1322 | if (isa<ConstantExpr>(Val: LHS) || isa<ConstantExpr>(Val: RHS)) |
1323 | if (Constant *C = SymbolicallyEvaluateBinop(Opc: Opcode, Op0: LHS, Op1: RHS, DL)) |
1324 | return C; |
1325 | |
1326 | if (ConstantExpr::isDesirableBinOp(Opcode)) |
1327 | return ConstantExpr::get(Opcode, C1: LHS, C2: RHS); |
1328 | return ConstantFoldBinaryInstruction(Opcode, V1: LHS, V2: RHS); |
1329 | } |
1330 | |
1331 | static ConstantFP *flushDenormalConstant(Type *Ty, const APFloat &APF, |
1332 | DenormalMode::DenormalModeKind Mode) { |
1333 | switch (Mode) { |
1334 | case DenormalMode::Dynamic: |
1335 | return nullptr; |
1336 | case DenormalMode::IEEE: |
1337 | return ConstantFP::get(Context&: Ty->getContext(), V: APF); |
1338 | case DenormalMode::PreserveSign: |
1339 | return ConstantFP::get( |
1340 | Context&: Ty->getContext(), |
1341 | V: APFloat::getZero(Sem: APF.getSemantics(), Negative: APF.isNegative())); |
1342 | case DenormalMode::PositiveZero: |
1343 | return ConstantFP::get(Context&: Ty->getContext(), |
1344 | V: APFloat::getZero(Sem: APF.getSemantics(), Negative: false)); |
1345 | default: |
1346 | break; |
1347 | } |
1348 | |
1349 | llvm_unreachable("unknown denormal mode" ); |
1350 | } |
1351 | |
1352 | /// Return the denormal mode that can be assumed when executing a floating point |
1353 | /// operation at \p CtxI. |
1354 | static DenormalMode getInstrDenormalMode(const Instruction *CtxI, Type *Ty) { |
1355 | if (!CtxI || !CtxI->getParent() || !CtxI->getFunction()) |
1356 | return DenormalMode::getDynamic(); |
1357 | return CtxI->getFunction()->getDenormalMode(FPType: Ty->getFltSemantics()); |
1358 | } |
1359 | |
1360 | static ConstantFP *flushDenormalConstantFP(ConstantFP *CFP, |
1361 | const Instruction *Inst, |
1362 | bool IsOutput) { |
1363 | const APFloat &APF = CFP->getValueAPF(); |
1364 | if (!APF.isDenormal()) |
1365 | return CFP; |
1366 | |
1367 | DenormalMode Mode = getInstrDenormalMode(CtxI: Inst, Ty: CFP->getType()); |
1368 | return flushDenormalConstant(Ty: CFP->getType(), APF, |
1369 | Mode: IsOutput ? Mode.Output : Mode.Input); |
1370 | } |
1371 | |
1372 | Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *Inst, |
1373 | bool IsOutput) { |
1374 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Operand)) |
1375 | return flushDenormalConstantFP(CFP, Inst, IsOutput); |
1376 | |
1377 | if (isa<ConstantAggregateZero, UndefValue, ConstantExpr>(Val: Operand)) |
1378 | return Operand; |
1379 | |
1380 | Type *Ty = Operand->getType(); |
1381 | VectorType *VecTy = dyn_cast<VectorType>(Val: Ty); |
1382 | if (VecTy) { |
1383 | if (auto *Splat = dyn_cast_or_null<ConstantFP>(Val: Operand->getSplatValue())) { |
1384 | ConstantFP *Folded = flushDenormalConstantFP(CFP: Splat, Inst, IsOutput); |
1385 | if (!Folded) |
1386 | return nullptr; |
1387 | return ConstantVector::getSplat(EC: VecTy->getElementCount(), Elt: Folded); |
1388 | } |
1389 | |
1390 | Ty = VecTy->getElementType(); |
1391 | } |
1392 | |
1393 | if (const auto *CV = dyn_cast<ConstantVector>(Val: Operand)) { |
1394 | SmallVector<Constant *, 16> NewElts; |
1395 | for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { |
1396 | Constant *Element = CV->getAggregateElement(Elt: i); |
1397 | if (isa<UndefValue>(Val: Element)) { |
1398 | NewElts.push_back(Elt: Element); |
1399 | continue; |
1400 | } |
1401 | |
1402 | ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Element); |
1403 | if (!CFP) |
1404 | return nullptr; |
1405 | |
1406 | ConstantFP *Folded = flushDenormalConstantFP(CFP, Inst, IsOutput); |
1407 | if (!Folded) |
1408 | return nullptr; |
1409 | NewElts.push_back(Elt: Folded); |
1410 | } |
1411 | |
1412 | return ConstantVector::get(V: NewElts); |
1413 | } |
1414 | |
1415 | if (const auto *CDV = dyn_cast<ConstantDataVector>(Val: Operand)) { |
1416 | SmallVector<Constant *, 16> NewElts; |
1417 | for (unsigned I = 0, E = CDV->getNumElements(); I < E; ++I) { |
1418 | const APFloat &Elt = CDV->getElementAsAPFloat(i: I); |
1419 | if (!Elt.isDenormal()) { |
1420 | NewElts.push_back(Elt: ConstantFP::get(Ty, V: Elt)); |
1421 | } else { |
1422 | DenormalMode Mode = getInstrDenormalMode(CtxI: Inst, Ty); |
1423 | ConstantFP *Folded = |
1424 | flushDenormalConstant(Ty, APF: Elt, Mode: IsOutput ? Mode.Output : Mode.Input); |
1425 | if (!Folded) |
1426 | return nullptr; |
1427 | NewElts.push_back(Elt: Folded); |
1428 | } |
1429 | } |
1430 | |
1431 | return ConstantVector::get(V: NewElts); |
1432 | } |
1433 | |
1434 | return nullptr; |
1435 | } |
1436 | |
1437 | Constant *llvm::ConstantFoldFPInstOperands(unsigned Opcode, Constant *LHS, |
1438 | Constant *RHS, const DataLayout &DL, |
1439 | const Instruction *I, |
1440 | bool AllowNonDeterministic) { |
1441 | if (Instruction::isBinaryOp(Opcode)) { |
1442 | // Flush denormal inputs if needed. |
1443 | Constant *Op0 = FlushFPConstant(Operand: LHS, Inst: I, /* IsOutput */ false); |
1444 | if (!Op0) |
1445 | return nullptr; |
1446 | Constant *Op1 = FlushFPConstant(Operand: RHS, Inst: I, /* IsOutput */ false); |
1447 | if (!Op1) |
1448 | return nullptr; |
1449 | |
1450 | // If nsz or an algebraic FMF flag is set, the result of the FP operation |
1451 | // may change due to future optimization. Don't constant fold them if |
1452 | // non-deterministic results are not allowed. |
1453 | if (!AllowNonDeterministic) |
1454 | if (auto *FP = dyn_cast_or_null<FPMathOperator>(Val: I)) |
1455 | if (FP->hasNoSignedZeros() || FP->hasAllowReassoc() || |
1456 | FP->hasAllowContract() || FP->hasAllowReciprocal()) |
1457 | return nullptr; |
1458 | |
1459 | // Calculate constant result. |
1460 | Constant *C = ConstantFoldBinaryOpOperands(Opcode, LHS: Op0, RHS: Op1, DL); |
1461 | if (!C) |
1462 | return nullptr; |
1463 | |
1464 | // Flush denormal output if needed. |
1465 | C = FlushFPConstant(Operand: C, Inst: I, /* IsOutput */ true); |
1466 | if (!C) |
1467 | return nullptr; |
1468 | |
1469 | // The precise NaN value is non-deterministic. |
1470 | if (!AllowNonDeterministic && C->isNaN()) |
1471 | return nullptr; |
1472 | |
1473 | return C; |
1474 | } |
1475 | // If instruction lacks a parent/function and the denormal mode cannot be |
1476 | // determined, use the default (IEEE). |
1477 | return ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL); |
1478 | } |
1479 | |
1480 | Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C, |
1481 | Type *DestTy, const DataLayout &DL) { |
1482 | assert(Instruction::isCast(Opcode)); |
1483 | switch (Opcode) { |
1484 | default: |
1485 | llvm_unreachable("Missing case" ); |
1486 | case Instruction::PtrToInt: |
1487 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
1488 | Constant *FoldedValue = nullptr; |
1489 | // If the input is a inttoptr, eliminate the pair. This requires knowing |
1490 | // the width of a pointer, so it can't be done in ConstantExpr::getCast. |
1491 | if (CE->getOpcode() == Instruction::IntToPtr) { |
1492 | // zext/trunc the inttoptr to pointer size. |
1493 | FoldedValue = ConstantFoldIntegerCast(C: CE->getOperand(i_nocapture: 0), |
1494 | DestTy: DL.getIntPtrType(CE->getType()), |
1495 | /*IsSigned=*/false, DL); |
1496 | } else if (auto *GEP = dyn_cast<GEPOperator>(Val: CE)) { |
1497 | // If we have GEP, we can perform the following folds: |
1498 | // (ptrtoint (gep null, x)) -> x |
1499 | // (ptrtoint (gep (gep null, x), y) -> x + y, etc. |
1500 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType()); |
1501 | APInt BaseOffset(BitWidth, 0); |
1502 | auto *Base = cast<Constant>(Val: GEP->stripAndAccumulateConstantOffsets( |
1503 | DL, Offset&: BaseOffset, /*AllowNonInbounds=*/true)); |
1504 | if (Base->isNullValue()) { |
1505 | FoldedValue = ConstantInt::get(Context&: CE->getContext(), V: BaseOffset); |
1506 | } else { |
1507 | // ptrtoint (gep i8, Ptr, (sub 0, V)) -> sub (ptrtoint Ptr), V |
1508 | if (GEP->getNumIndices() == 1 && |
1509 | GEP->getSourceElementType()->isIntegerTy(Bitwidth: 8)) { |
1510 | auto *Ptr = cast<Constant>(Val: GEP->getPointerOperand()); |
1511 | auto *Sub = dyn_cast<ConstantExpr>(Val: GEP->getOperand(i_nocapture: 1)); |
1512 | Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType()); |
1513 | if (Sub && Sub->getType() == IntIdxTy && |
1514 | Sub->getOpcode() == Instruction::Sub && |
1515 | Sub->getOperand(i_nocapture: 0)->isNullValue()) |
1516 | FoldedValue = ConstantExpr::getSub( |
1517 | C1: ConstantExpr::getPtrToInt(C: Ptr, Ty: IntIdxTy), C2: Sub->getOperand(i_nocapture: 1)); |
1518 | } |
1519 | } |
1520 | } |
1521 | if (FoldedValue) { |
1522 | // Do a zext or trunc to get to the ptrtoint dest size. |
1523 | return ConstantFoldIntegerCast(C: FoldedValue, DestTy, /*IsSigned=*/false, |
1524 | DL); |
1525 | } |
1526 | } |
1527 | break; |
1528 | case Instruction::IntToPtr: |
1529 | // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if |
1530 | // the int size is >= the ptr size and the address spaces are the same. |
1531 | // This requires knowing the width of a pointer, so it can't be done in |
1532 | // ConstantExpr::getCast. |
1533 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
1534 | if (CE->getOpcode() == Instruction::PtrToInt) { |
1535 | Constant *SrcPtr = CE->getOperand(i_nocapture: 0); |
1536 | unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType()); |
1537 | unsigned MidIntSize = CE->getType()->getScalarSizeInBits(); |
1538 | |
1539 | if (MidIntSize >= SrcPtrSize) { |
1540 | unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace(); |
1541 | if (SrcAS == DestTy->getPointerAddressSpace()) |
1542 | return FoldBitCast(C: CE->getOperand(i_nocapture: 0), DestTy, DL); |
1543 | } |
1544 | } |
1545 | } |
1546 | break; |
1547 | case Instruction::Trunc: |
1548 | case Instruction::ZExt: |
1549 | case Instruction::SExt: |
1550 | case Instruction::FPTrunc: |
1551 | case Instruction::FPExt: |
1552 | case Instruction::UIToFP: |
1553 | case Instruction::SIToFP: |
1554 | case Instruction::FPToUI: |
1555 | case Instruction::FPToSI: |
1556 | case Instruction::AddrSpaceCast: |
1557 | break; |
1558 | case Instruction::BitCast: |
1559 | return FoldBitCast(C, DestTy, DL); |
1560 | } |
1561 | |
1562 | if (ConstantExpr::isDesirableCastOp(Opcode)) |
1563 | return ConstantExpr::getCast(ops: Opcode, C, Ty: DestTy); |
1564 | return ConstantFoldCastInstruction(opcode: Opcode, V: C, DestTy); |
1565 | } |
1566 | |
1567 | Constant *llvm::ConstantFoldIntegerCast(Constant *C, Type *DestTy, |
1568 | bool IsSigned, const DataLayout &DL) { |
1569 | Type *SrcTy = C->getType(); |
1570 | if (SrcTy == DestTy) |
1571 | return C; |
1572 | if (SrcTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits()) |
1573 | return ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy, DL); |
1574 | if (IsSigned) |
1575 | return ConstantFoldCastOperand(Opcode: Instruction::SExt, C, DestTy, DL); |
1576 | return ConstantFoldCastOperand(Opcode: Instruction::ZExt, C, DestTy, DL); |
1577 | } |
1578 | |
1579 | //===----------------------------------------------------------------------===// |
1580 | // Constant Folding for Calls |
1581 | // |
1582 | |
1583 | bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) { |
1584 | if (Call->isNoBuiltin()) |
1585 | return false; |
1586 | if (Call->getFunctionType() != F->getFunctionType()) |
1587 | return false; |
1588 | |
1589 | // Allow FP calls (both libcalls and intrinsics) to avoid being folded. |
1590 | // This can be useful for GPU targets or in cross-compilation scenarios |
1591 | // when the exact target FP behaviour is required, and the host compiler's |
1592 | // behaviour may be slightly different from the device's run-time behaviour. |
1593 | if (DisableFPCallFolding && (F->getReturnType()->isFloatingPointTy() || |
1594 | any_of(Range: F->args(), P: [](const Argument &Arg) { |
1595 | return Arg.getType()->isFloatingPointTy(); |
1596 | }))) |
1597 | return false; |
1598 | |
1599 | switch (F->getIntrinsicID()) { |
1600 | // Operations that do not operate floating-point numbers and do not depend on |
1601 | // FP environment can be folded even in strictfp functions. |
1602 | case Intrinsic::bswap: |
1603 | case Intrinsic::ctpop: |
1604 | case Intrinsic::ctlz: |
1605 | case Intrinsic::cttz: |
1606 | case Intrinsic::fshl: |
1607 | case Intrinsic::fshr: |
1608 | case Intrinsic::launder_invariant_group: |
1609 | case Intrinsic::strip_invariant_group: |
1610 | case Intrinsic::masked_load: |
1611 | case Intrinsic::get_active_lane_mask: |
1612 | case Intrinsic::abs: |
1613 | case Intrinsic::smax: |
1614 | case Intrinsic::smin: |
1615 | case Intrinsic::umax: |
1616 | case Intrinsic::umin: |
1617 | case Intrinsic::scmp: |
1618 | case Intrinsic::ucmp: |
1619 | case Intrinsic::sadd_with_overflow: |
1620 | case Intrinsic::uadd_with_overflow: |
1621 | case Intrinsic::ssub_with_overflow: |
1622 | case Intrinsic::usub_with_overflow: |
1623 | case Intrinsic::smul_with_overflow: |
1624 | case Intrinsic::umul_with_overflow: |
1625 | case Intrinsic::sadd_sat: |
1626 | case Intrinsic::uadd_sat: |
1627 | case Intrinsic::ssub_sat: |
1628 | case Intrinsic::usub_sat: |
1629 | case Intrinsic::smul_fix: |
1630 | case Intrinsic::smul_fix_sat: |
1631 | case Intrinsic::bitreverse: |
1632 | case Intrinsic::is_constant: |
1633 | case Intrinsic::vector_reduce_add: |
1634 | case Intrinsic::vector_reduce_mul: |
1635 | case Intrinsic::vector_reduce_and: |
1636 | case Intrinsic::vector_reduce_or: |
1637 | case Intrinsic::vector_reduce_xor: |
1638 | case Intrinsic::vector_reduce_smin: |
1639 | case Intrinsic::vector_reduce_smax: |
1640 | case Intrinsic::vector_reduce_umin: |
1641 | case Intrinsic::vector_reduce_umax: |
1642 | case Intrinsic::vector_extract: |
1643 | case Intrinsic::vector_insert: |
1644 | case Intrinsic::vector_interleave2: |
1645 | case Intrinsic::vector_deinterleave2: |
1646 | // Target intrinsics |
1647 | case Intrinsic::amdgcn_perm: |
1648 | case Intrinsic::amdgcn_wave_reduce_umin: |
1649 | case Intrinsic::amdgcn_wave_reduce_umax: |
1650 | case Intrinsic::amdgcn_s_wqm: |
1651 | case Intrinsic::amdgcn_s_quadmask: |
1652 | case Intrinsic::amdgcn_s_bitreplicate: |
1653 | case Intrinsic::arm_mve_vctp8: |
1654 | case Intrinsic::arm_mve_vctp16: |
1655 | case Intrinsic::arm_mve_vctp32: |
1656 | case Intrinsic::arm_mve_vctp64: |
1657 | case Intrinsic::aarch64_sve_convert_from_svbool: |
1658 | // WebAssembly float semantics are always known |
1659 | case Intrinsic::wasm_trunc_signed: |
1660 | case Intrinsic::wasm_trunc_unsigned: |
1661 | return true; |
1662 | |
1663 | // Floating point operations cannot be folded in strictfp functions in |
1664 | // general case. They can be folded if FP environment is known to compiler. |
1665 | case Intrinsic::minnum: |
1666 | case Intrinsic::maxnum: |
1667 | case Intrinsic::minimum: |
1668 | case Intrinsic::maximum: |
1669 | case Intrinsic::minimumnum: |
1670 | case Intrinsic::maximumnum: |
1671 | case Intrinsic::log: |
1672 | case Intrinsic::log2: |
1673 | case Intrinsic::log10: |
1674 | case Intrinsic::exp: |
1675 | case Intrinsic::exp2: |
1676 | case Intrinsic::exp10: |
1677 | case Intrinsic::sqrt: |
1678 | case Intrinsic::sin: |
1679 | case Intrinsic::cos: |
1680 | case Intrinsic::sincos: |
1681 | case Intrinsic::sinh: |
1682 | case Intrinsic::cosh: |
1683 | case Intrinsic::atan: |
1684 | case Intrinsic::pow: |
1685 | case Intrinsic::powi: |
1686 | case Intrinsic::ldexp: |
1687 | case Intrinsic::fma: |
1688 | case Intrinsic::fmuladd: |
1689 | case Intrinsic::frexp: |
1690 | case Intrinsic::fptoui_sat: |
1691 | case Intrinsic::fptosi_sat: |
1692 | case Intrinsic::convert_from_fp16: |
1693 | case Intrinsic::convert_to_fp16: |
1694 | case Intrinsic::amdgcn_cos: |
1695 | case Intrinsic::amdgcn_cubeid: |
1696 | case Intrinsic::amdgcn_cubema: |
1697 | case Intrinsic::amdgcn_cubesc: |
1698 | case Intrinsic::amdgcn_cubetc: |
1699 | case Intrinsic::amdgcn_fmul_legacy: |
1700 | case Intrinsic::amdgcn_fma_legacy: |
1701 | case Intrinsic::amdgcn_fract: |
1702 | case Intrinsic::amdgcn_sin: |
1703 | // The intrinsics below depend on rounding mode in MXCSR. |
1704 | case Intrinsic::x86_sse_cvtss2si: |
1705 | case Intrinsic::x86_sse_cvtss2si64: |
1706 | case Intrinsic::x86_sse_cvttss2si: |
1707 | case Intrinsic::x86_sse_cvttss2si64: |
1708 | case Intrinsic::x86_sse2_cvtsd2si: |
1709 | case Intrinsic::x86_sse2_cvtsd2si64: |
1710 | case Intrinsic::x86_sse2_cvttsd2si: |
1711 | case Intrinsic::x86_sse2_cvttsd2si64: |
1712 | case Intrinsic::x86_avx512_vcvtss2si32: |
1713 | case Intrinsic::x86_avx512_vcvtss2si64: |
1714 | case Intrinsic::x86_avx512_cvttss2si: |
1715 | case Intrinsic::x86_avx512_cvttss2si64: |
1716 | case Intrinsic::x86_avx512_vcvtsd2si32: |
1717 | case Intrinsic::x86_avx512_vcvtsd2si64: |
1718 | case Intrinsic::x86_avx512_cvttsd2si: |
1719 | case Intrinsic::x86_avx512_cvttsd2si64: |
1720 | case Intrinsic::x86_avx512_vcvtss2usi32: |
1721 | case Intrinsic::x86_avx512_vcvtss2usi64: |
1722 | case Intrinsic::x86_avx512_cvttss2usi: |
1723 | case Intrinsic::x86_avx512_cvttss2usi64: |
1724 | case Intrinsic::x86_avx512_vcvtsd2usi32: |
1725 | case Intrinsic::x86_avx512_vcvtsd2usi64: |
1726 | case Intrinsic::x86_avx512_cvttsd2usi: |
1727 | case Intrinsic::x86_avx512_cvttsd2usi64: |
1728 | |
1729 | // NVVM FMax intrinsics |
1730 | case Intrinsic::nvvm_fmax_d: |
1731 | case Intrinsic::nvvm_fmax_f: |
1732 | case Intrinsic::nvvm_fmax_ftz_f: |
1733 | case Intrinsic::nvvm_fmax_ftz_nan_f: |
1734 | case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f: |
1735 | case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f: |
1736 | case Intrinsic::nvvm_fmax_nan_f: |
1737 | case Intrinsic::nvvm_fmax_nan_xorsign_abs_f: |
1738 | case Intrinsic::nvvm_fmax_xorsign_abs_f: |
1739 | |
1740 | // NVVM FMin intrinsics |
1741 | case Intrinsic::nvvm_fmin_d: |
1742 | case Intrinsic::nvvm_fmin_f: |
1743 | case Intrinsic::nvvm_fmin_ftz_f: |
1744 | case Intrinsic::nvvm_fmin_ftz_nan_f: |
1745 | case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f: |
1746 | case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f: |
1747 | case Intrinsic::nvvm_fmin_nan_f: |
1748 | case Intrinsic::nvvm_fmin_nan_xorsign_abs_f: |
1749 | case Intrinsic::nvvm_fmin_xorsign_abs_f: |
1750 | |
1751 | // NVVM float/double to int32/uint32 conversion intrinsics |
1752 | case Intrinsic::nvvm_f2i_rm: |
1753 | case Intrinsic::nvvm_f2i_rn: |
1754 | case Intrinsic::nvvm_f2i_rp: |
1755 | case Intrinsic::nvvm_f2i_rz: |
1756 | case Intrinsic::nvvm_f2i_rm_ftz: |
1757 | case Intrinsic::nvvm_f2i_rn_ftz: |
1758 | case Intrinsic::nvvm_f2i_rp_ftz: |
1759 | case Intrinsic::nvvm_f2i_rz_ftz: |
1760 | case Intrinsic::nvvm_f2ui_rm: |
1761 | case Intrinsic::nvvm_f2ui_rn: |
1762 | case Intrinsic::nvvm_f2ui_rp: |
1763 | case Intrinsic::nvvm_f2ui_rz: |
1764 | case Intrinsic::nvvm_f2ui_rm_ftz: |
1765 | case Intrinsic::nvvm_f2ui_rn_ftz: |
1766 | case Intrinsic::nvvm_f2ui_rp_ftz: |
1767 | case Intrinsic::nvvm_f2ui_rz_ftz: |
1768 | case Intrinsic::nvvm_d2i_rm: |
1769 | case Intrinsic::nvvm_d2i_rn: |
1770 | case Intrinsic::nvvm_d2i_rp: |
1771 | case Intrinsic::nvvm_d2i_rz: |
1772 | case Intrinsic::nvvm_d2ui_rm: |
1773 | case Intrinsic::nvvm_d2ui_rn: |
1774 | case Intrinsic::nvvm_d2ui_rp: |
1775 | case Intrinsic::nvvm_d2ui_rz: |
1776 | |
1777 | // NVVM float/double to int64/uint64 conversion intrinsics |
1778 | case Intrinsic::nvvm_f2ll_rm: |
1779 | case Intrinsic::nvvm_f2ll_rn: |
1780 | case Intrinsic::nvvm_f2ll_rp: |
1781 | case Intrinsic::nvvm_f2ll_rz: |
1782 | case Intrinsic::nvvm_f2ll_rm_ftz: |
1783 | case Intrinsic::nvvm_f2ll_rn_ftz: |
1784 | case Intrinsic::nvvm_f2ll_rp_ftz: |
1785 | case Intrinsic::nvvm_f2ll_rz_ftz: |
1786 | case Intrinsic::nvvm_f2ull_rm: |
1787 | case Intrinsic::nvvm_f2ull_rn: |
1788 | case Intrinsic::nvvm_f2ull_rp: |
1789 | case Intrinsic::nvvm_f2ull_rz: |
1790 | case Intrinsic::nvvm_f2ull_rm_ftz: |
1791 | case Intrinsic::nvvm_f2ull_rn_ftz: |
1792 | case Intrinsic::nvvm_f2ull_rp_ftz: |
1793 | case Intrinsic::nvvm_f2ull_rz_ftz: |
1794 | case Intrinsic::nvvm_d2ll_rm: |
1795 | case Intrinsic::nvvm_d2ll_rn: |
1796 | case Intrinsic::nvvm_d2ll_rp: |
1797 | case Intrinsic::nvvm_d2ll_rz: |
1798 | case Intrinsic::nvvm_d2ull_rm: |
1799 | case Intrinsic::nvvm_d2ull_rn: |
1800 | case Intrinsic::nvvm_d2ull_rp: |
1801 | case Intrinsic::nvvm_d2ull_rz: |
1802 | return !Call->isStrictFP(); |
1803 | |
1804 | // Sign operations are actually bitwise operations, they do not raise |
1805 | // exceptions even for SNANs. |
1806 | case Intrinsic::fabs: |
1807 | case Intrinsic::copysign: |
1808 | case Intrinsic::is_fpclass: |
1809 | // Non-constrained variants of rounding operations means default FP |
1810 | // environment, they can be folded in any case. |
1811 | case Intrinsic::ceil: |
1812 | case Intrinsic::floor: |
1813 | case Intrinsic::round: |
1814 | case Intrinsic::roundeven: |
1815 | case Intrinsic::trunc: |
1816 | case Intrinsic::nearbyint: |
1817 | case Intrinsic::rint: |
1818 | case Intrinsic::canonicalize: |
1819 | // Constrained intrinsics can be folded if FP environment is known |
1820 | // to compiler. |
1821 | case Intrinsic::experimental_constrained_fma: |
1822 | case Intrinsic::experimental_constrained_fmuladd: |
1823 | case Intrinsic::experimental_constrained_fadd: |
1824 | case Intrinsic::experimental_constrained_fsub: |
1825 | case Intrinsic::experimental_constrained_fmul: |
1826 | case Intrinsic::experimental_constrained_fdiv: |
1827 | case Intrinsic::experimental_constrained_frem: |
1828 | case Intrinsic::experimental_constrained_ceil: |
1829 | case Intrinsic::experimental_constrained_floor: |
1830 | case Intrinsic::experimental_constrained_round: |
1831 | case Intrinsic::experimental_constrained_roundeven: |
1832 | case Intrinsic::experimental_constrained_trunc: |
1833 | case Intrinsic::experimental_constrained_nearbyint: |
1834 | case Intrinsic::experimental_constrained_rint: |
1835 | case Intrinsic::experimental_constrained_fcmp: |
1836 | case Intrinsic::experimental_constrained_fcmps: |
1837 | return true; |
1838 | default: |
1839 | return false; |
1840 | case Intrinsic::not_intrinsic: break; |
1841 | } |
1842 | |
1843 | if (!F->hasName() || Call->isStrictFP()) |
1844 | return false; |
1845 | |
1846 | // In these cases, the check of the length is required. We don't want to |
1847 | // return true for a name like "cos\0blah" which strcmp would return equal to |
1848 | // "cos", but has length 8. |
1849 | StringRef Name = F->getName(); |
1850 | switch (Name[0]) { |
1851 | default: |
1852 | return false; |
1853 | case 'a': |
1854 | return Name == "acos" || Name == "acosf" || |
1855 | Name == "asin" || Name == "asinf" || |
1856 | Name == "atan" || Name == "atanf" || |
1857 | Name == "atan2" || Name == "atan2f" ; |
1858 | case 'c': |
1859 | return Name == "ceil" || Name == "ceilf" || |
1860 | Name == "cos" || Name == "cosf" || |
1861 | Name == "cosh" || Name == "coshf" ; |
1862 | case 'e': |
1863 | return Name == "exp" || Name == "expf" || Name == "exp2" || |
1864 | Name == "exp2f" || Name == "erf" || Name == "erff" ; |
1865 | case 'f': |
1866 | return Name == "fabs" || Name == "fabsf" || |
1867 | Name == "floor" || Name == "floorf" || |
1868 | Name == "fmod" || Name == "fmodf" ; |
1869 | case 'i': |
1870 | return Name == "ilogb" || Name == "ilogbf" ; |
1871 | case 'l': |
1872 | return Name == "log" || Name == "logf" || Name == "logl" || |
1873 | Name == "log2" || Name == "log2f" || Name == "log10" || |
1874 | Name == "log10f" || Name == "logb" || Name == "logbf" || |
1875 | Name == "log1p" || Name == "log1pf" ; |
1876 | case 'n': |
1877 | return Name == "nearbyint" || Name == "nearbyintf" ; |
1878 | case 'p': |
1879 | return Name == "pow" || Name == "powf" ; |
1880 | case 'r': |
1881 | return Name == "remainder" || Name == "remainderf" || |
1882 | Name == "rint" || Name == "rintf" || |
1883 | Name == "round" || Name == "roundf" ; |
1884 | case 's': |
1885 | return Name == "sin" || Name == "sinf" || |
1886 | Name == "sinh" || Name == "sinhf" || |
1887 | Name == "sqrt" || Name == "sqrtf" ; |
1888 | case 't': |
1889 | return Name == "tan" || Name == "tanf" || |
1890 | Name == "tanh" || Name == "tanhf" || |
1891 | Name == "trunc" || Name == "truncf" ; |
1892 | case '_': |
1893 | // Check for various function names that get used for the math functions |
1894 | // when the header files are preprocessed with the macro |
1895 | // __FINITE_MATH_ONLY__ enabled. |
1896 | // The '12' here is the length of the shortest name that can match. |
1897 | // We need to check the size before looking at Name[1] and Name[2] |
1898 | // so we may as well check a limit that will eliminate mismatches. |
1899 | if (Name.size() < 12 || Name[1] != '_') |
1900 | return false; |
1901 | switch (Name[2]) { |
1902 | default: |
1903 | return false; |
1904 | case 'a': |
1905 | return Name == "__acos_finite" || Name == "__acosf_finite" || |
1906 | Name == "__asin_finite" || Name == "__asinf_finite" || |
1907 | Name == "__atan2_finite" || Name == "__atan2f_finite" ; |
1908 | case 'c': |
1909 | return Name == "__cosh_finite" || Name == "__coshf_finite" ; |
1910 | case 'e': |
1911 | return Name == "__exp_finite" || Name == "__expf_finite" || |
1912 | Name == "__exp2_finite" || Name == "__exp2f_finite" ; |
1913 | case 'l': |
1914 | return Name == "__log_finite" || Name == "__logf_finite" || |
1915 | Name == "__log10_finite" || Name == "__log10f_finite" ; |
1916 | case 'p': |
1917 | return Name == "__pow_finite" || Name == "__powf_finite" ; |
1918 | case 's': |
1919 | return Name == "__sinh_finite" || Name == "__sinhf_finite" ; |
1920 | } |
1921 | } |
1922 | } |
1923 | |
1924 | namespace { |
1925 | |
1926 | Constant *GetConstantFoldFPValue(double V, Type *Ty) { |
1927 | if (Ty->isHalfTy() || Ty->isFloatTy()) { |
1928 | APFloat APF(V); |
1929 | bool unused; |
1930 | APF.convert(ToSemantics: Ty->getFltSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused); |
1931 | return ConstantFP::get(Context&: Ty->getContext(), V: APF); |
1932 | } |
1933 | if (Ty->isDoubleTy()) |
1934 | return ConstantFP::get(Context&: Ty->getContext(), V: APFloat(V)); |
1935 | llvm_unreachable("Can only constant fold half/float/double" ); |
1936 | } |
1937 | |
1938 | #if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128) |
1939 | Constant *GetConstantFoldFPValue128(float128 V, Type *Ty) { |
1940 | if (Ty->isFP128Ty()) |
1941 | return ConstantFP::get(Ty, V); |
1942 | llvm_unreachable("Can only constant fold fp128" ); |
1943 | } |
1944 | #endif |
1945 | |
1946 | /// Clear the floating-point exception state. |
1947 | inline void llvm_fenv_clearexcept() { |
1948 | #if HAVE_DECL_FE_ALL_EXCEPT |
1949 | feclearexcept(FE_ALL_EXCEPT); |
1950 | #endif |
1951 | errno = 0; |
1952 | } |
1953 | |
1954 | /// Test if a floating-point exception was raised. |
1955 | inline bool llvm_fenv_testexcept() { |
1956 | int errno_val = errno; |
1957 | if (errno_val == ERANGE || errno_val == EDOM) |
1958 | return true; |
1959 | #if HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT |
1960 | if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT)) |
1961 | return true; |
1962 | #endif |
1963 | return false; |
1964 | } |
1965 | |
1966 | static const APFloat FTZPreserveSign(const APFloat &V) { |
1967 | if (V.isDenormal()) |
1968 | return APFloat::getZero(Sem: V.getSemantics(), Negative: V.isNegative()); |
1969 | return V; |
1970 | } |
1971 | |
1972 | Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V, |
1973 | Type *Ty) { |
1974 | llvm_fenv_clearexcept(); |
1975 | double Result = NativeFP(V.convertToDouble()); |
1976 | if (llvm_fenv_testexcept()) { |
1977 | llvm_fenv_clearexcept(); |
1978 | return nullptr; |
1979 | } |
1980 | |
1981 | return GetConstantFoldFPValue(V: Result, Ty); |
1982 | } |
1983 | |
1984 | #if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128) |
1985 | Constant *ConstantFoldFP128(float128 (*NativeFP)(float128), const APFloat &V, |
1986 | Type *Ty) { |
1987 | llvm_fenv_clearexcept(); |
1988 | float128 Result = NativeFP(V.convertToQuad()); |
1989 | if (llvm_fenv_testexcept()) { |
1990 | llvm_fenv_clearexcept(); |
1991 | return nullptr; |
1992 | } |
1993 | |
1994 | return GetConstantFoldFPValue128(V: Result, Ty); |
1995 | } |
1996 | #endif |
1997 | |
1998 | Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), |
1999 | const APFloat &V, const APFloat &W, Type *Ty) { |
2000 | llvm_fenv_clearexcept(); |
2001 | double Result = NativeFP(V.convertToDouble(), W.convertToDouble()); |
2002 | if (llvm_fenv_testexcept()) { |
2003 | llvm_fenv_clearexcept(); |
2004 | return nullptr; |
2005 | } |
2006 | |
2007 | return GetConstantFoldFPValue(V: Result, Ty); |
2008 | } |
2009 | |
2010 | Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) { |
2011 | FixedVectorType *VT = dyn_cast<FixedVectorType>(Val: Op->getType()); |
2012 | if (!VT) |
2013 | return nullptr; |
2014 | |
2015 | // This isn't strictly necessary, but handle the special/common case of zero: |
2016 | // all integer reductions of a zero input produce zero. |
2017 | if (isa<ConstantAggregateZero>(Val: Op)) |
2018 | return ConstantInt::get(Ty: VT->getElementType(), V: 0); |
2019 | |
2020 | // This is the same as the underlying binops - poison propagates. |
2021 | if (isa<PoisonValue>(Val: Op) || Op->containsPoisonElement()) |
2022 | return PoisonValue::get(T: VT->getElementType()); |
2023 | |
2024 | // TODO: Handle undef. |
2025 | if (!isa<ConstantVector>(Val: Op) && !isa<ConstantDataVector>(Val: Op)) |
2026 | return nullptr; |
2027 | |
2028 | auto *EltC = dyn_cast<ConstantInt>(Val: Op->getAggregateElement(Elt: 0U)); |
2029 | if (!EltC) |
2030 | return nullptr; |
2031 | |
2032 | APInt Acc = EltC->getValue(); |
2033 | for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) { |
2034 | if (!(EltC = dyn_cast<ConstantInt>(Val: Op->getAggregateElement(Elt: I)))) |
2035 | return nullptr; |
2036 | const APInt &X = EltC->getValue(); |
2037 | switch (IID) { |
2038 | case Intrinsic::vector_reduce_add: |
2039 | Acc = Acc + X; |
2040 | break; |
2041 | case Intrinsic::vector_reduce_mul: |
2042 | Acc = Acc * X; |
2043 | break; |
2044 | case Intrinsic::vector_reduce_and: |
2045 | Acc = Acc & X; |
2046 | break; |
2047 | case Intrinsic::vector_reduce_or: |
2048 | Acc = Acc | X; |
2049 | break; |
2050 | case Intrinsic::vector_reduce_xor: |
2051 | Acc = Acc ^ X; |
2052 | break; |
2053 | case Intrinsic::vector_reduce_smin: |
2054 | Acc = APIntOps::smin(A: Acc, B: X); |
2055 | break; |
2056 | case Intrinsic::vector_reduce_smax: |
2057 | Acc = APIntOps::smax(A: Acc, B: X); |
2058 | break; |
2059 | case Intrinsic::vector_reduce_umin: |
2060 | Acc = APIntOps::umin(A: Acc, B: X); |
2061 | break; |
2062 | case Intrinsic::vector_reduce_umax: |
2063 | Acc = APIntOps::umax(A: Acc, B: X); |
2064 | break; |
2065 | } |
2066 | } |
2067 | |
2068 | return ConstantInt::get(Context&: Op->getContext(), V: Acc); |
2069 | } |
2070 | |
2071 | /// Attempt to fold an SSE floating point to integer conversion of a constant |
2072 | /// floating point. If roundTowardZero is false, the default IEEE rounding is |
2073 | /// used (toward nearest, ties to even). This matches the behavior of the |
2074 | /// non-truncating SSE instructions in the default rounding mode. The desired |
2075 | /// integer type Ty is used to select how many bits are available for the |
2076 | /// result. Returns null if the conversion cannot be performed, otherwise |
2077 | /// returns the Constant value resulting from the conversion. |
2078 | Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero, |
2079 | Type *Ty, bool IsSigned) { |
2080 | // All of these conversion intrinsics form an integer of at most 64bits. |
2081 | unsigned ResultWidth = Ty->getIntegerBitWidth(); |
2082 | assert(ResultWidth <= 64 && |
2083 | "Can only constant fold conversions to 64 and 32 bit ints" ); |
2084 | |
2085 | uint64_t UIntVal; |
2086 | bool isExact = false; |
2087 | APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero |
2088 | : APFloat::rmNearestTiesToEven; |
2089 | APFloat::opStatus status = |
2090 | Val.convertToInteger(Input: MutableArrayRef(UIntVal), Width: ResultWidth, |
2091 | IsSigned, RM: mode, IsExact: &isExact); |
2092 | if (status != APFloat::opOK && |
2093 | (!roundTowardZero || status != APFloat::opInexact)) |
2094 | return nullptr; |
2095 | return ConstantInt::get(Ty, V: UIntVal, IsSigned); |
2096 | } |
2097 | |
2098 | double getValueAsDouble(ConstantFP *Op) { |
2099 | Type *Ty = Op->getType(); |
2100 | |
2101 | if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy()) |
2102 | return Op->getValueAPF().convertToDouble(); |
2103 | |
2104 | bool unused; |
2105 | APFloat APF = Op->getValueAPF(); |
2106 | APF.convert(ToSemantics: APFloat::IEEEdouble(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused); |
2107 | return APF.convertToDouble(); |
2108 | } |
2109 | |
2110 | static bool getConstIntOrUndef(Value *Op, const APInt *&C) { |
2111 | if (auto *CI = dyn_cast<ConstantInt>(Val: Op)) { |
2112 | C = &CI->getValue(); |
2113 | return true; |
2114 | } |
2115 | if (isa<UndefValue>(Val: Op)) { |
2116 | C = nullptr; |
2117 | return true; |
2118 | } |
2119 | return false; |
2120 | } |
2121 | |
2122 | /// Checks if the given intrinsic call, which evaluates to constant, is allowed |
2123 | /// to be folded. |
2124 | /// |
2125 | /// \param CI Constrained intrinsic call. |
2126 | /// \param St Exception flags raised during constant evaluation. |
2127 | static bool mayFoldConstrained(ConstrainedFPIntrinsic *CI, |
2128 | APFloat::opStatus St) { |
2129 | std::optional<RoundingMode> ORM = CI->getRoundingMode(); |
2130 | std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior(); |
2131 | |
2132 | // If the operation does not change exception status flags, it is safe |
2133 | // to fold. |
2134 | if (St == APFloat::opStatus::opOK) |
2135 | return true; |
2136 | |
2137 | // If evaluation raised FP exception, the result can depend on rounding |
2138 | // mode. If the latter is unknown, folding is not possible. |
2139 | if (ORM == RoundingMode::Dynamic) |
2140 | return false; |
2141 | |
2142 | // If FP exceptions are ignored, fold the call, even if such exception is |
2143 | // raised. |
2144 | if (EB && *EB != fp::ExceptionBehavior::ebStrict) |
2145 | return true; |
2146 | |
2147 | // Leave the calculation for runtime so that exception flags be correctly set |
2148 | // in hardware. |
2149 | return false; |
2150 | } |
2151 | |
2152 | /// Returns the rounding mode that should be used for constant evaluation. |
2153 | static RoundingMode |
2154 | getEvaluationRoundingMode(const ConstrainedFPIntrinsic *CI) { |
2155 | std::optional<RoundingMode> ORM = CI->getRoundingMode(); |
2156 | if (!ORM || *ORM == RoundingMode::Dynamic) |
2157 | // Even if the rounding mode is unknown, try evaluating the operation. |
2158 | // If it does not raise inexact exception, rounding was not applied, |
2159 | // so the result is exact and does not depend on rounding mode. Whether |
2160 | // other FP exceptions are raised, it does not depend on rounding mode. |
2161 | return RoundingMode::NearestTiesToEven; |
2162 | return *ORM; |
2163 | } |
2164 | |
2165 | /// Try to constant fold llvm.canonicalize for the given caller and value. |
2166 | static Constant *constantFoldCanonicalize(const Type *Ty, const CallBase *CI, |
2167 | const APFloat &Src) { |
2168 | // Zero, positive and negative, is always OK to fold. |
2169 | if (Src.isZero()) { |
2170 | // Get a fresh 0, since ppc_fp128 does have non-canonical zeros. |
2171 | return ConstantFP::get( |
2172 | Context&: CI->getContext(), |
2173 | V: APFloat::getZero(Sem: Src.getSemantics(), Negative: Src.isNegative())); |
2174 | } |
2175 | |
2176 | if (!Ty->isIEEELikeFPTy()) |
2177 | return nullptr; |
2178 | |
2179 | // Zero is always canonical and the sign must be preserved. |
2180 | // |
2181 | // Denorms and nans may have special encodings, but it should be OK to fold a |
2182 | // totally average number. |
2183 | if (Src.isNormal() || Src.isInfinity()) |
2184 | return ConstantFP::get(Context&: CI->getContext(), V: Src); |
2185 | |
2186 | if (Src.isDenormal() && CI->getParent() && CI->getFunction()) { |
2187 | DenormalMode DenormMode = |
2188 | CI->getFunction()->getDenormalMode(FPType: Src.getSemantics()); |
2189 | |
2190 | if (DenormMode == DenormalMode::getIEEE()) |
2191 | return ConstantFP::get(Context&: CI->getContext(), V: Src); |
2192 | |
2193 | if (DenormMode.Input == DenormalMode::Dynamic) |
2194 | return nullptr; |
2195 | |
2196 | // If we know if either input or output is flushed, we can fold. |
2197 | if ((DenormMode.Input == DenormalMode::Dynamic && |
2198 | DenormMode.Output == DenormalMode::IEEE) || |
2199 | (DenormMode.Input == DenormalMode::IEEE && |
2200 | DenormMode.Output == DenormalMode::Dynamic)) |
2201 | return nullptr; |
2202 | |
2203 | bool IsPositive = |
2204 | (!Src.isNegative() || DenormMode.Input == DenormalMode::PositiveZero || |
2205 | (DenormMode.Output == DenormalMode::PositiveZero && |
2206 | DenormMode.Input == DenormalMode::IEEE)); |
2207 | |
2208 | return ConstantFP::get(Context&: CI->getContext(), |
2209 | V: APFloat::getZero(Sem: Src.getSemantics(), Negative: !IsPositive)); |
2210 | } |
2211 | |
2212 | return nullptr; |
2213 | } |
2214 | |
2215 | static Constant *ConstantFoldScalarCall1(StringRef Name, |
2216 | Intrinsic::ID IntrinsicID, |
2217 | Type *Ty, |
2218 | ArrayRef<Constant *> Operands, |
2219 | const TargetLibraryInfo *TLI, |
2220 | const CallBase *Call) { |
2221 | assert(Operands.size() == 1 && "Wrong number of operands." ); |
2222 | |
2223 | if (IntrinsicID == Intrinsic::is_constant) { |
2224 | // We know we have a "Constant" argument. But we want to only |
2225 | // return true for manifest constants, not those that depend on |
2226 | // constants with unknowable values, e.g. GlobalValue or BlockAddress. |
2227 | if (Operands[0]->isManifestConstant()) |
2228 | return ConstantInt::getTrue(Context&: Ty->getContext()); |
2229 | return nullptr; |
2230 | } |
2231 | |
2232 | if (isa<UndefValue>(Val: Operands[0])) { |
2233 | // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN. |
2234 | // ctpop() is between 0 and bitwidth, pick 0 for undef. |
2235 | // fptoui.sat and fptosi.sat can always fold to zero (for a zero input). |
2236 | if (IntrinsicID == Intrinsic::cos || |
2237 | IntrinsicID == Intrinsic::ctpop || |
2238 | IntrinsicID == Intrinsic::fptoui_sat || |
2239 | IntrinsicID == Intrinsic::fptosi_sat || |
2240 | IntrinsicID == Intrinsic::canonicalize) |
2241 | return Constant::getNullValue(Ty); |
2242 | if (IntrinsicID == Intrinsic::bswap || |
2243 | IntrinsicID == Intrinsic::bitreverse || |
2244 | IntrinsicID == Intrinsic::launder_invariant_group || |
2245 | IntrinsicID == Intrinsic::strip_invariant_group) |
2246 | return Operands[0]; |
2247 | } |
2248 | |
2249 | if (isa<ConstantPointerNull>(Val: Operands[0])) { |
2250 | // launder(null) == null == strip(null) iff in addrspace 0 |
2251 | if (IntrinsicID == Intrinsic::launder_invariant_group || |
2252 | IntrinsicID == Intrinsic::strip_invariant_group) { |
2253 | // If instruction is not yet put in a basic block (e.g. when cloning |
2254 | // a function during inlining), Call's caller may not be available. |
2255 | // So check Call's BB first before querying Call->getCaller. |
2256 | const Function *Caller = |
2257 | Call->getParent() ? Call->getCaller() : nullptr; |
2258 | if (Caller && |
2259 | !NullPointerIsDefined( |
2260 | F: Caller, AS: Operands[0]->getType()->getPointerAddressSpace())) { |
2261 | return Operands[0]; |
2262 | } |
2263 | return nullptr; |
2264 | } |
2265 | } |
2266 | |
2267 | if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[0])) { |
2268 | if (IntrinsicID == Intrinsic::convert_to_fp16) { |
2269 | APFloat Val(Op->getValueAPF()); |
2270 | |
2271 | bool lost = false; |
2272 | Val.convert(ToSemantics: APFloat::IEEEhalf(), RM: APFloat::rmNearestTiesToEven, losesInfo: &lost); |
2273 | |
2274 | return ConstantInt::get(Context&: Ty->getContext(), V: Val.bitcastToAPInt()); |
2275 | } |
2276 | |
2277 | APFloat U = Op->getValueAPF(); |
2278 | |
2279 | if (IntrinsicID == Intrinsic::wasm_trunc_signed || |
2280 | IntrinsicID == Intrinsic::wasm_trunc_unsigned) { |
2281 | bool Signed = IntrinsicID == Intrinsic::wasm_trunc_signed; |
2282 | |
2283 | if (U.isNaN()) |
2284 | return nullptr; |
2285 | |
2286 | unsigned Width = Ty->getIntegerBitWidth(); |
2287 | APSInt Int(Width, !Signed); |
2288 | bool IsExact = false; |
2289 | APFloat::opStatus Status = |
2290 | U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
2291 | |
2292 | if (Status == APFloat::opOK || Status == APFloat::opInexact) |
2293 | return ConstantInt::get(Ty, V: Int); |
2294 | |
2295 | return nullptr; |
2296 | } |
2297 | |
2298 | if (IntrinsicID == Intrinsic::fptoui_sat || |
2299 | IntrinsicID == Intrinsic::fptosi_sat) { |
2300 | // convertToInteger() already has the desired saturation semantics. |
2301 | APSInt Int(Ty->getIntegerBitWidth(), |
2302 | IntrinsicID == Intrinsic::fptoui_sat); |
2303 | bool IsExact; |
2304 | U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
2305 | return ConstantInt::get(Ty, V: Int); |
2306 | } |
2307 | |
2308 | if (IntrinsicID == Intrinsic::canonicalize) |
2309 | return constantFoldCanonicalize(Ty, CI: Call, Src: U); |
2310 | |
2311 | #if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128) |
2312 | if (Ty->isFP128Ty()) { |
2313 | if (IntrinsicID == Intrinsic::log) { |
2314 | float128 Result = logf128(Op->getValueAPF().convertToQuad()); |
2315 | return GetConstantFoldFPValue128(V: Result, Ty); |
2316 | } |
2317 | |
2318 | LibFunc Fp128Func = NotLibFunc; |
2319 | if (TLI && TLI->getLibFunc(funcName: Name, F&: Fp128Func) && TLI->has(F: Fp128Func) && |
2320 | Fp128Func == LibFunc_logl) |
2321 | return ConstantFoldFP128(logf128, Op->getValueAPF(), Ty); |
2322 | } |
2323 | #endif |
2324 | |
2325 | if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy() && |
2326 | !Ty->isIntegerTy()) |
2327 | return nullptr; |
2328 | |
2329 | // Use internal versions of these intrinsics. |
2330 | |
2331 | if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint) { |
2332 | U.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
2333 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2334 | } |
2335 | |
2336 | if (IntrinsicID == Intrinsic::round) { |
2337 | U.roundToIntegral(RM: APFloat::rmNearestTiesToAway); |
2338 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2339 | } |
2340 | |
2341 | if (IntrinsicID == Intrinsic::roundeven) { |
2342 | U.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
2343 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2344 | } |
2345 | |
2346 | if (IntrinsicID == Intrinsic::ceil) { |
2347 | U.roundToIntegral(RM: APFloat::rmTowardPositive); |
2348 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2349 | } |
2350 | |
2351 | if (IntrinsicID == Intrinsic::floor) { |
2352 | U.roundToIntegral(RM: APFloat::rmTowardNegative); |
2353 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2354 | } |
2355 | |
2356 | if (IntrinsicID == Intrinsic::trunc) { |
2357 | U.roundToIntegral(RM: APFloat::rmTowardZero); |
2358 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2359 | } |
2360 | |
2361 | if (IntrinsicID == Intrinsic::fabs) { |
2362 | U.clearSign(); |
2363 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2364 | } |
2365 | |
2366 | if (IntrinsicID == Intrinsic::amdgcn_fract) { |
2367 | // The v_fract instruction behaves like the OpenCL spec, which defines |
2368 | // fract(x) as fmin(x - floor(x), 0x1.fffffep-1f): "The min() operator is |
2369 | // there to prevent fract(-small) from returning 1.0. It returns the |
2370 | // largest positive floating-point number less than 1.0." |
2371 | APFloat FloorU(U); |
2372 | FloorU.roundToIntegral(RM: APFloat::rmTowardNegative); |
2373 | APFloat FractU(U - FloorU); |
2374 | APFloat AlmostOne(U.getSemantics(), 1); |
2375 | AlmostOne.next(/*nextDown*/ true); |
2376 | return ConstantFP::get(Context&: Ty->getContext(), V: minimum(A: FractU, B: AlmostOne)); |
2377 | } |
2378 | |
2379 | // Rounding operations (floor, trunc, ceil, round and nearbyint) do not |
2380 | // raise FP exceptions, unless the argument is signaling NaN. |
2381 | |
2382 | std::optional<APFloat::roundingMode> RM; |
2383 | switch (IntrinsicID) { |
2384 | default: |
2385 | break; |
2386 | case Intrinsic::experimental_constrained_nearbyint: |
2387 | case Intrinsic::experimental_constrained_rint: { |
2388 | auto CI = cast<ConstrainedFPIntrinsic>(Val: Call); |
2389 | RM = CI->getRoundingMode(); |
2390 | if (!RM || *RM == RoundingMode::Dynamic) |
2391 | return nullptr; |
2392 | break; |
2393 | } |
2394 | case Intrinsic::experimental_constrained_round: |
2395 | RM = APFloat::rmNearestTiesToAway; |
2396 | break; |
2397 | case Intrinsic::experimental_constrained_ceil: |
2398 | RM = APFloat::rmTowardPositive; |
2399 | break; |
2400 | case Intrinsic::experimental_constrained_floor: |
2401 | RM = APFloat::rmTowardNegative; |
2402 | break; |
2403 | case Intrinsic::experimental_constrained_trunc: |
2404 | RM = APFloat::rmTowardZero; |
2405 | break; |
2406 | } |
2407 | if (RM) { |
2408 | auto CI = cast<ConstrainedFPIntrinsic>(Val: Call); |
2409 | if (U.isFinite()) { |
2410 | APFloat::opStatus St = U.roundToIntegral(RM: *RM); |
2411 | if (IntrinsicID == Intrinsic::experimental_constrained_rint && |
2412 | St == APFloat::opInexact) { |
2413 | std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior(); |
2414 | if (EB == fp::ebStrict) |
2415 | return nullptr; |
2416 | } |
2417 | } else if (U.isSignaling()) { |
2418 | std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior(); |
2419 | if (EB && *EB != fp::ebIgnore) |
2420 | return nullptr; |
2421 | U = APFloat::getQNaN(Sem: U.getSemantics()); |
2422 | } |
2423 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2424 | } |
2425 | |
2426 | // NVVM float/double to signed/unsigned int32/int64 conversions: |
2427 | switch (IntrinsicID) { |
2428 | // f2i |
2429 | case Intrinsic::nvvm_f2i_rm: |
2430 | case Intrinsic::nvvm_f2i_rn: |
2431 | case Intrinsic::nvvm_f2i_rp: |
2432 | case Intrinsic::nvvm_f2i_rz: |
2433 | case Intrinsic::nvvm_f2i_rm_ftz: |
2434 | case Intrinsic::nvvm_f2i_rn_ftz: |
2435 | case Intrinsic::nvvm_f2i_rp_ftz: |
2436 | case Intrinsic::nvvm_f2i_rz_ftz: |
2437 | // f2ui |
2438 | case Intrinsic::nvvm_f2ui_rm: |
2439 | case Intrinsic::nvvm_f2ui_rn: |
2440 | case Intrinsic::nvvm_f2ui_rp: |
2441 | case Intrinsic::nvvm_f2ui_rz: |
2442 | case Intrinsic::nvvm_f2ui_rm_ftz: |
2443 | case Intrinsic::nvvm_f2ui_rn_ftz: |
2444 | case Intrinsic::nvvm_f2ui_rp_ftz: |
2445 | case Intrinsic::nvvm_f2ui_rz_ftz: |
2446 | // d2i |
2447 | case Intrinsic::nvvm_d2i_rm: |
2448 | case Intrinsic::nvvm_d2i_rn: |
2449 | case Intrinsic::nvvm_d2i_rp: |
2450 | case Intrinsic::nvvm_d2i_rz: |
2451 | // d2ui |
2452 | case Intrinsic::nvvm_d2ui_rm: |
2453 | case Intrinsic::nvvm_d2ui_rn: |
2454 | case Intrinsic::nvvm_d2ui_rp: |
2455 | case Intrinsic::nvvm_d2ui_rz: |
2456 | // f2ll |
2457 | case Intrinsic::nvvm_f2ll_rm: |
2458 | case Intrinsic::nvvm_f2ll_rn: |
2459 | case Intrinsic::nvvm_f2ll_rp: |
2460 | case Intrinsic::nvvm_f2ll_rz: |
2461 | case Intrinsic::nvvm_f2ll_rm_ftz: |
2462 | case Intrinsic::nvvm_f2ll_rn_ftz: |
2463 | case Intrinsic::nvvm_f2ll_rp_ftz: |
2464 | case Intrinsic::nvvm_f2ll_rz_ftz: |
2465 | // f2ull |
2466 | case Intrinsic::nvvm_f2ull_rm: |
2467 | case Intrinsic::nvvm_f2ull_rn: |
2468 | case Intrinsic::nvvm_f2ull_rp: |
2469 | case Intrinsic::nvvm_f2ull_rz: |
2470 | case Intrinsic::nvvm_f2ull_rm_ftz: |
2471 | case Intrinsic::nvvm_f2ull_rn_ftz: |
2472 | case Intrinsic::nvvm_f2ull_rp_ftz: |
2473 | case Intrinsic::nvvm_f2ull_rz_ftz: |
2474 | // d2ll |
2475 | case Intrinsic::nvvm_d2ll_rm: |
2476 | case Intrinsic::nvvm_d2ll_rn: |
2477 | case Intrinsic::nvvm_d2ll_rp: |
2478 | case Intrinsic::nvvm_d2ll_rz: |
2479 | // d2ull |
2480 | case Intrinsic::nvvm_d2ull_rm: |
2481 | case Intrinsic::nvvm_d2ull_rn: |
2482 | case Intrinsic::nvvm_d2ull_rp: |
2483 | case Intrinsic::nvvm_d2ull_rz: { |
2484 | // In float-to-integer conversion, NaN inputs are converted to 0. |
2485 | if (U.isNaN()) |
2486 | return ConstantInt::get(Ty, V: 0); |
2487 | |
2488 | APFloat::roundingMode RMode = |
2489 | nvvm::GetFPToIntegerRoundingMode(IntrinsicID); |
2490 | bool IsFTZ = nvvm::FPToIntegerIntrinsicShouldFTZ(IntrinsicID); |
2491 | bool IsSigned = nvvm::FPToIntegerIntrinsicResultIsSigned(IntrinsicID); |
2492 | |
2493 | APSInt ResInt(Ty->getIntegerBitWidth(), !IsSigned); |
2494 | auto FloatToRound = IsFTZ ? FTZPreserveSign(V: U) : U; |
2495 | |
2496 | bool IsExact = false; |
2497 | APFloat::opStatus Status = |
2498 | FloatToRound.convertToInteger(Result&: ResInt, RM: RMode, IsExact: &IsExact); |
2499 | |
2500 | if (Status != APFloat::opInvalidOp) |
2501 | return ConstantInt::get(Ty, V: ResInt); |
2502 | return nullptr; |
2503 | } |
2504 | } |
2505 | |
2506 | /// We only fold functions with finite arguments. Folding NaN and inf is |
2507 | /// likely to be aborted with an exception anyway, and some host libms |
2508 | /// have known errors raising exceptions. |
2509 | if (!U.isFinite()) |
2510 | return nullptr; |
2511 | |
2512 | /// Currently APFloat versions of these functions do not exist, so we use |
2513 | /// the host native double versions. Float versions are not called |
2514 | /// directly but for all these it is true (float)(f((double)arg)) == |
2515 | /// f(arg). Long double not supported yet. |
2516 | const APFloat &APF = Op->getValueAPF(); |
2517 | |
2518 | switch (IntrinsicID) { |
2519 | default: break; |
2520 | case Intrinsic::log: |
2521 | return ConstantFoldFP(NativeFP: log, V: APF, Ty); |
2522 | case Intrinsic::log2: |
2523 | // TODO: What about hosts that lack a C99 library? |
2524 | return ConstantFoldFP(NativeFP: log2, V: APF, Ty); |
2525 | case Intrinsic::log10: |
2526 | // TODO: What about hosts that lack a C99 library? |
2527 | return ConstantFoldFP(NativeFP: log10, V: APF, Ty); |
2528 | case Intrinsic::exp: |
2529 | return ConstantFoldFP(NativeFP: exp, V: APF, Ty); |
2530 | case Intrinsic::exp2: |
2531 | // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library. |
2532 | return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty); |
2533 | case Intrinsic::exp10: |
2534 | // Fold exp10(x) as pow(10, x), in case the host lacks a C99 library. |
2535 | return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(10.0), W: APF, Ty); |
2536 | case Intrinsic::sin: |
2537 | return ConstantFoldFP(NativeFP: sin, V: APF, Ty); |
2538 | case Intrinsic::cos: |
2539 | return ConstantFoldFP(NativeFP: cos, V: APF, Ty); |
2540 | case Intrinsic::sinh: |
2541 | return ConstantFoldFP(NativeFP: sinh, V: APF, Ty); |
2542 | case Intrinsic::cosh: |
2543 | return ConstantFoldFP(NativeFP: cosh, V: APF, Ty); |
2544 | case Intrinsic::atan: |
2545 | // Implement optional behavior from C's Annex F for +/-0.0. |
2546 | if (U.isZero()) |
2547 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2548 | return ConstantFoldFP(NativeFP: atan, V: APF, Ty); |
2549 | case Intrinsic::sqrt: |
2550 | return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty); |
2551 | case Intrinsic::amdgcn_cos: |
2552 | case Intrinsic::amdgcn_sin: { |
2553 | double V = getValueAsDouble(Op); |
2554 | if (V < -256.0 || V > 256.0) |
2555 | // The gfx8 and gfx9 architectures handle arguments outside the range |
2556 | // [-256, 256] differently. This should be a rare case so bail out |
2557 | // rather than trying to handle the difference. |
2558 | return nullptr; |
2559 | bool IsCos = IntrinsicID == Intrinsic::amdgcn_cos; |
2560 | double V4 = V * 4.0; |
2561 | if (V4 == floor(x: V4)) { |
2562 | // Force exact results for quarter-integer inputs. |
2563 | const double SinVals[4] = { 0.0, 1.0, 0.0, -1.0 }; |
2564 | V = SinVals[((int)V4 + (IsCos ? 1 : 0)) & 3]; |
2565 | } else { |
2566 | if (IsCos) |
2567 | V = cos(x: V * 2.0 * numbers::pi); |
2568 | else |
2569 | V = sin(x: V * 2.0 * numbers::pi); |
2570 | } |
2571 | return GetConstantFoldFPValue(V, Ty); |
2572 | } |
2573 | } |
2574 | |
2575 | if (!TLI) |
2576 | return nullptr; |
2577 | |
2578 | LibFunc Func = NotLibFunc; |
2579 | if (!TLI->getLibFunc(funcName: Name, F&: Func)) |
2580 | return nullptr; |
2581 | |
2582 | switch (Func) { |
2583 | default: |
2584 | break; |
2585 | case LibFunc_acos: |
2586 | case LibFunc_acosf: |
2587 | case LibFunc_acos_finite: |
2588 | case LibFunc_acosf_finite: |
2589 | if (TLI->has(F: Func)) |
2590 | return ConstantFoldFP(NativeFP: acos, V: APF, Ty); |
2591 | break; |
2592 | case LibFunc_asin: |
2593 | case LibFunc_asinf: |
2594 | case LibFunc_asin_finite: |
2595 | case LibFunc_asinf_finite: |
2596 | if (TLI->has(F: Func)) |
2597 | return ConstantFoldFP(NativeFP: asin, V: APF, Ty); |
2598 | break; |
2599 | case LibFunc_atan: |
2600 | case LibFunc_atanf: |
2601 | // Implement optional behavior from C's Annex F for +/-0.0. |
2602 | if (U.isZero()) |
2603 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2604 | if (TLI->has(F: Func)) |
2605 | return ConstantFoldFP(NativeFP: atan, V: APF, Ty); |
2606 | break; |
2607 | case LibFunc_ceil: |
2608 | case LibFunc_ceilf: |
2609 | if (TLI->has(F: Func)) { |
2610 | U.roundToIntegral(RM: APFloat::rmTowardPositive); |
2611 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2612 | } |
2613 | break; |
2614 | case LibFunc_cos: |
2615 | case LibFunc_cosf: |
2616 | if (TLI->has(F: Func)) |
2617 | return ConstantFoldFP(NativeFP: cos, V: APF, Ty); |
2618 | break; |
2619 | case LibFunc_cosh: |
2620 | case LibFunc_coshf: |
2621 | case LibFunc_cosh_finite: |
2622 | case LibFunc_coshf_finite: |
2623 | if (TLI->has(F: Func)) |
2624 | return ConstantFoldFP(NativeFP: cosh, V: APF, Ty); |
2625 | break; |
2626 | case LibFunc_exp: |
2627 | case LibFunc_expf: |
2628 | case LibFunc_exp_finite: |
2629 | case LibFunc_expf_finite: |
2630 | if (TLI->has(F: Func)) |
2631 | return ConstantFoldFP(NativeFP: exp, V: APF, Ty); |
2632 | break; |
2633 | case LibFunc_exp2: |
2634 | case LibFunc_exp2f: |
2635 | case LibFunc_exp2_finite: |
2636 | case LibFunc_exp2f_finite: |
2637 | if (TLI->has(F: Func)) |
2638 | // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library. |
2639 | return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty); |
2640 | break; |
2641 | case LibFunc_fabs: |
2642 | case LibFunc_fabsf: |
2643 | if (TLI->has(F: Func)) { |
2644 | U.clearSign(); |
2645 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2646 | } |
2647 | break; |
2648 | case LibFunc_floor: |
2649 | case LibFunc_floorf: |
2650 | if (TLI->has(F: Func)) { |
2651 | U.roundToIntegral(RM: APFloat::rmTowardNegative); |
2652 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2653 | } |
2654 | break; |
2655 | case LibFunc_log: |
2656 | case LibFunc_logf: |
2657 | case LibFunc_log_finite: |
2658 | case LibFunc_logf_finite: |
2659 | if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func)) |
2660 | return ConstantFoldFP(NativeFP: log, V: APF, Ty); |
2661 | break; |
2662 | case LibFunc_log2: |
2663 | case LibFunc_log2f: |
2664 | case LibFunc_log2_finite: |
2665 | case LibFunc_log2f_finite: |
2666 | if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func)) |
2667 | // TODO: What about hosts that lack a C99 library? |
2668 | return ConstantFoldFP(NativeFP: log2, V: APF, Ty); |
2669 | break; |
2670 | case LibFunc_log10: |
2671 | case LibFunc_log10f: |
2672 | case LibFunc_log10_finite: |
2673 | case LibFunc_log10f_finite: |
2674 | if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func)) |
2675 | // TODO: What about hosts that lack a C99 library? |
2676 | return ConstantFoldFP(NativeFP: log10, V: APF, Ty); |
2677 | break; |
2678 | case LibFunc_ilogb: |
2679 | case LibFunc_ilogbf: |
2680 | if (!APF.isZero() && TLI->has(F: Func)) |
2681 | return ConstantInt::get(Ty, V: ilogb(Arg: APF), IsSigned: true); |
2682 | break; |
2683 | case LibFunc_logb: |
2684 | case LibFunc_logbf: |
2685 | if (!APF.isZero() && TLI->has(F: Func)) |
2686 | return ConstantFoldFP(NativeFP: logb, V: APF, Ty); |
2687 | break; |
2688 | case LibFunc_log1p: |
2689 | case LibFunc_log1pf: |
2690 | // Implement optional behavior from C's Annex F for +/-0.0. |
2691 | if (U.isZero()) |
2692 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2693 | if (APF > APFloat::getOne(Sem: APF.getSemantics(), Negative: true) && TLI->has(F: Func)) |
2694 | return ConstantFoldFP(NativeFP: log1p, V: APF, Ty); |
2695 | break; |
2696 | case LibFunc_logl: |
2697 | return nullptr; |
2698 | case LibFunc_erf: |
2699 | case LibFunc_erff: |
2700 | if (TLI->has(F: Func)) |
2701 | return ConstantFoldFP(NativeFP: erf, V: APF, Ty); |
2702 | break; |
2703 | case LibFunc_nearbyint: |
2704 | case LibFunc_nearbyintf: |
2705 | case LibFunc_rint: |
2706 | case LibFunc_rintf: |
2707 | if (TLI->has(F: Func)) { |
2708 | U.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
2709 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2710 | } |
2711 | break; |
2712 | case LibFunc_round: |
2713 | case LibFunc_roundf: |
2714 | if (TLI->has(F: Func)) { |
2715 | U.roundToIntegral(RM: APFloat::rmNearestTiesToAway); |
2716 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2717 | } |
2718 | break; |
2719 | case LibFunc_sin: |
2720 | case LibFunc_sinf: |
2721 | if (TLI->has(F: Func)) |
2722 | return ConstantFoldFP(NativeFP: sin, V: APF, Ty); |
2723 | break; |
2724 | case LibFunc_sinh: |
2725 | case LibFunc_sinhf: |
2726 | case LibFunc_sinh_finite: |
2727 | case LibFunc_sinhf_finite: |
2728 | if (TLI->has(F: Func)) |
2729 | return ConstantFoldFP(NativeFP: sinh, V: APF, Ty); |
2730 | break; |
2731 | case LibFunc_sqrt: |
2732 | case LibFunc_sqrtf: |
2733 | if (!APF.isNegative() && TLI->has(F: Func)) |
2734 | return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty); |
2735 | break; |
2736 | case LibFunc_tan: |
2737 | case LibFunc_tanf: |
2738 | if (TLI->has(F: Func)) |
2739 | return ConstantFoldFP(NativeFP: tan, V: APF, Ty); |
2740 | break; |
2741 | case LibFunc_tanh: |
2742 | case LibFunc_tanhf: |
2743 | if (TLI->has(F: Func)) |
2744 | return ConstantFoldFP(NativeFP: tanh, V: APF, Ty); |
2745 | break; |
2746 | case LibFunc_trunc: |
2747 | case LibFunc_truncf: |
2748 | if (TLI->has(F: Func)) { |
2749 | U.roundToIntegral(RM: APFloat::rmTowardZero); |
2750 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2751 | } |
2752 | break; |
2753 | } |
2754 | return nullptr; |
2755 | } |
2756 | |
2757 | if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) { |
2758 | switch (IntrinsicID) { |
2759 | case Intrinsic::bswap: |
2760 | return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().byteSwap()); |
2761 | case Intrinsic::ctpop: |
2762 | return ConstantInt::get(Ty, V: Op->getValue().popcount()); |
2763 | case Intrinsic::bitreverse: |
2764 | return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().reverseBits()); |
2765 | case Intrinsic::convert_from_fp16: { |
2766 | APFloat Val(APFloat::IEEEhalf(), Op->getValue()); |
2767 | |
2768 | bool lost = false; |
2769 | APFloat::opStatus status = Val.convert( |
2770 | ToSemantics: Ty->getFltSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &lost); |
2771 | |
2772 | // Conversion is always precise. |
2773 | (void)status; |
2774 | assert(status != APFloat::opInexact && !lost && |
2775 | "Precision lost during fp16 constfolding" ); |
2776 | |
2777 | return ConstantFP::get(Context&: Ty->getContext(), V: Val); |
2778 | } |
2779 | |
2780 | case Intrinsic::amdgcn_s_wqm: { |
2781 | uint64_t Val = Op->getZExtValue(); |
2782 | Val |= (Val & 0x5555555555555555ULL) << 1 | |
2783 | ((Val >> 1) & 0x5555555555555555ULL); |
2784 | Val |= (Val & 0x3333333333333333ULL) << 2 | |
2785 | ((Val >> 2) & 0x3333333333333333ULL); |
2786 | return ConstantInt::get(Ty, V: Val); |
2787 | } |
2788 | |
2789 | case Intrinsic::amdgcn_s_quadmask: { |
2790 | uint64_t Val = Op->getZExtValue(); |
2791 | uint64_t QuadMask = 0; |
2792 | for (unsigned I = 0; I < Op->getBitWidth() / 4; ++I, Val >>= 4) { |
2793 | if (!(Val & 0xF)) |
2794 | continue; |
2795 | |
2796 | QuadMask |= (1ULL << I); |
2797 | } |
2798 | return ConstantInt::get(Ty, V: QuadMask); |
2799 | } |
2800 | |
2801 | case Intrinsic::amdgcn_s_bitreplicate: { |
2802 | uint64_t Val = Op->getZExtValue(); |
2803 | Val = (Val & 0x000000000000FFFFULL) | (Val & 0x00000000FFFF0000ULL) << 16; |
2804 | Val = (Val & 0x000000FF000000FFULL) | (Val & 0x0000FF000000FF00ULL) << 8; |
2805 | Val = (Val & 0x000F000F000F000FULL) | (Val & 0x00F000F000F000F0ULL) << 4; |
2806 | Val = (Val & 0x0303030303030303ULL) | (Val & 0x0C0C0C0C0C0C0C0CULL) << 2; |
2807 | Val = (Val & 0x1111111111111111ULL) | (Val & 0x2222222222222222ULL) << 1; |
2808 | Val = Val | Val << 1; |
2809 | return ConstantInt::get(Ty, V: Val); |
2810 | } |
2811 | |
2812 | default: |
2813 | return nullptr; |
2814 | } |
2815 | } |
2816 | |
2817 | switch (IntrinsicID) { |
2818 | default: break; |
2819 | case Intrinsic::vector_reduce_add: |
2820 | case Intrinsic::vector_reduce_mul: |
2821 | case Intrinsic::vector_reduce_and: |
2822 | case Intrinsic::vector_reduce_or: |
2823 | case Intrinsic::vector_reduce_xor: |
2824 | case Intrinsic::vector_reduce_smin: |
2825 | case Intrinsic::vector_reduce_smax: |
2826 | case Intrinsic::vector_reduce_umin: |
2827 | case Intrinsic::vector_reduce_umax: |
2828 | if (Constant *C = constantFoldVectorReduce(IID: IntrinsicID, Op: Operands[0])) |
2829 | return C; |
2830 | break; |
2831 | } |
2832 | |
2833 | // Support ConstantVector in case we have an Undef in the top. |
2834 | if (isa<ConstantVector>(Val: Operands[0]) || |
2835 | isa<ConstantDataVector>(Val: Operands[0])) { |
2836 | auto *Op = cast<Constant>(Val: Operands[0]); |
2837 | switch (IntrinsicID) { |
2838 | default: break; |
2839 | case Intrinsic::x86_sse_cvtss2si: |
2840 | case Intrinsic::x86_sse_cvtss2si64: |
2841 | case Intrinsic::x86_sse2_cvtsd2si: |
2842 | case Intrinsic::x86_sse2_cvtsd2si64: |
2843 | if (ConstantFP *FPOp = |
2844 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2845 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2846 | /*roundTowardZero=*/false, Ty, |
2847 | /*IsSigned*/true); |
2848 | break; |
2849 | case Intrinsic::x86_sse_cvttss2si: |
2850 | case Intrinsic::x86_sse_cvttss2si64: |
2851 | case Intrinsic::x86_sse2_cvttsd2si: |
2852 | case Intrinsic::x86_sse2_cvttsd2si64: |
2853 | if (ConstantFP *FPOp = |
2854 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2855 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2856 | /*roundTowardZero=*/true, Ty, |
2857 | /*IsSigned*/true); |
2858 | break; |
2859 | } |
2860 | } |
2861 | |
2862 | return nullptr; |
2863 | } |
2864 | |
2865 | static Constant *evaluateCompare(const APFloat &Op1, const APFloat &Op2, |
2866 | const ConstrainedFPIntrinsic *Call) { |
2867 | APFloat::opStatus St = APFloat::opOK; |
2868 | auto *FCmp = cast<ConstrainedFPCmpIntrinsic>(Val: Call); |
2869 | FCmpInst::Predicate Cond = FCmp->getPredicate(); |
2870 | if (FCmp->isSignaling()) { |
2871 | if (Op1.isNaN() || Op2.isNaN()) |
2872 | St = APFloat::opInvalidOp; |
2873 | } else { |
2874 | if (Op1.isSignaling() || Op2.isSignaling()) |
2875 | St = APFloat::opInvalidOp; |
2876 | } |
2877 | bool Result = FCmpInst::compare(LHS: Op1, RHS: Op2, Pred: Cond); |
2878 | if (mayFoldConstrained(CI: const_cast<ConstrainedFPCmpIntrinsic *>(FCmp), St)) |
2879 | return ConstantInt::get(Ty: Call->getType()->getScalarType(), V: Result); |
2880 | return nullptr; |
2881 | } |
2882 | |
2883 | static Constant *ConstantFoldLibCall2(StringRef Name, Type *Ty, |
2884 | ArrayRef<Constant *> Operands, |
2885 | const TargetLibraryInfo *TLI) { |
2886 | if (!TLI) |
2887 | return nullptr; |
2888 | |
2889 | LibFunc Func = NotLibFunc; |
2890 | if (!TLI->getLibFunc(funcName: Name, F&: Func)) |
2891 | return nullptr; |
2892 | |
2893 | const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0]); |
2894 | if (!Op1) |
2895 | return nullptr; |
2896 | |
2897 | const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1]); |
2898 | if (!Op2) |
2899 | return nullptr; |
2900 | |
2901 | const APFloat &Op1V = Op1->getValueAPF(); |
2902 | const APFloat &Op2V = Op2->getValueAPF(); |
2903 | |
2904 | switch (Func) { |
2905 | default: |
2906 | break; |
2907 | case LibFunc_pow: |
2908 | case LibFunc_powf: |
2909 | case LibFunc_pow_finite: |
2910 | case LibFunc_powf_finite: |
2911 | if (TLI->has(F: Func)) |
2912 | return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty); |
2913 | break; |
2914 | case LibFunc_fmod: |
2915 | case LibFunc_fmodf: |
2916 | if (TLI->has(F: Func)) { |
2917 | APFloat V = Op1->getValueAPF(); |
2918 | if (APFloat::opStatus::opOK == V.mod(RHS: Op2->getValueAPF())) |
2919 | return ConstantFP::get(Context&: Ty->getContext(), V); |
2920 | } |
2921 | break; |
2922 | case LibFunc_remainder: |
2923 | case LibFunc_remainderf: |
2924 | if (TLI->has(F: Func)) { |
2925 | APFloat V = Op1->getValueAPF(); |
2926 | if (APFloat::opStatus::opOK == V.remainder(RHS: Op2->getValueAPF())) |
2927 | return ConstantFP::get(Context&: Ty->getContext(), V); |
2928 | } |
2929 | break; |
2930 | case LibFunc_atan2: |
2931 | case LibFunc_atan2f: |
2932 | // atan2(+/-0.0, +/-0.0) is known to raise an exception on some libm |
2933 | // (Solaris), so we do not assume a known result for that. |
2934 | if (Op1V.isZero() && Op2V.isZero()) |
2935 | return nullptr; |
2936 | [[fallthrough]]; |
2937 | case LibFunc_atan2_finite: |
2938 | case LibFunc_atan2f_finite: |
2939 | if (TLI->has(F: Func)) |
2940 | return ConstantFoldBinaryFP(NativeFP: atan2, V: Op1V, W: Op2V, Ty); |
2941 | break; |
2942 | } |
2943 | |
2944 | return nullptr; |
2945 | } |
2946 | |
2947 | static Constant *ConstantFoldIntrinsicCall2(Intrinsic::ID IntrinsicID, Type *Ty, |
2948 | ArrayRef<Constant *> Operands, |
2949 | const CallBase *Call) { |
2950 | assert(Operands.size() == 2 && "Wrong number of operands." ); |
2951 | |
2952 | if (Ty->isFloatingPointTy()) { |
2953 | // TODO: We should have undef handling for all of the FP intrinsics that |
2954 | // are attempted to be folded in this function. |
2955 | bool IsOp0Undef = isa<UndefValue>(Val: Operands[0]); |
2956 | bool IsOp1Undef = isa<UndefValue>(Val: Operands[1]); |
2957 | switch (IntrinsicID) { |
2958 | case Intrinsic::maxnum: |
2959 | case Intrinsic::minnum: |
2960 | case Intrinsic::maximum: |
2961 | case Intrinsic::minimum: |
2962 | case Intrinsic::maximumnum: |
2963 | case Intrinsic::minimumnum: |
2964 | case Intrinsic::nvvm_fmax_d: |
2965 | case Intrinsic::nvvm_fmin_d: |
2966 | // If one argument is undef, return the other argument. |
2967 | if (IsOp0Undef) |
2968 | return Operands[1]; |
2969 | if (IsOp1Undef) |
2970 | return Operands[0]; |
2971 | break; |
2972 | |
2973 | case Intrinsic::nvvm_fmax_f: |
2974 | case Intrinsic::nvvm_fmax_ftz_f: |
2975 | case Intrinsic::nvvm_fmax_ftz_nan_f: |
2976 | case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f: |
2977 | case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f: |
2978 | case Intrinsic::nvvm_fmax_nan_f: |
2979 | case Intrinsic::nvvm_fmax_nan_xorsign_abs_f: |
2980 | case Intrinsic::nvvm_fmax_xorsign_abs_f: |
2981 | |
2982 | case Intrinsic::nvvm_fmin_f: |
2983 | case Intrinsic::nvvm_fmin_ftz_f: |
2984 | case Intrinsic::nvvm_fmin_ftz_nan_f: |
2985 | case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f: |
2986 | case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f: |
2987 | case Intrinsic::nvvm_fmin_nan_f: |
2988 | case Intrinsic::nvvm_fmin_nan_xorsign_abs_f: |
2989 | case Intrinsic::nvvm_fmin_xorsign_abs_f: |
2990 | // If one arg is undef, the other arg can be returned only if it is |
2991 | // constant, as we may need to flush it to sign-preserving zero or |
2992 | // canonicalize the NaN. |
2993 | if (!IsOp0Undef && !IsOp1Undef) |
2994 | break; |
2995 | if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[IsOp0Undef ? 1 : 0])) { |
2996 | if (Op->isNaN()) { |
2997 | APInt NVCanonicalNaN(32, 0x7fffffff); |
2998 | return ConstantFP::get( |
2999 | Ty, V: APFloat(Ty->getFltSemantics(), NVCanonicalNaN)); |
3000 | } |
3001 | if (nvvm::FMinFMaxShouldFTZ(IntrinsicID)) |
3002 | return ConstantFP::get(Ty, V: FTZPreserveSign(V: Op->getValueAPF())); |
3003 | else |
3004 | return Op; |
3005 | } |
3006 | break; |
3007 | } |
3008 | } |
3009 | |
3010 | if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) { |
3011 | const APFloat &Op1V = Op1->getValueAPF(); |
3012 | |
3013 | if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) { |
3014 | if (Op2->getType() != Op1->getType()) |
3015 | return nullptr; |
3016 | const APFloat &Op2V = Op2->getValueAPF(); |
3017 | |
3018 | if (const auto *ConstrIntr = |
3019 | dyn_cast_if_present<ConstrainedFPIntrinsic>(Val: Call)) { |
3020 | RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr); |
3021 | APFloat Res = Op1V; |
3022 | APFloat::opStatus St; |
3023 | switch (IntrinsicID) { |
3024 | default: |
3025 | return nullptr; |
3026 | case Intrinsic::experimental_constrained_fadd: |
3027 | St = Res.add(RHS: Op2V, RM); |
3028 | break; |
3029 | case Intrinsic::experimental_constrained_fsub: |
3030 | St = Res.subtract(RHS: Op2V, RM); |
3031 | break; |
3032 | case Intrinsic::experimental_constrained_fmul: |
3033 | St = Res.multiply(RHS: Op2V, RM); |
3034 | break; |
3035 | case Intrinsic::experimental_constrained_fdiv: |
3036 | St = Res.divide(RHS: Op2V, RM); |
3037 | break; |
3038 | case Intrinsic::experimental_constrained_frem: |
3039 | St = Res.mod(RHS: Op2V); |
3040 | break; |
3041 | case Intrinsic::experimental_constrained_fcmp: |
3042 | case Intrinsic::experimental_constrained_fcmps: |
3043 | return evaluateCompare(Op1: Op1V, Op2: Op2V, Call: ConstrIntr); |
3044 | } |
3045 | if (mayFoldConstrained(CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), |
3046 | St)) |
3047 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
3048 | return nullptr; |
3049 | } |
3050 | |
3051 | switch (IntrinsicID) { |
3052 | default: |
3053 | break; |
3054 | case Intrinsic::copysign: |
3055 | return ConstantFP::get(Context&: Ty->getContext(), V: APFloat::copySign(Value: Op1V, Sign: Op2V)); |
3056 | case Intrinsic::minnum: |
3057 | return ConstantFP::get(Context&: Ty->getContext(), V: minnum(A: Op1V, B: Op2V)); |
3058 | case Intrinsic::maxnum: |
3059 | return ConstantFP::get(Context&: Ty->getContext(), V: maxnum(A: Op1V, B: Op2V)); |
3060 | case Intrinsic::minimum: |
3061 | return ConstantFP::get(Context&: Ty->getContext(), V: minimum(A: Op1V, B: Op2V)); |
3062 | case Intrinsic::maximum: |
3063 | return ConstantFP::get(Context&: Ty->getContext(), V: maximum(A: Op1V, B: Op2V)); |
3064 | case Intrinsic::minimumnum: |
3065 | return ConstantFP::get(Context&: Ty->getContext(), V: minimumnum(A: Op1V, B: Op2V)); |
3066 | case Intrinsic::maximumnum: |
3067 | return ConstantFP::get(Context&: Ty->getContext(), V: maximumnum(A: Op1V, B: Op2V)); |
3068 | |
3069 | case Intrinsic::nvvm_fmax_d: |
3070 | case Intrinsic::nvvm_fmax_f: |
3071 | case Intrinsic::nvvm_fmax_ftz_f: |
3072 | case Intrinsic::nvvm_fmax_ftz_nan_f: |
3073 | case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f: |
3074 | case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f: |
3075 | case Intrinsic::nvvm_fmax_nan_f: |
3076 | case Intrinsic::nvvm_fmax_nan_xorsign_abs_f: |
3077 | case Intrinsic::nvvm_fmax_xorsign_abs_f: |
3078 | |
3079 | case Intrinsic::nvvm_fmin_d: |
3080 | case Intrinsic::nvvm_fmin_f: |
3081 | case Intrinsic::nvvm_fmin_ftz_f: |
3082 | case Intrinsic::nvvm_fmin_ftz_nan_f: |
3083 | case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f: |
3084 | case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f: |
3085 | case Intrinsic::nvvm_fmin_nan_f: |
3086 | case Intrinsic::nvvm_fmin_nan_xorsign_abs_f: |
3087 | case Intrinsic::nvvm_fmin_xorsign_abs_f: { |
3088 | |
3089 | bool ShouldCanonicalizeNaNs = !(IntrinsicID == Intrinsic::nvvm_fmax_d || |
3090 | IntrinsicID == Intrinsic::nvvm_fmin_d); |
3091 | bool IsFTZ = nvvm::FMinFMaxShouldFTZ(IntrinsicID); |
3092 | bool IsNaNPropagating = nvvm::FMinFMaxPropagatesNaNs(IntrinsicID); |
3093 | bool IsXorSignAbs = nvvm::FMinFMaxIsXorSignAbs(IntrinsicID); |
3094 | |
3095 | APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V; |
3096 | APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V; |
3097 | |
3098 | bool XorSign = false; |
3099 | if (IsXorSignAbs) { |
3100 | XorSign = A.isNegative() ^ B.isNegative(); |
3101 | A = abs(X: A); |
3102 | B = abs(X: B); |
3103 | } |
3104 | |
3105 | bool IsFMax = false; |
3106 | switch (IntrinsicID) { |
3107 | case Intrinsic::nvvm_fmax_d: |
3108 | case Intrinsic::nvvm_fmax_f: |
3109 | case Intrinsic::nvvm_fmax_ftz_f: |
3110 | case Intrinsic::nvvm_fmax_ftz_nan_f: |
3111 | case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f: |
3112 | case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f: |
3113 | case Intrinsic::nvvm_fmax_nan_f: |
3114 | case Intrinsic::nvvm_fmax_nan_xorsign_abs_f: |
3115 | case Intrinsic::nvvm_fmax_xorsign_abs_f: |
3116 | IsFMax = true; |
3117 | break; |
3118 | } |
3119 | APFloat Res = IsFMax ? maximum(A, B) : minimum(A, B); |
3120 | |
3121 | if (ShouldCanonicalizeNaNs) { |
3122 | APFloat NVCanonicalNaN(Res.getSemantics(), APInt(32, 0x7fffffff)); |
3123 | if (A.isNaN() && B.isNaN()) |
3124 | return ConstantFP::get(Ty, V: NVCanonicalNaN); |
3125 | else if (IsNaNPropagating && (A.isNaN() || B.isNaN())) |
3126 | return ConstantFP::get(Ty, V: NVCanonicalNaN); |
3127 | } |
3128 | |
3129 | if (A.isNaN() && B.isNaN()) |
3130 | return Operands[1]; |
3131 | else if (A.isNaN()) |
3132 | Res = B; |
3133 | else if (B.isNaN()) |
3134 | Res = A; |
3135 | |
3136 | if (IsXorSignAbs && XorSign != Res.isNegative()) |
3137 | Res.changeSign(); |
3138 | |
3139 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
3140 | } |
3141 | } |
3142 | |
3143 | if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) |
3144 | return nullptr; |
3145 | |
3146 | switch (IntrinsicID) { |
3147 | default: |
3148 | break; |
3149 | case Intrinsic::pow: |
3150 | return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty); |
3151 | case Intrinsic::amdgcn_fmul_legacy: |
3152 | // The legacy behaviour is that multiplying +/- 0.0 by anything, even |
3153 | // NaN or infinity, gives +0.0. |
3154 | if (Op1V.isZero() || Op2V.isZero()) |
3155 | return ConstantFP::getZero(Ty); |
3156 | return ConstantFP::get(Context&: Ty->getContext(), V: Op1V * Op2V); |
3157 | } |
3158 | |
3159 | } else if (auto *Op2C = dyn_cast<ConstantInt>(Val: Operands[1])) { |
3160 | switch (IntrinsicID) { |
3161 | case Intrinsic::ldexp: { |
3162 | return ConstantFP::get( |
3163 | Context&: Ty->getContext(), |
3164 | V: scalbn(X: Op1V, Exp: Op2C->getSExtValue(), RM: APFloat::rmNearestTiesToEven)); |
3165 | } |
3166 | case Intrinsic::is_fpclass: { |
3167 | FPClassTest Mask = static_cast<FPClassTest>(Op2C->getZExtValue()); |
3168 | bool Result = |
3169 | ((Mask & fcSNan) && Op1V.isNaN() && Op1V.isSignaling()) || |
3170 | ((Mask & fcQNan) && Op1V.isNaN() && !Op1V.isSignaling()) || |
3171 | ((Mask & fcNegInf) && Op1V.isNegInfinity()) || |
3172 | ((Mask & fcNegNormal) && Op1V.isNormal() && Op1V.isNegative()) || |
3173 | ((Mask & fcNegSubnormal) && Op1V.isDenormal() && Op1V.isNegative()) || |
3174 | ((Mask & fcNegZero) && Op1V.isZero() && Op1V.isNegative()) || |
3175 | ((Mask & fcPosZero) && Op1V.isZero() && !Op1V.isNegative()) || |
3176 | ((Mask & fcPosSubnormal) && Op1V.isDenormal() && !Op1V.isNegative()) || |
3177 | ((Mask & fcPosNormal) && Op1V.isNormal() && !Op1V.isNegative()) || |
3178 | ((Mask & fcPosInf) && Op1V.isPosInfinity()); |
3179 | return ConstantInt::get(Ty, V: Result); |
3180 | } |
3181 | case Intrinsic::powi: { |
3182 | int Exp = static_cast<int>(Op2C->getSExtValue()); |
3183 | switch (Ty->getTypeID()) { |
3184 | case Type::HalfTyID: |
3185 | case Type::FloatTyID: { |
3186 | APFloat Res(static_cast<float>(std::pow(x: Op1V.convertToFloat(), y: Exp))); |
3187 | if (Ty->isHalfTy()) { |
3188 | bool Unused; |
3189 | Res.convert(ToSemantics: APFloat::IEEEhalf(), RM: APFloat::rmNearestTiesToEven, |
3190 | losesInfo: &Unused); |
3191 | } |
3192 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
3193 | } |
3194 | case Type::DoubleTyID: |
3195 | return ConstantFP::get(Ty, V: std::pow(x: Op1V.convertToDouble(), y: Exp)); |
3196 | default: |
3197 | return nullptr; |
3198 | } |
3199 | } |
3200 | default: |
3201 | break; |
3202 | } |
3203 | } |
3204 | return nullptr; |
3205 | } |
3206 | |
3207 | if (Operands[0]->getType()->isIntegerTy() && |
3208 | Operands[1]->getType()->isIntegerTy()) { |
3209 | const APInt *C0, *C1; |
3210 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3211 | !getConstIntOrUndef(Op: Operands[1], C&: C1)) |
3212 | return nullptr; |
3213 | |
3214 | switch (IntrinsicID) { |
3215 | default: break; |
3216 | case Intrinsic::smax: |
3217 | case Intrinsic::smin: |
3218 | case Intrinsic::umax: |
3219 | case Intrinsic::umin: |
3220 | if (!C0 && !C1) |
3221 | return UndefValue::get(T: Ty); |
3222 | if (!C0 || !C1) |
3223 | return MinMaxIntrinsic::getSaturationPoint(ID: IntrinsicID, Ty); |
3224 | return ConstantInt::get( |
3225 | Ty, V: ICmpInst::compare(LHS: *C0, RHS: *C1, |
3226 | Pred: MinMaxIntrinsic::getPredicate(ID: IntrinsicID)) |
3227 | ? *C0 |
3228 | : *C1); |
3229 | |
3230 | case Intrinsic::scmp: |
3231 | case Intrinsic::ucmp: |
3232 | if (!C0 || !C1) |
3233 | return ConstantInt::get(Ty, V: 0); |
3234 | |
3235 | int Res; |
3236 | if (IntrinsicID == Intrinsic::scmp) |
3237 | Res = C0->sgt(RHS: *C1) ? 1 : C0->slt(RHS: *C1) ? -1 : 0; |
3238 | else |
3239 | Res = C0->ugt(RHS: *C1) ? 1 : C0->ult(RHS: *C1) ? -1 : 0; |
3240 | return ConstantInt::get(Ty, V: Res, /*IsSigned=*/true); |
3241 | |
3242 | case Intrinsic::usub_with_overflow: |
3243 | case Intrinsic::ssub_with_overflow: |
3244 | // X - undef -> { 0, false } |
3245 | // undef - X -> { 0, false } |
3246 | if (!C0 || !C1) |
3247 | return Constant::getNullValue(Ty); |
3248 | [[fallthrough]]; |
3249 | case Intrinsic::uadd_with_overflow: |
3250 | case Intrinsic::sadd_with_overflow: |
3251 | // X + undef -> { -1, false } |
3252 | // undef + x -> { -1, false } |
3253 | if (!C0 || !C1) { |
3254 | return ConstantStruct::get( |
3255 | T: cast<StructType>(Val: Ty), |
3256 | V: {Constant::getAllOnesValue(Ty: Ty->getStructElementType(N: 0)), |
3257 | Constant::getNullValue(Ty: Ty->getStructElementType(N: 1))}); |
3258 | } |
3259 | [[fallthrough]]; |
3260 | case Intrinsic::smul_with_overflow: |
3261 | case Intrinsic::umul_with_overflow: { |
3262 | // undef * X -> { 0, false } |
3263 | // X * undef -> { 0, false } |
3264 | if (!C0 || !C1) |
3265 | return Constant::getNullValue(Ty); |
3266 | |
3267 | APInt Res; |
3268 | bool Overflow; |
3269 | switch (IntrinsicID) { |
3270 | default: llvm_unreachable("Invalid case" ); |
3271 | case Intrinsic::sadd_with_overflow: |
3272 | Res = C0->sadd_ov(RHS: *C1, Overflow); |
3273 | break; |
3274 | case Intrinsic::uadd_with_overflow: |
3275 | Res = C0->uadd_ov(RHS: *C1, Overflow); |
3276 | break; |
3277 | case Intrinsic::ssub_with_overflow: |
3278 | Res = C0->ssub_ov(RHS: *C1, Overflow); |
3279 | break; |
3280 | case Intrinsic::usub_with_overflow: |
3281 | Res = C0->usub_ov(RHS: *C1, Overflow); |
3282 | break; |
3283 | case Intrinsic::smul_with_overflow: |
3284 | Res = C0->smul_ov(RHS: *C1, Overflow); |
3285 | break; |
3286 | case Intrinsic::umul_with_overflow: |
3287 | Res = C0->umul_ov(RHS: *C1, Overflow); |
3288 | break; |
3289 | } |
3290 | Constant *Ops[] = { |
3291 | ConstantInt::get(Context&: Ty->getContext(), V: Res), |
3292 | ConstantInt::get(Ty: Type::getInt1Ty(C&: Ty->getContext()), V: Overflow) |
3293 | }; |
3294 | return ConstantStruct::get(T: cast<StructType>(Val: Ty), V: Ops); |
3295 | } |
3296 | case Intrinsic::uadd_sat: |
3297 | case Intrinsic::sadd_sat: |
3298 | if (!C0 && !C1) |
3299 | return UndefValue::get(T: Ty); |
3300 | if (!C0 || !C1) |
3301 | return Constant::getAllOnesValue(Ty); |
3302 | if (IntrinsicID == Intrinsic::uadd_sat) |
3303 | return ConstantInt::get(Ty, V: C0->uadd_sat(RHS: *C1)); |
3304 | else |
3305 | return ConstantInt::get(Ty, V: C0->sadd_sat(RHS: *C1)); |
3306 | case Intrinsic::usub_sat: |
3307 | case Intrinsic::ssub_sat: |
3308 | if (!C0 && !C1) |
3309 | return UndefValue::get(T: Ty); |
3310 | if (!C0 || !C1) |
3311 | return Constant::getNullValue(Ty); |
3312 | if (IntrinsicID == Intrinsic::usub_sat) |
3313 | return ConstantInt::get(Ty, V: C0->usub_sat(RHS: *C1)); |
3314 | else |
3315 | return ConstantInt::get(Ty, V: C0->ssub_sat(RHS: *C1)); |
3316 | case Intrinsic::cttz: |
3317 | case Intrinsic::ctlz: |
3318 | assert(C1 && "Must be constant int" ); |
3319 | |
3320 | // cttz(0, 1) and ctlz(0, 1) are poison. |
3321 | if (C1->isOne() && (!C0 || C0->isZero())) |
3322 | return PoisonValue::get(T: Ty); |
3323 | if (!C0) |
3324 | return Constant::getNullValue(Ty); |
3325 | if (IntrinsicID == Intrinsic::cttz) |
3326 | return ConstantInt::get(Ty, V: C0->countr_zero()); |
3327 | else |
3328 | return ConstantInt::get(Ty, V: C0->countl_zero()); |
3329 | |
3330 | case Intrinsic::abs: |
3331 | assert(C1 && "Must be constant int" ); |
3332 | assert((C1->isOne() || C1->isZero()) && "Must be 0 or 1" ); |
3333 | |
3334 | // Undef or minimum val operand with poison min --> poison |
3335 | if (C1->isOne() && (!C0 || C0->isMinSignedValue())) |
3336 | return PoisonValue::get(T: Ty); |
3337 | |
3338 | // Undef operand with no poison min --> 0 (sign bit must be clear) |
3339 | if (!C0) |
3340 | return Constant::getNullValue(Ty); |
3341 | |
3342 | return ConstantInt::get(Ty, V: C0->abs()); |
3343 | case Intrinsic::amdgcn_wave_reduce_umin: |
3344 | case Intrinsic::amdgcn_wave_reduce_umax: |
3345 | return dyn_cast<Constant>(Val: Operands[0]); |
3346 | } |
3347 | |
3348 | return nullptr; |
3349 | } |
3350 | |
3351 | // Support ConstantVector in case we have an Undef in the top. |
3352 | if ((isa<ConstantVector>(Val: Operands[0]) || |
3353 | isa<ConstantDataVector>(Val: Operands[0])) && |
3354 | // Check for default rounding mode. |
3355 | // FIXME: Support other rounding modes? |
3356 | isa<ConstantInt>(Val: Operands[1]) && |
3357 | cast<ConstantInt>(Val: Operands[1])->getValue() == 4) { |
3358 | auto *Op = cast<Constant>(Val: Operands[0]); |
3359 | switch (IntrinsicID) { |
3360 | default: break; |
3361 | case Intrinsic::x86_avx512_vcvtss2si32: |
3362 | case Intrinsic::x86_avx512_vcvtss2si64: |
3363 | case Intrinsic::x86_avx512_vcvtsd2si32: |
3364 | case Intrinsic::x86_avx512_vcvtsd2si64: |
3365 | if (ConstantFP *FPOp = |
3366 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
3367 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
3368 | /*roundTowardZero=*/false, Ty, |
3369 | /*IsSigned*/true); |
3370 | break; |
3371 | case Intrinsic::x86_avx512_vcvtss2usi32: |
3372 | case Intrinsic::x86_avx512_vcvtss2usi64: |
3373 | case Intrinsic::x86_avx512_vcvtsd2usi32: |
3374 | case Intrinsic::x86_avx512_vcvtsd2usi64: |
3375 | if (ConstantFP *FPOp = |
3376 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
3377 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
3378 | /*roundTowardZero=*/false, Ty, |
3379 | /*IsSigned*/false); |
3380 | break; |
3381 | case Intrinsic::x86_avx512_cvttss2si: |
3382 | case Intrinsic::x86_avx512_cvttss2si64: |
3383 | case Intrinsic::x86_avx512_cvttsd2si: |
3384 | case Intrinsic::x86_avx512_cvttsd2si64: |
3385 | if (ConstantFP *FPOp = |
3386 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
3387 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
3388 | /*roundTowardZero=*/true, Ty, |
3389 | /*IsSigned*/true); |
3390 | break; |
3391 | case Intrinsic::x86_avx512_cvttss2usi: |
3392 | case Intrinsic::x86_avx512_cvttss2usi64: |
3393 | case Intrinsic::x86_avx512_cvttsd2usi: |
3394 | case Intrinsic::x86_avx512_cvttsd2usi64: |
3395 | if (ConstantFP *FPOp = |
3396 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
3397 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
3398 | /*roundTowardZero=*/true, Ty, |
3399 | /*IsSigned*/false); |
3400 | break; |
3401 | } |
3402 | } |
3403 | return nullptr; |
3404 | } |
3405 | |
3406 | static APFloat ConstantFoldAMDGCNCubeIntrinsic(Intrinsic::ID IntrinsicID, |
3407 | const APFloat &S0, |
3408 | const APFloat &S1, |
3409 | const APFloat &S2) { |
3410 | unsigned ID; |
3411 | const fltSemantics &Sem = S0.getSemantics(); |
3412 | APFloat MA(Sem), SC(Sem), TC(Sem); |
3413 | if (abs(X: S2) >= abs(X: S0) && abs(X: S2) >= abs(X: S1)) { |
3414 | if (S2.isNegative() && S2.isNonZero() && !S2.isNaN()) { |
3415 | // S2 < 0 |
3416 | ID = 5; |
3417 | SC = -S0; |
3418 | } else { |
3419 | ID = 4; |
3420 | SC = S0; |
3421 | } |
3422 | MA = S2; |
3423 | TC = -S1; |
3424 | } else if (abs(X: S1) >= abs(X: S0)) { |
3425 | if (S1.isNegative() && S1.isNonZero() && !S1.isNaN()) { |
3426 | // S1 < 0 |
3427 | ID = 3; |
3428 | TC = -S2; |
3429 | } else { |
3430 | ID = 2; |
3431 | TC = S2; |
3432 | } |
3433 | MA = S1; |
3434 | SC = S0; |
3435 | } else { |
3436 | if (S0.isNegative() && S0.isNonZero() && !S0.isNaN()) { |
3437 | // S0 < 0 |
3438 | ID = 1; |
3439 | SC = S2; |
3440 | } else { |
3441 | ID = 0; |
3442 | SC = -S2; |
3443 | } |
3444 | MA = S0; |
3445 | TC = -S1; |
3446 | } |
3447 | switch (IntrinsicID) { |
3448 | default: |
3449 | llvm_unreachable("unhandled amdgcn cube intrinsic" ); |
3450 | case Intrinsic::amdgcn_cubeid: |
3451 | return APFloat(Sem, ID); |
3452 | case Intrinsic::amdgcn_cubema: |
3453 | return MA + MA; |
3454 | case Intrinsic::amdgcn_cubesc: |
3455 | return SC; |
3456 | case Intrinsic::amdgcn_cubetc: |
3457 | return TC; |
3458 | } |
3459 | } |
3460 | |
3461 | static Constant *ConstantFoldAMDGCNPermIntrinsic(ArrayRef<Constant *> Operands, |
3462 | Type *Ty) { |
3463 | const APInt *C0, *C1, *C2; |
3464 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3465 | !getConstIntOrUndef(Op: Operands[1], C&: C1) || |
3466 | !getConstIntOrUndef(Op: Operands[2], C&: C2)) |
3467 | return nullptr; |
3468 | |
3469 | if (!C2) |
3470 | return UndefValue::get(T: Ty); |
3471 | |
3472 | APInt Val(32, 0); |
3473 | unsigned NumUndefBytes = 0; |
3474 | for (unsigned I = 0; I < 32; I += 8) { |
3475 | unsigned Sel = C2->extractBitsAsZExtValue(numBits: 8, bitPosition: I); |
3476 | unsigned B = 0; |
3477 | |
3478 | if (Sel >= 13) |
3479 | B = 0xff; |
3480 | else if (Sel == 12) |
3481 | B = 0x00; |
3482 | else { |
3483 | const APInt *Src = ((Sel & 10) == 10 || (Sel & 12) == 4) ? C0 : C1; |
3484 | if (!Src) |
3485 | ++NumUndefBytes; |
3486 | else if (Sel < 8) |
3487 | B = Src->extractBitsAsZExtValue(numBits: 8, bitPosition: (Sel & 3) * 8); |
3488 | else |
3489 | B = Src->extractBitsAsZExtValue(numBits: 1, bitPosition: (Sel & 1) ? 31 : 15) * 0xff; |
3490 | } |
3491 | |
3492 | Val.insertBits(SubBits: B, bitPosition: I, numBits: 8); |
3493 | } |
3494 | |
3495 | if (NumUndefBytes == 4) |
3496 | return UndefValue::get(T: Ty); |
3497 | |
3498 | return ConstantInt::get(Ty, V: Val); |
3499 | } |
3500 | |
3501 | static Constant *ConstantFoldScalarCall3(StringRef Name, |
3502 | Intrinsic::ID IntrinsicID, |
3503 | Type *Ty, |
3504 | ArrayRef<Constant *> Operands, |
3505 | const TargetLibraryInfo *TLI, |
3506 | const CallBase *Call) { |
3507 | assert(Operands.size() == 3 && "Wrong number of operands." ); |
3508 | |
3509 | if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) { |
3510 | if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) { |
3511 | if (const auto *Op3 = dyn_cast<ConstantFP>(Val: Operands[2])) { |
3512 | const APFloat &C1 = Op1->getValueAPF(); |
3513 | const APFloat &C2 = Op2->getValueAPF(); |
3514 | const APFloat &C3 = Op3->getValueAPF(); |
3515 | |
3516 | if (const auto *ConstrIntr = dyn_cast<ConstrainedFPIntrinsic>(Val: Call)) { |
3517 | RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr); |
3518 | APFloat Res = C1; |
3519 | APFloat::opStatus St; |
3520 | switch (IntrinsicID) { |
3521 | default: |
3522 | return nullptr; |
3523 | case Intrinsic::experimental_constrained_fma: |
3524 | case Intrinsic::experimental_constrained_fmuladd: |
3525 | St = Res.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM); |
3526 | break; |
3527 | } |
3528 | if (mayFoldConstrained( |
3529 | CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), St)) |
3530 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
3531 | return nullptr; |
3532 | } |
3533 | |
3534 | switch (IntrinsicID) { |
3535 | default: break; |
3536 | case Intrinsic::amdgcn_fma_legacy: { |
3537 | // The legacy behaviour is that multiplying +/- 0.0 by anything, even |
3538 | // NaN or infinity, gives +0.0. |
3539 | if (C1.isZero() || C2.isZero()) { |
3540 | // It's tempting to just return C3 here, but that would give the |
3541 | // wrong result if C3 was -0.0. |
3542 | return ConstantFP::get(Context&: Ty->getContext(), V: APFloat(0.0f) + C3); |
3543 | } |
3544 | [[fallthrough]]; |
3545 | } |
3546 | case Intrinsic::fma: |
3547 | case Intrinsic::fmuladd: { |
3548 | APFloat V = C1; |
3549 | V.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM: APFloat::rmNearestTiesToEven); |
3550 | return ConstantFP::get(Context&: Ty->getContext(), V); |
3551 | } |
3552 | case Intrinsic::amdgcn_cubeid: |
3553 | case Intrinsic::amdgcn_cubema: |
3554 | case Intrinsic::amdgcn_cubesc: |
3555 | case Intrinsic::amdgcn_cubetc: { |
3556 | APFloat V = ConstantFoldAMDGCNCubeIntrinsic(IntrinsicID, S0: C1, S1: C2, S2: C3); |
3557 | return ConstantFP::get(Context&: Ty->getContext(), V); |
3558 | } |
3559 | } |
3560 | } |
3561 | } |
3562 | } |
3563 | |
3564 | if (IntrinsicID == Intrinsic::smul_fix || |
3565 | IntrinsicID == Intrinsic::smul_fix_sat) { |
3566 | const APInt *C0, *C1; |
3567 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3568 | !getConstIntOrUndef(Op: Operands[1], C&: C1)) |
3569 | return nullptr; |
3570 | |
3571 | // undef * C -> 0 |
3572 | // C * undef -> 0 |
3573 | if (!C0 || !C1) |
3574 | return Constant::getNullValue(Ty); |
3575 | |
3576 | // This code performs rounding towards negative infinity in case the result |
3577 | // cannot be represented exactly for the given scale. Targets that do care |
3578 | // about rounding should use a target hook for specifying how rounding |
3579 | // should be done, and provide their own folding to be consistent with |
3580 | // rounding. This is the same approach as used by |
3581 | // DAGTypeLegalizer::ExpandIntRes_MULFIX. |
3582 | unsigned Scale = cast<ConstantInt>(Val: Operands[2])->getZExtValue(); |
3583 | unsigned Width = C0->getBitWidth(); |
3584 | assert(Scale < Width && "Illegal scale." ); |
3585 | unsigned ExtendedWidth = Width * 2; |
3586 | APInt Product = |
3587 | (C0->sext(width: ExtendedWidth) * C1->sext(width: ExtendedWidth)).ashr(ShiftAmt: Scale); |
3588 | if (IntrinsicID == Intrinsic::smul_fix_sat) { |
3589 | APInt Max = APInt::getSignedMaxValue(numBits: Width).sext(width: ExtendedWidth); |
3590 | APInt Min = APInt::getSignedMinValue(numBits: Width).sext(width: ExtendedWidth); |
3591 | Product = APIntOps::smin(A: Product, B: Max); |
3592 | Product = APIntOps::smax(A: Product, B: Min); |
3593 | } |
3594 | return ConstantInt::get(Context&: Ty->getContext(), V: Product.sextOrTrunc(width: Width)); |
3595 | } |
3596 | |
3597 | if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) { |
3598 | const APInt *C0, *C1, *C2; |
3599 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3600 | !getConstIntOrUndef(Op: Operands[1], C&: C1) || |
3601 | !getConstIntOrUndef(Op: Operands[2], C&: C2)) |
3602 | return nullptr; |
3603 | |
3604 | bool IsRight = IntrinsicID == Intrinsic::fshr; |
3605 | if (!C2) |
3606 | return Operands[IsRight ? 1 : 0]; |
3607 | if (!C0 && !C1) |
3608 | return UndefValue::get(T: Ty); |
3609 | |
3610 | // The shift amount is interpreted as modulo the bitwidth. If the shift |
3611 | // amount is effectively 0, avoid UB due to oversized inverse shift below. |
3612 | unsigned BitWidth = C2->getBitWidth(); |
3613 | unsigned ShAmt = C2->urem(RHS: BitWidth); |
3614 | if (!ShAmt) |
3615 | return Operands[IsRight ? 1 : 0]; |
3616 | |
3617 | // (C0 << ShlAmt) | (C1 >> LshrAmt) |
3618 | unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt; |
3619 | unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt; |
3620 | if (!C0) |
3621 | return ConstantInt::get(Ty, V: C1->lshr(shiftAmt: LshrAmt)); |
3622 | if (!C1) |
3623 | return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt)); |
3624 | return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt) | C1->lshr(shiftAmt: LshrAmt)); |
3625 | } |
3626 | |
3627 | if (IntrinsicID == Intrinsic::amdgcn_perm) |
3628 | return ConstantFoldAMDGCNPermIntrinsic(Operands, Ty); |
3629 | |
3630 | return nullptr; |
3631 | } |
3632 | |
3633 | static Constant *ConstantFoldScalarCall(StringRef Name, |
3634 | Intrinsic::ID IntrinsicID, |
3635 | Type *Ty, |
3636 | ArrayRef<Constant *> Operands, |
3637 | const TargetLibraryInfo *TLI, |
3638 | const CallBase *Call) { |
3639 | if (IntrinsicID != Intrinsic::not_intrinsic && |
3640 | any_of(Range&: Operands, P: IsaPred<PoisonValue>) && |
3641 | intrinsicPropagatesPoison(IID: IntrinsicID)) |
3642 | return PoisonValue::get(T: Ty); |
3643 | |
3644 | if (Operands.size() == 1) |
3645 | return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call); |
3646 | |
3647 | if (Operands.size() == 2) { |
3648 | if (Constant *FoldedLibCall = |
3649 | ConstantFoldLibCall2(Name, Ty, Operands, TLI)) { |
3650 | return FoldedLibCall; |
3651 | } |
3652 | return ConstantFoldIntrinsicCall2(IntrinsicID, Ty, Operands, Call); |
3653 | } |
3654 | |
3655 | if (Operands.size() == 3) |
3656 | return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call); |
3657 | |
3658 | return nullptr; |
3659 | } |
3660 | |
3661 | static Constant *ConstantFoldFixedVectorCall( |
3662 | StringRef Name, Intrinsic::ID IntrinsicID, FixedVectorType *FVTy, |
3663 | ArrayRef<Constant *> Operands, const DataLayout &DL, |
3664 | const TargetLibraryInfo *TLI, const CallBase *Call) { |
3665 | SmallVector<Constant *, 4> Result(FVTy->getNumElements()); |
3666 | SmallVector<Constant *, 4> Lane(Operands.size()); |
3667 | Type *Ty = FVTy->getElementType(); |
3668 | |
3669 | switch (IntrinsicID) { |
3670 | case Intrinsic::masked_load: { |
3671 | auto *SrcPtr = Operands[0]; |
3672 | auto *Mask = Operands[2]; |
3673 | auto *Passthru = Operands[3]; |
3674 | |
3675 | Constant *VecData = ConstantFoldLoadFromConstPtr(C: SrcPtr, Ty: FVTy, DL); |
3676 | |
3677 | SmallVector<Constant *, 32> NewElements; |
3678 | for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) { |
3679 | auto *MaskElt = Mask->getAggregateElement(Elt: I); |
3680 | if (!MaskElt) |
3681 | break; |
3682 | auto *PassthruElt = Passthru->getAggregateElement(Elt: I); |
3683 | auto *VecElt = VecData ? VecData->getAggregateElement(Elt: I) : nullptr; |
3684 | if (isa<UndefValue>(Val: MaskElt)) { |
3685 | if (PassthruElt) |
3686 | NewElements.push_back(Elt: PassthruElt); |
3687 | else if (VecElt) |
3688 | NewElements.push_back(Elt: VecElt); |
3689 | else |
3690 | return nullptr; |
3691 | } |
3692 | if (MaskElt->isNullValue()) { |
3693 | if (!PassthruElt) |
3694 | return nullptr; |
3695 | NewElements.push_back(Elt: PassthruElt); |
3696 | } else if (MaskElt->isOneValue()) { |
3697 | if (!VecElt) |
3698 | return nullptr; |
3699 | NewElements.push_back(Elt: VecElt); |
3700 | } else { |
3701 | return nullptr; |
3702 | } |
3703 | } |
3704 | if (NewElements.size() != FVTy->getNumElements()) |
3705 | return nullptr; |
3706 | return ConstantVector::get(V: NewElements); |
3707 | } |
3708 | case Intrinsic::arm_mve_vctp8: |
3709 | case Intrinsic::arm_mve_vctp16: |
3710 | case Intrinsic::arm_mve_vctp32: |
3711 | case Intrinsic::arm_mve_vctp64: { |
3712 | if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) { |
3713 | unsigned Lanes = FVTy->getNumElements(); |
3714 | uint64_t Limit = Op->getZExtValue(); |
3715 | |
3716 | SmallVector<Constant *, 16> NCs; |
3717 | for (unsigned i = 0; i < Lanes; i++) { |
3718 | if (i < Limit) |
3719 | NCs.push_back(Elt: ConstantInt::getTrue(Ty)); |
3720 | else |
3721 | NCs.push_back(Elt: ConstantInt::getFalse(Ty)); |
3722 | } |
3723 | return ConstantVector::get(V: NCs); |
3724 | } |
3725 | return nullptr; |
3726 | } |
3727 | case Intrinsic::get_active_lane_mask: { |
3728 | auto *Op0 = dyn_cast<ConstantInt>(Val: Operands[0]); |
3729 | auto *Op1 = dyn_cast<ConstantInt>(Val: Operands[1]); |
3730 | if (Op0 && Op1) { |
3731 | unsigned Lanes = FVTy->getNumElements(); |
3732 | uint64_t Base = Op0->getZExtValue(); |
3733 | uint64_t Limit = Op1->getZExtValue(); |
3734 | |
3735 | SmallVector<Constant *, 16> NCs; |
3736 | for (unsigned i = 0; i < Lanes; i++) { |
3737 | if (Base + i < Limit) |
3738 | NCs.push_back(Elt: ConstantInt::getTrue(Ty)); |
3739 | else |
3740 | NCs.push_back(Elt: ConstantInt::getFalse(Ty)); |
3741 | } |
3742 | return ConstantVector::get(V: NCs); |
3743 | } |
3744 | return nullptr; |
3745 | } |
3746 | case Intrinsic::vector_extract: { |
3747 | auto *Idx = dyn_cast<ConstantInt>(Val: Operands[1]); |
3748 | Constant *Vec = Operands[0]; |
3749 | if (!Idx || !isa<FixedVectorType>(Val: Vec->getType())) |
3750 | return nullptr; |
3751 | |
3752 | unsigned NumElements = FVTy->getNumElements(); |
3753 | unsigned VecNumElements = |
3754 | cast<FixedVectorType>(Val: Vec->getType())->getNumElements(); |
3755 | unsigned StartingIndex = Idx->getZExtValue(); |
3756 | |
3757 | // Extracting entire vector is nop |
3758 | if (NumElements == VecNumElements && StartingIndex == 0) |
3759 | return Vec; |
3760 | |
3761 | for (unsigned I = StartingIndex, E = StartingIndex + NumElements; I < E; |
3762 | ++I) { |
3763 | Constant *Elt = Vec->getAggregateElement(Elt: I); |
3764 | if (!Elt) |
3765 | return nullptr; |
3766 | Result[I - StartingIndex] = Elt; |
3767 | } |
3768 | |
3769 | return ConstantVector::get(V: Result); |
3770 | } |
3771 | case Intrinsic::vector_insert: { |
3772 | Constant *Vec = Operands[0]; |
3773 | Constant *SubVec = Operands[1]; |
3774 | auto *Idx = dyn_cast<ConstantInt>(Val: Operands[2]); |
3775 | if (!Idx || !isa<FixedVectorType>(Val: Vec->getType())) |
3776 | return nullptr; |
3777 | |
3778 | unsigned SubVecNumElements = |
3779 | cast<FixedVectorType>(Val: SubVec->getType())->getNumElements(); |
3780 | unsigned VecNumElements = |
3781 | cast<FixedVectorType>(Val: Vec->getType())->getNumElements(); |
3782 | unsigned IdxN = Idx->getZExtValue(); |
3783 | // Replacing entire vector with a subvec is nop |
3784 | if (SubVecNumElements == VecNumElements && IdxN == 0) |
3785 | return SubVec; |
3786 | |
3787 | for (unsigned I = 0; I < VecNumElements; ++I) { |
3788 | Constant *Elt; |
3789 | if (I < IdxN + SubVecNumElements) |
3790 | Elt = SubVec->getAggregateElement(Elt: I - IdxN); |
3791 | else |
3792 | Elt = Vec->getAggregateElement(Elt: I); |
3793 | if (!Elt) |
3794 | return nullptr; |
3795 | Result[I] = Elt; |
3796 | } |
3797 | return ConstantVector::get(V: Result); |
3798 | } |
3799 | case Intrinsic::vector_interleave2: { |
3800 | unsigned NumElements = |
3801 | cast<FixedVectorType>(Val: Operands[0]->getType())->getNumElements(); |
3802 | for (unsigned I = 0; I < NumElements; ++I) { |
3803 | Constant *Elt0 = Operands[0]->getAggregateElement(Elt: I); |
3804 | Constant *Elt1 = Operands[1]->getAggregateElement(Elt: I); |
3805 | if (!Elt0 || !Elt1) |
3806 | return nullptr; |
3807 | Result[2 * I] = Elt0; |
3808 | Result[2 * I + 1] = Elt1; |
3809 | } |
3810 | return ConstantVector::get(V: Result); |
3811 | } |
3812 | default: |
3813 | break; |
3814 | } |
3815 | |
3816 | for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) { |
3817 | // Gather a column of constants. |
3818 | for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) { |
3819 | // Some intrinsics use a scalar type for certain arguments. |
3820 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: J, /*TTI=*/nullptr)) { |
3821 | Lane[J] = Operands[J]; |
3822 | continue; |
3823 | } |
3824 | |
3825 | Constant *Agg = Operands[J]->getAggregateElement(Elt: I); |
3826 | if (!Agg) |
3827 | return nullptr; |
3828 | |
3829 | Lane[J] = Agg; |
3830 | } |
3831 | |
3832 | // Use the regular scalar folding to simplify this column. |
3833 | Constant *Folded = |
3834 | ConstantFoldScalarCall(Name, IntrinsicID, Ty, Operands: Lane, TLI, Call); |
3835 | if (!Folded) |
3836 | return nullptr; |
3837 | Result[I] = Folded; |
3838 | } |
3839 | |
3840 | return ConstantVector::get(V: Result); |
3841 | } |
3842 | |
3843 | static Constant *ConstantFoldScalableVectorCall( |
3844 | StringRef Name, Intrinsic::ID IntrinsicID, ScalableVectorType *SVTy, |
3845 | ArrayRef<Constant *> Operands, const DataLayout &DL, |
3846 | const TargetLibraryInfo *TLI, const CallBase *Call) { |
3847 | switch (IntrinsicID) { |
3848 | case Intrinsic::aarch64_sve_convert_from_svbool: { |
3849 | auto *Src = dyn_cast<Constant>(Val: Operands[0]); |
3850 | if (!Src || !Src->isNullValue()) |
3851 | break; |
3852 | |
3853 | return ConstantInt::getFalse(Ty: SVTy); |
3854 | } |
3855 | default: |
3856 | break; |
3857 | } |
3858 | |
3859 | // If trivially vectorizable, try folding it via the scalar call if all |
3860 | // operands are splats. |
3861 | |
3862 | // TODO: ConstantFoldFixedVectorCall should probably check this too? |
3863 | if (!isTriviallyVectorizable(ID: IntrinsicID)) |
3864 | return nullptr; |
3865 | |
3866 | SmallVector<Constant *, 4> SplatOps; |
3867 | for (auto [I, Op] : enumerate(First&: Operands)) { |
3868 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: I, /*TTI=*/nullptr)) { |
3869 | SplatOps.push_back(Elt: Op); |
3870 | continue; |
3871 | } |
3872 | Constant *Splat = Op->getSplatValue(); |
3873 | if (!Splat) |
3874 | return nullptr; |
3875 | SplatOps.push_back(Elt: Splat); |
3876 | } |
3877 | Constant *Folded = ConstantFoldScalarCall( |
3878 | Name, IntrinsicID, Ty: SVTy->getElementType(), Operands: SplatOps, TLI, Call); |
3879 | if (!Folded) |
3880 | return nullptr; |
3881 | return ConstantVector::getSplat(EC: SVTy->getElementCount(), Elt: Folded); |
3882 | } |
3883 | |
3884 | static std::pair<Constant *, Constant *> |
3885 | ConstantFoldScalarFrexpCall(Constant *Op, Type *IntTy) { |
3886 | if (isa<PoisonValue>(Val: Op)) |
3887 | return {Op, PoisonValue::get(T: IntTy)}; |
3888 | |
3889 | auto *ConstFP = dyn_cast<ConstantFP>(Val: Op); |
3890 | if (!ConstFP) |
3891 | return {}; |
3892 | |
3893 | const APFloat &U = ConstFP->getValueAPF(); |
3894 | int FrexpExp; |
3895 | APFloat FrexpMant = frexp(X: U, Exp&: FrexpExp, RM: APFloat::rmNearestTiesToEven); |
3896 | Constant *Result0 = ConstantFP::get(Ty: ConstFP->getType(), V: FrexpMant); |
3897 | |
3898 | // The exponent is an "unspecified value" for inf/nan. We use zero to avoid |
3899 | // using undef. |
3900 | Constant *Result1 = FrexpMant.isFinite() |
3901 | ? ConstantInt::getSigned(Ty: IntTy, V: FrexpExp) |
3902 | : ConstantInt::getNullValue(Ty: IntTy); |
3903 | return {Result0, Result1}; |
3904 | } |
3905 | |
3906 | /// Handle intrinsics that return tuples, which may be tuples of vectors. |
3907 | static Constant * |
3908 | ConstantFoldStructCall(StringRef Name, Intrinsic::ID IntrinsicID, |
3909 | StructType *StTy, ArrayRef<Constant *> Operands, |
3910 | const DataLayout &DL, const TargetLibraryInfo *TLI, |
3911 | const CallBase *Call) { |
3912 | |
3913 | switch (IntrinsicID) { |
3914 | case Intrinsic::frexp: { |
3915 | Type *Ty0 = StTy->getContainedType(i: 0); |
3916 | Type *Ty1 = StTy->getContainedType(i: 1)->getScalarType(); |
3917 | |
3918 | if (auto *FVTy0 = dyn_cast<FixedVectorType>(Val: Ty0)) { |
3919 | SmallVector<Constant *, 4> Results0(FVTy0->getNumElements()); |
3920 | SmallVector<Constant *, 4> Results1(FVTy0->getNumElements()); |
3921 | |
3922 | for (unsigned I = 0, E = FVTy0->getNumElements(); I != E; ++I) { |
3923 | Constant *Lane = Operands[0]->getAggregateElement(Elt: I); |
3924 | std::tie(args&: Results0[I], args&: Results1[I]) = |
3925 | ConstantFoldScalarFrexpCall(Op: Lane, IntTy: Ty1); |
3926 | if (!Results0[I]) |
3927 | return nullptr; |
3928 | } |
3929 | |
3930 | return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: Results0), |
3931 | Vs: ConstantVector::get(V: Results1)); |
3932 | } |
3933 | |
3934 | auto [Result0, Result1] = ConstantFoldScalarFrexpCall(Op: Operands[0], IntTy: Ty1); |
3935 | if (!Result0) |
3936 | return nullptr; |
3937 | return ConstantStruct::get(T: StTy, Vs: Result0, Vs: Result1); |
3938 | } |
3939 | case Intrinsic::sincos: { |
3940 | Type *Ty = StTy->getContainedType(i: 0); |
3941 | Type *TyScalar = Ty->getScalarType(); |
3942 | |
3943 | auto ConstantFoldScalarSincosCall = |
3944 | [&](Constant *Op) -> std::pair<Constant *, Constant *> { |
3945 | Constant *SinResult = |
3946 | ConstantFoldScalarCall(Name, IntrinsicID: Intrinsic::sin, Ty: TyScalar, Operands: Op, TLI, Call); |
3947 | Constant *CosResult = |
3948 | ConstantFoldScalarCall(Name, IntrinsicID: Intrinsic::cos, Ty: TyScalar, Operands: Op, TLI, Call); |
3949 | return std::make_pair(x&: SinResult, y&: CosResult); |
3950 | }; |
3951 | |
3952 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) { |
3953 | SmallVector<Constant *> SinResults(FVTy->getNumElements()); |
3954 | SmallVector<Constant *> CosResults(FVTy->getNumElements()); |
3955 | |
3956 | for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) { |
3957 | Constant *Lane = Operands[0]->getAggregateElement(Elt: I); |
3958 | std::tie(args&: SinResults[I], args&: CosResults[I]) = |
3959 | ConstantFoldScalarSincosCall(Lane); |
3960 | if (!SinResults[I] || !CosResults[I]) |
3961 | return nullptr; |
3962 | } |
3963 | |
3964 | return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: SinResults), |
3965 | Vs: ConstantVector::get(V: CosResults)); |
3966 | } |
3967 | |
3968 | auto [SinResult, CosResult] = ConstantFoldScalarSincosCall(Operands[0]); |
3969 | if (!SinResult || !CosResult) |
3970 | return nullptr; |
3971 | return ConstantStruct::get(T: StTy, Vs: SinResult, Vs: CosResult); |
3972 | } |
3973 | case Intrinsic::vector_deinterleave2: { |
3974 | auto *Vec = Operands[0]; |
3975 | auto *VecTy = cast<VectorType>(Val: Vec->getType()); |
3976 | |
3977 | if (auto *EltC = Vec->getSplatValue()) { |
3978 | ElementCount HalfEC = VecTy->getElementCount().divideCoefficientBy(RHS: 2); |
3979 | auto *HalfVec = ConstantVector::getSplat(EC: HalfEC, Elt: EltC); |
3980 | return ConstantStruct::get(T: StTy, Vs: HalfVec, Vs: HalfVec); |
3981 | } |
3982 | |
3983 | if (!isa<FixedVectorType>(Val: Vec->getType())) |
3984 | return nullptr; |
3985 | |
3986 | unsigned NumElements = VecTy->getElementCount().getFixedValue() / 2; |
3987 | SmallVector<Constant *, 4> Res0(NumElements), Res1(NumElements); |
3988 | for (unsigned I = 0; I < NumElements; ++I) { |
3989 | Constant *Elt0 = Vec->getAggregateElement(Elt: 2 * I); |
3990 | Constant *Elt1 = Vec->getAggregateElement(Elt: 2 * I + 1); |
3991 | if (!Elt0 || !Elt1) |
3992 | return nullptr; |
3993 | Res0[I] = Elt0; |
3994 | Res1[I] = Elt1; |
3995 | } |
3996 | return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: Res0), |
3997 | Vs: ConstantVector::get(V: Res1)); |
3998 | } |
3999 | default: |
4000 | // TODO: Constant folding of vector intrinsics that fall through here does |
4001 | // not work (e.g. overflow intrinsics) |
4002 | return ConstantFoldScalarCall(Name, IntrinsicID, Ty: StTy, Operands, TLI, Call); |
4003 | } |
4004 | |
4005 | return nullptr; |
4006 | } |
4007 | |
4008 | } // end anonymous namespace |
4009 | |
4010 | Constant *llvm::ConstantFoldBinaryIntrinsic(Intrinsic::ID ID, Constant *LHS, |
4011 | Constant *RHS, Type *Ty, |
4012 | Instruction *FMFSource) { |
4013 | auto *Call = dyn_cast_if_present<CallBase>(Val: FMFSource); |
4014 | // Ensure we check flags like StrictFP that might prevent this from getting |
4015 | // folded before generating a result. |
4016 | if (Call && !canConstantFoldCallTo(Call, F: Call->getCalledFunction())) |
4017 | return nullptr; |
4018 | return ConstantFoldIntrinsicCall2(IntrinsicID: ID, Ty, Operands: {LHS, RHS}, Call); |
4019 | } |
4020 | |
4021 | Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F, |
4022 | ArrayRef<Constant *> Operands, |
4023 | const TargetLibraryInfo *TLI, |
4024 | bool AllowNonDeterministic) { |
4025 | if (Call->isNoBuiltin()) |
4026 | return nullptr; |
4027 | if (!F->hasName()) |
4028 | return nullptr; |
4029 | |
4030 | // If this is not an intrinsic and not recognized as a library call, bail out. |
4031 | Intrinsic::ID IID = F->getIntrinsicID(); |
4032 | if (IID == Intrinsic::not_intrinsic) { |
4033 | if (!TLI) |
4034 | return nullptr; |
4035 | LibFunc LibF; |
4036 | if (!TLI->getLibFunc(FDecl: *F, F&: LibF)) |
4037 | return nullptr; |
4038 | } |
4039 | |
4040 | // Conservatively assume that floating-point libcalls may be |
4041 | // non-deterministic. |
4042 | Type *Ty = F->getReturnType(); |
4043 | if (!AllowNonDeterministic && Ty->isFPOrFPVectorTy()) |
4044 | return nullptr; |
4045 | |
4046 | StringRef Name = F->getName(); |
4047 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) |
4048 | return ConstantFoldFixedVectorCall( |
4049 | Name, IntrinsicID: IID, FVTy, Operands, DL: F->getDataLayout(), TLI, Call); |
4050 | |
4051 | if (auto *SVTy = dyn_cast<ScalableVectorType>(Val: Ty)) |
4052 | return ConstantFoldScalableVectorCall( |
4053 | Name, IntrinsicID: IID, SVTy, Operands, DL: F->getDataLayout(), TLI, Call); |
4054 | |
4055 | if (auto *StTy = dyn_cast<StructType>(Val: Ty)) |
4056 | return ConstantFoldStructCall(Name, IntrinsicID: IID, StTy, Operands, |
4057 | DL: F->getDataLayout(), TLI, Call); |
4058 | |
4059 | // TODO: If this is a library function, we already discovered that above, |
4060 | // so we should pass the LibFunc, not the name (and it might be better |
4061 | // still to separate intrinsic handling from libcalls). |
4062 | return ConstantFoldScalarCall(Name, IntrinsicID: IID, Ty, Operands, TLI, Call); |
4063 | } |
4064 | |
4065 | bool llvm::isMathLibCallNoop(const CallBase *Call, |
4066 | const TargetLibraryInfo *TLI) { |
4067 | // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap |
4068 | // (and to some extent ConstantFoldScalarCall). |
4069 | if (Call->isNoBuiltin() || Call->isStrictFP()) |
4070 | return false; |
4071 | Function *F = Call->getCalledFunction(); |
4072 | if (!F) |
4073 | return false; |
4074 | |
4075 | LibFunc Func; |
4076 | if (!TLI || !TLI->getLibFunc(FDecl: *F, F&: Func)) |
4077 | return false; |
4078 | |
4079 | if (Call->arg_size() == 1) { |
4080 | if (ConstantFP *OpC = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0))) { |
4081 | const APFloat &Op = OpC->getValueAPF(); |
4082 | switch (Func) { |
4083 | case LibFunc_logl: |
4084 | case LibFunc_log: |
4085 | case LibFunc_logf: |
4086 | case LibFunc_log2l: |
4087 | case LibFunc_log2: |
4088 | case LibFunc_log2f: |
4089 | case LibFunc_log10l: |
4090 | case LibFunc_log10: |
4091 | case LibFunc_log10f: |
4092 | return Op.isNaN() || (!Op.isZero() && !Op.isNegative()); |
4093 | |
4094 | case LibFunc_ilogb: |
4095 | return !Op.isNaN() && !Op.isZero() && !Op.isInfinity(); |
4096 | |
4097 | case LibFunc_expl: |
4098 | case LibFunc_exp: |
4099 | case LibFunc_expf: |
4100 | // FIXME: These boundaries are slightly conservative. |
4101 | if (OpC->getType()->isDoubleTy()) |
4102 | return !(Op < APFloat(-745.0) || Op > APFloat(709.0)); |
4103 | if (OpC->getType()->isFloatTy()) |
4104 | return !(Op < APFloat(-103.0f) || Op > APFloat(88.0f)); |
4105 | break; |
4106 | |
4107 | case LibFunc_exp2l: |
4108 | case LibFunc_exp2: |
4109 | case LibFunc_exp2f: |
4110 | // FIXME: These boundaries are slightly conservative. |
4111 | if (OpC->getType()->isDoubleTy()) |
4112 | return !(Op < APFloat(-1074.0) || Op > APFloat(1023.0)); |
4113 | if (OpC->getType()->isFloatTy()) |
4114 | return !(Op < APFloat(-149.0f) || Op > APFloat(127.0f)); |
4115 | break; |
4116 | |
4117 | case LibFunc_sinl: |
4118 | case LibFunc_sin: |
4119 | case LibFunc_sinf: |
4120 | case LibFunc_cosl: |
4121 | case LibFunc_cos: |
4122 | case LibFunc_cosf: |
4123 | return !Op.isInfinity(); |
4124 | |
4125 | case LibFunc_tanl: |
4126 | case LibFunc_tan: |
4127 | case LibFunc_tanf: { |
4128 | // FIXME: Stop using the host math library. |
4129 | // FIXME: The computation isn't done in the right precision. |
4130 | Type *Ty = OpC->getType(); |
4131 | if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) |
4132 | return ConstantFoldFP(NativeFP: tan, V: OpC->getValueAPF(), Ty) != nullptr; |
4133 | break; |
4134 | } |
4135 | |
4136 | case LibFunc_atan: |
4137 | case LibFunc_atanf: |
4138 | case LibFunc_atanl: |
4139 | // Per POSIX, this MAY fail if Op is denormal. We choose not failing. |
4140 | return true; |
4141 | |
4142 | case LibFunc_asinl: |
4143 | case LibFunc_asin: |
4144 | case LibFunc_asinf: |
4145 | case LibFunc_acosl: |
4146 | case LibFunc_acos: |
4147 | case LibFunc_acosf: |
4148 | return !(Op < APFloat::getOne(Sem: Op.getSemantics(), Negative: true) || |
4149 | Op > APFloat::getOne(Sem: Op.getSemantics())); |
4150 | |
4151 | case LibFunc_sinh: |
4152 | case LibFunc_cosh: |
4153 | case LibFunc_sinhf: |
4154 | case LibFunc_coshf: |
4155 | case LibFunc_sinhl: |
4156 | case LibFunc_coshl: |
4157 | // FIXME: These boundaries are slightly conservative. |
4158 | if (OpC->getType()->isDoubleTy()) |
4159 | return !(Op < APFloat(-710.0) || Op > APFloat(710.0)); |
4160 | if (OpC->getType()->isFloatTy()) |
4161 | return !(Op < APFloat(-89.0f) || Op > APFloat(89.0f)); |
4162 | break; |
4163 | |
4164 | case LibFunc_sqrtl: |
4165 | case LibFunc_sqrt: |
4166 | case LibFunc_sqrtf: |
4167 | return Op.isNaN() || Op.isZero() || !Op.isNegative(); |
4168 | |
4169 | // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p, |
4170 | // maybe others? |
4171 | default: |
4172 | break; |
4173 | } |
4174 | } |
4175 | } |
4176 | |
4177 | if (Call->arg_size() == 2) { |
4178 | ConstantFP *Op0C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0)); |
4179 | ConstantFP *Op1C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 1)); |
4180 | if (Op0C && Op1C) { |
4181 | const APFloat &Op0 = Op0C->getValueAPF(); |
4182 | const APFloat &Op1 = Op1C->getValueAPF(); |
4183 | |
4184 | switch (Func) { |
4185 | case LibFunc_powl: |
4186 | case LibFunc_pow: |
4187 | case LibFunc_powf: { |
4188 | // FIXME: Stop using the host math library. |
4189 | // FIXME: The computation isn't done in the right precision. |
4190 | Type *Ty = Op0C->getType(); |
4191 | if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) { |
4192 | if (Ty == Op1C->getType()) |
4193 | return ConstantFoldBinaryFP(NativeFP: pow, V: Op0, W: Op1, Ty) != nullptr; |
4194 | } |
4195 | break; |
4196 | } |
4197 | |
4198 | case LibFunc_fmodl: |
4199 | case LibFunc_fmod: |
4200 | case LibFunc_fmodf: |
4201 | case LibFunc_remainderl: |
4202 | case LibFunc_remainder: |
4203 | case LibFunc_remainderf: |
4204 | return Op0.isNaN() || Op1.isNaN() || |
4205 | (!Op0.isInfinity() && !Op1.isZero()); |
4206 | |
4207 | case LibFunc_atan2: |
4208 | case LibFunc_atan2f: |
4209 | case LibFunc_atan2l: |
4210 | // Although IEEE-754 says atan2(+/-0.0, +/-0.0) are well-defined, and |
4211 | // GLIBC and MSVC do not appear to raise an error on those, we |
4212 | // cannot rely on that behavior. POSIX and C11 say that a domain error |
4213 | // may occur, so allow for that possibility. |
4214 | return !Op0.isZero() || !Op1.isZero(); |
4215 | |
4216 | default: |
4217 | break; |
4218 | } |
4219 | } |
4220 | } |
4221 | |
4222 | return false; |
4223 | } |
4224 | |
4225 | void TargetFolder::anchor() {} |
4226 | |