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/IntrinsicsWebAssembly.h" |
49 | #include "llvm/IR/IntrinsicsX86.h" |
50 | #include "llvm/IR/Operator.h" |
51 | #include "llvm/IR/Type.h" |
52 | #include "llvm/IR/Value.h" |
53 | #include "llvm/Support/Casting.h" |
54 | #include "llvm/Support/ErrorHandling.h" |
55 | #include "llvm/Support/KnownBits.h" |
56 | #include "llvm/Support/MathExtras.h" |
57 | #include <cassert> |
58 | #include <cerrno> |
59 | #include <cfenv> |
60 | #include <cmath> |
61 | #include <cstdint> |
62 | |
63 | using namespace llvm; |
64 | |
65 | namespace { |
66 | |
67 | //===----------------------------------------------------------------------===// |
68 | // Constant Folding internal helper functions |
69 | //===----------------------------------------------------------------------===// |
70 | |
71 | static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy, |
72 | Constant *C, Type *SrcEltTy, |
73 | unsigned NumSrcElts, |
74 | const DataLayout &DL) { |
75 | // Now that we know that the input value is a vector of integers, just shift |
76 | // and insert them into our result. |
77 | unsigned BitShift = DL.getTypeSizeInBits(Ty: SrcEltTy); |
78 | for (unsigned i = 0; i != NumSrcElts; ++i) { |
79 | Constant *Element; |
80 | if (DL.isLittleEndian()) |
81 | Element = C->getAggregateElement(Elt: NumSrcElts - i - 1); |
82 | else |
83 | Element = C->getAggregateElement(Elt: i); |
84 | |
85 | if (Element && isa<UndefValue>(Val: Element)) { |
86 | Result <<= BitShift; |
87 | continue; |
88 | } |
89 | |
90 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Val: Element); |
91 | if (!ElementCI) |
92 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
93 | |
94 | Result <<= BitShift; |
95 | Result |= ElementCI->getValue().zext(width: Result.getBitWidth()); |
96 | } |
97 | |
98 | return nullptr; |
99 | } |
100 | |
101 | /// Constant fold bitcast, symbolically evaluating it with DataLayout. |
102 | /// This always returns a non-null constant, but it may be a |
103 | /// ConstantExpr if unfoldable. |
104 | Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) { |
105 | assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) && |
106 | "Invalid constantexpr bitcast!" ); |
107 | |
108 | // Catch the obvious splat cases. |
109 | if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL)) |
110 | return Res; |
111 | |
112 | if (auto *VTy = dyn_cast<VectorType>(Val: C->getType())) { |
113 | // Handle a vector->scalar integer/fp cast. |
114 | if (isa<IntegerType>(Val: DestTy) || DestTy->isFloatingPointTy()) { |
115 | unsigned NumSrcElts = cast<FixedVectorType>(Val: VTy)->getNumElements(); |
116 | Type *SrcEltTy = VTy->getElementType(); |
117 | |
118 | // If the vector is a vector of floating point, convert it to vector of int |
119 | // to simplify things. |
120 | if (SrcEltTy->isFloatingPointTy()) { |
121 | unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); |
122 | auto *SrcIVTy = FixedVectorType::get( |
123 | ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElts); |
124 | // Ask IR to do the conversion now that #elts line up. |
125 | C = ConstantExpr::getBitCast(C, Ty: SrcIVTy); |
126 | } |
127 | |
128 | APInt Result(DL.getTypeSizeInBits(Ty: DestTy), 0); |
129 | if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C, |
130 | SrcEltTy, NumSrcElts, DL)) |
131 | return CE; |
132 | |
133 | if (isa<IntegerType>(Val: DestTy)) |
134 | return ConstantInt::get(Ty: DestTy, V: Result); |
135 | |
136 | APFloat FP(DestTy->getFltSemantics(), Result); |
137 | return ConstantFP::get(Context&: DestTy->getContext(), V: FP); |
138 | } |
139 | } |
140 | |
141 | // The code below only handles casts to vectors currently. |
142 | auto *DestVTy = dyn_cast<VectorType>(Val: DestTy); |
143 | if (!DestVTy) |
144 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
145 | |
146 | // If this is a scalar -> vector cast, convert the input into a <1 x scalar> |
147 | // vector so the code below can handle it uniformly. |
148 | if (isa<ConstantFP>(Val: C) || isa<ConstantInt>(Val: C)) { |
149 | Constant *Ops = C; // don't take the address of C! |
150 | return FoldBitCast(C: ConstantVector::get(V: Ops), DestTy, DL); |
151 | } |
152 | |
153 | // If this is a bitcast from constant vector -> vector, fold it. |
154 | if (!isa<ConstantDataVector>(Val: C) && !isa<ConstantVector>(Val: C)) |
155 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
156 | |
157 | // If the element types match, IR can fold it. |
158 | unsigned NumDstElt = cast<FixedVectorType>(Val: DestVTy)->getNumElements(); |
159 | unsigned NumSrcElt = cast<FixedVectorType>(Val: C->getType())->getNumElements(); |
160 | if (NumDstElt == NumSrcElt) |
161 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
162 | |
163 | Type *SrcEltTy = cast<VectorType>(Val: C->getType())->getElementType(); |
164 | Type *DstEltTy = DestVTy->getElementType(); |
165 | |
166 | // Otherwise, we're changing the number of elements in a vector, which |
167 | // requires endianness information to do the right thing. For example, |
168 | // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) |
169 | // folds to (little endian): |
170 | // <4 x i32> <i32 0, i32 0, i32 1, i32 0> |
171 | // and to (big endian): |
172 | // <4 x i32> <i32 0, i32 0, i32 0, i32 1> |
173 | |
174 | // First thing is first. We only want to think about integer here, so if |
175 | // we have something in FP form, recast it as integer. |
176 | if (DstEltTy->isFloatingPointTy()) { |
177 | // Fold to an vector of integers with same size as our FP type. |
178 | unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits(); |
179 | auto *DestIVTy = FixedVectorType::get( |
180 | ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumDstElt); |
181 | // Recursively handle this integer conversion, if possible. |
182 | C = FoldBitCast(C, DestTy: DestIVTy, DL); |
183 | |
184 | // Finally, IR can handle this now that #elts line up. |
185 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
186 | } |
187 | |
188 | // Okay, we know the destination is integer, if the input is FP, convert |
189 | // it to integer first. |
190 | if (SrcEltTy->isFloatingPointTy()) { |
191 | unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits(); |
192 | auto *SrcIVTy = FixedVectorType::get( |
193 | ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElt); |
194 | // Ask IR to do the conversion now that #elts line up. |
195 | C = ConstantExpr::getBitCast(C, Ty: SrcIVTy); |
196 | // If IR wasn't able to fold it, bail out. |
197 | if (!isa<ConstantVector>(Val: C) && // FIXME: Remove ConstantVector. |
198 | !isa<ConstantDataVector>(Val: C)) |
199 | return C; |
200 | } |
201 | |
202 | // Now we know that the input and output vectors are both integer vectors |
203 | // of the same size, and that their #elements is not the same. Do the |
204 | // conversion here, which depends on whether the input or output has |
205 | // more elements. |
206 | bool isLittleEndian = DL.isLittleEndian(); |
207 | |
208 | SmallVector<Constant*, 32> Result; |
209 | if (NumDstElt < NumSrcElt) { |
210 | // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>) |
211 | Constant *Zero = Constant::getNullValue(Ty: DstEltTy); |
212 | unsigned Ratio = NumSrcElt/NumDstElt; |
213 | unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits(); |
214 | unsigned SrcElt = 0; |
215 | for (unsigned i = 0; i != NumDstElt; ++i) { |
216 | // Build each element of the result. |
217 | Constant *Elt = Zero; |
218 | unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1); |
219 | for (unsigned j = 0; j != Ratio; ++j) { |
220 | Constant *Src = C->getAggregateElement(Elt: SrcElt++); |
221 | if (Src && isa<UndefValue>(Val: Src)) |
222 | Src = Constant::getNullValue( |
223 | Ty: cast<VectorType>(Val: C->getType())->getElementType()); |
224 | else |
225 | Src = dyn_cast_or_null<ConstantInt>(Val: Src); |
226 | if (!Src) // Reject constantexpr elements. |
227 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
228 | |
229 | // Zero extend the element to the right size. |
230 | Src = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: Src, DestTy: Elt->getType(), |
231 | DL); |
232 | assert(Src && "Constant folding cannot fail on plain integers" ); |
233 | |
234 | // Shift it to the right place, depending on endianness. |
235 | Src = ConstantFoldBinaryOpOperands( |
236 | Opcode: Instruction::Shl, LHS: Src, RHS: ConstantInt::get(Ty: Src->getType(), V: ShiftAmt), |
237 | DL); |
238 | assert(Src && "Constant folding cannot fail on plain integers" ); |
239 | |
240 | ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize; |
241 | |
242 | // Mix it in. |
243 | Elt = ConstantFoldBinaryOpOperands(Opcode: Instruction::Or, LHS: Elt, RHS: Src, DL); |
244 | assert(Elt && "Constant folding cannot fail on plain integers" ); |
245 | } |
246 | Result.push_back(Elt); |
247 | } |
248 | return ConstantVector::get(V: Result); |
249 | } |
250 | |
251 | // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>) |
252 | unsigned Ratio = NumDstElt/NumSrcElt; |
253 | unsigned DstBitSize = DL.getTypeSizeInBits(Ty: DstEltTy); |
254 | |
255 | // Loop over each source value, expanding into multiple results. |
256 | for (unsigned i = 0; i != NumSrcElt; ++i) { |
257 | auto *Element = C->getAggregateElement(Elt: i); |
258 | |
259 | if (!Element) // Reject constantexpr elements. |
260 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
261 | |
262 | if (isa<UndefValue>(Val: Element)) { |
263 | // Correctly Propagate undef values. |
264 | Result.append(NumInputs: Ratio, Elt: UndefValue::get(T: DstEltTy)); |
265 | continue; |
266 | } |
267 | |
268 | auto *Src = dyn_cast<ConstantInt>(Val: Element); |
269 | if (!Src) |
270 | return ConstantExpr::getBitCast(C, Ty: DestTy); |
271 | |
272 | unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1); |
273 | for (unsigned j = 0; j != Ratio; ++j) { |
274 | // Shift the piece of the value into the right place, depending on |
275 | // endianness. |
276 | APInt Elt = Src->getValue().lshr(shiftAmt: ShiftAmt); |
277 | ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize; |
278 | |
279 | // Truncate and remember this piece. |
280 | Result.push_back(Elt: ConstantInt::get(Ty: DstEltTy, V: Elt.trunc(width: DstBitSize))); |
281 | } |
282 | } |
283 | |
284 | return ConstantVector::get(V: Result); |
285 | } |
286 | |
287 | } // end anonymous namespace |
288 | |
289 | /// If this constant is a constant offset from a global, return the global and |
290 | /// the constant. Because of constantexprs, this function is recursive. |
291 | bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV, |
292 | APInt &Offset, const DataLayout &DL, |
293 | DSOLocalEquivalent **DSOEquiv) { |
294 | if (DSOEquiv) |
295 | *DSOEquiv = nullptr; |
296 | |
297 | // Trivial case, constant is the global. |
298 | if ((GV = dyn_cast<GlobalValue>(Val: C))) { |
299 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType()); |
300 | Offset = APInt(BitWidth, 0); |
301 | return true; |
302 | } |
303 | |
304 | if (auto *FoundDSOEquiv = dyn_cast<DSOLocalEquivalent>(Val: C)) { |
305 | if (DSOEquiv) |
306 | *DSOEquiv = FoundDSOEquiv; |
307 | GV = FoundDSOEquiv->getGlobalValue(); |
308 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType()); |
309 | Offset = APInt(BitWidth, 0); |
310 | return true; |
311 | } |
312 | |
313 | // Otherwise, if this isn't a constant expr, bail out. |
314 | auto *CE = dyn_cast<ConstantExpr>(Val: C); |
315 | if (!CE) return false; |
316 | |
317 | // Look through ptr->int and ptr->ptr casts. |
318 | if (CE->getOpcode() == Instruction::PtrToInt || |
319 | CE->getOpcode() == Instruction::BitCast) |
320 | return IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset, DL, |
321 | DSOEquiv); |
322 | |
323 | // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5) |
324 | auto *GEP = dyn_cast<GEPOperator>(Val: CE); |
325 | if (!GEP) |
326 | return false; |
327 | |
328 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType()); |
329 | APInt TmpOffset(BitWidth, 0); |
330 | |
331 | // If the base isn't a global+constant, we aren't either. |
332 | if (!IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset&: TmpOffset, DL, |
333 | DSOEquiv)) |
334 | return false; |
335 | |
336 | // Otherwise, add any offset that our operands provide. |
337 | if (!GEP->accumulateConstantOffset(DL, Offset&: TmpOffset)) |
338 | return false; |
339 | |
340 | Offset = TmpOffset; |
341 | return true; |
342 | } |
343 | |
344 | Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy, |
345 | const DataLayout &DL) { |
346 | do { |
347 | Type *SrcTy = C->getType(); |
348 | if (SrcTy == DestTy) |
349 | return C; |
350 | |
351 | TypeSize DestSize = DL.getTypeSizeInBits(Ty: DestTy); |
352 | TypeSize SrcSize = DL.getTypeSizeInBits(Ty: SrcTy); |
353 | if (!TypeSize::isKnownGE(LHS: SrcSize, RHS: DestSize)) |
354 | return nullptr; |
355 | |
356 | // Catch the obvious splat cases (since all-zeros can coerce non-integral |
357 | // pointers legally). |
358 | if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL)) |
359 | return Res; |
360 | |
361 | // If the type sizes are the same and a cast is legal, just directly |
362 | // cast the constant. |
363 | // But be careful not to coerce non-integral pointers illegally. |
364 | if (SrcSize == DestSize && |
365 | DL.isNonIntegralPointerType(Ty: SrcTy->getScalarType()) == |
366 | DL.isNonIntegralPointerType(Ty: DestTy->getScalarType())) { |
367 | Instruction::CastOps Cast = Instruction::BitCast; |
368 | // If we are going from a pointer to int or vice versa, we spell the cast |
369 | // differently. |
370 | if (SrcTy->isIntegerTy() && DestTy->isPointerTy()) |
371 | Cast = Instruction::IntToPtr; |
372 | else if (SrcTy->isPointerTy() && DestTy->isIntegerTy()) |
373 | Cast = Instruction::PtrToInt; |
374 | |
375 | if (CastInst::castIsValid(op: Cast, S: C, DstTy: DestTy)) |
376 | return ConstantFoldCastOperand(Opcode: Cast, C, DestTy, DL); |
377 | } |
378 | |
379 | // If this isn't an aggregate type, there is nothing we can do to drill down |
380 | // and find a bitcastable constant. |
381 | if (!SrcTy->isAggregateType() && !SrcTy->isVectorTy()) |
382 | return nullptr; |
383 | |
384 | // We're simulating a load through a pointer that was bitcast to point to |
385 | // a different type, so we can try to walk down through the initial |
386 | // elements of an aggregate to see if some part of the aggregate is |
387 | // castable to implement the "load" semantic model. |
388 | if (SrcTy->isStructTy()) { |
389 | // Struct types might have leading zero-length elements like [0 x i32], |
390 | // which are certainly not what we are looking for, so skip them. |
391 | unsigned Elem = 0; |
392 | Constant *ElemC; |
393 | do { |
394 | ElemC = C->getAggregateElement(Elt: Elem++); |
395 | } while (ElemC && DL.getTypeSizeInBits(Ty: ElemC->getType()).isZero()); |
396 | C = ElemC; |
397 | } else { |
398 | // For non-byte-sized vector elements, the first element is not |
399 | // necessarily located at the vector base address. |
400 | if (auto *VT = dyn_cast<VectorType>(Val: SrcTy)) |
401 | if (!DL.typeSizeEqualsStoreSize(Ty: VT->getElementType())) |
402 | return nullptr; |
403 | |
404 | C = C->getAggregateElement(Elt: 0u); |
405 | } |
406 | } while (C); |
407 | |
408 | return nullptr; |
409 | } |
410 | |
411 | namespace { |
412 | |
413 | /// Recursive helper to read bits out of global. C is the constant being copied |
414 | /// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy |
415 | /// results into and BytesLeft is the number of bytes left in |
416 | /// the CurPtr buffer. DL is the DataLayout. |
417 | bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr, |
418 | unsigned BytesLeft, const DataLayout &DL) { |
419 | assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) && |
420 | "Out of range access" ); |
421 | |
422 | // If this element is zero or undefined, we can just return since *CurPtr is |
423 | // zero initialized. |
424 | if (isa<ConstantAggregateZero>(Val: C) || isa<UndefValue>(Val: C)) |
425 | return true; |
426 | |
427 | if (auto *CI = dyn_cast<ConstantInt>(Val: C)) { |
428 | if ((CI->getBitWidth() & 7) != 0) |
429 | return false; |
430 | const APInt &Val = CI->getValue(); |
431 | unsigned IntBytes = unsigned(CI->getBitWidth()/8); |
432 | |
433 | for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) { |
434 | unsigned n = ByteOffset; |
435 | if (!DL.isLittleEndian()) |
436 | n = IntBytes - n - 1; |
437 | CurPtr[i] = Val.extractBits(numBits: 8, bitPosition: n * 8).getZExtValue(); |
438 | ++ByteOffset; |
439 | } |
440 | return true; |
441 | } |
442 | |
443 | if (auto *CFP = dyn_cast<ConstantFP>(Val: C)) { |
444 | if (CFP->getType()->isDoubleTy()) { |
445 | C = FoldBitCast(C, DestTy: Type::getInt64Ty(C&: C->getContext()), DL); |
446 | return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); |
447 | } |
448 | if (CFP->getType()->isFloatTy()){ |
449 | C = FoldBitCast(C, DestTy: Type::getInt32Ty(C&: C->getContext()), DL); |
450 | return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); |
451 | } |
452 | if (CFP->getType()->isHalfTy()){ |
453 | C = FoldBitCast(C, DestTy: Type::getInt16Ty(C&: C->getContext()), DL); |
454 | return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL); |
455 | } |
456 | return false; |
457 | } |
458 | |
459 | if (auto *CS = dyn_cast<ConstantStruct>(Val: C)) { |
460 | const StructLayout *SL = DL.getStructLayout(Ty: CS->getType()); |
461 | unsigned Index = SL->getElementContainingOffset(FixedOffset: ByteOffset); |
462 | uint64_t CurEltOffset = SL->getElementOffset(Idx: Index); |
463 | ByteOffset -= CurEltOffset; |
464 | |
465 | while (true) { |
466 | // If the element access is to the element itself and not to tail padding, |
467 | // read the bytes from the element. |
468 | uint64_t EltSize = DL.getTypeAllocSize(Ty: CS->getOperand(i_nocapture: Index)->getType()); |
469 | |
470 | if (ByteOffset < EltSize && |
471 | !ReadDataFromGlobal(C: CS->getOperand(i_nocapture: Index), ByteOffset, CurPtr, |
472 | BytesLeft, DL)) |
473 | return false; |
474 | |
475 | ++Index; |
476 | |
477 | // Check to see if we read from the last struct element, if so we're done. |
478 | if (Index == CS->getType()->getNumElements()) |
479 | return true; |
480 | |
481 | // If we read all of the bytes we needed from this element we're done. |
482 | uint64_t NextEltOffset = SL->getElementOffset(Idx: Index); |
483 | |
484 | if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset) |
485 | return true; |
486 | |
487 | // Move to the next element of the struct. |
488 | CurPtr += NextEltOffset - CurEltOffset - ByteOffset; |
489 | BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset; |
490 | ByteOffset = 0; |
491 | CurEltOffset = NextEltOffset; |
492 | } |
493 | // not reached. |
494 | } |
495 | |
496 | if (isa<ConstantArray>(Val: C) || isa<ConstantVector>(Val: C) || |
497 | isa<ConstantDataSequential>(Val: C)) { |
498 | uint64_t NumElts, EltSize; |
499 | Type *EltTy; |
500 | if (auto *AT = dyn_cast<ArrayType>(Val: C->getType())) { |
501 | NumElts = AT->getNumElements(); |
502 | EltTy = AT->getElementType(); |
503 | EltSize = DL.getTypeAllocSize(Ty: EltTy); |
504 | } else { |
505 | NumElts = cast<FixedVectorType>(Val: C->getType())->getNumElements(); |
506 | EltTy = cast<FixedVectorType>(Val: C->getType())->getElementType(); |
507 | // TODO: For non-byte-sized vectors, current implementation assumes there is |
508 | // padding to the next byte boundary between elements. |
509 | if (!DL.typeSizeEqualsStoreSize(Ty: EltTy)) |
510 | return false; |
511 | |
512 | EltSize = DL.getTypeStoreSize(Ty: EltTy); |
513 | } |
514 | uint64_t Index = ByteOffset / EltSize; |
515 | uint64_t Offset = ByteOffset - Index * EltSize; |
516 | |
517 | for (; Index != NumElts; ++Index) { |
518 | if (!ReadDataFromGlobal(C: C->getAggregateElement(Elt: Index), ByteOffset: Offset, CurPtr, |
519 | BytesLeft, DL)) |
520 | return false; |
521 | |
522 | uint64_t BytesWritten = EltSize - Offset; |
523 | assert(BytesWritten <= EltSize && "Not indexing into this element?" ); |
524 | if (BytesWritten >= BytesLeft) |
525 | return true; |
526 | |
527 | Offset = 0; |
528 | BytesLeft -= BytesWritten; |
529 | CurPtr += BytesWritten; |
530 | } |
531 | return true; |
532 | } |
533 | |
534 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
535 | if (CE->getOpcode() == Instruction::IntToPtr && |
536 | CE->getOperand(i_nocapture: 0)->getType() == DL.getIntPtrType(CE->getType())) { |
537 | return ReadDataFromGlobal(C: CE->getOperand(i_nocapture: 0), ByteOffset, CurPtr, |
538 | BytesLeft, DL); |
539 | } |
540 | } |
541 | |
542 | // Otherwise, unknown initializer type. |
543 | return false; |
544 | } |
545 | |
546 | Constant *FoldReinterpretLoadFromConst(Constant *C, Type *LoadTy, |
547 | int64_t Offset, const DataLayout &DL) { |
548 | // Bail out early. Not expect to load from scalable global variable. |
549 | if (isa<ScalableVectorType>(Val: LoadTy)) |
550 | return nullptr; |
551 | |
552 | auto *IntType = dyn_cast<IntegerType>(Val: LoadTy); |
553 | |
554 | // If this isn't an integer load we can't fold it directly. |
555 | if (!IntType) { |
556 | // If this is a non-integer load, we can try folding it as an int load and |
557 | // then bitcast the result. This can be useful for union cases. Note |
558 | // that address spaces don't matter here since we're not going to result in |
559 | // an actual new load. |
560 | if (!LoadTy->isFloatingPointTy() && !LoadTy->isPointerTy() && |
561 | !LoadTy->isVectorTy()) |
562 | return nullptr; |
563 | |
564 | Type *MapTy = Type::getIntNTy(C&: C->getContext(), |
565 | N: DL.getTypeSizeInBits(Ty: LoadTy).getFixedValue()); |
566 | if (Constant *Res = FoldReinterpretLoadFromConst(C, LoadTy: MapTy, Offset, DL)) { |
567 | if (Res->isNullValue() && !LoadTy->isX86_MMXTy() && |
568 | !LoadTy->isX86_AMXTy()) |
569 | // Materializing a zero can be done trivially without a bitcast |
570 | return Constant::getNullValue(Ty: LoadTy); |
571 | Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy; |
572 | Res = FoldBitCast(C: Res, DestTy: CastTy, DL); |
573 | if (LoadTy->isPtrOrPtrVectorTy()) { |
574 | // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr |
575 | if (Res->isNullValue() && !LoadTy->isX86_MMXTy() && |
576 | !LoadTy->isX86_AMXTy()) |
577 | return Constant::getNullValue(Ty: LoadTy); |
578 | if (DL.isNonIntegralPointerType(Ty: LoadTy->getScalarType())) |
579 | // Be careful not to replace a load of an addrspace value with an inttoptr here |
580 | return nullptr; |
581 | Res = ConstantExpr::getIntToPtr(C: Res, Ty: LoadTy); |
582 | } |
583 | return Res; |
584 | } |
585 | return nullptr; |
586 | } |
587 | |
588 | unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8; |
589 | if (BytesLoaded > 32 || BytesLoaded == 0) |
590 | return nullptr; |
591 | |
592 | // If we're not accessing anything in this constant, the result is undefined. |
593 | if (Offset <= -1 * static_cast<int64_t>(BytesLoaded)) |
594 | return PoisonValue::get(T: IntType); |
595 | |
596 | // TODO: We should be able to support scalable types. |
597 | TypeSize InitializerSize = DL.getTypeAllocSize(Ty: C->getType()); |
598 | if (InitializerSize.isScalable()) |
599 | return nullptr; |
600 | |
601 | // If we're not accessing anything in this constant, the result is undefined. |
602 | if (Offset >= (int64_t)InitializerSize.getFixedValue()) |
603 | return PoisonValue::get(T: IntType); |
604 | |
605 | unsigned char RawBytes[32] = {0}; |
606 | unsigned char *CurPtr = RawBytes; |
607 | unsigned BytesLeft = BytesLoaded; |
608 | |
609 | // If we're loading off the beginning of the global, some bytes may be valid. |
610 | if (Offset < 0) { |
611 | CurPtr += -Offset; |
612 | BytesLeft += Offset; |
613 | Offset = 0; |
614 | } |
615 | |
616 | if (!ReadDataFromGlobal(C, ByteOffset: Offset, CurPtr, BytesLeft, DL)) |
617 | return nullptr; |
618 | |
619 | APInt ResultVal = APInt(IntType->getBitWidth(), 0); |
620 | if (DL.isLittleEndian()) { |
621 | ResultVal = RawBytes[BytesLoaded - 1]; |
622 | for (unsigned i = 1; i != BytesLoaded; ++i) { |
623 | ResultVal <<= 8; |
624 | ResultVal |= RawBytes[BytesLoaded - 1 - i]; |
625 | } |
626 | } else { |
627 | ResultVal = RawBytes[0]; |
628 | for (unsigned i = 1; i != BytesLoaded; ++i) { |
629 | ResultVal <<= 8; |
630 | ResultVal |= RawBytes[i]; |
631 | } |
632 | } |
633 | |
634 | return ConstantInt::get(Context&: IntType->getContext(), V: ResultVal); |
635 | } |
636 | |
637 | } // anonymous namespace |
638 | |
639 | // If GV is a constant with an initializer read its representation starting |
640 | // at Offset and return it as a constant array of unsigned char. Otherwise |
641 | // return null. |
642 | Constant *llvm::ReadByteArrayFromGlobal(const GlobalVariable *GV, |
643 | uint64_t Offset) { |
644 | if (!GV->isConstant() || !GV->hasDefinitiveInitializer()) |
645 | return nullptr; |
646 | |
647 | const DataLayout &DL = GV->getDataLayout(); |
648 | Constant *Init = const_cast<Constant *>(GV->getInitializer()); |
649 | TypeSize InitSize = DL.getTypeAllocSize(Ty: Init->getType()); |
650 | if (InitSize < Offset) |
651 | return nullptr; |
652 | |
653 | uint64_t NBytes = InitSize - Offset; |
654 | if (NBytes > UINT16_MAX) |
655 | // Bail for large initializers in excess of 64K to avoid allocating |
656 | // too much memory. |
657 | // Offset is assumed to be less than or equal than InitSize (this |
658 | // is enforced in ReadDataFromGlobal). |
659 | return nullptr; |
660 | |
661 | SmallVector<unsigned char, 256> RawBytes(static_cast<size_t>(NBytes)); |
662 | unsigned char *CurPtr = RawBytes.data(); |
663 | |
664 | if (!ReadDataFromGlobal(C: Init, ByteOffset: Offset, CurPtr, BytesLeft: NBytes, DL)) |
665 | return nullptr; |
666 | |
667 | return ConstantDataArray::get(Context&: GV->getContext(), Elts&: RawBytes); |
668 | } |
669 | |
670 | /// If this Offset points exactly to the start of an aggregate element, return |
671 | /// that element, otherwise return nullptr. |
672 | Constant *getConstantAtOffset(Constant *Base, APInt Offset, |
673 | const DataLayout &DL) { |
674 | if (Offset.isZero()) |
675 | return Base; |
676 | |
677 | if (!isa<ConstantAggregate>(Val: Base) && !isa<ConstantDataSequential>(Val: Base)) |
678 | return nullptr; |
679 | |
680 | Type *ElemTy = Base->getType(); |
681 | SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset); |
682 | if (!Offset.isZero() || !Indices[0].isZero()) |
683 | return nullptr; |
684 | |
685 | Constant *C = Base; |
686 | for (const APInt &Index : drop_begin(RangeOrContainer&: Indices)) { |
687 | if (Index.isNegative() || Index.getActiveBits() >= 32) |
688 | return nullptr; |
689 | |
690 | C = C->getAggregateElement(Elt: Index.getZExtValue()); |
691 | if (!C) |
692 | return nullptr; |
693 | } |
694 | |
695 | return C; |
696 | } |
697 | |
698 | Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty, |
699 | const APInt &Offset, |
700 | const DataLayout &DL) { |
701 | if (Constant *AtOffset = getConstantAtOffset(Base: C, Offset, DL)) |
702 | if (Constant *Result = ConstantFoldLoadThroughBitcast(C: AtOffset, DestTy: Ty, DL)) |
703 | return Result; |
704 | |
705 | // Explicitly check for out-of-bounds access, so we return poison even if the |
706 | // constant is a uniform value. |
707 | TypeSize Size = DL.getTypeAllocSize(Ty: C->getType()); |
708 | if (!Size.isScalable() && Offset.sge(RHS: Size.getFixedValue())) |
709 | return PoisonValue::get(T: Ty); |
710 | |
711 | // Try an offset-independent fold of a uniform value. |
712 | if (Constant *Result = ConstantFoldLoadFromUniformValue(C, Ty, DL)) |
713 | return Result; |
714 | |
715 | // Try hard to fold loads from bitcasted strange and non-type-safe things. |
716 | if (Offset.getSignificantBits() <= 64) |
717 | if (Constant *Result = |
718 | FoldReinterpretLoadFromConst(C, LoadTy: Ty, Offset: Offset.getSExtValue(), DL)) |
719 | return Result; |
720 | |
721 | return nullptr; |
722 | } |
723 | |
724 | Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty, |
725 | const DataLayout &DL) { |
726 | return ConstantFoldLoadFromConst(C, Ty, Offset: APInt(64, 0), DL); |
727 | } |
728 | |
729 | Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, |
730 | APInt Offset, |
731 | const DataLayout &DL) { |
732 | // We can only fold loads from constant globals with a definitive initializer. |
733 | // Check this upfront, to skip expensive offset calculations. |
734 | auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: C)); |
735 | if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) |
736 | return nullptr; |
737 | |
738 | C = cast<Constant>(Val: C->stripAndAccumulateConstantOffsets( |
739 | DL, Offset, /* AllowNonInbounds */ true)); |
740 | |
741 | if (C == GV) |
742 | if (Constant *Result = ConstantFoldLoadFromConst(C: GV->getInitializer(), Ty, |
743 | Offset, DL)) |
744 | return Result; |
745 | |
746 | // If this load comes from anywhere in a uniform constant global, the value |
747 | // is always the same, regardless of the loaded offset. |
748 | return ConstantFoldLoadFromUniformValue(C: GV->getInitializer(), Ty, DL); |
749 | } |
750 | |
751 | Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, |
752 | const DataLayout &DL) { |
753 | APInt Offset(DL.getIndexTypeSizeInBits(Ty: C->getType()), 0); |
754 | return ConstantFoldLoadFromConstPtr(C, Ty, Offset: std::move(Offset), DL); |
755 | } |
756 | |
757 | Constant *llvm::ConstantFoldLoadFromUniformValue(Constant *C, Type *Ty, |
758 | const DataLayout &DL) { |
759 | if (isa<PoisonValue>(Val: C)) |
760 | return PoisonValue::get(T: Ty); |
761 | if (isa<UndefValue>(Val: C)) |
762 | return UndefValue::get(T: Ty); |
763 | // If padding is needed when storing C to memory, then it isn't considered as |
764 | // uniform. |
765 | if (!DL.typeSizeEqualsStoreSize(Ty: C->getType())) |
766 | return nullptr; |
767 | if (C->isNullValue() && !Ty->isX86_MMXTy() && !Ty->isX86_AMXTy()) |
768 | return Constant::getNullValue(Ty); |
769 | if (C->isAllOnesValue() && |
770 | (Ty->isIntOrIntVectorTy() || Ty->isFPOrFPVectorTy())) |
771 | return Constant::getAllOnesValue(Ty); |
772 | return nullptr; |
773 | } |
774 | |
775 | namespace { |
776 | |
777 | /// One of Op0/Op1 is a constant expression. |
778 | /// Attempt to symbolically evaluate the result of a binary operator merging |
779 | /// these together. If target data info is available, it is provided as DL, |
780 | /// otherwise DL is null. |
781 | Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1, |
782 | const DataLayout &DL) { |
783 | // SROA |
784 | |
785 | // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl. |
786 | // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute |
787 | // bits. |
788 | |
789 | if (Opc == Instruction::And) { |
790 | KnownBits Known0 = computeKnownBits(V: Op0, DL); |
791 | KnownBits Known1 = computeKnownBits(V: Op1, DL); |
792 | if ((Known1.One | Known0.Zero).isAllOnes()) { |
793 | // All the bits of Op0 that the 'and' could be masking are already zero. |
794 | return Op0; |
795 | } |
796 | if ((Known0.One | Known1.Zero).isAllOnes()) { |
797 | // All the bits of Op1 that the 'and' could be masking are already zero. |
798 | return Op1; |
799 | } |
800 | |
801 | Known0 &= Known1; |
802 | if (Known0.isConstant()) |
803 | return ConstantInt::get(Ty: Op0->getType(), V: Known0.getConstant()); |
804 | } |
805 | |
806 | // If the constant expr is something like &A[123] - &A[4].f, fold this into a |
807 | // constant. This happens frequently when iterating over a global array. |
808 | if (Opc == Instruction::Sub) { |
809 | GlobalValue *GV1, *GV2; |
810 | APInt Offs1, Offs2; |
811 | |
812 | if (IsConstantOffsetFromGlobal(C: Op0, GV&: GV1, Offset&: Offs1, DL)) |
813 | if (IsConstantOffsetFromGlobal(C: Op1, GV&: GV2, Offset&: Offs2, DL) && GV1 == GV2) { |
814 | unsigned OpSize = DL.getTypeSizeInBits(Ty: Op0->getType()); |
815 | |
816 | // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow. |
817 | // PtrToInt may change the bitwidth so we have convert to the right size |
818 | // first. |
819 | return ConstantInt::get(Ty: Op0->getType(), V: Offs1.zextOrTrunc(width: OpSize) - |
820 | Offs2.zextOrTrunc(width: OpSize)); |
821 | } |
822 | } |
823 | |
824 | return nullptr; |
825 | } |
826 | |
827 | /// If array indices are not pointer-sized integers, explicitly cast them so |
828 | /// that they aren't implicitly casted by the getelementptr. |
829 | Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops, |
830 | Type *ResultTy, GEPNoWrapFlags NW, |
831 | std::optional<ConstantRange> InRange, |
832 | const DataLayout &DL, const TargetLibraryInfo *TLI) { |
833 | Type *IntIdxTy = DL.getIndexType(PtrTy: ResultTy); |
834 | Type *IntIdxScalarTy = IntIdxTy->getScalarType(); |
835 | |
836 | bool Any = false; |
837 | SmallVector<Constant*, 32> NewIdxs; |
838 | for (unsigned i = 1, e = Ops.size(); i != e; ++i) { |
839 | if ((i == 1 || |
840 | !isa<StructType>(Val: GetElementPtrInst::getIndexedType( |
841 | Ty: SrcElemTy, IdxList: Ops.slice(N: 1, M: i - 1)))) && |
842 | Ops[i]->getType()->getScalarType() != IntIdxScalarTy) { |
843 | Any = true; |
844 | Type *NewType = |
845 | Ops[i]->getType()->isVectorTy() ? IntIdxTy : IntIdxScalarTy; |
846 | Constant *NewIdx = ConstantFoldCastOperand( |
847 | Opcode: CastInst::getCastOpcode(Val: Ops[i], SrcIsSigned: true, Ty: NewType, DstIsSigned: true), C: Ops[i], DestTy: NewType, |
848 | DL); |
849 | if (!NewIdx) |
850 | return nullptr; |
851 | NewIdxs.push_back(Elt: NewIdx); |
852 | } else |
853 | NewIdxs.push_back(Elt: Ops[i]); |
854 | } |
855 | |
856 | if (!Any) |
857 | return nullptr; |
858 | |
859 | Constant *C = |
860 | ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: NewIdxs, NW, InRange); |
861 | return ConstantFoldConstant(C, DL, TLI); |
862 | } |
863 | |
864 | /// If we can symbolically evaluate the GEP constant expression, do so. |
865 | Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP, |
866 | ArrayRef<Constant *> Ops, |
867 | const DataLayout &DL, |
868 | const TargetLibraryInfo *TLI) { |
869 | Type *SrcElemTy = GEP->getSourceElementType(); |
870 | Type *ResTy = GEP->getType(); |
871 | if (!SrcElemTy->isSized() || isa<ScalableVectorType>(Val: SrcElemTy)) |
872 | return nullptr; |
873 | |
874 | if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResultTy: ResTy, NW: GEP->getNoWrapFlags(), |
875 | InRange: GEP->getInRange(), DL, TLI)) |
876 | return C; |
877 | |
878 | Constant *Ptr = Ops[0]; |
879 | if (!Ptr->getType()->isPointerTy()) |
880 | return nullptr; |
881 | |
882 | Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType()); |
883 | |
884 | for (unsigned i = 1, e = Ops.size(); i != e; ++i) |
885 | if (!isa<ConstantInt>(Val: Ops[i])) |
886 | return nullptr; |
887 | |
888 | unsigned BitWidth = DL.getTypeSizeInBits(Ty: IntIdxTy); |
889 | APInt Offset = APInt( |
890 | BitWidth, |
891 | DL.getIndexedOffsetInType( |
892 | ElemTy: SrcElemTy, Indices: ArrayRef((Value *const *)Ops.data() + 1, Ops.size() - 1))); |
893 | |
894 | std::optional<ConstantRange> InRange = GEP->getInRange(); |
895 | if (InRange) |
896 | InRange = InRange->sextOrTrunc(BitWidth); |
897 | |
898 | // If this is a GEP of a GEP, fold it all into a single GEP. |
899 | GEPNoWrapFlags NW = GEP->getNoWrapFlags(); |
900 | bool Overflow = false; |
901 | while (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr)) { |
902 | NW &= GEP->getNoWrapFlags(); |
903 | |
904 | SmallVector<Value *, 4> NestedOps(llvm::drop_begin(RangeOrContainer: GEP->operands())); |
905 | |
906 | // Do not try the incorporate the sub-GEP if some index is not a number. |
907 | bool AllConstantInt = true; |
908 | for (Value *NestedOp : NestedOps) |
909 | if (!isa<ConstantInt>(Val: NestedOp)) { |
910 | AllConstantInt = false; |
911 | break; |
912 | } |
913 | if (!AllConstantInt) |
914 | break; |
915 | |
916 | // TODO: Try to intersect two inrange attributes? |
917 | if (!InRange) { |
918 | InRange = GEP->getInRange(); |
919 | if (InRange) |
920 | // Adjust inrange by offset until now. |
921 | InRange = InRange->sextOrTrunc(BitWidth).subtract(CI: Offset); |
922 | } |
923 | |
924 | Ptr = cast<Constant>(Val: GEP->getOperand(i_nocapture: 0)); |
925 | SrcElemTy = GEP->getSourceElementType(); |
926 | Offset = Offset.sadd_ov( |
927 | RHS: APInt(BitWidth, DL.getIndexedOffsetInType(ElemTy: SrcElemTy, Indices: NestedOps)), |
928 | Overflow); |
929 | } |
930 | |
931 | // Preserving nusw (without inbounds) also requires that the offset |
932 | // additions did not overflow. |
933 | if (NW.hasNoUnsignedSignedWrap() && !NW.isInBounds() && Overflow) |
934 | NW = NW.withoutNoUnsignedSignedWrap(); |
935 | |
936 | // If the base value for this address is a literal integer value, fold the |
937 | // getelementptr to the resulting integer value casted to the pointer type. |
938 | APInt BasePtr(BitWidth, 0); |
939 | if (auto *CE = dyn_cast<ConstantExpr>(Val: Ptr)) { |
940 | if (CE->getOpcode() == Instruction::IntToPtr) { |
941 | if (auto *Base = dyn_cast<ConstantInt>(Val: CE->getOperand(i_nocapture: 0))) |
942 | BasePtr = Base->getValue().zextOrTrunc(width: BitWidth); |
943 | } |
944 | } |
945 | |
946 | auto *PTy = cast<PointerType>(Val: Ptr->getType()); |
947 | if ((Ptr->isNullValue() || BasePtr != 0) && |
948 | !DL.isNonIntegralPointerType(PT: PTy)) { |
949 | Constant *C = ConstantInt::get(Context&: Ptr->getContext(), V: Offset + BasePtr); |
950 | return ConstantExpr::getIntToPtr(C, Ty: ResTy); |
951 | } |
952 | |
953 | // Try to infer inbounds for GEPs of globals. |
954 | // TODO(gep_nowrap): Also infer nuw flag. |
955 | if (!NW.isInBounds() && Offset.isNonNegative()) { |
956 | bool CanBeNull, CanBeFreed; |
957 | uint64_t DerefBytes = |
958 | Ptr->getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed); |
959 | if (DerefBytes != 0 && !CanBeNull && Offset.sle(RHS: DerefBytes)) |
960 | NW |= GEPNoWrapFlags::inBounds(); |
961 | } |
962 | |
963 | // Otherwise canonicalize this to a single ptradd. |
964 | LLVMContext &Ctx = Ptr->getContext(); |
965 | return ConstantExpr::getGetElementPtr(Ty: Type::getInt8Ty(C&: Ctx), C: Ptr, |
966 | Idx: ConstantInt::get(Context&: Ctx, V: Offset), NW, |
967 | InRange); |
968 | } |
969 | |
970 | /// Attempt to constant fold an instruction with the |
971 | /// specified opcode and operands. If successful, the constant result is |
972 | /// returned, if not, null is returned. Note that this function can fail when |
973 | /// attempting to fold instructions like loads and stores, which have no |
974 | /// constant expression form. |
975 | Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode, |
976 | ArrayRef<Constant *> Ops, |
977 | const DataLayout &DL, |
978 | const TargetLibraryInfo *TLI, |
979 | bool AllowNonDeterministic) { |
980 | Type *DestTy = InstOrCE->getType(); |
981 | |
982 | if (Instruction::isUnaryOp(Opcode)) |
983 | return ConstantFoldUnaryOpOperand(Opcode, Op: Ops[0], DL); |
984 | |
985 | if (Instruction::isBinaryOp(Opcode)) { |
986 | switch (Opcode) { |
987 | default: |
988 | break; |
989 | case Instruction::FAdd: |
990 | case Instruction::FSub: |
991 | case Instruction::FMul: |
992 | case Instruction::FDiv: |
993 | case Instruction::FRem: |
994 | // Handle floating point instructions separately to account for denormals |
995 | // TODO: If a constant expression is being folded rather than an |
996 | // instruction, denormals will not be flushed/treated as zero |
997 | if (const auto *I = dyn_cast<Instruction>(Val: InstOrCE)) { |
998 | return ConstantFoldFPInstOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL, I, |
999 | AllowNonDeterministic); |
1000 | } |
1001 | } |
1002 | return ConstantFoldBinaryOpOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL); |
1003 | } |
1004 | |
1005 | if (Instruction::isCast(Opcode)) |
1006 | return ConstantFoldCastOperand(Opcode, C: Ops[0], DestTy, DL); |
1007 | |
1008 | if (auto *GEP = dyn_cast<GEPOperator>(Val: InstOrCE)) { |
1009 | Type *SrcElemTy = GEP->getSourceElementType(); |
1010 | if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy)) |
1011 | return nullptr; |
1012 | |
1013 | if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI)) |
1014 | return C; |
1015 | |
1016 | return ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: Ops.slice(N: 1), |
1017 | NW: GEP->getNoWrapFlags(), |
1018 | InRange: GEP->getInRange()); |
1019 | } |
1020 | |
1021 | if (auto *CE = dyn_cast<ConstantExpr>(Val: InstOrCE)) |
1022 | return CE->getWithOperands(Ops); |
1023 | |
1024 | switch (Opcode) { |
1025 | default: return nullptr; |
1026 | case Instruction::ICmp: |
1027 | case Instruction::FCmp: { |
1028 | auto *C = cast<CmpInst>(Val: InstOrCE); |
1029 | return ConstantFoldCompareInstOperands(Predicate: C->getPredicate(), LHS: Ops[0], RHS: Ops[1], |
1030 | DL, TLI, I: C); |
1031 | } |
1032 | case Instruction::Freeze: |
1033 | return isGuaranteedNotToBeUndefOrPoison(V: Ops[0]) ? Ops[0] : nullptr; |
1034 | case Instruction::Call: |
1035 | if (auto *F = dyn_cast<Function>(Val: Ops.back())) { |
1036 | const auto *Call = cast<CallBase>(Val: InstOrCE); |
1037 | if (canConstantFoldCallTo(Call, F)) |
1038 | return ConstantFoldCall(Call, F, Operands: Ops.slice(N: 0, M: Ops.size() - 1), TLI, |
1039 | AllowNonDeterministic); |
1040 | } |
1041 | return nullptr; |
1042 | case Instruction::Select: |
1043 | return ConstantFoldSelectInstruction(Cond: Ops[0], V1: Ops[1], V2: Ops[2]); |
1044 | case Instruction::ExtractElement: |
1045 | return ConstantExpr::getExtractElement(Vec: Ops[0], Idx: Ops[1]); |
1046 | case Instruction::ExtractValue: |
1047 | return ConstantFoldExtractValueInstruction( |
1048 | Agg: Ops[0], Idxs: cast<ExtractValueInst>(Val: InstOrCE)->getIndices()); |
1049 | case Instruction::InsertElement: |
1050 | return ConstantExpr::getInsertElement(Vec: Ops[0], Elt: Ops[1], Idx: Ops[2]); |
1051 | case Instruction::InsertValue: |
1052 | return ConstantFoldInsertValueInstruction( |
1053 | Agg: Ops[0], Val: Ops[1], Idxs: cast<InsertValueInst>(Val: InstOrCE)->getIndices()); |
1054 | case Instruction::ShuffleVector: |
1055 | return ConstantExpr::getShuffleVector( |
1056 | V1: Ops[0], V2: Ops[1], Mask: cast<ShuffleVectorInst>(Val: InstOrCE)->getShuffleMask()); |
1057 | case Instruction::Load: { |
1058 | const auto *LI = dyn_cast<LoadInst>(Val: InstOrCE); |
1059 | if (LI->isVolatile()) |
1060 | return nullptr; |
1061 | return ConstantFoldLoadFromConstPtr(C: Ops[0], Ty: LI->getType(), DL); |
1062 | } |
1063 | } |
1064 | } |
1065 | |
1066 | } // end anonymous namespace |
1067 | |
1068 | //===----------------------------------------------------------------------===// |
1069 | // Constant Folding public APIs |
1070 | //===----------------------------------------------------------------------===// |
1071 | |
1072 | namespace { |
1073 | |
1074 | Constant * |
1075 | ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL, |
1076 | const TargetLibraryInfo *TLI, |
1077 | SmallDenseMap<Constant *, Constant *> &FoldedOps) { |
1078 | if (!isa<ConstantVector>(Val: C) && !isa<ConstantExpr>(Val: C)) |
1079 | return const_cast<Constant *>(C); |
1080 | |
1081 | SmallVector<Constant *, 8> Ops; |
1082 | for (const Use &OldU : C->operands()) { |
1083 | Constant *OldC = cast<Constant>(Val: &OldU); |
1084 | Constant *NewC = OldC; |
1085 | // Recursively fold the ConstantExpr's operands. If we have already folded |
1086 | // a ConstantExpr, we don't have to process it again. |
1087 | if (isa<ConstantVector>(Val: OldC) || isa<ConstantExpr>(Val: OldC)) { |
1088 | auto It = FoldedOps.find(Val: OldC); |
1089 | if (It == FoldedOps.end()) { |
1090 | NewC = ConstantFoldConstantImpl(C: OldC, DL, TLI, FoldedOps); |
1091 | FoldedOps.insert(KV: {OldC, NewC}); |
1092 | } else { |
1093 | NewC = It->second; |
1094 | } |
1095 | } |
1096 | Ops.push_back(Elt: NewC); |
1097 | } |
1098 | |
1099 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
1100 | if (Constant *Res = ConstantFoldInstOperandsImpl( |
1101 | InstOrCE: CE, Opcode: CE->getOpcode(), Ops, DL, TLI, /*AllowNonDeterministic=*/true)) |
1102 | return Res; |
1103 | return const_cast<Constant *>(C); |
1104 | } |
1105 | |
1106 | assert(isa<ConstantVector>(C)); |
1107 | return ConstantVector::get(V: Ops); |
1108 | } |
1109 | |
1110 | } // end anonymous namespace |
1111 | |
1112 | Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL, |
1113 | const TargetLibraryInfo *TLI) { |
1114 | // Handle PHI nodes quickly here... |
1115 | if (auto *PN = dyn_cast<PHINode>(Val: I)) { |
1116 | Constant *CommonValue = nullptr; |
1117 | |
1118 | SmallDenseMap<Constant *, Constant *> FoldedOps; |
1119 | for (Value *Incoming : PN->incoming_values()) { |
1120 | // If the incoming value is undef then skip it. Note that while we could |
1121 | // skip the value if it is equal to the phi node itself we choose not to |
1122 | // because that would break the rule that constant folding only applies if |
1123 | // all operands are constants. |
1124 | if (isa<UndefValue>(Val: Incoming)) |
1125 | continue; |
1126 | // If the incoming value is not a constant, then give up. |
1127 | auto *C = dyn_cast<Constant>(Val: Incoming); |
1128 | if (!C) |
1129 | return nullptr; |
1130 | // Fold the PHI's operands. |
1131 | C = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps); |
1132 | // If the incoming value is a different constant to |
1133 | // the one we saw previously, then give up. |
1134 | if (CommonValue && C != CommonValue) |
1135 | return nullptr; |
1136 | CommonValue = C; |
1137 | } |
1138 | |
1139 | // If we reach here, all incoming values are the same constant or undef. |
1140 | return CommonValue ? CommonValue : UndefValue::get(T: PN->getType()); |
1141 | } |
1142 | |
1143 | // Scan the operand list, checking to see if they are all constants, if so, |
1144 | // hand off to ConstantFoldInstOperandsImpl. |
1145 | if (!all_of(Range: I->operands(), P: [](Use &U) { return isa<Constant>(Val: U); })) |
1146 | return nullptr; |
1147 | |
1148 | SmallDenseMap<Constant *, Constant *> FoldedOps; |
1149 | SmallVector<Constant *, 8> Ops; |
1150 | for (const Use &OpU : I->operands()) { |
1151 | auto *Op = cast<Constant>(Val: &OpU); |
1152 | // Fold the Instruction's operands. |
1153 | Op = ConstantFoldConstantImpl(C: Op, DL, TLI, FoldedOps); |
1154 | Ops.push_back(Elt: Op); |
1155 | } |
1156 | |
1157 | return ConstantFoldInstOperands(I, Ops, DL, TLI); |
1158 | } |
1159 | |
1160 | Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL, |
1161 | const TargetLibraryInfo *TLI) { |
1162 | SmallDenseMap<Constant *, Constant *> FoldedOps; |
1163 | return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps); |
1164 | } |
1165 | |
1166 | Constant *llvm::ConstantFoldInstOperands(Instruction *I, |
1167 | ArrayRef<Constant *> Ops, |
1168 | const DataLayout &DL, |
1169 | const TargetLibraryInfo *TLI, |
1170 | bool AllowNonDeterministic) { |
1171 | return ConstantFoldInstOperandsImpl(InstOrCE: I, Opcode: I->getOpcode(), Ops, DL, TLI, |
1172 | AllowNonDeterministic); |
1173 | } |
1174 | |
1175 | Constant *llvm::ConstantFoldCompareInstOperands( |
1176 | unsigned IntPredicate, Constant *Ops0, Constant *Ops1, const DataLayout &DL, |
1177 | const TargetLibraryInfo *TLI, const Instruction *I) { |
1178 | CmpInst::Predicate Predicate = (CmpInst::Predicate)IntPredicate; |
1179 | // fold: icmp (inttoptr x), null -> icmp x, 0 |
1180 | // fold: icmp null, (inttoptr x) -> icmp 0, x |
1181 | // fold: icmp (ptrtoint x), 0 -> icmp x, null |
1182 | // fold: icmp 0, (ptrtoint x) -> icmp null, x |
1183 | // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y |
1184 | // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y |
1185 | // |
1186 | // FIXME: The following comment is out of data and the DataLayout is here now. |
1187 | // ConstantExpr::getCompare cannot do this, because it doesn't have DL |
1188 | // around to know if bit truncation is happening. |
1189 | if (auto *CE0 = dyn_cast<ConstantExpr>(Val: Ops0)) { |
1190 | if (Ops1->isNullValue()) { |
1191 | if (CE0->getOpcode() == Instruction::IntToPtr) { |
1192 | Type *IntPtrTy = DL.getIntPtrType(CE0->getType()); |
1193 | // Convert the integer value to the right size to ensure we get the |
1194 | // proper extension or truncation. |
1195 | if (Constant *C = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy, |
1196 | /*IsSigned*/ false, DL)) { |
1197 | Constant *Null = Constant::getNullValue(Ty: C->getType()); |
1198 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI); |
1199 | } |
1200 | } |
1201 | |
1202 | // Only do this transformation if the int is intptrty in size, otherwise |
1203 | // there is a truncation or extension that we aren't modeling. |
1204 | if (CE0->getOpcode() == Instruction::PtrToInt) { |
1205 | Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(i_nocapture: 0)->getType()); |
1206 | if (CE0->getType() == IntPtrTy) { |
1207 | Constant *C = CE0->getOperand(i_nocapture: 0); |
1208 | Constant *Null = Constant::getNullValue(Ty: C->getType()); |
1209 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI); |
1210 | } |
1211 | } |
1212 | } |
1213 | |
1214 | if (auto *CE1 = dyn_cast<ConstantExpr>(Val: Ops1)) { |
1215 | if (CE0->getOpcode() == CE1->getOpcode()) { |
1216 | if (CE0->getOpcode() == Instruction::IntToPtr) { |
1217 | Type *IntPtrTy = DL.getIntPtrType(CE0->getType()); |
1218 | |
1219 | // Convert the integer value to the right size to ensure we get the |
1220 | // proper extension or truncation. |
1221 | Constant *C0 = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy, |
1222 | /*IsSigned*/ false, DL); |
1223 | Constant *C1 = ConstantFoldIntegerCast(C: CE1->getOperand(i_nocapture: 0), DestTy: IntPtrTy, |
1224 | /*IsSigned*/ false, DL); |
1225 | if (C0 && C1) |
1226 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C0, Ops1: C1, DL, TLI); |
1227 | } |
1228 | |
1229 | // Only do this transformation if the int is intptrty in size, otherwise |
1230 | // there is a truncation or extension that we aren't modeling. |
1231 | if (CE0->getOpcode() == Instruction::PtrToInt) { |
1232 | Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(i_nocapture: 0)->getType()); |
1233 | if (CE0->getType() == IntPtrTy && |
1234 | CE0->getOperand(i_nocapture: 0)->getType() == CE1->getOperand(i_nocapture: 0)->getType()) { |
1235 | return ConstantFoldCompareInstOperands( |
1236 | IntPredicate: Predicate, Ops0: CE0->getOperand(i_nocapture: 0), Ops1: CE1->getOperand(i_nocapture: 0), DL, TLI); |
1237 | } |
1238 | } |
1239 | } |
1240 | } |
1241 | |
1242 | // Convert pointer comparison (base+offset1) pred (base+offset2) into |
1243 | // offset1 pred offset2, for the case where the offset is inbounds. This |
1244 | // only works for equality and unsigned comparison, as inbounds permits |
1245 | // crossing the sign boundary. However, the offset comparison itself is |
1246 | // signed. |
1247 | if (Ops0->getType()->isPointerTy() && !ICmpInst::isSigned(predicate: Predicate)) { |
1248 | unsigned IndexWidth = DL.getIndexTypeSizeInBits(Ty: Ops0->getType()); |
1249 | APInt Offset0(IndexWidth, 0); |
1250 | Value *Stripped0 = |
1251 | Ops0->stripAndAccumulateInBoundsConstantOffsets(DL, Offset&: Offset0); |
1252 | APInt Offset1(IndexWidth, 0); |
1253 | Value *Stripped1 = |
1254 | Ops1->stripAndAccumulateInBoundsConstantOffsets(DL, Offset&: Offset1); |
1255 | if (Stripped0 == Stripped1) |
1256 | return ConstantInt::getBool( |
1257 | Context&: Ops0->getContext(), |
1258 | V: ICmpInst::compare(LHS: Offset0, RHS: Offset1, |
1259 | Pred: ICmpInst::getSignedPredicate(pred: Predicate))); |
1260 | } |
1261 | } else if (isa<ConstantExpr>(Val: Ops1)) { |
1262 | // If RHS is a constant expression, but the left side isn't, swap the |
1263 | // operands and try again. |
1264 | Predicate = ICmpInst::getSwappedPredicate(pred: Predicate); |
1265 | return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: Ops1, Ops1: Ops0, DL, TLI); |
1266 | } |
1267 | |
1268 | // Flush any denormal constant float input according to denormal handling |
1269 | // mode. |
1270 | Ops0 = FlushFPConstant(Operand: Ops0, I, /* IsOutput */ false); |
1271 | if (!Ops0) |
1272 | return nullptr; |
1273 | Ops1 = FlushFPConstant(Operand: Ops1, I, /* IsOutput */ false); |
1274 | if (!Ops1) |
1275 | return nullptr; |
1276 | |
1277 | return ConstantFoldCompareInstruction(Predicate, C1: Ops0, C2: Ops1); |
1278 | } |
1279 | |
1280 | Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op, |
1281 | const DataLayout &DL) { |
1282 | assert(Instruction::isUnaryOp(Opcode)); |
1283 | |
1284 | return ConstantFoldUnaryInstruction(Opcode, V: Op); |
1285 | } |
1286 | |
1287 | Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, |
1288 | Constant *RHS, |
1289 | const DataLayout &DL) { |
1290 | assert(Instruction::isBinaryOp(Opcode)); |
1291 | if (isa<ConstantExpr>(Val: LHS) || isa<ConstantExpr>(Val: RHS)) |
1292 | if (Constant *C = SymbolicallyEvaluateBinop(Opc: Opcode, Op0: LHS, Op1: RHS, DL)) |
1293 | return C; |
1294 | |
1295 | if (ConstantExpr::isDesirableBinOp(Opcode)) |
1296 | return ConstantExpr::get(Opcode, C1: LHS, C2: RHS); |
1297 | return ConstantFoldBinaryInstruction(Opcode, V1: LHS, V2: RHS); |
1298 | } |
1299 | |
1300 | Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *I, |
1301 | bool IsOutput) { |
1302 | if (!I || !I->getParent() || !I->getFunction()) |
1303 | return Operand; |
1304 | |
1305 | ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Operand); |
1306 | if (!CFP) |
1307 | return Operand; |
1308 | |
1309 | const APFloat &APF = CFP->getValueAPF(); |
1310 | // TODO: Should this canonicalize nans? |
1311 | if (!APF.isDenormal()) |
1312 | return Operand; |
1313 | |
1314 | Type *Ty = CFP->getType(); |
1315 | DenormalMode DenormMode = |
1316 | I->getFunction()->getDenormalMode(FPType: Ty->getFltSemantics()); |
1317 | DenormalMode::DenormalModeKind Mode = |
1318 | IsOutput ? DenormMode.Output : DenormMode.Input; |
1319 | switch (Mode) { |
1320 | default: |
1321 | llvm_unreachable("unknown denormal mode" ); |
1322 | case DenormalMode::Dynamic: |
1323 | return nullptr; |
1324 | case DenormalMode::IEEE: |
1325 | return Operand; |
1326 | case DenormalMode::PreserveSign: |
1327 | if (APF.isDenormal()) { |
1328 | return ConstantFP::get( |
1329 | Context&: Ty->getContext(), |
1330 | V: APFloat::getZero(Sem: Ty->getFltSemantics(), Negative: APF.isNegative())); |
1331 | } |
1332 | return Operand; |
1333 | case DenormalMode::PositiveZero: |
1334 | if (APF.isDenormal()) { |
1335 | return ConstantFP::get(Context&: Ty->getContext(), |
1336 | V: APFloat::getZero(Sem: Ty->getFltSemantics(), Negative: false)); |
1337 | } |
1338 | return Operand; |
1339 | } |
1340 | return Operand; |
1341 | } |
1342 | |
1343 | Constant *llvm::ConstantFoldFPInstOperands(unsigned Opcode, Constant *LHS, |
1344 | Constant *RHS, const DataLayout &DL, |
1345 | const Instruction *I, |
1346 | bool AllowNonDeterministic) { |
1347 | if (Instruction::isBinaryOp(Opcode)) { |
1348 | // Flush denormal inputs if needed. |
1349 | Constant *Op0 = FlushFPConstant(Operand: LHS, I, /* IsOutput */ false); |
1350 | if (!Op0) |
1351 | return nullptr; |
1352 | Constant *Op1 = FlushFPConstant(Operand: RHS, I, /* IsOutput */ false); |
1353 | if (!Op1) |
1354 | return nullptr; |
1355 | |
1356 | // If nsz or an algebraic FMF flag is set, the result of the FP operation |
1357 | // may change due to future optimization. Don't constant fold them if |
1358 | // non-deterministic results are not allowed. |
1359 | if (!AllowNonDeterministic) |
1360 | if (auto *FP = dyn_cast_or_null<FPMathOperator>(Val: I)) |
1361 | if (FP->hasNoSignedZeros() || FP->hasAllowReassoc() || |
1362 | FP->hasAllowContract() || FP->hasAllowReciprocal()) |
1363 | return nullptr; |
1364 | |
1365 | // Calculate constant result. |
1366 | Constant *C = ConstantFoldBinaryOpOperands(Opcode, LHS: Op0, RHS: Op1, DL); |
1367 | if (!C) |
1368 | return nullptr; |
1369 | |
1370 | // Flush denormal output if needed. |
1371 | C = FlushFPConstant(Operand: C, I, /* IsOutput */ true); |
1372 | if (!C) |
1373 | return nullptr; |
1374 | |
1375 | // The precise NaN value is non-deterministic. |
1376 | if (!AllowNonDeterministic && C->isNaN()) |
1377 | return nullptr; |
1378 | |
1379 | return C; |
1380 | } |
1381 | // If instruction lacks a parent/function and the denormal mode cannot be |
1382 | // determined, use the default (IEEE). |
1383 | return ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL); |
1384 | } |
1385 | |
1386 | Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C, |
1387 | Type *DestTy, const DataLayout &DL) { |
1388 | assert(Instruction::isCast(Opcode)); |
1389 | switch (Opcode) { |
1390 | default: |
1391 | llvm_unreachable("Missing case" ); |
1392 | case Instruction::PtrToInt: |
1393 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
1394 | Constant *FoldedValue = nullptr; |
1395 | // If the input is a inttoptr, eliminate the pair. This requires knowing |
1396 | // the width of a pointer, so it can't be done in ConstantExpr::getCast. |
1397 | if (CE->getOpcode() == Instruction::IntToPtr) { |
1398 | // zext/trunc the inttoptr to pointer size. |
1399 | FoldedValue = ConstantFoldIntegerCast(C: CE->getOperand(i_nocapture: 0), |
1400 | DestTy: DL.getIntPtrType(CE->getType()), |
1401 | /*IsSigned=*/false, DL); |
1402 | } else if (auto *GEP = dyn_cast<GEPOperator>(Val: CE)) { |
1403 | // If we have GEP, we can perform the following folds: |
1404 | // (ptrtoint (gep null, x)) -> x |
1405 | // (ptrtoint (gep (gep null, x), y) -> x + y, etc. |
1406 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType()); |
1407 | APInt BaseOffset(BitWidth, 0); |
1408 | auto *Base = cast<Constant>(Val: GEP->stripAndAccumulateConstantOffsets( |
1409 | DL, Offset&: BaseOffset, /*AllowNonInbounds=*/true)); |
1410 | if (Base->isNullValue()) { |
1411 | FoldedValue = ConstantInt::get(Context&: CE->getContext(), V: BaseOffset); |
1412 | } else { |
1413 | // ptrtoint (gep i8, Ptr, (sub 0, V)) -> sub (ptrtoint Ptr), V |
1414 | if (GEP->getNumIndices() == 1 && |
1415 | GEP->getSourceElementType()->isIntegerTy(Bitwidth: 8)) { |
1416 | auto *Ptr = cast<Constant>(Val: GEP->getPointerOperand()); |
1417 | auto *Sub = dyn_cast<ConstantExpr>(Val: GEP->getOperand(i_nocapture: 1)); |
1418 | Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType()); |
1419 | if (Sub && Sub->getType() == IntIdxTy && |
1420 | Sub->getOpcode() == Instruction::Sub && |
1421 | Sub->getOperand(i_nocapture: 0)->isNullValue()) |
1422 | FoldedValue = ConstantExpr::getSub( |
1423 | C1: ConstantExpr::getPtrToInt(C: Ptr, Ty: IntIdxTy), C2: Sub->getOperand(i_nocapture: 1)); |
1424 | } |
1425 | } |
1426 | } |
1427 | if (FoldedValue) { |
1428 | // Do a zext or trunc to get to the ptrtoint dest size. |
1429 | return ConstantFoldIntegerCast(C: FoldedValue, DestTy, /*IsSigned=*/false, |
1430 | DL); |
1431 | } |
1432 | } |
1433 | break; |
1434 | case Instruction::IntToPtr: |
1435 | // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if |
1436 | // the int size is >= the ptr size and the address spaces are the same. |
1437 | // This requires knowing the width of a pointer, so it can't be done in |
1438 | // ConstantExpr::getCast. |
1439 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
1440 | if (CE->getOpcode() == Instruction::PtrToInt) { |
1441 | Constant *SrcPtr = CE->getOperand(i_nocapture: 0); |
1442 | unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType()); |
1443 | unsigned MidIntSize = CE->getType()->getScalarSizeInBits(); |
1444 | |
1445 | if (MidIntSize >= SrcPtrSize) { |
1446 | unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace(); |
1447 | if (SrcAS == DestTy->getPointerAddressSpace()) |
1448 | return FoldBitCast(C: CE->getOperand(i_nocapture: 0), DestTy, DL); |
1449 | } |
1450 | } |
1451 | } |
1452 | break; |
1453 | case Instruction::Trunc: |
1454 | case Instruction::ZExt: |
1455 | case Instruction::SExt: |
1456 | case Instruction::FPTrunc: |
1457 | case Instruction::FPExt: |
1458 | case Instruction::UIToFP: |
1459 | case Instruction::SIToFP: |
1460 | case Instruction::FPToUI: |
1461 | case Instruction::FPToSI: |
1462 | case Instruction::AddrSpaceCast: |
1463 | break; |
1464 | case Instruction::BitCast: |
1465 | return FoldBitCast(C, DestTy, DL); |
1466 | } |
1467 | |
1468 | if (ConstantExpr::isDesirableCastOp(Opcode)) |
1469 | return ConstantExpr::getCast(ops: Opcode, C, Ty: DestTy); |
1470 | return ConstantFoldCastInstruction(opcode: Opcode, V: C, DestTy); |
1471 | } |
1472 | |
1473 | Constant *llvm::ConstantFoldIntegerCast(Constant *C, Type *DestTy, |
1474 | bool IsSigned, const DataLayout &DL) { |
1475 | Type *SrcTy = C->getType(); |
1476 | if (SrcTy == DestTy) |
1477 | return C; |
1478 | if (SrcTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits()) |
1479 | return ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy, DL); |
1480 | if (IsSigned) |
1481 | return ConstantFoldCastOperand(Opcode: Instruction::SExt, C, DestTy, DL); |
1482 | return ConstantFoldCastOperand(Opcode: Instruction::ZExt, C, DestTy, DL); |
1483 | } |
1484 | |
1485 | //===----------------------------------------------------------------------===// |
1486 | // Constant Folding for Calls |
1487 | // |
1488 | |
1489 | bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) { |
1490 | if (Call->isNoBuiltin()) |
1491 | return false; |
1492 | if (Call->getFunctionType() != F->getFunctionType()) |
1493 | return false; |
1494 | switch (F->getIntrinsicID()) { |
1495 | // Operations that do not operate floating-point numbers and do not depend on |
1496 | // FP environment can be folded even in strictfp functions. |
1497 | case Intrinsic::bswap: |
1498 | case Intrinsic::ctpop: |
1499 | case Intrinsic::ctlz: |
1500 | case Intrinsic::cttz: |
1501 | case Intrinsic::fshl: |
1502 | case Intrinsic::fshr: |
1503 | case Intrinsic::launder_invariant_group: |
1504 | case Intrinsic::strip_invariant_group: |
1505 | case Intrinsic::masked_load: |
1506 | case Intrinsic::get_active_lane_mask: |
1507 | case Intrinsic::abs: |
1508 | case Intrinsic::smax: |
1509 | case Intrinsic::smin: |
1510 | case Intrinsic::umax: |
1511 | case Intrinsic::umin: |
1512 | case Intrinsic::scmp: |
1513 | case Intrinsic::ucmp: |
1514 | case Intrinsic::sadd_with_overflow: |
1515 | case Intrinsic::uadd_with_overflow: |
1516 | case Intrinsic::ssub_with_overflow: |
1517 | case Intrinsic::usub_with_overflow: |
1518 | case Intrinsic::smul_with_overflow: |
1519 | case Intrinsic::umul_with_overflow: |
1520 | case Intrinsic::sadd_sat: |
1521 | case Intrinsic::uadd_sat: |
1522 | case Intrinsic::ssub_sat: |
1523 | case Intrinsic::usub_sat: |
1524 | case Intrinsic::smul_fix: |
1525 | case Intrinsic::smul_fix_sat: |
1526 | case Intrinsic::bitreverse: |
1527 | case Intrinsic::is_constant: |
1528 | case Intrinsic::vector_reduce_add: |
1529 | case Intrinsic::vector_reduce_mul: |
1530 | case Intrinsic::vector_reduce_and: |
1531 | case Intrinsic::vector_reduce_or: |
1532 | case Intrinsic::vector_reduce_xor: |
1533 | case Intrinsic::vector_reduce_smin: |
1534 | case Intrinsic::vector_reduce_smax: |
1535 | case Intrinsic::vector_reduce_umin: |
1536 | case Intrinsic::vector_reduce_umax: |
1537 | // Target intrinsics |
1538 | case Intrinsic::amdgcn_perm: |
1539 | case Intrinsic::amdgcn_wave_reduce_umin: |
1540 | case Intrinsic::amdgcn_wave_reduce_umax: |
1541 | case Intrinsic::amdgcn_s_wqm: |
1542 | case Intrinsic::amdgcn_s_quadmask: |
1543 | case Intrinsic::amdgcn_s_bitreplicate: |
1544 | case Intrinsic::arm_mve_vctp8: |
1545 | case Intrinsic::arm_mve_vctp16: |
1546 | case Intrinsic::arm_mve_vctp32: |
1547 | case Intrinsic::arm_mve_vctp64: |
1548 | case Intrinsic::aarch64_sve_convert_from_svbool: |
1549 | // WebAssembly float semantics are always known |
1550 | case Intrinsic::wasm_trunc_signed: |
1551 | case Intrinsic::wasm_trunc_unsigned: |
1552 | return true; |
1553 | |
1554 | // Floating point operations cannot be folded in strictfp functions in |
1555 | // general case. They can be folded if FP environment is known to compiler. |
1556 | case Intrinsic::minnum: |
1557 | case Intrinsic::maxnum: |
1558 | case Intrinsic::minimum: |
1559 | case Intrinsic::maximum: |
1560 | case Intrinsic::log: |
1561 | case Intrinsic::log2: |
1562 | case Intrinsic::log10: |
1563 | case Intrinsic::exp: |
1564 | case Intrinsic::exp2: |
1565 | case Intrinsic::exp10: |
1566 | case Intrinsic::sqrt: |
1567 | case Intrinsic::sin: |
1568 | case Intrinsic::cos: |
1569 | case Intrinsic::pow: |
1570 | case Intrinsic::powi: |
1571 | case Intrinsic::ldexp: |
1572 | case Intrinsic::fma: |
1573 | case Intrinsic::fmuladd: |
1574 | case Intrinsic::frexp: |
1575 | case Intrinsic::fptoui_sat: |
1576 | case Intrinsic::fptosi_sat: |
1577 | case Intrinsic::convert_from_fp16: |
1578 | case Intrinsic::convert_to_fp16: |
1579 | case Intrinsic::amdgcn_cos: |
1580 | case Intrinsic::amdgcn_cubeid: |
1581 | case Intrinsic::amdgcn_cubema: |
1582 | case Intrinsic::amdgcn_cubesc: |
1583 | case Intrinsic::amdgcn_cubetc: |
1584 | case Intrinsic::amdgcn_fmul_legacy: |
1585 | case Intrinsic::amdgcn_fma_legacy: |
1586 | case Intrinsic::amdgcn_fract: |
1587 | case Intrinsic::amdgcn_sin: |
1588 | // The intrinsics below depend on rounding mode in MXCSR. |
1589 | case Intrinsic::x86_sse_cvtss2si: |
1590 | case Intrinsic::x86_sse_cvtss2si64: |
1591 | case Intrinsic::x86_sse_cvttss2si: |
1592 | case Intrinsic::x86_sse_cvttss2si64: |
1593 | case Intrinsic::x86_sse2_cvtsd2si: |
1594 | case Intrinsic::x86_sse2_cvtsd2si64: |
1595 | case Intrinsic::x86_sse2_cvttsd2si: |
1596 | case Intrinsic::x86_sse2_cvttsd2si64: |
1597 | case Intrinsic::x86_avx512_vcvtss2si32: |
1598 | case Intrinsic::x86_avx512_vcvtss2si64: |
1599 | case Intrinsic::x86_avx512_cvttss2si: |
1600 | case Intrinsic::x86_avx512_cvttss2si64: |
1601 | case Intrinsic::x86_avx512_vcvtsd2si32: |
1602 | case Intrinsic::x86_avx512_vcvtsd2si64: |
1603 | case Intrinsic::x86_avx512_cvttsd2si: |
1604 | case Intrinsic::x86_avx512_cvttsd2si64: |
1605 | case Intrinsic::x86_avx512_vcvtss2usi32: |
1606 | case Intrinsic::x86_avx512_vcvtss2usi64: |
1607 | case Intrinsic::x86_avx512_cvttss2usi: |
1608 | case Intrinsic::x86_avx512_cvttss2usi64: |
1609 | case Intrinsic::x86_avx512_vcvtsd2usi32: |
1610 | case Intrinsic::x86_avx512_vcvtsd2usi64: |
1611 | case Intrinsic::x86_avx512_cvttsd2usi: |
1612 | case Intrinsic::x86_avx512_cvttsd2usi64: |
1613 | return !Call->isStrictFP(); |
1614 | |
1615 | // Sign operations are actually bitwise operations, they do not raise |
1616 | // exceptions even for SNANs. |
1617 | case Intrinsic::fabs: |
1618 | case Intrinsic::copysign: |
1619 | case Intrinsic::is_fpclass: |
1620 | // Non-constrained variants of rounding operations means default FP |
1621 | // environment, they can be folded in any case. |
1622 | case Intrinsic::ceil: |
1623 | case Intrinsic::floor: |
1624 | case Intrinsic::round: |
1625 | case Intrinsic::roundeven: |
1626 | case Intrinsic::trunc: |
1627 | case Intrinsic::nearbyint: |
1628 | case Intrinsic::rint: |
1629 | case Intrinsic::canonicalize: |
1630 | // Constrained intrinsics can be folded if FP environment is known |
1631 | // to compiler. |
1632 | case Intrinsic::experimental_constrained_fma: |
1633 | case Intrinsic::experimental_constrained_fmuladd: |
1634 | case Intrinsic::experimental_constrained_fadd: |
1635 | case Intrinsic::experimental_constrained_fsub: |
1636 | case Intrinsic::experimental_constrained_fmul: |
1637 | case Intrinsic::experimental_constrained_fdiv: |
1638 | case Intrinsic::experimental_constrained_frem: |
1639 | case Intrinsic::experimental_constrained_ceil: |
1640 | case Intrinsic::experimental_constrained_floor: |
1641 | case Intrinsic::experimental_constrained_round: |
1642 | case Intrinsic::experimental_constrained_roundeven: |
1643 | case Intrinsic::experimental_constrained_trunc: |
1644 | case Intrinsic::experimental_constrained_nearbyint: |
1645 | case Intrinsic::experimental_constrained_rint: |
1646 | case Intrinsic::experimental_constrained_fcmp: |
1647 | case Intrinsic::experimental_constrained_fcmps: |
1648 | return true; |
1649 | default: |
1650 | return false; |
1651 | case Intrinsic::not_intrinsic: break; |
1652 | } |
1653 | |
1654 | if (!F->hasName() || Call->isStrictFP()) |
1655 | return false; |
1656 | |
1657 | // In these cases, the check of the length is required. We don't want to |
1658 | // return true for a name like "cos\0blah" which strcmp would return equal to |
1659 | // "cos", but has length 8. |
1660 | StringRef Name = F->getName(); |
1661 | switch (Name[0]) { |
1662 | default: |
1663 | return false; |
1664 | case 'a': |
1665 | return Name == "acos" || Name == "acosf" || |
1666 | Name == "asin" || Name == "asinf" || |
1667 | Name == "atan" || Name == "atanf" || |
1668 | Name == "atan2" || Name == "atan2f" ; |
1669 | case 'c': |
1670 | return Name == "ceil" || Name == "ceilf" || |
1671 | Name == "cos" || Name == "cosf" || |
1672 | Name == "cosh" || Name == "coshf" ; |
1673 | case 'e': |
1674 | return Name == "exp" || Name == "expf" || |
1675 | Name == "exp2" || Name == "exp2f" ; |
1676 | case 'f': |
1677 | return Name == "fabs" || Name == "fabsf" || |
1678 | Name == "floor" || Name == "floorf" || |
1679 | Name == "fmod" || Name == "fmodf" ; |
1680 | case 'l': |
1681 | return Name == "log" || Name == "logf" || Name == "log2" || |
1682 | Name == "log2f" || Name == "log10" || Name == "log10f" || |
1683 | Name == "logl" ; |
1684 | case 'n': |
1685 | return Name == "nearbyint" || Name == "nearbyintf" ; |
1686 | case 'p': |
1687 | return Name == "pow" || Name == "powf" ; |
1688 | case 'r': |
1689 | return Name == "remainder" || Name == "remainderf" || |
1690 | Name == "rint" || Name == "rintf" || |
1691 | Name == "round" || Name == "roundf" ; |
1692 | case 's': |
1693 | return Name == "sin" || Name == "sinf" || |
1694 | Name == "sinh" || Name == "sinhf" || |
1695 | Name == "sqrt" || Name == "sqrtf" ; |
1696 | case 't': |
1697 | return Name == "tan" || Name == "tanf" || |
1698 | Name == "tanh" || Name == "tanhf" || |
1699 | Name == "trunc" || Name == "truncf" ; |
1700 | case '_': |
1701 | // Check for various function names that get used for the math functions |
1702 | // when the header files are preprocessed with the macro |
1703 | // __FINITE_MATH_ONLY__ enabled. |
1704 | // The '12' here is the length of the shortest name that can match. |
1705 | // We need to check the size before looking at Name[1] and Name[2] |
1706 | // so we may as well check a limit that will eliminate mismatches. |
1707 | if (Name.size() < 12 || Name[1] != '_') |
1708 | return false; |
1709 | switch (Name[2]) { |
1710 | default: |
1711 | return false; |
1712 | case 'a': |
1713 | return Name == "__acos_finite" || Name == "__acosf_finite" || |
1714 | Name == "__asin_finite" || Name == "__asinf_finite" || |
1715 | Name == "__atan2_finite" || Name == "__atan2f_finite" ; |
1716 | case 'c': |
1717 | return Name == "__cosh_finite" || Name == "__coshf_finite" ; |
1718 | case 'e': |
1719 | return Name == "__exp_finite" || Name == "__expf_finite" || |
1720 | Name == "__exp2_finite" || Name == "__exp2f_finite" ; |
1721 | case 'l': |
1722 | return Name == "__log_finite" || Name == "__logf_finite" || |
1723 | Name == "__log10_finite" || Name == "__log10f_finite" ; |
1724 | case 'p': |
1725 | return Name == "__pow_finite" || Name == "__powf_finite" ; |
1726 | case 's': |
1727 | return Name == "__sinh_finite" || Name == "__sinhf_finite" ; |
1728 | } |
1729 | } |
1730 | } |
1731 | |
1732 | namespace { |
1733 | |
1734 | Constant *GetConstantFoldFPValue(double V, Type *Ty) { |
1735 | if (Ty->isHalfTy() || Ty->isFloatTy()) { |
1736 | APFloat APF(V); |
1737 | bool unused; |
1738 | APF.convert(ToSemantics: Ty->getFltSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused); |
1739 | return ConstantFP::get(Context&: Ty->getContext(), V: APF); |
1740 | } |
1741 | if (Ty->isDoubleTy()) |
1742 | return ConstantFP::get(Context&: Ty->getContext(), V: APFloat(V)); |
1743 | llvm_unreachable("Can only constant fold half/float/double" ); |
1744 | } |
1745 | |
1746 | #if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128) |
1747 | Constant *GetConstantFoldFPValue128(float128 V, Type *Ty) { |
1748 | if (Ty->isFP128Ty()) |
1749 | return ConstantFP::get(Ty, V); |
1750 | llvm_unreachable("Can only constant fold fp128" ); |
1751 | } |
1752 | #endif |
1753 | |
1754 | /// Clear the floating-point exception state. |
1755 | inline void llvm_fenv_clearexcept() { |
1756 | #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT |
1757 | feclearexcept(FE_ALL_EXCEPT); |
1758 | #endif |
1759 | errno = 0; |
1760 | } |
1761 | |
1762 | /// Test if a floating-point exception was raised. |
1763 | inline bool llvm_fenv_testexcept() { |
1764 | int errno_val = errno; |
1765 | if (errno_val == ERANGE || errno_val == EDOM) |
1766 | return true; |
1767 | #if defined(HAVE_FENV_H) && HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT |
1768 | if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT)) |
1769 | return true; |
1770 | #endif |
1771 | return false; |
1772 | } |
1773 | |
1774 | Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V, |
1775 | Type *Ty) { |
1776 | llvm_fenv_clearexcept(); |
1777 | double Result = NativeFP(V.convertToDouble()); |
1778 | if (llvm_fenv_testexcept()) { |
1779 | llvm_fenv_clearexcept(); |
1780 | return nullptr; |
1781 | } |
1782 | |
1783 | return GetConstantFoldFPValue(V: Result, Ty); |
1784 | } |
1785 | |
1786 | #if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128) |
1787 | Constant *ConstantFoldFP128(long double (*NativeFP)(long double), |
1788 | const APFloat &V, Type *Ty) { |
1789 | llvm_fenv_clearexcept(); |
1790 | float128 Result = NativeFP(V.convertToQuad()); |
1791 | if (llvm_fenv_testexcept()) { |
1792 | llvm_fenv_clearexcept(); |
1793 | return nullptr; |
1794 | } |
1795 | |
1796 | return GetConstantFoldFPValue128(V: Result, Ty); |
1797 | } |
1798 | #endif |
1799 | |
1800 | Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double), |
1801 | const APFloat &V, const APFloat &W, Type *Ty) { |
1802 | llvm_fenv_clearexcept(); |
1803 | double Result = NativeFP(V.convertToDouble(), W.convertToDouble()); |
1804 | if (llvm_fenv_testexcept()) { |
1805 | llvm_fenv_clearexcept(); |
1806 | return nullptr; |
1807 | } |
1808 | |
1809 | return GetConstantFoldFPValue(V: Result, Ty); |
1810 | } |
1811 | |
1812 | Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) { |
1813 | FixedVectorType *VT = dyn_cast<FixedVectorType>(Val: Op->getType()); |
1814 | if (!VT) |
1815 | return nullptr; |
1816 | |
1817 | // This isn't strictly necessary, but handle the special/common case of zero: |
1818 | // all integer reductions of a zero input produce zero. |
1819 | if (isa<ConstantAggregateZero>(Val: Op)) |
1820 | return ConstantInt::get(Ty: VT->getElementType(), V: 0); |
1821 | |
1822 | // This is the same as the underlying binops - poison propagates. |
1823 | if (isa<PoisonValue>(Val: Op) || Op->containsPoisonElement()) |
1824 | return PoisonValue::get(T: VT->getElementType()); |
1825 | |
1826 | // TODO: Handle undef. |
1827 | if (!isa<ConstantVector>(Val: Op) && !isa<ConstantDataVector>(Val: Op)) |
1828 | return nullptr; |
1829 | |
1830 | auto *EltC = dyn_cast<ConstantInt>(Val: Op->getAggregateElement(Elt: 0U)); |
1831 | if (!EltC) |
1832 | return nullptr; |
1833 | |
1834 | APInt Acc = EltC->getValue(); |
1835 | for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) { |
1836 | if (!(EltC = dyn_cast<ConstantInt>(Val: Op->getAggregateElement(Elt: I)))) |
1837 | return nullptr; |
1838 | const APInt &X = EltC->getValue(); |
1839 | switch (IID) { |
1840 | case Intrinsic::vector_reduce_add: |
1841 | Acc = Acc + X; |
1842 | break; |
1843 | case Intrinsic::vector_reduce_mul: |
1844 | Acc = Acc * X; |
1845 | break; |
1846 | case Intrinsic::vector_reduce_and: |
1847 | Acc = Acc & X; |
1848 | break; |
1849 | case Intrinsic::vector_reduce_or: |
1850 | Acc = Acc | X; |
1851 | break; |
1852 | case Intrinsic::vector_reduce_xor: |
1853 | Acc = Acc ^ X; |
1854 | break; |
1855 | case Intrinsic::vector_reduce_smin: |
1856 | Acc = APIntOps::smin(A: Acc, B: X); |
1857 | break; |
1858 | case Intrinsic::vector_reduce_smax: |
1859 | Acc = APIntOps::smax(A: Acc, B: X); |
1860 | break; |
1861 | case Intrinsic::vector_reduce_umin: |
1862 | Acc = APIntOps::umin(A: Acc, B: X); |
1863 | break; |
1864 | case Intrinsic::vector_reduce_umax: |
1865 | Acc = APIntOps::umax(A: Acc, B: X); |
1866 | break; |
1867 | } |
1868 | } |
1869 | |
1870 | return ConstantInt::get(Context&: Op->getContext(), V: Acc); |
1871 | } |
1872 | |
1873 | /// Attempt to fold an SSE floating point to integer conversion of a constant |
1874 | /// floating point. If roundTowardZero is false, the default IEEE rounding is |
1875 | /// used (toward nearest, ties to even). This matches the behavior of the |
1876 | /// non-truncating SSE instructions in the default rounding mode. The desired |
1877 | /// integer type Ty is used to select how many bits are available for the |
1878 | /// result. Returns null if the conversion cannot be performed, otherwise |
1879 | /// returns the Constant value resulting from the conversion. |
1880 | Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero, |
1881 | Type *Ty, bool IsSigned) { |
1882 | // All of these conversion intrinsics form an integer of at most 64bits. |
1883 | unsigned ResultWidth = Ty->getIntegerBitWidth(); |
1884 | assert(ResultWidth <= 64 && |
1885 | "Can only constant fold conversions to 64 and 32 bit ints" ); |
1886 | |
1887 | uint64_t UIntVal; |
1888 | bool isExact = false; |
1889 | APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero |
1890 | : APFloat::rmNearestTiesToEven; |
1891 | APFloat::opStatus status = |
1892 | Val.convertToInteger(Input: MutableArrayRef(UIntVal), Width: ResultWidth, |
1893 | IsSigned, RM: mode, IsExact: &isExact); |
1894 | if (status != APFloat::opOK && |
1895 | (!roundTowardZero || status != APFloat::opInexact)) |
1896 | return nullptr; |
1897 | return ConstantInt::get(Ty, V: UIntVal, IsSigned); |
1898 | } |
1899 | |
1900 | double getValueAsDouble(ConstantFP *Op) { |
1901 | Type *Ty = Op->getType(); |
1902 | |
1903 | if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy()) |
1904 | return Op->getValueAPF().convertToDouble(); |
1905 | |
1906 | bool unused; |
1907 | APFloat APF = Op->getValueAPF(); |
1908 | APF.convert(ToSemantics: APFloat::IEEEdouble(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused); |
1909 | return APF.convertToDouble(); |
1910 | } |
1911 | |
1912 | static bool getConstIntOrUndef(Value *Op, const APInt *&C) { |
1913 | if (auto *CI = dyn_cast<ConstantInt>(Val: Op)) { |
1914 | C = &CI->getValue(); |
1915 | return true; |
1916 | } |
1917 | if (isa<UndefValue>(Val: Op)) { |
1918 | C = nullptr; |
1919 | return true; |
1920 | } |
1921 | return false; |
1922 | } |
1923 | |
1924 | /// Checks if the given intrinsic call, which evaluates to constant, is allowed |
1925 | /// to be folded. |
1926 | /// |
1927 | /// \param CI Constrained intrinsic call. |
1928 | /// \param St Exception flags raised during constant evaluation. |
1929 | static bool mayFoldConstrained(ConstrainedFPIntrinsic *CI, |
1930 | APFloat::opStatus St) { |
1931 | std::optional<RoundingMode> ORM = CI->getRoundingMode(); |
1932 | std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior(); |
1933 | |
1934 | // If the operation does not change exception status flags, it is safe |
1935 | // to fold. |
1936 | if (St == APFloat::opStatus::opOK) |
1937 | return true; |
1938 | |
1939 | // If evaluation raised FP exception, the result can depend on rounding |
1940 | // mode. If the latter is unknown, folding is not possible. |
1941 | if (ORM && *ORM == RoundingMode::Dynamic) |
1942 | return false; |
1943 | |
1944 | // If FP exceptions are ignored, fold the call, even if such exception is |
1945 | // raised. |
1946 | if (EB && *EB != fp::ExceptionBehavior::ebStrict) |
1947 | return true; |
1948 | |
1949 | // Leave the calculation for runtime so that exception flags be correctly set |
1950 | // in hardware. |
1951 | return false; |
1952 | } |
1953 | |
1954 | /// Returns the rounding mode that should be used for constant evaluation. |
1955 | static RoundingMode |
1956 | getEvaluationRoundingMode(const ConstrainedFPIntrinsic *CI) { |
1957 | std::optional<RoundingMode> ORM = CI->getRoundingMode(); |
1958 | if (!ORM || *ORM == RoundingMode::Dynamic) |
1959 | // Even if the rounding mode is unknown, try evaluating the operation. |
1960 | // If it does not raise inexact exception, rounding was not applied, |
1961 | // so the result is exact and does not depend on rounding mode. Whether |
1962 | // other FP exceptions are raised, it does not depend on rounding mode. |
1963 | return RoundingMode::NearestTiesToEven; |
1964 | return *ORM; |
1965 | } |
1966 | |
1967 | /// Try to constant fold llvm.canonicalize for the given caller and value. |
1968 | static Constant *constantFoldCanonicalize(const Type *Ty, const CallBase *CI, |
1969 | const APFloat &Src) { |
1970 | // Zero, positive and negative, is always OK to fold. |
1971 | if (Src.isZero()) { |
1972 | // Get a fresh 0, since ppc_fp128 does have non-canonical zeros. |
1973 | return ConstantFP::get( |
1974 | Context&: CI->getContext(), |
1975 | V: APFloat::getZero(Sem: Src.getSemantics(), Negative: Src.isNegative())); |
1976 | } |
1977 | |
1978 | if (!Ty->isIEEELikeFPTy()) |
1979 | return nullptr; |
1980 | |
1981 | // Zero is always canonical and the sign must be preserved. |
1982 | // |
1983 | // Denorms and nans may have special encodings, but it should be OK to fold a |
1984 | // totally average number. |
1985 | if (Src.isNormal() || Src.isInfinity()) |
1986 | return ConstantFP::get(Context&: CI->getContext(), V: Src); |
1987 | |
1988 | if (Src.isDenormal() && CI->getParent() && CI->getFunction()) { |
1989 | DenormalMode DenormMode = |
1990 | CI->getFunction()->getDenormalMode(FPType: Src.getSemantics()); |
1991 | |
1992 | if (DenormMode == DenormalMode::getIEEE()) |
1993 | return ConstantFP::get(Context&: CI->getContext(), V: Src); |
1994 | |
1995 | if (DenormMode.Input == DenormalMode::Dynamic) |
1996 | return nullptr; |
1997 | |
1998 | // If we know if either input or output is flushed, we can fold. |
1999 | if ((DenormMode.Input == DenormalMode::Dynamic && |
2000 | DenormMode.Output == DenormalMode::IEEE) || |
2001 | (DenormMode.Input == DenormalMode::IEEE && |
2002 | DenormMode.Output == DenormalMode::Dynamic)) |
2003 | return nullptr; |
2004 | |
2005 | bool IsPositive = |
2006 | (!Src.isNegative() || DenormMode.Input == DenormalMode::PositiveZero || |
2007 | (DenormMode.Output == DenormalMode::PositiveZero && |
2008 | DenormMode.Input == DenormalMode::IEEE)); |
2009 | |
2010 | return ConstantFP::get(Context&: CI->getContext(), |
2011 | V: APFloat::getZero(Sem: Src.getSemantics(), Negative: !IsPositive)); |
2012 | } |
2013 | |
2014 | return nullptr; |
2015 | } |
2016 | |
2017 | static Constant *ConstantFoldScalarCall1(StringRef Name, |
2018 | Intrinsic::ID IntrinsicID, |
2019 | Type *Ty, |
2020 | ArrayRef<Constant *> Operands, |
2021 | const TargetLibraryInfo *TLI, |
2022 | const CallBase *Call) { |
2023 | assert(Operands.size() == 1 && "Wrong number of operands." ); |
2024 | |
2025 | if (IntrinsicID == Intrinsic::is_constant) { |
2026 | // We know we have a "Constant" argument. But we want to only |
2027 | // return true for manifest constants, not those that depend on |
2028 | // constants with unknowable values, e.g. GlobalValue or BlockAddress. |
2029 | if (Operands[0]->isManifestConstant()) |
2030 | return ConstantInt::getTrue(Context&: Ty->getContext()); |
2031 | return nullptr; |
2032 | } |
2033 | |
2034 | if (isa<PoisonValue>(Val: Operands[0])) { |
2035 | // TODO: All of these operations should probably propagate poison. |
2036 | if (IntrinsicID == Intrinsic::canonicalize) |
2037 | return PoisonValue::get(T: Ty); |
2038 | } |
2039 | |
2040 | if (isa<UndefValue>(Val: Operands[0])) { |
2041 | // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN. |
2042 | // ctpop() is between 0 and bitwidth, pick 0 for undef. |
2043 | // fptoui.sat and fptosi.sat can always fold to zero (for a zero input). |
2044 | if (IntrinsicID == Intrinsic::cos || |
2045 | IntrinsicID == Intrinsic::ctpop || |
2046 | IntrinsicID == Intrinsic::fptoui_sat || |
2047 | IntrinsicID == Intrinsic::fptosi_sat || |
2048 | IntrinsicID == Intrinsic::canonicalize) |
2049 | return Constant::getNullValue(Ty); |
2050 | if (IntrinsicID == Intrinsic::bswap || |
2051 | IntrinsicID == Intrinsic::bitreverse || |
2052 | IntrinsicID == Intrinsic::launder_invariant_group || |
2053 | IntrinsicID == Intrinsic::strip_invariant_group) |
2054 | return Operands[0]; |
2055 | } |
2056 | |
2057 | if (isa<ConstantPointerNull>(Val: Operands[0])) { |
2058 | // launder(null) == null == strip(null) iff in addrspace 0 |
2059 | if (IntrinsicID == Intrinsic::launder_invariant_group || |
2060 | IntrinsicID == Intrinsic::strip_invariant_group) { |
2061 | // If instruction is not yet put in a basic block (e.g. when cloning |
2062 | // a function during inlining), Call's caller may not be available. |
2063 | // So check Call's BB first before querying Call->getCaller. |
2064 | const Function *Caller = |
2065 | Call->getParent() ? Call->getCaller() : nullptr; |
2066 | if (Caller && |
2067 | !NullPointerIsDefined( |
2068 | F: Caller, AS: Operands[0]->getType()->getPointerAddressSpace())) { |
2069 | return Operands[0]; |
2070 | } |
2071 | return nullptr; |
2072 | } |
2073 | } |
2074 | |
2075 | if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[0])) { |
2076 | if (IntrinsicID == Intrinsic::convert_to_fp16) { |
2077 | APFloat Val(Op->getValueAPF()); |
2078 | |
2079 | bool lost = false; |
2080 | Val.convert(ToSemantics: APFloat::IEEEhalf(), RM: APFloat::rmNearestTiesToEven, losesInfo: &lost); |
2081 | |
2082 | return ConstantInt::get(Context&: Ty->getContext(), V: Val.bitcastToAPInt()); |
2083 | } |
2084 | |
2085 | APFloat U = Op->getValueAPF(); |
2086 | |
2087 | if (IntrinsicID == Intrinsic::wasm_trunc_signed || |
2088 | IntrinsicID == Intrinsic::wasm_trunc_unsigned) { |
2089 | bool Signed = IntrinsicID == Intrinsic::wasm_trunc_signed; |
2090 | |
2091 | if (U.isNaN()) |
2092 | return nullptr; |
2093 | |
2094 | unsigned Width = Ty->getIntegerBitWidth(); |
2095 | APSInt Int(Width, !Signed); |
2096 | bool IsExact = false; |
2097 | APFloat::opStatus Status = |
2098 | U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
2099 | |
2100 | if (Status == APFloat::opOK || Status == APFloat::opInexact) |
2101 | return ConstantInt::get(Ty, V: Int); |
2102 | |
2103 | return nullptr; |
2104 | } |
2105 | |
2106 | if (IntrinsicID == Intrinsic::fptoui_sat || |
2107 | IntrinsicID == Intrinsic::fptosi_sat) { |
2108 | // convertToInteger() already has the desired saturation semantics. |
2109 | APSInt Int(Ty->getIntegerBitWidth(), |
2110 | IntrinsicID == Intrinsic::fptoui_sat); |
2111 | bool IsExact; |
2112 | U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
2113 | return ConstantInt::get(Ty, V: Int); |
2114 | } |
2115 | |
2116 | if (IntrinsicID == Intrinsic::canonicalize) |
2117 | return constantFoldCanonicalize(Ty, CI: Call, Src: U); |
2118 | |
2119 | #if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128) |
2120 | if (Ty->isFP128Ty()) { |
2121 | if (IntrinsicID == Intrinsic::log) { |
2122 | float128 Result = logf128(Op->getValueAPF().convertToQuad()); |
2123 | return GetConstantFoldFPValue128(V: Result, Ty); |
2124 | } |
2125 | |
2126 | LibFunc Fp128Func = NotLibFunc; |
2127 | if (TLI->getLibFunc(funcName: Name, F&: Fp128Func) && TLI->has(F: Fp128Func) && |
2128 | Fp128Func == LibFunc_logl) |
2129 | return ConstantFoldFP128(logf128, Op->getValueAPF(), Ty); |
2130 | } |
2131 | #endif |
2132 | |
2133 | if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) |
2134 | return nullptr; |
2135 | |
2136 | // Use internal versions of these intrinsics. |
2137 | |
2138 | if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint) { |
2139 | U.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
2140 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2141 | } |
2142 | |
2143 | if (IntrinsicID == Intrinsic::round) { |
2144 | U.roundToIntegral(RM: APFloat::rmNearestTiesToAway); |
2145 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2146 | } |
2147 | |
2148 | if (IntrinsicID == Intrinsic::roundeven) { |
2149 | U.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
2150 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2151 | } |
2152 | |
2153 | if (IntrinsicID == Intrinsic::ceil) { |
2154 | U.roundToIntegral(RM: APFloat::rmTowardPositive); |
2155 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2156 | } |
2157 | |
2158 | if (IntrinsicID == Intrinsic::floor) { |
2159 | U.roundToIntegral(RM: APFloat::rmTowardNegative); |
2160 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2161 | } |
2162 | |
2163 | if (IntrinsicID == Intrinsic::trunc) { |
2164 | U.roundToIntegral(RM: APFloat::rmTowardZero); |
2165 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2166 | } |
2167 | |
2168 | if (IntrinsicID == Intrinsic::fabs) { |
2169 | U.clearSign(); |
2170 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2171 | } |
2172 | |
2173 | if (IntrinsicID == Intrinsic::amdgcn_fract) { |
2174 | // The v_fract instruction behaves like the OpenCL spec, which defines |
2175 | // fract(x) as fmin(x - floor(x), 0x1.fffffep-1f): "The min() operator is |
2176 | // there to prevent fract(-small) from returning 1.0. It returns the |
2177 | // largest positive floating-point number less than 1.0." |
2178 | APFloat FloorU(U); |
2179 | FloorU.roundToIntegral(RM: APFloat::rmTowardNegative); |
2180 | APFloat FractU(U - FloorU); |
2181 | APFloat AlmostOne(U.getSemantics(), 1); |
2182 | AlmostOne.next(/*nextDown*/ true); |
2183 | return ConstantFP::get(Context&: Ty->getContext(), V: minimum(A: FractU, B: AlmostOne)); |
2184 | } |
2185 | |
2186 | // Rounding operations (floor, trunc, ceil, round and nearbyint) do not |
2187 | // raise FP exceptions, unless the argument is signaling NaN. |
2188 | |
2189 | std::optional<APFloat::roundingMode> RM; |
2190 | switch (IntrinsicID) { |
2191 | default: |
2192 | break; |
2193 | case Intrinsic::experimental_constrained_nearbyint: |
2194 | case Intrinsic::experimental_constrained_rint: { |
2195 | auto CI = cast<ConstrainedFPIntrinsic>(Val: Call); |
2196 | RM = CI->getRoundingMode(); |
2197 | if (!RM || *RM == RoundingMode::Dynamic) |
2198 | return nullptr; |
2199 | break; |
2200 | } |
2201 | case Intrinsic::experimental_constrained_round: |
2202 | RM = APFloat::rmNearestTiesToAway; |
2203 | break; |
2204 | case Intrinsic::experimental_constrained_ceil: |
2205 | RM = APFloat::rmTowardPositive; |
2206 | break; |
2207 | case Intrinsic::experimental_constrained_floor: |
2208 | RM = APFloat::rmTowardNegative; |
2209 | break; |
2210 | case Intrinsic::experimental_constrained_trunc: |
2211 | RM = APFloat::rmTowardZero; |
2212 | break; |
2213 | } |
2214 | if (RM) { |
2215 | auto CI = cast<ConstrainedFPIntrinsic>(Val: Call); |
2216 | if (U.isFinite()) { |
2217 | APFloat::opStatus St = U.roundToIntegral(RM: *RM); |
2218 | if (IntrinsicID == Intrinsic::experimental_constrained_rint && |
2219 | St == APFloat::opInexact) { |
2220 | std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior(); |
2221 | if (EB && *EB == fp::ebStrict) |
2222 | return nullptr; |
2223 | } |
2224 | } else if (U.isSignaling()) { |
2225 | std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior(); |
2226 | if (EB && *EB != fp::ebIgnore) |
2227 | return nullptr; |
2228 | U = APFloat::getQNaN(Sem: U.getSemantics()); |
2229 | } |
2230 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2231 | } |
2232 | |
2233 | /// We only fold functions with finite arguments. Folding NaN and inf is |
2234 | /// likely to be aborted with an exception anyway, and some host libms |
2235 | /// have known errors raising exceptions. |
2236 | if (!U.isFinite()) |
2237 | return nullptr; |
2238 | |
2239 | /// Currently APFloat versions of these functions do not exist, so we use |
2240 | /// the host native double versions. Float versions are not called |
2241 | /// directly but for all these it is true (float)(f((double)arg)) == |
2242 | /// f(arg). Long double not supported yet. |
2243 | const APFloat &APF = Op->getValueAPF(); |
2244 | |
2245 | switch (IntrinsicID) { |
2246 | default: break; |
2247 | case Intrinsic::log: |
2248 | return ConstantFoldFP(NativeFP: log, V: APF, Ty); |
2249 | case Intrinsic::log2: |
2250 | // TODO: What about hosts that lack a C99 library? |
2251 | return ConstantFoldFP(NativeFP: log2, V: APF, Ty); |
2252 | case Intrinsic::log10: |
2253 | // TODO: What about hosts that lack a C99 library? |
2254 | return ConstantFoldFP(NativeFP: log10, V: APF, Ty); |
2255 | case Intrinsic::exp: |
2256 | return ConstantFoldFP(NativeFP: exp, V: APF, Ty); |
2257 | case Intrinsic::exp2: |
2258 | // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library. |
2259 | return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty); |
2260 | case Intrinsic::exp10: |
2261 | // Fold exp10(x) as pow(10, x), in case the host lacks a C99 library. |
2262 | return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(10.0), W: APF, Ty); |
2263 | case Intrinsic::sin: |
2264 | return ConstantFoldFP(NativeFP: sin, V: APF, Ty); |
2265 | case Intrinsic::cos: |
2266 | return ConstantFoldFP(NativeFP: cos, V: APF, Ty); |
2267 | case Intrinsic::sqrt: |
2268 | return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty); |
2269 | case Intrinsic::amdgcn_cos: |
2270 | case Intrinsic::amdgcn_sin: { |
2271 | double V = getValueAsDouble(Op); |
2272 | if (V < -256.0 || V > 256.0) |
2273 | // The gfx8 and gfx9 architectures handle arguments outside the range |
2274 | // [-256, 256] differently. This should be a rare case so bail out |
2275 | // rather than trying to handle the difference. |
2276 | return nullptr; |
2277 | bool IsCos = IntrinsicID == Intrinsic::amdgcn_cos; |
2278 | double V4 = V * 4.0; |
2279 | if (V4 == floor(x: V4)) { |
2280 | // Force exact results for quarter-integer inputs. |
2281 | const double SinVals[4] = { 0.0, 1.0, 0.0, -1.0 }; |
2282 | V = SinVals[((int)V4 + (IsCos ? 1 : 0)) & 3]; |
2283 | } else { |
2284 | if (IsCos) |
2285 | V = cos(x: V * 2.0 * numbers::pi); |
2286 | else |
2287 | V = sin(x: V * 2.0 * numbers::pi); |
2288 | } |
2289 | return GetConstantFoldFPValue(V, Ty); |
2290 | } |
2291 | } |
2292 | |
2293 | if (!TLI) |
2294 | return nullptr; |
2295 | |
2296 | LibFunc Func = NotLibFunc; |
2297 | if (!TLI->getLibFunc(funcName: Name, F&: Func)) |
2298 | return nullptr; |
2299 | |
2300 | switch (Func) { |
2301 | default: |
2302 | break; |
2303 | case LibFunc_acos: |
2304 | case LibFunc_acosf: |
2305 | case LibFunc_acos_finite: |
2306 | case LibFunc_acosf_finite: |
2307 | if (TLI->has(F: Func)) |
2308 | return ConstantFoldFP(NativeFP: acos, V: APF, Ty); |
2309 | break; |
2310 | case LibFunc_asin: |
2311 | case LibFunc_asinf: |
2312 | case LibFunc_asin_finite: |
2313 | case LibFunc_asinf_finite: |
2314 | if (TLI->has(F: Func)) |
2315 | return ConstantFoldFP(NativeFP: asin, V: APF, Ty); |
2316 | break; |
2317 | case LibFunc_atan: |
2318 | case LibFunc_atanf: |
2319 | if (TLI->has(F: Func)) |
2320 | return ConstantFoldFP(NativeFP: atan, V: APF, Ty); |
2321 | break; |
2322 | case LibFunc_ceil: |
2323 | case LibFunc_ceilf: |
2324 | if (TLI->has(F: Func)) { |
2325 | U.roundToIntegral(RM: APFloat::rmTowardPositive); |
2326 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2327 | } |
2328 | break; |
2329 | case LibFunc_cos: |
2330 | case LibFunc_cosf: |
2331 | if (TLI->has(F: Func)) |
2332 | return ConstantFoldFP(NativeFP: cos, V: APF, Ty); |
2333 | break; |
2334 | case LibFunc_cosh: |
2335 | case LibFunc_coshf: |
2336 | case LibFunc_cosh_finite: |
2337 | case LibFunc_coshf_finite: |
2338 | if (TLI->has(F: Func)) |
2339 | return ConstantFoldFP(NativeFP: cosh, V: APF, Ty); |
2340 | break; |
2341 | case LibFunc_exp: |
2342 | case LibFunc_expf: |
2343 | case LibFunc_exp_finite: |
2344 | case LibFunc_expf_finite: |
2345 | if (TLI->has(F: Func)) |
2346 | return ConstantFoldFP(NativeFP: exp, V: APF, Ty); |
2347 | break; |
2348 | case LibFunc_exp2: |
2349 | case LibFunc_exp2f: |
2350 | case LibFunc_exp2_finite: |
2351 | case LibFunc_exp2f_finite: |
2352 | if (TLI->has(F: Func)) |
2353 | // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library. |
2354 | return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty); |
2355 | break; |
2356 | case LibFunc_fabs: |
2357 | case LibFunc_fabsf: |
2358 | if (TLI->has(F: Func)) { |
2359 | U.clearSign(); |
2360 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2361 | } |
2362 | break; |
2363 | case LibFunc_floor: |
2364 | case LibFunc_floorf: |
2365 | if (TLI->has(F: Func)) { |
2366 | U.roundToIntegral(RM: APFloat::rmTowardNegative); |
2367 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2368 | } |
2369 | break; |
2370 | case LibFunc_log: |
2371 | case LibFunc_logf: |
2372 | case LibFunc_log_finite: |
2373 | case LibFunc_logf_finite: |
2374 | if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func)) |
2375 | return ConstantFoldFP(NativeFP: log, V: APF, Ty); |
2376 | break; |
2377 | case LibFunc_log2: |
2378 | case LibFunc_log2f: |
2379 | case LibFunc_log2_finite: |
2380 | case LibFunc_log2f_finite: |
2381 | if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func)) |
2382 | // TODO: What about hosts that lack a C99 library? |
2383 | return ConstantFoldFP(NativeFP: log2, V: APF, Ty); |
2384 | break; |
2385 | case LibFunc_log10: |
2386 | case LibFunc_log10f: |
2387 | case LibFunc_log10_finite: |
2388 | case LibFunc_log10f_finite: |
2389 | if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func)) |
2390 | // TODO: What about hosts that lack a C99 library? |
2391 | return ConstantFoldFP(NativeFP: log10, V: APF, Ty); |
2392 | break; |
2393 | case LibFunc_logl: |
2394 | return nullptr; |
2395 | case LibFunc_nearbyint: |
2396 | case LibFunc_nearbyintf: |
2397 | case LibFunc_rint: |
2398 | case LibFunc_rintf: |
2399 | if (TLI->has(F: Func)) { |
2400 | U.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
2401 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2402 | } |
2403 | break; |
2404 | case LibFunc_round: |
2405 | case LibFunc_roundf: |
2406 | if (TLI->has(F: Func)) { |
2407 | U.roundToIntegral(RM: APFloat::rmNearestTiesToAway); |
2408 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2409 | } |
2410 | break; |
2411 | case LibFunc_sin: |
2412 | case LibFunc_sinf: |
2413 | if (TLI->has(F: Func)) |
2414 | return ConstantFoldFP(NativeFP: sin, V: APF, Ty); |
2415 | break; |
2416 | case LibFunc_sinh: |
2417 | case LibFunc_sinhf: |
2418 | case LibFunc_sinh_finite: |
2419 | case LibFunc_sinhf_finite: |
2420 | if (TLI->has(F: Func)) |
2421 | return ConstantFoldFP(NativeFP: sinh, V: APF, Ty); |
2422 | break; |
2423 | case LibFunc_sqrt: |
2424 | case LibFunc_sqrtf: |
2425 | if (!APF.isNegative() && TLI->has(F: Func)) |
2426 | return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty); |
2427 | break; |
2428 | case LibFunc_tan: |
2429 | case LibFunc_tanf: |
2430 | if (TLI->has(F: Func)) |
2431 | return ConstantFoldFP(NativeFP: tan, V: APF, Ty); |
2432 | break; |
2433 | case LibFunc_tanh: |
2434 | case LibFunc_tanhf: |
2435 | if (TLI->has(F: Func)) |
2436 | return ConstantFoldFP(NativeFP: tanh, V: APF, Ty); |
2437 | break; |
2438 | case LibFunc_trunc: |
2439 | case LibFunc_truncf: |
2440 | if (TLI->has(F: Func)) { |
2441 | U.roundToIntegral(RM: APFloat::rmTowardZero); |
2442 | return ConstantFP::get(Context&: Ty->getContext(), V: U); |
2443 | } |
2444 | break; |
2445 | } |
2446 | return nullptr; |
2447 | } |
2448 | |
2449 | if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) { |
2450 | switch (IntrinsicID) { |
2451 | case Intrinsic::bswap: |
2452 | return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().byteSwap()); |
2453 | case Intrinsic::ctpop: |
2454 | return ConstantInt::get(Ty, V: Op->getValue().popcount()); |
2455 | case Intrinsic::bitreverse: |
2456 | return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().reverseBits()); |
2457 | case Intrinsic::convert_from_fp16: { |
2458 | APFloat Val(APFloat::IEEEhalf(), Op->getValue()); |
2459 | |
2460 | bool lost = false; |
2461 | APFloat::opStatus status = Val.convert( |
2462 | ToSemantics: Ty->getFltSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &lost); |
2463 | |
2464 | // Conversion is always precise. |
2465 | (void)status; |
2466 | assert(status != APFloat::opInexact && !lost && |
2467 | "Precision lost during fp16 constfolding" ); |
2468 | |
2469 | return ConstantFP::get(Context&: Ty->getContext(), V: Val); |
2470 | } |
2471 | |
2472 | case Intrinsic::amdgcn_s_wqm: { |
2473 | uint64_t Val = Op->getZExtValue(); |
2474 | Val |= (Val & 0x5555555555555555ULL) << 1 | |
2475 | ((Val >> 1) & 0x5555555555555555ULL); |
2476 | Val |= (Val & 0x3333333333333333ULL) << 2 | |
2477 | ((Val >> 2) & 0x3333333333333333ULL); |
2478 | return ConstantInt::get(Ty, V: Val); |
2479 | } |
2480 | |
2481 | case Intrinsic::amdgcn_s_quadmask: { |
2482 | uint64_t Val = Op->getZExtValue(); |
2483 | uint64_t QuadMask = 0; |
2484 | for (unsigned I = 0; I < Op->getBitWidth() / 4; ++I, Val >>= 4) { |
2485 | if (!(Val & 0xF)) |
2486 | continue; |
2487 | |
2488 | QuadMask |= (1ULL << I); |
2489 | } |
2490 | return ConstantInt::get(Ty, V: QuadMask); |
2491 | } |
2492 | |
2493 | case Intrinsic::amdgcn_s_bitreplicate: { |
2494 | uint64_t Val = Op->getZExtValue(); |
2495 | Val = (Val & 0x000000000000FFFFULL) | (Val & 0x00000000FFFF0000ULL) << 16; |
2496 | Val = (Val & 0x000000FF000000FFULL) | (Val & 0x0000FF000000FF00ULL) << 8; |
2497 | Val = (Val & 0x000F000F000F000FULL) | (Val & 0x00F000F000F000F0ULL) << 4; |
2498 | Val = (Val & 0x0303030303030303ULL) | (Val & 0x0C0C0C0C0C0C0C0CULL) << 2; |
2499 | Val = (Val & 0x1111111111111111ULL) | (Val & 0x2222222222222222ULL) << 1; |
2500 | Val = Val | Val << 1; |
2501 | return ConstantInt::get(Ty, V: Val); |
2502 | } |
2503 | |
2504 | default: |
2505 | return nullptr; |
2506 | } |
2507 | } |
2508 | |
2509 | switch (IntrinsicID) { |
2510 | default: break; |
2511 | case Intrinsic::vector_reduce_add: |
2512 | case Intrinsic::vector_reduce_mul: |
2513 | case Intrinsic::vector_reduce_and: |
2514 | case Intrinsic::vector_reduce_or: |
2515 | case Intrinsic::vector_reduce_xor: |
2516 | case Intrinsic::vector_reduce_smin: |
2517 | case Intrinsic::vector_reduce_smax: |
2518 | case Intrinsic::vector_reduce_umin: |
2519 | case Intrinsic::vector_reduce_umax: |
2520 | if (Constant *C = constantFoldVectorReduce(IID: IntrinsicID, Op: Operands[0])) |
2521 | return C; |
2522 | break; |
2523 | } |
2524 | |
2525 | // Support ConstantVector in case we have an Undef in the top. |
2526 | if (isa<ConstantVector>(Val: Operands[0]) || |
2527 | isa<ConstantDataVector>(Val: Operands[0])) { |
2528 | auto *Op = cast<Constant>(Val: Operands[0]); |
2529 | switch (IntrinsicID) { |
2530 | default: break; |
2531 | case Intrinsic::x86_sse_cvtss2si: |
2532 | case Intrinsic::x86_sse_cvtss2si64: |
2533 | case Intrinsic::x86_sse2_cvtsd2si: |
2534 | case Intrinsic::x86_sse2_cvtsd2si64: |
2535 | if (ConstantFP *FPOp = |
2536 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2537 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2538 | /*roundTowardZero=*/false, Ty, |
2539 | /*IsSigned*/true); |
2540 | break; |
2541 | case Intrinsic::x86_sse_cvttss2si: |
2542 | case Intrinsic::x86_sse_cvttss2si64: |
2543 | case Intrinsic::x86_sse2_cvttsd2si: |
2544 | case Intrinsic::x86_sse2_cvttsd2si64: |
2545 | if (ConstantFP *FPOp = |
2546 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2547 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2548 | /*roundTowardZero=*/true, Ty, |
2549 | /*IsSigned*/true); |
2550 | break; |
2551 | } |
2552 | } |
2553 | |
2554 | return nullptr; |
2555 | } |
2556 | |
2557 | static Constant *evaluateCompare(const APFloat &Op1, const APFloat &Op2, |
2558 | const ConstrainedFPIntrinsic *Call) { |
2559 | APFloat::opStatus St = APFloat::opOK; |
2560 | auto *FCmp = cast<ConstrainedFPCmpIntrinsic>(Val: Call); |
2561 | FCmpInst::Predicate Cond = FCmp->getPredicate(); |
2562 | if (FCmp->isSignaling()) { |
2563 | if (Op1.isNaN() || Op2.isNaN()) |
2564 | St = APFloat::opInvalidOp; |
2565 | } else { |
2566 | if (Op1.isSignaling() || Op2.isSignaling()) |
2567 | St = APFloat::opInvalidOp; |
2568 | } |
2569 | bool Result = FCmpInst::compare(LHS: Op1, RHS: Op2, Pred: Cond); |
2570 | if (mayFoldConstrained(CI: const_cast<ConstrainedFPCmpIntrinsic *>(FCmp), St)) |
2571 | return ConstantInt::get(Ty: Call->getType()->getScalarType(), V: Result); |
2572 | return nullptr; |
2573 | } |
2574 | |
2575 | static Constant *ConstantFoldLibCall2(StringRef Name, Type *Ty, |
2576 | ArrayRef<Constant *> Operands, |
2577 | const TargetLibraryInfo *TLI) { |
2578 | if (!TLI) |
2579 | return nullptr; |
2580 | |
2581 | LibFunc Func = NotLibFunc; |
2582 | if (!TLI->getLibFunc(funcName: Name, F&: Func)) |
2583 | return nullptr; |
2584 | |
2585 | const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0]); |
2586 | if (!Op1) |
2587 | return nullptr; |
2588 | |
2589 | const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1]); |
2590 | if (!Op2) |
2591 | return nullptr; |
2592 | |
2593 | const APFloat &Op1V = Op1->getValueAPF(); |
2594 | const APFloat &Op2V = Op2->getValueAPF(); |
2595 | |
2596 | switch (Func) { |
2597 | default: |
2598 | break; |
2599 | case LibFunc_pow: |
2600 | case LibFunc_powf: |
2601 | case LibFunc_pow_finite: |
2602 | case LibFunc_powf_finite: |
2603 | if (TLI->has(F: Func)) |
2604 | return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty); |
2605 | break; |
2606 | case LibFunc_fmod: |
2607 | case LibFunc_fmodf: |
2608 | if (TLI->has(F: Func)) { |
2609 | APFloat V = Op1->getValueAPF(); |
2610 | if (APFloat::opStatus::opOK == V.mod(RHS: Op2->getValueAPF())) |
2611 | return ConstantFP::get(Context&: Ty->getContext(), V); |
2612 | } |
2613 | break; |
2614 | case LibFunc_remainder: |
2615 | case LibFunc_remainderf: |
2616 | if (TLI->has(F: Func)) { |
2617 | APFloat V = Op1->getValueAPF(); |
2618 | if (APFloat::opStatus::opOK == V.remainder(RHS: Op2->getValueAPF())) |
2619 | return ConstantFP::get(Context&: Ty->getContext(), V); |
2620 | } |
2621 | break; |
2622 | case LibFunc_atan2: |
2623 | case LibFunc_atan2f: |
2624 | // atan2(+/-0.0, +/-0.0) is known to raise an exception on some libm |
2625 | // (Solaris), so we do not assume a known result for that. |
2626 | if (Op1V.isZero() && Op2V.isZero()) |
2627 | return nullptr; |
2628 | [[fallthrough]]; |
2629 | case LibFunc_atan2_finite: |
2630 | case LibFunc_atan2f_finite: |
2631 | if (TLI->has(F: Func)) |
2632 | return ConstantFoldBinaryFP(NativeFP: atan2, V: Op1V, W: Op2V, Ty); |
2633 | break; |
2634 | } |
2635 | |
2636 | return nullptr; |
2637 | } |
2638 | |
2639 | static Constant *ConstantFoldIntrinsicCall2(Intrinsic::ID IntrinsicID, Type *Ty, |
2640 | ArrayRef<Constant *> Operands, |
2641 | const CallBase *Call) { |
2642 | assert(Operands.size() == 2 && "Wrong number of operands." ); |
2643 | |
2644 | if (Ty->isFloatingPointTy()) { |
2645 | // TODO: We should have undef handling for all of the FP intrinsics that |
2646 | // are attempted to be folded in this function. |
2647 | bool IsOp0Undef = isa<UndefValue>(Val: Operands[0]); |
2648 | bool IsOp1Undef = isa<UndefValue>(Val: Operands[1]); |
2649 | switch (IntrinsicID) { |
2650 | case Intrinsic::maxnum: |
2651 | case Intrinsic::minnum: |
2652 | case Intrinsic::maximum: |
2653 | case Intrinsic::minimum: |
2654 | // If one argument is undef, return the other argument. |
2655 | if (IsOp0Undef) |
2656 | return Operands[1]; |
2657 | if (IsOp1Undef) |
2658 | return Operands[0]; |
2659 | break; |
2660 | } |
2661 | } |
2662 | |
2663 | if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) { |
2664 | const APFloat &Op1V = Op1->getValueAPF(); |
2665 | |
2666 | if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) { |
2667 | if (Op2->getType() != Op1->getType()) |
2668 | return nullptr; |
2669 | const APFloat &Op2V = Op2->getValueAPF(); |
2670 | |
2671 | if (const auto *ConstrIntr = |
2672 | dyn_cast_if_present<ConstrainedFPIntrinsic>(Val: Call)) { |
2673 | RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr); |
2674 | APFloat Res = Op1V; |
2675 | APFloat::opStatus St; |
2676 | switch (IntrinsicID) { |
2677 | default: |
2678 | return nullptr; |
2679 | case Intrinsic::experimental_constrained_fadd: |
2680 | St = Res.add(RHS: Op2V, RM); |
2681 | break; |
2682 | case Intrinsic::experimental_constrained_fsub: |
2683 | St = Res.subtract(RHS: Op2V, RM); |
2684 | break; |
2685 | case Intrinsic::experimental_constrained_fmul: |
2686 | St = Res.multiply(RHS: Op2V, RM); |
2687 | break; |
2688 | case Intrinsic::experimental_constrained_fdiv: |
2689 | St = Res.divide(RHS: Op2V, RM); |
2690 | break; |
2691 | case Intrinsic::experimental_constrained_frem: |
2692 | St = Res.mod(RHS: Op2V); |
2693 | break; |
2694 | case Intrinsic::experimental_constrained_fcmp: |
2695 | case Intrinsic::experimental_constrained_fcmps: |
2696 | return evaluateCompare(Op1: Op1V, Op2: Op2V, Call: ConstrIntr); |
2697 | } |
2698 | if (mayFoldConstrained(CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), |
2699 | St)) |
2700 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
2701 | return nullptr; |
2702 | } |
2703 | |
2704 | switch (IntrinsicID) { |
2705 | default: |
2706 | break; |
2707 | case Intrinsic::copysign: |
2708 | return ConstantFP::get(Context&: Ty->getContext(), V: APFloat::copySign(Value: Op1V, Sign: Op2V)); |
2709 | case Intrinsic::minnum: |
2710 | return ConstantFP::get(Context&: Ty->getContext(), V: minnum(A: Op1V, B: Op2V)); |
2711 | case Intrinsic::maxnum: |
2712 | return ConstantFP::get(Context&: Ty->getContext(), V: maxnum(A: Op1V, B: Op2V)); |
2713 | case Intrinsic::minimum: |
2714 | return ConstantFP::get(Context&: Ty->getContext(), V: minimum(A: Op1V, B: Op2V)); |
2715 | case Intrinsic::maximum: |
2716 | return ConstantFP::get(Context&: Ty->getContext(), V: maximum(A: Op1V, B: Op2V)); |
2717 | } |
2718 | |
2719 | if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy()) |
2720 | return nullptr; |
2721 | |
2722 | switch (IntrinsicID) { |
2723 | default: |
2724 | break; |
2725 | case Intrinsic::pow: |
2726 | return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty); |
2727 | case Intrinsic::amdgcn_fmul_legacy: |
2728 | // The legacy behaviour is that multiplying +/- 0.0 by anything, even |
2729 | // NaN or infinity, gives +0.0. |
2730 | if (Op1V.isZero() || Op2V.isZero()) |
2731 | return ConstantFP::getZero(Ty); |
2732 | return ConstantFP::get(Context&: Ty->getContext(), V: Op1V * Op2V); |
2733 | } |
2734 | |
2735 | } else if (auto *Op2C = dyn_cast<ConstantInt>(Val: Operands[1])) { |
2736 | switch (IntrinsicID) { |
2737 | case Intrinsic::ldexp: { |
2738 | return ConstantFP::get( |
2739 | Context&: Ty->getContext(), |
2740 | V: scalbn(X: Op1V, Exp: Op2C->getSExtValue(), RM: APFloat::rmNearestTiesToEven)); |
2741 | } |
2742 | case Intrinsic::is_fpclass: { |
2743 | FPClassTest Mask = static_cast<FPClassTest>(Op2C->getZExtValue()); |
2744 | bool Result = |
2745 | ((Mask & fcSNan) && Op1V.isNaN() && Op1V.isSignaling()) || |
2746 | ((Mask & fcQNan) && Op1V.isNaN() && !Op1V.isSignaling()) || |
2747 | ((Mask & fcNegInf) && Op1V.isNegInfinity()) || |
2748 | ((Mask & fcNegNormal) && Op1V.isNormal() && Op1V.isNegative()) || |
2749 | ((Mask & fcNegSubnormal) && Op1V.isDenormal() && Op1V.isNegative()) || |
2750 | ((Mask & fcNegZero) && Op1V.isZero() && Op1V.isNegative()) || |
2751 | ((Mask & fcPosZero) && Op1V.isZero() && !Op1V.isNegative()) || |
2752 | ((Mask & fcPosSubnormal) && Op1V.isDenormal() && !Op1V.isNegative()) || |
2753 | ((Mask & fcPosNormal) && Op1V.isNormal() && !Op1V.isNegative()) || |
2754 | ((Mask & fcPosInf) && Op1V.isPosInfinity()); |
2755 | return ConstantInt::get(Ty, V: Result); |
2756 | } |
2757 | case Intrinsic::powi: { |
2758 | int Exp = static_cast<int>(Op2C->getSExtValue()); |
2759 | switch (Ty->getTypeID()) { |
2760 | case Type::HalfTyID: |
2761 | case Type::FloatTyID: { |
2762 | APFloat Res(static_cast<float>(std::pow(x: Op1V.convertToFloat(), y: Exp))); |
2763 | if (Ty->isHalfTy()) { |
2764 | bool Unused; |
2765 | Res.convert(ToSemantics: APFloat::IEEEhalf(), RM: APFloat::rmNearestTiesToEven, |
2766 | losesInfo: &Unused); |
2767 | } |
2768 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
2769 | } |
2770 | case Type::DoubleTyID: |
2771 | return ConstantFP::get(Ty, V: std::pow(x: Op1V.convertToDouble(), y: Exp)); |
2772 | default: |
2773 | return nullptr; |
2774 | } |
2775 | } |
2776 | default: |
2777 | break; |
2778 | } |
2779 | } |
2780 | return nullptr; |
2781 | } |
2782 | |
2783 | if (Operands[0]->getType()->isIntegerTy() && |
2784 | Operands[1]->getType()->isIntegerTy()) { |
2785 | const APInt *C0, *C1; |
2786 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
2787 | !getConstIntOrUndef(Op: Operands[1], C&: C1)) |
2788 | return nullptr; |
2789 | |
2790 | switch (IntrinsicID) { |
2791 | default: break; |
2792 | case Intrinsic::smax: |
2793 | case Intrinsic::smin: |
2794 | case Intrinsic::umax: |
2795 | case Intrinsic::umin: |
2796 | // This is the same as for binary ops - poison propagates. |
2797 | // TODO: Poison handling should be consolidated. |
2798 | if (isa<PoisonValue>(Val: Operands[0]) || isa<PoisonValue>(Val: Operands[1])) |
2799 | return PoisonValue::get(T: Ty); |
2800 | |
2801 | if (!C0 && !C1) |
2802 | return UndefValue::get(T: Ty); |
2803 | if (!C0 || !C1) |
2804 | return MinMaxIntrinsic::getSaturationPoint(ID: IntrinsicID, Ty); |
2805 | return ConstantInt::get( |
2806 | Ty, V: ICmpInst::compare(LHS: *C0, RHS: *C1, |
2807 | Pred: MinMaxIntrinsic::getPredicate(ID: IntrinsicID)) |
2808 | ? *C0 |
2809 | : *C1); |
2810 | |
2811 | case Intrinsic::scmp: |
2812 | case Intrinsic::ucmp: |
2813 | if (isa<PoisonValue>(Val: Operands[0]) || isa<PoisonValue>(Val: Operands[1])) |
2814 | return PoisonValue::get(T: Ty); |
2815 | |
2816 | if (!C0 || !C1) |
2817 | return ConstantInt::get(Ty, V: 0); |
2818 | |
2819 | int Res; |
2820 | if (IntrinsicID == Intrinsic::scmp) |
2821 | Res = C0->sgt(RHS: *C1) ? 1 : C0->slt(RHS: *C1) ? -1 : 0; |
2822 | else |
2823 | Res = C0->ugt(RHS: *C1) ? 1 : C0->ult(RHS: *C1) ? -1 : 0; |
2824 | return ConstantInt::get(Ty, V: Res, /*IsSigned=*/true); |
2825 | |
2826 | case Intrinsic::usub_with_overflow: |
2827 | case Intrinsic::ssub_with_overflow: |
2828 | // X - undef -> { 0, false } |
2829 | // undef - X -> { 0, false } |
2830 | if (!C0 || !C1) |
2831 | return Constant::getNullValue(Ty); |
2832 | [[fallthrough]]; |
2833 | case Intrinsic::uadd_with_overflow: |
2834 | case Intrinsic::sadd_with_overflow: |
2835 | // X + undef -> { -1, false } |
2836 | // undef + x -> { -1, false } |
2837 | if (!C0 || !C1) { |
2838 | return ConstantStruct::get( |
2839 | T: cast<StructType>(Val: Ty), |
2840 | V: {Constant::getAllOnesValue(Ty: Ty->getStructElementType(N: 0)), |
2841 | Constant::getNullValue(Ty: Ty->getStructElementType(N: 1))}); |
2842 | } |
2843 | [[fallthrough]]; |
2844 | case Intrinsic::smul_with_overflow: |
2845 | case Intrinsic::umul_with_overflow: { |
2846 | // undef * X -> { 0, false } |
2847 | // X * undef -> { 0, false } |
2848 | if (!C0 || !C1) |
2849 | return Constant::getNullValue(Ty); |
2850 | |
2851 | APInt Res; |
2852 | bool Overflow; |
2853 | switch (IntrinsicID) { |
2854 | default: llvm_unreachable("Invalid case" ); |
2855 | case Intrinsic::sadd_with_overflow: |
2856 | Res = C0->sadd_ov(RHS: *C1, Overflow); |
2857 | break; |
2858 | case Intrinsic::uadd_with_overflow: |
2859 | Res = C0->uadd_ov(RHS: *C1, Overflow); |
2860 | break; |
2861 | case Intrinsic::ssub_with_overflow: |
2862 | Res = C0->ssub_ov(RHS: *C1, Overflow); |
2863 | break; |
2864 | case Intrinsic::usub_with_overflow: |
2865 | Res = C0->usub_ov(RHS: *C1, Overflow); |
2866 | break; |
2867 | case Intrinsic::smul_with_overflow: |
2868 | Res = C0->smul_ov(RHS: *C1, Overflow); |
2869 | break; |
2870 | case Intrinsic::umul_with_overflow: |
2871 | Res = C0->umul_ov(RHS: *C1, Overflow); |
2872 | break; |
2873 | } |
2874 | Constant *Ops[] = { |
2875 | ConstantInt::get(Context&: Ty->getContext(), V: Res), |
2876 | ConstantInt::get(Ty: Type::getInt1Ty(C&: Ty->getContext()), V: Overflow) |
2877 | }; |
2878 | return ConstantStruct::get(T: cast<StructType>(Val: Ty), V: Ops); |
2879 | } |
2880 | case Intrinsic::uadd_sat: |
2881 | case Intrinsic::sadd_sat: |
2882 | // This is the same as for binary ops - poison propagates. |
2883 | // TODO: Poison handling should be consolidated. |
2884 | if (isa<PoisonValue>(Val: Operands[0]) || isa<PoisonValue>(Val: Operands[1])) |
2885 | return PoisonValue::get(T: Ty); |
2886 | |
2887 | if (!C0 && !C1) |
2888 | return UndefValue::get(T: Ty); |
2889 | if (!C0 || !C1) |
2890 | return Constant::getAllOnesValue(Ty); |
2891 | if (IntrinsicID == Intrinsic::uadd_sat) |
2892 | return ConstantInt::get(Ty, V: C0->uadd_sat(RHS: *C1)); |
2893 | else |
2894 | return ConstantInt::get(Ty, V: C0->sadd_sat(RHS: *C1)); |
2895 | case Intrinsic::usub_sat: |
2896 | case Intrinsic::ssub_sat: |
2897 | // This is the same as for binary ops - poison propagates. |
2898 | // TODO: Poison handling should be consolidated. |
2899 | if (isa<PoisonValue>(Val: Operands[0]) || isa<PoisonValue>(Val: Operands[1])) |
2900 | return PoisonValue::get(T: Ty); |
2901 | |
2902 | if (!C0 && !C1) |
2903 | return UndefValue::get(T: Ty); |
2904 | if (!C0 || !C1) |
2905 | return Constant::getNullValue(Ty); |
2906 | if (IntrinsicID == Intrinsic::usub_sat) |
2907 | return ConstantInt::get(Ty, V: C0->usub_sat(RHS: *C1)); |
2908 | else |
2909 | return ConstantInt::get(Ty, V: C0->ssub_sat(RHS: *C1)); |
2910 | case Intrinsic::cttz: |
2911 | case Intrinsic::ctlz: |
2912 | assert(C1 && "Must be constant int" ); |
2913 | |
2914 | // cttz(0, 1) and ctlz(0, 1) are poison. |
2915 | if (C1->isOne() && (!C0 || C0->isZero())) |
2916 | return PoisonValue::get(T: Ty); |
2917 | if (!C0) |
2918 | return Constant::getNullValue(Ty); |
2919 | if (IntrinsicID == Intrinsic::cttz) |
2920 | return ConstantInt::get(Ty, V: C0->countr_zero()); |
2921 | else |
2922 | return ConstantInt::get(Ty, V: C0->countl_zero()); |
2923 | |
2924 | case Intrinsic::abs: |
2925 | assert(C1 && "Must be constant int" ); |
2926 | assert((C1->isOne() || C1->isZero()) && "Must be 0 or 1" ); |
2927 | |
2928 | // Undef or minimum val operand with poison min --> undef |
2929 | if (C1->isOne() && (!C0 || C0->isMinSignedValue())) |
2930 | return UndefValue::get(T: Ty); |
2931 | |
2932 | // Undef operand with no poison min --> 0 (sign bit must be clear) |
2933 | if (!C0) |
2934 | return Constant::getNullValue(Ty); |
2935 | |
2936 | return ConstantInt::get(Ty, V: C0->abs()); |
2937 | case Intrinsic::amdgcn_wave_reduce_umin: |
2938 | case Intrinsic::amdgcn_wave_reduce_umax: |
2939 | return dyn_cast<Constant>(Val: Operands[0]); |
2940 | } |
2941 | |
2942 | return nullptr; |
2943 | } |
2944 | |
2945 | // Support ConstantVector in case we have an Undef in the top. |
2946 | if ((isa<ConstantVector>(Val: Operands[0]) || |
2947 | isa<ConstantDataVector>(Val: Operands[0])) && |
2948 | // Check for default rounding mode. |
2949 | // FIXME: Support other rounding modes? |
2950 | isa<ConstantInt>(Val: Operands[1]) && |
2951 | cast<ConstantInt>(Val: Operands[1])->getValue() == 4) { |
2952 | auto *Op = cast<Constant>(Val: Operands[0]); |
2953 | switch (IntrinsicID) { |
2954 | default: break; |
2955 | case Intrinsic::x86_avx512_vcvtss2si32: |
2956 | case Intrinsic::x86_avx512_vcvtss2si64: |
2957 | case Intrinsic::x86_avx512_vcvtsd2si32: |
2958 | case Intrinsic::x86_avx512_vcvtsd2si64: |
2959 | if (ConstantFP *FPOp = |
2960 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2961 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2962 | /*roundTowardZero=*/false, Ty, |
2963 | /*IsSigned*/true); |
2964 | break; |
2965 | case Intrinsic::x86_avx512_vcvtss2usi32: |
2966 | case Intrinsic::x86_avx512_vcvtss2usi64: |
2967 | case Intrinsic::x86_avx512_vcvtsd2usi32: |
2968 | case Intrinsic::x86_avx512_vcvtsd2usi64: |
2969 | if (ConstantFP *FPOp = |
2970 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2971 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2972 | /*roundTowardZero=*/false, Ty, |
2973 | /*IsSigned*/false); |
2974 | break; |
2975 | case Intrinsic::x86_avx512_cvttss2si: |
2976 | case Intrinsic::x86_avx512_cvttss2si64: |
2977 | case Intrinsic::x86_avx512_cvttsd2si: |
2978 | case Intrinsic::x86_avx512_cvttsd2si64: |
2979 | if (ConstantFP *FPOp = |
2980 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2981 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2982 | /*roundTowardZero=*/true, Ty, |
2983 | /*IsSigned*/true); |
2984 | break; |
2985 | case Intrinsic::x86_avx512_cvttss2usi: |
2986 | case Intrinsic::x86_avx512_cvttss2usi64: |
2987 | case Intrinsic::x86_avx512_cvttsd2usi: |
2988 | case Intrinsic::x86_avx512_cvttsd2usi64: |
2989 | if (ConstantFP *FPOp = |
2990 | dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U))) |
2991 | return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(), |
2992 | /*roundTowardZero=*/true, Ty, |
2993 | /*IsSigned*/false); |
2994 | break; |
2995 | } |
2996 | } |
2997 | return nullptr; |
2998 | } |
2999 | |
3000 | static APFloat ConstantFoldAMDGCNCubeIntrinsic(Intrinsic::ID IntrinsicID, |
3001 | const APFloat &S0, |
3002 | const APFloat &S1, |
3003 | const APFloat &S2) { |
3004 | unsigned ID; |
3005 | const fltSemantics &Sem = S0.getSemantics(); |
3006 | APFloat MA(Sem), SC(Sem), TC(Sem); |
3007 | if (abs(X: S2) >= abs(X: S0) && abs(X: S2) >= abs(X: S1)) { |
3008 | if (S2.isNegative() && S2.isNonZero() && !S2.isNaN()) { |
3009 | // S2 < 0 |
3010 | ID = 5; |
3011 | SC = -S0; |
3012 | } else { |
3013 | ID = 4; |
3014 | SC = S0; |
3015 | } |
3016 | MA = S2; |
3017 | TC = -S1; |
3018 | } else if (abs(X: S1) >= abs(X: S0)) { |
3019 | if (S1.isNegative() && S1.isNonZero() && !S1.isNaN()) { |
3020 | // S1 < 0 |
3021 | ID = 3; |
3022 | TC = -S2; |
3023 | } else { |
3024 | ID = 2; |
3025 | TC = S2; |
3026 | } |
3027 | MA = S1; |
3028 | SC = S0; |
3029 | } else { |
3030 | if (S0.isNegative() && S0.isNonZero() && !S0.isNaN()) { |
3031 | // S0 < 0 |
3032 | ID = 1; |
3033 | SC = S2; |
3034 | } else { |
3035 | ID = 0; |
3036 | SC = -S2; |
3037 | } |
3038 | MA = S0; |
3039 | TC = -S1; |
3040 | } |
3041 | switch (IntrinsicID) { |
3042 | default: |
3043 | llvm_unreachable("unhandled amdgcn cube intrinsic" ); |
3044 | case Intrinsic::amdgcn_cubeid: |
3045 | return APFloat(Sem, ID); |
3046 | case Intrinsic::amdgcn_cubema: |
3047 | return MA + MA; |
3048 | case Intrinsic::amdgcn_cubesc: |
3049 | return SC; |
3050 | case Intrinsic::amdgcn_cubetc: |
3051 | return TC; |
3052 | } |
3053 | } |
3054 | |
3055 | static Constant *ConstantFoldAMDGCNPermIntrinsic(ArrayRef<Constant *> Operands, |
3056 | Type *Ty) { |
3057 | const APInt *C0, *C1, *C2; |
3058 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3059 | !getConstIntOrUndef(Op: Operands[1], C&: C1) || |
3060 | !getConstIntOrUndef(Op: Operands[2], C&: C2)) |
3061 | return nullptr; |
3062 | |
3063 | if (!C2) |
3064 | return UndefValue::get(T: Ty); |
3065 | |
3066 | APInt Val(32, 0); |
3067 | unsigned NumUndefBytes = 0; |
3068 | for (unsigned I = 0; I < 32; I += 8) { |
3069 | unsigned Sel = C2->extractBitsAsZExtValue(numBits: 8, bitPosition: I); |
3070 | unsigned B = 0; |
3071 | |
3072 | if (Sel >= 13) |
3073 | B = 0xff; |
3074 | else if (Sel == 12) |
3075 | B = 0x00; |
3076 | else { |
3077 | const APInt *Src = ((Sel & 10) == 10 || (Sel & 12) == 4) ? C0 : C1; |
3078 | if (!Src) |
3079 | ++NumUndefBytes; |
3080 | else if (Sel < 8) |
3081 | B = Src->extractBitsAsZExtValue(numBits: 8, bitPosition: (Sel & 3) * 8); |
3082 | else |
3083 | B = Src->extractBitsAsZExtValue(numBits: 1, bitPosition: (Sel & 1) ? 31 : 15) * 0xff; |
3084 | } |
3085 | |
3086 | Val.insertBits(SubBits: B, bitPosition: I, numBits: 8); |
3087 | } |
3088 | |
3089 | if (NumUndefBytes == 4) |
3090 | return UndefValue::get(T: Ty); |
3091 | |
3092 | return ConstantInt::get(Ty, V: Val); |
3093 | } |
3094 | |
3095 | static Constant *ConstantFoldScalarCall3(StringRef Name, |
3096 | Intrinsic::ID IntrinsicID, |
3097 | Type *Ty, |
3098 | ArrayRef<Constant *> Operands, |
3099 | const TargetLibraryInfo *TLI, |
3100 | const CallBase *Call) { |
3101 | assert(Operands.size() == 3 && "Wrong number of operands." ); |
3102 | |
3103 | if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) { |
3104 | if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) { |
3105 | if (const auto *Op3 = dyn_cast<ConstantFP>(Val: Operands[2])) { |
3106 | const APFloat &C1 = Op1->getValueAPF(); |
3107 | const APFloat &C2 = Op2->getValueAPF(); |
3108 | const APFloat &C3 = Op3->getValueAPF(); |
3109 | |
3110 | if (const auto *ConstrIntr = dyn_cast<ConstrainedFPIntrinsic>(Val: Call)) { |
3111 | RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr); |
3112 | APFloat Res = C1; |
3113 | APFloat::opStatus St; |
3114 | switch (IntrinsicID) { |
3115 | default: |
3116 | return nullptr; |
3117 | case Intrinsic::experimental_constrained_fma: |
3118 | case Intrinsic::experimental_constrained_fmuladd: |
3119 | St = Res.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM); |
3120 | break; |
3121 | } |
3122 | if (mayFoldConstrained( |
3123 | CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), St)) |
3124 | return ConstantFP::get(Context&: Ty->getContext(), V: Res); |
3125 | return nullptr; |
3126 | } |
3127 | |
3128 | switch (IntrinsicID) { |
3129 | default: break; |
3130 | case Intrinsic::amdgcn_fma_legacy: { |
3131 | // The legacy behaviour is that multiplying +/- 0.0 by anything, even |
3132 | // NaN or infinity, gives +0.0. |
3133 | if (C1.isZero() || C2.isZero()) { |
3134 | // It's tempting to just return C3 here, but that would give the |
3135 | // wrong result if C3 was -0.0. |
3136 | return ConstantFP::get(Context&: Ty->getContext(), V: APFloat(0.0f) + C3); |
3137 | } |
3138 | [[fallthrough]]; |
3139 | } |
3140 | case Intrinsic::fma: |
3141 | case Intrinsic::fmuladd: { |
3142 | APFloat V = C1; |
3143 | V.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM: APFloat::rmNearestTiesToEven); |
3144 | return ConstantFP::get(Context&: Ty->getContext(), V); |
3145 | } |
3146 | case Intrinsic::amdgcn_cubeid: |
3147 | case Intrinsic::amdgcn_cubema: |
3148 | case Intrinsic::amdgcn_cubesc: |
3149 | case Intrinsic::amdgcn_cubetc: { |
3150 | APFloat V = ConstantFoldAMDGCNCubeIntrinsic(IntrinsicID, S0: C1, S1: C2, S2: C3); |
3151 | return ConstantFP::get(Context&: Ty->getContext(), V); |
3152 | } |
3153 | } |
3154 | } |
3155 | } |
3156 | } |
3157 | |
3158 | if (IntrinsicID == Intrinsic::smul_fix || |
3159 | IntrinsicID == Intrinsic::smul_fix_sat) { |
3160 | // poison * C -> poison |
3161 | // C * poison -> poison |
3162 | if (isa<PoisonValue>(Val: Operands[0]) || isa<PoisonValue>(Val: Operands[1])) |
3163 | return PoisonValue::get(T: Ty); |
3164 | |
3165 | const APInt *C0, *C1; |
3166 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3167 | !getConstIntOrUndef(Op: Operands[1], C&: C1)) |
3168 | return nullptr; |
3169 | |
3170 | // undef * C -> 0 |
3171 | // C * undef -> 0 |
3172 | if (!C0 || !C1) |
3173 | return Constant::getNullValue(Ty); |
3174 | |
3175 | // This code performs rounding towards negative infinity in case the result |
3176 | // cannot be represented exactly for the given scale. Targets that do care |
3177 | // about rounding should use a target hook for specifying how rounding |
3178 | // should be done, and provide their own folding to be consistent with |
3179 | // rounding. This is the same approach as used by |
3180 | // DAGTypeLegalizer::ExpandIntRes_MULFIX. |
3181 | unsigned Scale = cast<ConstantInt>(Val: Operands[2])->getZExtValue(); |
3182 | unsigned Width = C0->getBitWidth(); |
3183 | assert(Scale < Width && "Illegal scale." ); |
3184 | unsigned ExtendedWidth = Width * 2; |
3185 | APInt Product = |
3186 | (C0->sext(width: ExtendedWidth) * C1->sext(width: ExtendedWidth)).ashr(ShiftAmt: Scale); |
3187 | if (IntrinsicID == Intrinsic::smul_fix_sat) { |
3188 | APInt Max = APInt::getSignedMaxValue(numBits: Width).sext(width: ExtendedWidth); |
3189 | APInt Min = APInt::getSignedMinValue(numBits: Width).sext(width: ExtendedWidth); |
3190 | Product = APIntOps::smin(A: Product, B: Max); |
3191 | Product = APIntOps::smax(A: Product, B: Min); |
3192 | } |
3193 | return ConstantInt::get(Context&: Ty->getContext(), V: Product.sextOrTrunc(width: Width)); |
3194 | } |
3195 | |
3196 | if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) { |
3197 | const APInt *C0, *C1, *C2; |
3198 | if (!getConstIntOrUndef(Op: Operands[0], C&: C0) || |
3199 | !getConstIntOrUndef(Op: Operands[1], C&: C1) || |
3200 | !getConstIntOrUndef(Op: Operands[2], C&: C2)) |
3201 | return nullptr; |
3202 | |
3203 | bool IsRight = IntrinsicID == Intrinsic::fshr; |
3204 | if (!C2) |
3205 | return Operands[IsRight ? 1 : 0]; |
3206 | if (!C0 && !C1) |
3207 | return UndefValue::get(T: Ty); |
3208 | |
3209 | // The shift amount is interpreted as modulo the bitwidth. If the shift |
3210 | // amount is effectively 0, avoid UB due to oversized inverse shift below. |
3211 | unsigned BitWidth = C2->getBitWidth(); |
3212 | unsigned ShAmt = C2->urem(RHS: BitWidth); |
3213 | if (!ShAmt) |
3214 | return Operands[IsRight ? 1 : 0]; |
3215 | |
3216 | // (C0 << ShlAmt) | (C1 >> LshrAmt) |
3217 | unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt; |
3218 | unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt; |
3219 | if (!C0) |
3220 | return ConstantInt::get(Ty, V: C1->lshr(shiftAmt: LshrAmt)); |
3221 | if (!C1) |
3222 | return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt)); |
3223 | return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt) | C1->lshr(shiftAmt: LshrAmt)); |
3224 | } |
3225 | |
3226 | if (IntrinsicID == Intrinsic::amdgcn_perm) |
3227 | return ConstantFoldAMDGCNPermIntrinsic(Operands, Ty); |
3228 | |
3229 | return nullptr; |
3230 | } |
3231 | |
3232 | static Constant *ConstantFoldScalarCall(StringRef Name, |
3233 | Intrinsic::ID IntrinsicID, |
3234 | Type *Ty, |
3235 | ArrayRef<Constant *> Operands, |
3236 | const TargetLibraryInfo *TLI, |
3237 | const CallBase *Call) { |
3238 | if (Operands.size() == 1) |
3239 | return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call); |
3240 | |
3241 | if (Operands.size() == 2) { |
3242 | if (Constant *FoldedLibCall = |
3243 | ConstantFoldLibCall2(Name, Ty, Operands, TLI)) { |
3244 | return FoldedLibCall; |
3245 | } |
3246 | return ConstantFoldIntrinsicCall2(IntrinsicID, Ty, Operands, Call); |
3247 | } |
3248 | |
3249 | if (Operands.size() == 3) |
3250 | return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call); |
3251 | |
3252 | return nullptr; |
3253 | } |
3254 | |
3255 | static Constant *ConstantFoldFixedVectorCall( |
3256 | StringRef Name, Intrinsic::ID IntrinsicID, FixedVectorType *FVTy, |
3257 | ArrayRef<Constant *> Operands, const DataLayout &DL, |
3258 | const TargetLibraryInfo *TLI, const CallBase *Call) { |
3259 | SmallVector<Constant *, 4> Result(FVTy->getNumElements()); |
3260 | SmallVector<Constant *, 4> Lane(Operands.size()); |
3261 | Type *Ty = FVTy->getElementType(); |
3262 | |
3263 | switch (IntrinsicID) { |
3264 | case Intrinsic::masked_load: { |
3265 | auto *SrcPtr = Operands[0]; |
3266 | auto *Mask = Operands[2]; |
3267 | auto *Passthru = Operands[3]; |
3268 | |
3269 | Constant *VecData = ConstantFoldLoadFromConstPtr(C: SrcPtr, Ty: FVTy, DL); |
3270 | |
3271 | SmallVector<Constant *, 32> NewElements; |
3272 | for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) { |
3273 | auto *MaskElt = Mask->getAggregateElement(Elt: I); |
3274 | if (!MaskElt) |
3275 | break; |
3276 | auto *PassthruElt = Passthru->getAggregateElement(Elt: I); |
3277 | auto *VecElt = VecData ? VecData->getAggregateElement(Elt: I) : nullptr; |
3278 | if (isa<UndefValue>(Val: MaskElt)) { |
3279 | if (PassthruElt) |
3280 | NewElements.push_back(Elt: PassthruElt); |
3281 | else if (VecElt) |
3282 | NewElements.push_back(Elt: VecElt); |
3283 | else |
3284 | return nullptr; |
3285 | } |
3286 | if (MaskElt->isNullValue()) { |
3287 | if (!PassthruElt) |
3288 | return nullptr; |
3289 | NewElements.push_back(Elt: PassthruElt); |
3290 | } else if (MaskElt->isOneValue()) { |
3291 | if (!VecElt) |
3292 | return nullptr; |
3293 | NewElements.push_back(Elt: VecElt); |
3294 | } else { |
3295 | return nullptr; |
3296 | } |
3297 | } |
3298 | if (NewElements.size() != FVTy->getNumElements()) |
3299 | return nullptr; |
3300 | return ConstantVector::get(V: NewElements); |
3301 | } |
3302 | case Intrinsic::arm_mve_vctp8: |
3303 | case Intrinsic::arm_mve_vctp16: |
3304 | case Intrinsic::arm_mve_vctp32: |
3305 | case Intrinsic::arm_mve_vctp64: { |
3306 | if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) { |
3307 | unsigned Lanes = FVTy->getNumElements(); |
3308 | uint64_t Limit = Op->getZExtValue(); |
3309 | |
3310 | SmallVector<Constant *, 16> NCs; |
3311 | for (unsigned i = 0; i < Lanes; i++) { |
3312 | if (i < Limit) |
3313 | NCs.push_back(Elt: ConstantInt::getTrue(Ty)); |
3314 | else |
3315 | NCs.push_back(Elt: ConstantInt::getFalse(Ty)); |
3316 | } |
3317 | return ConstantVector::get(V: NCs); |
3318 | } |
3319 | return nullptr; |
3320 | } |
3321 | case Intrinsic::get_active_lane_mask: { |
3322 | auto *Op0 = dyn_cast<ConstantInt>(Val: Operands[0]); |
3323 | auto *Op1 = dyn_cast<ConstantInt>(Val: Operands[1]); |
3324 | if (Op0 && Op1) { |
3325 | unsigned Lanes = FVTy->getNumElements(); |
3326 | uint64_t Base = Op0->getZExtValue(); |
3327 | uint64_t Limit = Op1->getZExtValue(); |
3328 | |
3329 | SmallVector<Constant *, 16> NCs; |
3330 | for (unsigned i = 0; i < Lanes; i++) { |
3331 | if (Base + i < Limit) |
3332 | NCs.push_back(Elt: ConstantInt::getTrue(Ty)); |
3333 | else |
3334 | NCs.push_back(Elt: ConstantInt::getFalse(Ty)); |
3335 | } |
3336 | return ConstantVector::get(V: NCs); |
3337 | } |
3338 | return nullptr; |
3339 | } |
3340 | default: |
3341 | break; |
3342 | } |
3343 | |
3344 | for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) { |
3345 | // Gather a column of constants. |
3346 | for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) { |
3347 | // Some intrinsics use a scalar type for certain arguments. |
3348 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: J)) { |
3349 | Lane[J] = Operands[J]; |
3350 | continue; |
3351 | } |
3352 | |
3353 | Constant *Agg = Operands[J]->getAggregateElement(Elt: I); |
3354 | if (!Agg) |
3355 | return nullptr; |
3356 | |
3357 | Lane[J] = Agg; |
3358 | } |
3359 | |
3360 | // Use the regular scalar folding to simplify this column. |
3361 | Constant *Folded = |
3362 | ConstantFoldScalarCall(Name, IntrinsicID, Ty, Operands: Lane, TLI, Call); |
3363 | if (!Folded) |
3364 | return nullptr; |
3365 | Result[I] = Folded; |
3366 | } |
3367 | |
3368 | return ConstantVector::get(V: Result); |
3369 | } |
3370 | |
3371 | static Constant *ConstantFoldScalableVectorCall( |
3372 | StringRef Name, Intrinsic::ID IntrinsicID, ScalableVectorType *SVTy, |
3373 | ArrayRef<Constant *> Operands, const DataLayout &DL, |
3374 | const TargetLibraryInfo *TLI, const CallBase *Call) { |
3375 | switch (IntrinsicID) { |
3376 | case Intrinsic::aarch64_sve_convert_from_svbool: { |
3377 | auto *Src = dyn_cast<Constant>(Val: Operands[0]); |
3378 | if (!Src || !Src->isNullValue()) |
3379 | break; |
3380 | |
3381 | return ConstantInt::getFalse(Ty: SVTy); |
3382 | } |
3383 | default: |
3384 | break; |
3385 | } |
3386 | return nullptr; |
3387 | } |
3388 | |
3389 | static std::pair<Constant *, Constant *> |
3390 | ConstantFoldScalarFrexpCall(Constant *Op, Type *IntTy) { |
3391 | if (isa<PoisonValue>(Val: Op)) |
3392 | return {Op, PoisonValue::get(T: IntTy)}; |
3393 | |
3394 | auto *ConstFP = dyn_cast<ConstantFP>(Val: Op); |
3395 | if (!ConstFP) |
3396 | return {}; |
3397 | |
3398 | const APFloat &U = ConstFP->getValueAPF(); |
3399 | int FrexpExp; |
3400 | APFloat FrexpMant = frexp(X: U, Exp&: FrexpExp, RM: APFloat::rmNearestTiesToEven); |
3401 | Constant *Result0 = ConstantFP::get(Ty: ConstFP->getType(), V: FrexpMant); |
3402 | |
3403 | // The exponent is an "unspecified value" for inf/nan. We use zero to avoid |
3404 | // using undef. |
3405 | Constant *Result1 = FrexpMant.isFinite() ? ConstantInt::get(Ty: IntTy, V: FrexpExp) |
3406 | : ConstantInt::getNullValue(Ty: IntTy); |
3407 | return {Result0, Result1}; |
3408 | } |
3409 | |
3410 | /// Handle intrinsics that return tuples, which may be tuples of vectors. |
3411 | static Constant * |
3412 | ConstantFoldStructCall(StringRef Name, Intrinsic::ID IntrinsicID, |
3413 | StructType *StTy, ArrayRef<Constant *> Operands, |
3414 | const DataLayout &DL, const TargetLibraryInfo *TLI, |
3415 | const CallBase *Call) { |
3416 | |
3417 | switch (IntrinsicID) { |
3418 | case Intrinsic::frexp: { |
3419 | Type *Ty0 = StTy->getContainedType(i: 0); |
3420 | Type *Ty1 = StTy->getContainedType(i: 1)->getScalarType(); |
3421 | |
3422 | if (auto *FVTy0 = dyn_cast<FixedVectorType>(Val: Ty0)) { |
3423 | SmallVector<Constant *, 4> Results0(FVTy0->getNumElements()); |
3424 | SmallVector<Constant *, 4> Results1(FVTy0->getNumElements()); |
3425 | |
3426 | for (unsigned I = 0, E = FVTy0->getNumElements(); I != E; ++I) { |
3427 | Constant *Lane = Operands[0]->getAggregateElement(Elt: I); |
3428 | std::tie(args&: Results0[I], args&: Results1[I]) = |
3429 | ConstantFoldScalarFrexpCall(Op: Lane, IntTy: Ty1); |
3430 | if (!Results0[I]) |
3431 | return nullptr; |
3432 | } |
3433 | |
3434 | return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: Results0), |
3435 | Vs: ConstantVector::get(V: Results1)); |
3436 | } |
3437 | |
3438 | auto [Result0, Result1] = ConstantFoldScalarFrexpCall(Op: Operands[0], IntTy: Ty1); |
3439 | if (!Result0) |
3440 | return nullptr; |
3441 | return ConstantStruct::get(T: StTy, Vs: Result0, Vs: Result1); |
3442 | } |
3443 | default: |
3444 | // TODO: Constant folding of vector intrinsics that fall through here does |
3445 | // not work (e.g. overflow intrinsics) |
3446 | return ConstantFoldScalarCall(Name, IntrinsicID, Ty: StTy, Operands, TLI, Call); |
3447 | } |
3448 | |
3449 | return nullptr; |
3450 | } |
3451 | |
3452 | } // end anonymous namespace |
3453 | |
3454 | Constant *llvm::ConstantFoldBinaryIntrinsic(Intrinsic::ID ID, Constant *LHS, |
3455 | Constant *RHS, Type *Ty, |
3456 | Instruction *FMFSource) { |
3457 | return ConstantFoldIntrinsicCall2(IntrinsicID: ID, Ty, Operands: {LHS, RHS}, |
3458 | Call: dyn_cast_if_present<CallBase>(Val: FMFSource)); |
3459 | } |
3460 | |
3461 | Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F, |
3462 | ArrayRef<Constant *> Operands, |
3463 | const TargetLibraryInfo *TLI, |
3464 | bool AllowNonDeterministic) { |
3465 | if (Call->isNoBuiltin()) |
3466 | return nullptr; |
3467 | if (!F->hasName()) |
3468 | return nullptr; |
3469 | |
3470 | // If this is not an intrinsic and not recognized as a library call, bail out. |
3471 | Intrinsic::ID IID = F->getIntrinsicID(); |
3472 | if (IID == Intrinsic::not_intrinsic) { |
3473 | if (!TLI) |
3474 | return nullptr; |
3475 | LibFunc LibF; |
3476 | if (!TLI->getLibFunc(FDecl: *F, F&: LibF)) |
3477 | return nullptr; |
3478 | } |
3479 | |
3480 | // Conservatively assume that floating-point libcalls may be |
3481 | // non-deterministic. |
3482 | Type *Ty = F->getReturnType(); |
3483 | if (!AllowNonDeterministic && Ty->isFPOrFPVectorTy()) |
3484 | return nullptr; |
3485 | |
3486 | StringRef Name = F->getName(); |
3487 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) |
3488 | return ConstantFoldFixedVectorCall( |
3489 | Name, IntrinsicID: IID, FVTy, Operands, DL: F->getDataLayout(), TLI, Call); |
3490 | |
3491 | if (auto *SVTy = dyn_cast<ScalableVectorType>(Val: Ty)) |
3492 | return ConstantFoldScalableVectorCall( |
3493 | Name, IntrinsicID: IID, SVTy, Operands, DL: F->getDataLayout(), TLI, Call); |
3494 | |
3495 | if (auto *StTy = dyn_cast<StructType>(Val: Ty)) |
3496 | return ConstantFoldStructCall(Name, IntrinsicID: IID, StTy, Operands, |
3497 | DL: F->getDataLayout(), TLI, Call); |
3498 | |
3499 | // TODO: If this is a library function, we already discovered that above, |
3500 | // so we should pass the LibFunc, not the name (and it might be better |
3501 | // still to separate intrinsic handling from libcalls). |
3502 | return ConstantFoldScalarCall(Name, IntrinsicID: IID, Ty, Operands, TLI, Call); |
3503 | } |
3504 | |
3505 | bool llvm::isMathLibCallNoop(const CallBase *Call, |
3506 | const TargetLibraryInfo *TLI) { |
3507 | // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap |
3508 | // (and to some extent ConstantFoldScalarCall). |
3509 | if (Call->isNoBuiltin() || Call->isStrictFP()) |
3510 | return false; |
3511 | Function *F = Call->getCalledFunction(); |
3512 | if (!F) |
3513 | return false; |
3514 | |
3515 | LibFunc Func; |
3516 | if (!TLI || !TLI->getLibFunc(FDecl: *F, F&: Func)) |
3517 | return false; |
3518 | |
3519 | if (Call->arg_size() == 1) { |
3520 | if (ConstantFP *OpC = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0))) { |
3521 | const APFloat &Op = OpC->getValueAPF(); |
3522 | switch (Func) { |
3523 | case LibFunc_logl: |
3524 | case LibFunc_log: |
3525 | case LibFunc_logf: |
3526 | case LibFunc_log2l: |
3527 | case LibFunc_log2: |
3528 | case LibFunc_log2f: |
3529 | case LibFunc_log10l: |
3530 | case LibFunc_log10: |
3531 | case LibFunc_log10f: |
3532 | return Op.isNaN() || (!Op.isZero() && !Op.isNegative()); |
3533 | |
3534 | case LibFunc_expl: |
3535 | case LibFunc_exp: |
3536 | case LibFunc_expf: |
3537 | // FIXME: These boundaries are slightly conservative. |
3538 | if (OpC->getType()->isDoubleTy()) |
3539 | return !(Op < APFloat(-745.0) || Op > APFloat(709.0)); |
3540 | if (OpC->getType()->isFloatTy()) |
3541 | return !(Op < APFloat(-103.0f) || Op > APFloat(88.0f)); |
3542 | break; |
3543 | |
3544 | case LibFunc_exp2l: |
3545 | case LibFunc_exp2: |
3546 | case LibFunc_exp2f: |
3547 | // FIXME: These boundaries are slightly conservative. |
3548 | if (OpC->getType()->isDoubleTy()) |
3549 | return !(Op < APFloat(-1074.0) || Op > APFloat(1023.0)); |
3550 | if (OpC->getType()->isFloatTy()) |
3551 | return !(Op < APFloat(-149.0f) || Op > APFloat(127.0f)); |
3552 | break; |
3553 | |
3554 | case LibFunc_sinl: |
3555 | case LibFunc_sin: |
3556 | case LibFunc_sinf: |
3557 | case LibFunc_cosl: |
3558 | case LibFunc_cos: |
3559 | case LibFunc_cosf: |
3560 | return !Op.isInfinity(); |
3561 | |
3562 | case LibFunc_tanl: |
3563 | case LibFunc_tan: |
3564 | case LibFunc_tanf: { |
3565 | // FIXME: Stop using the host math library. |
3566 | // FIXME: The computation isn't done in the right precision. |
3567 | Type *Ty = OpC->getType(); |
3568 | if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) |
3569 | return ConstantFoldFP(NativeFP: tan, V: OpC->getValueAPF(), Ty) != nullptr; |
3570 | break; |
3571 | } |
3572 | |
3573 | case LibFunc_atan: |
3574 | case LibFunc_atanf: |
3575 | case LibFunc_atanl: |
3576 | // Per POSIX, this MAY fail if Op is denormal. We choose not failing. |
3577 | return true; |
3578 | |
3579 | |
3580 | case LibFunc_asinl: |
3581 | case LibFunc_asin: |
3582 | case LibFunc_asinf: |
3583 | case LibFunc_acosl: |
3584 | case LibFunc_acos: |
3585 | case LibFunc_acosf: |
3586 | return !(Op < APFloat(Op.getSemantics(), "-1" ) || |
3587 | Op > APFloat(Op.getSemantics(), "1" )); |
3588 | |
3589 | case LibFunc_sinh: |
3590 | case LibFunc_cosh: |
3591 | case LibFunc_sinhf: |
3592 | case LibFunc_coshf: |
3593 | case LibFunc_sinhl: |
3594 | case LibFunc_coshl: |
3595 | // FIXME: These boundaries are slightly conservative. |
3596 | if (OpC->getType()->isDoubleTy()) |
3597 | return !(Op < APFloat(-710.0) || Op > APFloat(710.0)); |
3598 | if (OpC->getType()->isFloatTy()) |
3599 | return !(Op < APFloat(-89.0f) || Op > APFloat(89.0f)); |
3600 | break; |
3601 | |
3602 | case LibFunc_sqrtl: |
3603 | case LibFunc_sqrt: |
3604 | case LibFunc_sqrtf: |
3605 | return Op.isNaN() || Op.isZero() || !Op.isNegative(); |
3606 | |
3607 | // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p, |
3608 | // maybe others? |
3609 | default: |
3610 | break; |
3611 | } |
3612 | } |
3613 | } |
3614 | |
3615 | if (Call->arg_size() == 2) { |
3616 | ConstantFP *Op0C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0)); |
3617 | ConstantFP *Op1C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 1)); |
3618 | if (Op0C && Op1C) { |
3619 | const APFloat &Op0 = Op0C->getValueAPF(); |
3620 | const APFloat &Op1 = Op1C->getValueAPF(); |
3621 | |
3622 | switch (Func) { |
3623 | case LibFunc_powl: |
3624 | case LibFunc_pow: |
3625 | case LibFunc_powf: { |
3626 | // FIXME: Stop using the host math library. |
3627 | // FIXME: The computation isn't done in the right precision. |
3628 | Type *Ty = Op0C->getType(); |
3629 | if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) { |
3630 | if (Ty == Op1C->getType()) |
3631 | return ConstantFoldBinaryFP(NativeFP: pow, V: Op0, W: Op1, Ty) != nullptr; |
3632 | } |
3633 | break; |
3634 | } |
3635 | |
3636 | case LibFunc_fmodl: |
3637 | case LibFunc_fmod: |
3638 | case LibFunc_fmodf: |
3639 | case LibFunc_remainderl: |
3640 | case LibFunc_remainder: |
3641 | case LibFunc_remainderf: |
3642 | return Op0.isNaN() || Op1.isNaN() || |
3643 | (!Op0.isInfinity() && !Op1.isZero()); |
3644 | |
3645 | case LibFunc_atan2: |
3646 | case LibFunc_atan2f: |
3647 | case LibFunc_atan2l: |
3648 | // Although IEEE-754 says atan2(+/-0.0, +/-0.0) are well-defined, and |
3649 | // GLIBC and MSVC do not appear to raise an error on those, we |
3650 | // cannot rely on that behavior. POSIX and C11 say that a domain error |
3651 | // may occur, so allow for that possibility. |
3652 | return !Op0.isZero() || !Op1.isZero(); |
3653 | |
3654 | default: |
3655 | break; |
3656 | } |
3657 | } |
3658 | } |
3659 | |
3660 | return false; |
3661 | } |
3662 | |
3663 | void TargetFolder::anchor() {} |
3664 | |