1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines routines for folding instructions into constants.
10//
11// Also, to supplement the basic IR ConstantExpr simplifications,
12// this file defines some additional folding routines that can make use of
13// DataLayout information. These functions cannot go in IR due to library
14// dependency issues.
15//
16//===----------------------------------------------------------------------===//
17
18#include "llvm/Analysis/ConstantFolding.h"
19#include "llvm/ADT/APFloat.h"
20#include "llvm/ADT/APInt.h"
21#include "llvm/ADT/APSInt.h"
22#include "llvm/ADT/ArrayRef.h"
23#include "llvm/ADT/DenseMap.h"
24#include "llvm/ADT/STLExtras.h"
25#include "llvm/ADT/SmallVector.h"
26#include "llvm/ADT/StringRef.h"
27#include "llvm/Analysis/TargetFolder.h"
28#include "llvm/Analysis/TargetLibraryInfo.h"
29#include "llvm/Analysis/ValueTracking.h"
30#include "llvm/Analysis/VectorUtils.h"
31#include "llvm/Config/config.h"
32#include "llvm/IR/Constant.h"
33#include "llvm/IR/ConstantFold.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DataLayout.h"
36#include "llvm/IR/DerivedTypes.h"
37#include "llvm/IR/Function.h"
38#include "llvm/IR/GlobalValue.h"
39#include "llvm/IR/GlobalVariable.h"
40#include "llvm/IR/InstrTypes.h"
41#include "llvm/IR/Instruction.h"
42#include "llvm/IR/Instructions.h"
43#include "llvm/IR/IntrinsicInst.h"
44#include "llvm/IR/Intrinsics.h"
45#include "llvm/IR/IntrinsicsAArch64.h"
46#include "llvm/IR/IntrinsicsAMDGPU.h"
47#include "llvm/IR/IntrinsicsARM.h"
48#include "llvm/IR/IntrinsicsNVPTX.h"
49#include "llvm/IR/IntrinsicsWebAssembly.h"
50#include "llvm/IR/IntrinsicsX86.h"
51#include "llvm/IR/NVVMIntrinsicUtils.h"
52#include "llvm/IR/Operator.h"
53#include "llvm/IR/Type.h"
54#include "llvm/IR/Value.h"
55#include "llvm/Support/Casting.h"
56#include "llvm/Support/ErrorHandling.h"
57#include "llvm/Support/KnownBits.h"
58#include "llvm/Support/MathExtras.h"
59#include <cassert>
60#include <cerrno>
61#include <cfenv>
62#include <cmath>
63#include <cstdint>
64
65using namespace llvm;
66
67static cl::opt<bool> DisableFPCallFolding(
68 "disable-fp-call-folding",
69 cl::desc("Disable constant-folding of FP intrinsics and libcalls."),
70 cl::init(Val: false), cl::Hidden);
71
72namespace {
73
74//===----------------------------------------------------------------------===//
75// Constant Folding internal helper functions
76//===----------------------------------------------------------------------===//
77
78static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
79 Constant *C, Type *SrcEltTy,
80 unsigned NumSrcElts,
81 const DataLayout &DL) {
82 // Now that we know that the input value is a vector of integers, just shift
83 // and insert them into our result.
84 unsigned BitShift = DL.getTypeSizeInBits(Ty: SrcEltTy);
85 for (unsigned i = 0; i != NumSrcElts; ++i) {
86 Constant *Element;
87 if (DL.isLittleEndian())
88 Element = C->getAggregateElement(Elt: NumSrcElts - i - 1);
89 else
90 Element = C->getAggregateElement(Elt: i);
91
92 if (isa_and_nonnull<UndefValue>(Val: Element)) {
93 Result <<= BitShift;
94 continue;
95 }
96
97 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Val: Element);
98 if (!ElementCI)
99 return ConstantExpr::getBitCast(C, Ty: DestTy);
100
101 Result <<= BitShift;
102 Result |= ElementCI->getValue().zext(width: Result.getBitWidth());
103 }
104
105 return nullptr;
106}
107
108/// Constant fold bitcast, symbolically evaluating it with DataLayout.
109/// This always returns a non-null constant, but it may be a
110/// ConstantExpr if unfoldable.
111Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
112 assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) &&
113 "Invalid constantexpr bitcast!");
114
115 // Catch the obvious splat cases.
116 if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL))
117 return Res;
118
119 if (auto *VTy = dyn_cast<VectorType>(Val: C->getType())) {
120 // Handle a vector->scalar integer/fp cast.
121 if (isa<IntegerType>(Val: DestTy) || DestTy->isFloatingPointTy()) {
122 unsigned NumSrcElts = cast<FixedVectorType>(Val: VTy)->getNumElements();
123 Type *SrcEltTy = VTy->getElementType();
124
125 // If the vector is a vector of floating point, convert it to vector of int
126 // to simplify things.
127 if (SrcEltTy->isFloatingPointTy()) {
128 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
129 auto *SrcIVTy = FixedVectorType::get(
130 ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElts);
131 // Ask IR to do the conversion now that #elts line up.
132 C = ConstantExpr::getBitCast(C, Ty: SrcIVTy);
133 }
134
135 APInt Result(DL.getTypeSizeInBits(Ty: DestTy), 0);
136 if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
137 SrcEltTy, NumSrcElts, DL))
138 return CE;
139
140 if (isa<IntegerType>(Val: DestTy))
141 return ConstantInt::get(Ty: DestTy, V: Result);
142
143 APFloat FP(DestTy->getFltSemantics(), Result);
144 return ConstantFP::get(Context&: DestTy->getContext(), V: FP);
145 }
146 }
147
148 // The code below only handles casts to vectors currently.
149 auto *DestVTy = dyn_cast<VectorType>(Val: DestTy);
150 if (!DestVTy)
151 return ConstantExpr::getBitCast(C, Ty: DestTy);
152
153 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
154 // vector so the code below can handle it uniformly.
155 if (!isa<VectorType>(Val: C->getType()) &&
156 (isa<ConstantFP>(Val: C) || isa<ConstantInt>(Val: C))) {
157 Constant *Ops = C; // don't take the address of C!
158 return FoldBitCast(C: ConstantVector::get(V: Ops), DestTy, DL);
159 }
160
161 // Some of what follows may extend to cover scalable vectors but the current
162 // implementation is fixed length specific.
163 if (!isa<FixedVectorType>(Val: C->getType()))
164 return ConstantExpr::getBitCast(C, Ty: DestTy);
165
166 // If this is a bitcast from constant vector -> vector, fold it.
167 if (!isa<ConstantDataVector>(Val: C) && !isa<ConstantVector>(Val: C) &&
168 !isa<ConstantInt>(Val: C) && !isa<ConstantFP>(Val: C))
169 return ConstantExpr::getBitCast(C, Ty: DestTy);
170
171 // If the element types match, IR can fold it.
172 unsigned NumDstElt = cast<FixedVectorType>(Val: DestVTy)->getNumElements();
173 unsigned NumSrcElt = cast<FixedVectorType>(Val: C->getType())->getNumElements();
174 if (NumDstElt == NumSrcElt)
175 return ConstantExpr::getBitCast(C, Ty: DestTy);
176
177 Type *SrcEltTy = cast<VectorType>(Val: C->getType())->getElementType();
178 Type *DstEltTy = DestVTy->getElementType();
179
180 // Otherwise, we're changing the number of elements in a vector, which
181 // requires endianness information to do the right thing. For example,
182 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
183 // folds to (little endian):
184 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
185 // and to (big endian):
186 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
187
188 // First thing is first. We only want to think about integer here, so if
189 // we have something in FP form, recast it as integer.
190 if (DstEltTy->isFloatingPointTy()) {
191 // Fold to an vector of integers with same size as our FP type.
192 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
193 auto *DestIVTy = FixedVectorType::get(
194 ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumDstElt);
195 // Recursively handle this integer conversion, if possible.
196 C = FoldBitCast(C, DestTy: DestIVTy, DL);
197
198 // Finally, IR can handle this now that #elts line up.
199 return ConstantExpr::getBitCast(C, Ty: DestTy);
200 }
201
202 // Okay, we know the destination is integer, if the input is FP, convert
203 // it to integer first.
204 if (SrcEltTy->isFloatingPointTy()) {
205 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
206 auto *SrcIVTy = FixedVectorType::get(
207 ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElt);
208 // Ask IR to do the conversion now that #elts line up.
209 C = ConstantExpr::getBitCast(C, Ty: SrcIVTy);
210 assert((isa<ConstantVector>(C) || // FIXME: Remove ConstantVector.
211 isa<ConstantDataVector>(C) || isa<ConstantInt>(C)) &&
212 "Constant folding cannot fail for plain fp->int bitcast!");
213 }
214
215 // Now we know that the input and output vectors are both integer vectors
216 // of the same size, and that their #elements is not the same. Do the
217 // conversion here, which depends on whether the input or output has
218 // more elements.
219 bool isLittleEndian = DL.isLittleEndian();
220
221 SmallVector<Constant*, 32> Result;
222 if (NumDstElt < NumSrcElt) {
223 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
224 Constant *Zero = Constant::getNullValue(Ty: DstEltTy);
225 unsigned Ratio = NumSrcElt/NumDstElt;
226 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
227 unsigned SrcElt = 0;
228 for (unsigned i = 0; i != NumDstElt; ++i) {
229 // Build each element of the result.
230 Constant *Elt = Zero;
231 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
232 for (unsigned j = 0; j != Ratio; ++j) {
233 Constant *Src = C->getAggregateElement(Elt: SrcElt++);
234 if (isa_and_nonnull<UndefValue>(Val: Src))
235 Src = Constant::getNullValue(
236 Ty: cast<VectorType>(Val: C->getType())->getElementType());
237 else
238 Src = dyn_cast_or_null<ConstantInt>(Val: Src);
239 if (!Src) // Reject constantexpr elements.
240 return ConstantExpr::getBitCast(C, Ty: DestTy);
241
242 // Zero extend the element to the right size.
243 Src = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: Src, DestTy: Elt->getType(),
244 DL);
245 assert(Src && "Constant folding cannot fail on plain integers");
246
247 // Shift it to the right place, depending on endianness.
248 Src = ConstantFoldBinaryOpOperands(
249 Opcode: Instruction::Shl, LHS: Src, RHS: ConstantInt::get(Ty: Src->getType(), V: ShiftAmt),
250 DL);
251 assert(Src && "Constant folding cannot fail on plain integers");
252
253 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
254
255 // Mix it in.
256 Elt = ConstantFoldBinaryOpOperands(Opcode: Instruction::Or, LHS: Elt, RHS: Src, DL);
257 assert(Elt && "Constant folding cannot fail on plain integers");
258 }
259 Result.push_back(Elt);
260 }
261 return ConstantVector::get(V: Result);
262 }
263
264 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
265 unsigned Ratio = NumDstElt/NumSrcElt;
266 unsigned DstBitSize = DL.getTypeSizeInBits(Ty: DstEltTy);
267
268 // Loop over each source value, expanding into multiple results.
269 for (unsigned i = 0; i != NumSrcElt; ++i) {
270 auto *Element = C->getAggregateElement(Elt: i);
271
272 if (!Element) // Reject constantexpr elements.
273 return ConstantExpr::getBitCast(C, Ty: DestTy);
274
275 if (isa<UndefValue>(Val: Element)) {
276 // Correctly Propagate undef values.
277 Result.append(NumInputs: Ratio, Elt: UndefValue::get(T: DstEltTy));
278 continue;
279 }
280
281 auto *Src = dyn_cast<ConstantInt>(Val: Element);
282 if (!Src)
283 return ConstantExpr::getBitCast(C, Ty: DestTy);
284
285 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
286 for (unsigned j = 0; j != Ratio; ++j) {
287 // Shift the piece of the value into the right place, depending on
288 // endianness.
289 APInt Elt = Src->getValue().lshr(shiftAmt: ShiftAmt);
290 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
291
292 // Truncate and remember this piece.
293 Result.push_back(Elt: ConstantInt::get(Ty: DstEltTy, V: Elt.trunc(width: DstBitSize)));
294 }
295 }
296
297 return ConstantVector::get(V: Result);
298}
299
300} // end anonymous namespace
301
302/// If this constant is a constant offset from a global, return the global and
303/// the constant. Because of constantexprs, this function is recursive.
304bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
305 APInt &Offset, const DataLayout &DL,
306 DSOLocalEquivalent **DSOEquiv) {
307 if (DSOEquiv)
308 *DSOEquiv = nullptr;
309
310 // Trivial case, constant is the global.
311 if ((GV = dyn_cast<GlobalValue>(Val: C))) {
312 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType());
313 Offset = APInt(BitWidth, 0);
314 return true;
315 }
316
317 if (auto *FoundDSOEquiv = dyn_cast<DSOLocalEquivalent>(Val: C)) {
318 if (DSOEquiv)
319 *DSOEquiv = FoundDSOEquiv;
320 GV = FoundDSOEquiv->getGlobalValue();
321 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType());
322 Offset = APInt(BitWidth, 0);
323 return true;
324 }
325
326 // Otherwise, if this isn't a constant expr, bail out.
327 auto *CE = dyn_cast<ConstantExpr>(Val: C);
328 if (!CE) return false;
329
330 // Look through ptr->int and ptr->ptr casts.
331 if (CE->getOpcode() == Instruction::PtrToInt ||
332 CE->getOpcode() == Instruction::PtrToAddr)
333 return IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset, DL,
334 DSOEquiv);
335
336 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
337 auto *GEP = dyn_cast<GEPOperator>(Val: CE);
338 if (!GEP)
339 return false;
340
341 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType());
342 APInt TmpOffset(BitWidth, 0);
343
344 // If the base isn't a global+constant, we aren't either.
345 if (!IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset&: TmpOffset, DL,
346 DSOEquiv))
347 return false;
348
349 // Otherwise, add any offset that our operands provide.
350 if (!GEP->accumulateConstantOffset(DL, Offset&: TmpOffset))
351 return false;
352
353 Offset = TmpOffset;
354 return true;
355}
356
357Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
358 const DataLayout &DL) {
359 do {
360 Type *SrcTy = C->getType();
361 if (SrcTy == DestTy)
362 return C;
363
364 TypeSize DestSize = DL.getTypeSizeInBits(Ty: DestTy);
365 TypeSize SrcSize = DL.getTypeSizeInBits(Ty: SrcTy);
366 if (!TypeSize::isKnownGE(LHS: SrcSize, RHS: DestSize))
367 return nullptr;
368
369 // Catch the obvious splat cases (since all-zeros can coerce non-integral
370 // pointers legally).
371 if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL))
372 return Res;
373
374 // If the type sizes are the same and a cast is legal, just directly
375 // cast the constant.
376 // But be careful not to coerce non-integral pointers illegally.
377 if (SrcSize == DestSize &&
378 DL.isNonIntegralPointerType(Ty: SrcTy->getScalarType()) ==
379 DL.isNonIntegralPointerType(Ty: DestTy->getScalarType())) {
380 Instruction::CastOps Cast = Instruction::BitCast;
381 // If we are going from a pointer to int or vice versa, we spell the cast
382 // differently.
383 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
384 Cast = Instruction::IntToPtr;
385 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
386 Cast = Instruction::PtrToInt;
387
388 if (CastInst::castIsValid(op: Cast, S: C, DstTy: DestTy))
389 return ConstantFoldCastOperand(Opcode: Cast, C, DestTy, DL);
390 }
391
392 // If this isn't an aggregate type, there is nothing we can do to drill down
393 // and find a bitcastable constant.
394 if (!SrcTy->isAggregateType() && !SrcTy->isVectorTy())
395 return nullptr;
396
397 // We're simulating a load through a pointer that was bitcast to point to
398 // a different type, so we can try to walk down through the initial
399 // elements of an aggregate to see if some part of the aggregate is
400 // castable to implement the "load" semantic model.
401 if (SrcTy->isStructTy()) {
402 // Struct types might have leading zero-length elements like [0 x i32],
403 // which are certainly not what we are looking for, so skip them.
404 unsigned Elem = 0;
405 Constant *ElemC;
406 do {
407 ElemC = C->getAggregateElement(Elt: Elem++);
408 } while (ElemC && DL.getTypeSizeInBits(Ty: ElemC->getType()).isZero());
409 C = ElemC;
410 } else {
411 // For non-byte-sized vector elements, the first element is not
412 // necessarily located at the vector base address.
413 if (auto *VT = dyn_cast<VectorType>(Val: SrcTy))
414 if (!DL.typeSizeEqualsStoreSize(Ty: VT->getElementType()))
415 return nullptr;
416
417 C = C->getAggregateElement(Elt: 0u);
418 }
419 } while (C);
420
421 return nullptr;
422}
423
424namespace {
425
426/// Recursive helper to read bits out of global. C is the constant being copied
427/// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
428/// results into and BytesLeft is the number of bytes left in
429/// the CurPtr buffer. DL is the DataLayout.
430bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
431 unsigned BytesLeft, const DataLayout &DL) {
432 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
433 "Out of range access");
434
435 // Reading type padding, return zero.
436 if (ByteOffset >= DL.getTypeStoreSize(Ty: C->getType()))
437 return true;
438
439 // If this element is zero or undefined, we can just return since *CurPtr is
440 // zero initialized.
441 if (isa<ConstantAggregateZero>(Val: C) || isa<UndefValue>(Val: C))
442 return true;
443
444 auto *CI = dyn_cast<ConstantInt>(Val: C);
445 if (CI && CI->getType()->isIntegerTy()) {
446 if ((CI->getBitWidth() & 7) != 0)
447 return false;
448 const APInt &Val = CI->getValue();
449 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
450
451 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
452 unsigned n = ByteOffset;
453 if (!DL.isLittleEndian())
454 n = IntBytes - n - 1;
455 CurPtr[i] = Val.extractBits(numBits: 8, bitPosition: n * 8).getZExtValue();
456 ++ByteOffset;
457 }
458 return true;
459 }
460
461 auto *CFP = dyn_cast<ConstantFP>(Val: C);
462 if (CFP && CFP->getType()->isFloatingPointTy()) {
463 if (CFP->getType()->isDoubleTy()) {
464 C = FoldBitCast(C, DestTy: Type::getInt64Ty(C&: C->getContext()), DL);
465 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
466 }
467 if (CFP->getType()->isFloatTy()){
468 C = FoldBitCast(C, DestTy: Type::getInt32Ty(C&: C->getContext()), DL);
469 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
470 }
471 if (CFP->getType()->isHalfTy()){
472 C = FoldBitCast(C, DestTy: Type::getInt16Ty(C&: C->getContext()), DL);
473 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
474 }
475 return false;
476 }
477
478 if (auto *CS = dyn_cast<ConstantStruct>(Val: C)) {
479 const StructLayout *SL = DL.getStructLayout(Ty: CS->getType());
480 unsigned Index = SL->getElementContainingOffset(FixedOffset: ByteOffset);
481 uint64_t CurEltOffset = SL->getElementOffset(Idx: Index);
482 ByteOffset -= CurEltOffset;
483
484 while (true) {
485 // If the element access is to the element itself and not to tail padding,
486 // read the bytes from the element.
487 uint64_t EltSize = DL.getTypeAllocSize(Ty: CS->getOperand(i_nocapture: Index)->getType());
488
489 if (ByteOffset < EltSize &&
490 !ReadDataFromGlobal(C: CS->getOperand(i_nocapture: Index), ByteOffset, CurPtr,
491 BytesLeft, DL))
492 return false;
493
494 ++Index;
495
496 // Check to see if we read from the last struct element, if so we're done.
497 if (Index == CS->getType()->getNumElements())
498 return true;
499
500 // If we read all of the bytes we needed from this element we're done.
501 uint64_t NextEltOffset = SL->getElementOffset(Idx: Index);
502
503 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
504 return true;
505
506 // Move to the next element of the struct.
507 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
508 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
509 ByteOffset = 0;
510 CurEltOffset = NextEltOffset;
511 }
512 // not reached.
513 }
514
515 if (isa<ConstantArray>(Val: C) || isa<ConstantVector>(Val: C) ||
516 isa<ConstantDataSequential>(Val: C) || isa<ConstantInt>(Val: C) ||
517 isa<ConstantFP>(Val: C)) {
518 uint64_t NumElts, EltSize;
519 Type *EltTy;
520 if (auto *AT = dyn_cast<ArrayType>(Val: C->getType())) {
521 NumElts = AT->getNumElements();
522 EltTy = AT->getElementType();
523 EltSize = DL.getTypeAllocSize(Ty: EltTy);
524 } else {
525 NumElts = cast<FixedVectorType>(Val: C->getType())->getNumElements();
526 EltTy = cast<FixedVectorType>(Val: C->getType())->getElementType();
527 // TODO: For non-byte-sized vectors, current implementation assumes there is
528 // padding to the next byte boundary between elements.
529 if (!DL.typeSizeEqualsStoreSize(Ty: EltTy))
530 return false;
531
532 EltSize = DL.getTypeStoreSize(Ty: EltTy);
533 }
534 uint64_t Index = ByteOffset / EltSize;
535 uint64_t Offset = ByteOffset - Index * EltSize;
536
537 for (; Index != NumElts; ++Index) {
538 if (!ReadDataFromGlobal(C: C->getAggregateElement(Elt: Index), ByteOffset: Offset, CurPtr,
539 BytesLeft, DL))
540 return false;
541
542 uint64_t BytesWritten = EltSize - Offset;
543 assert(BytesWritten <= EltSize && "Not indexing into this element?");
544 if (BytesWritten >= BytesLeft)
545 return true;
546
547 Offset = 0;
548 BytesLeft -= BytesWritten;
549 CurPtr += BytesWritten;
550 }
551 return true;
552 }
553
554 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
555 if (CE->getOpcode() == Instruction::IntToPtr &&
556 CE->getOperand(i_nocapture: 0)->getType() == DL.getIntPtrType(CE->getType())) {
557 return ReadDataFromGlobal(C: CE->getOperand(i_nocapture: 0), ByteOffset, CurPtr,
558 BytesLeft, DL);
559 }
560 }
561
562 // Otherwise, unknown initializer type.
563 return false;
564}
565
566Constant *FoldReinterpretLoadFromConst(Constant *C, Type *LoadTy,
567 int64_t Offset, const DataLayout &DL) {
568 // Bail out early. Not expect to load from scalable global variable.
569 if (isa<ScalableVectorType>(Val: LoadTy))
570 return nullptr;
571
572 auto *IntType = dyn_cast<IntegerType>(Val: LoadTy);
573
574 // If this isn't an integer load we can't fold it directly.
575 if (!IntType) {
576 // If this is a non-integer load, we can try folding it as an int load and
577 // then bitcast the result. This can be useful for union cases. Note
578 // that address spaces don't matter here since we're not going to result in
579 // an actual new load.
580 if (!LoadTy->isFloatingPointTy() && !LoadTy->isPointerTy() &&
581 !LoadTy->isVectorTy())
582 return nullptr;
583
584 Type *MapTy = Type::getIntNTy(C&: C->getContext(),
585 N: DL.getTypeSizeInBits(Ty: LoadTy).getFixedValue());
586 if (Constant *Res = FoldReinterpretLoadFromConst(C, LoadTy: MapTy, Offset, DL)) {
587 if (Res->isNullValue() && !LoadTy->isX86_AMXTy())
588 // Materializing a zero can be done trivially without a bitcast
589 return Constant::getNullValue(Ty: LoadTy);
590 Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy;
591 Res = FoldBitCast(C: Res, DestTy: CastTy, DL);
592 if (LoadTy->isPtrOrPtrVectorTy()) {
593 // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr
594 if (Res->isNullValue() && !LoadTy->isX86_AMXTy())
595 return Constant::getNullValue(Ty: LoadTy);
596 if (DL.isNonIntegralPointerType(Ty: LoadTy->getScalarType()))
597 // Be careful not to replace a load of an addrspace value with an inttoptr here
598 return nullptr;
599 Res = ConstantExpr::getIntToPtr(C: Res, Ty: LoadTy);
600 }
601 return Res;
602 }
603 return nullptr;
604 }
605
606 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
607 if (BytesLoaded > 32 || BytesLoaded == 0)
608 return nullptr;
609
610 // If we're not accessing anything in this constant, the result is undefined.
611 if (Offset <= -1 * static_cast<int64_t>(BytesLoaded))
612 return PoisonValue::get(T: IntType);
613
614 // TODO: We should be able to support scalable types.
615 TypeSize InitializerSize = DL.getTypeAllocSize(Ty: C->getType());
616 if (InitializerSize.isScalable())
617 return nullptr;
618
619 // If we're not accessing anything in this constant, the result is undefined.
620 if (Offset >= (int64_t)InitializerSize.getFixedValue())
621 return PoisonValue::get(T: IntType);
622
623 unsigned char RawBytes[32] = {0};
624 unsigned char *CurPtr = RawBytes;
625 unsigned BytesLeft = BytesLoaded;
626
627 // If we're loading off the beginning of the global, some bytes may be valid.
628 if (Offset < 0) {
629 CurPtr += -Offset;
630 BytesLeft += Offset;
631 Offset = 0;
632 }
633
634 if (!ReadDataFromGlobal(C, ByteOffset: Offset, CurPtr, BytesLeft, DL))
635 return nullptr;
636
637 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
638 if (DL.isLittleEndian()) {
639 ResultVal = RawBytes[BytesLoaded - 1];
640 for (unsigned i = 1; i != BytesLoaded; ++i) {
641 ResultVal <<= 8;
642 ResultVal |= RawBytes[BytesLoaded - 1 - i];
643 }
644 } else {
645 ResultVal = RawBytes[0];
646 for (unsigned i = 1; i != BytesLoaded; ++i) {
647 ResultVal <<= 8;
648 ResultVal |= RawBytes[i];
649 }
650 }
651
652 return ConstantInt::get(Context&: IntType->getContext(), V: ResultVal);
653}
654
655} // anonymous namespace
656
657// If GV is a constant with an initializer read its representation starting
658// at Offset and return it as a constant array of unsigned char. Otherwise
659// return null.
660Constant *llvm::ReadByteArrayFromGlobal(const GlobalVariable *GV,
661 uint64_t Offset) {
662 if (!GV->isConstant() || !GV->hasDefinitiveInitializer())
663 return nullptr;
664
665 const DataLayout &DL = GV->getDataLayout();
666 Constant *Init = const_cast<Constant *>(GV->getInitializer());
667 TypeSize InitSize = DL.getTypeAllocSize(Ty: Init->getType());
668 if (InitSize < Offset)
669 return nullptr;
670
671 uint64_t NBytes = InitSize - Offset;
672 if (NBytes > UINT16_MAX)
673 // Bail for large initializers in excess of 64K to avoid allocating
674 // too much memory.
675 // Offset is assumed to be less than or equal than InitSize (this
676 // is enforced in ReadDataFromGlobal).
677 return nullptr;
678
679 SmallVector<unsigned char, 256> RawBytes(static_cast<size_t>(NBytes));
680 unsigned char *CurPtr = RawBytes.data();
681
682 if (!ReadDataFromGlobal(C: Init, ByteOffset: Offset, CurPtr, BytesLeft: NBytes, DL))
683 return nullptr;
684
685 return ConstantDataArray::get(Context&: GV->getContext(), Elts&: RawBytes);
686}
687
688/// If this Offset points exactly to the start of an aggregate element, return
689/// that element, otherwise return nullptr.
690Constant *getConstantAtOffset(Constant *Base, APInt Offset,
691 const DataLayout &DL) {
692 if (Offset.isZero())
693 return Base;
694
695 if (!isa<ConstantAggregate>(Val: Base) && !isa<ConstantDataSequential>(Val: Base))
696 return nullptr;
697
698 Type *ElemTy = Base->getType();
699 SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset);
700 if (!Offset.isZero() || !Indices[0].isZero())
701 return nullptr;
702
703 Constant *C = Base;
704 for (const APInt &Index : drop_begin(RangeOrContainer&: Indices)) {
705 if (Index.isNegative() || Index.getActiveBits() >= 32)
706 return nullptr;
707
708 C = C->getAggregateElement(Elt: Index.getZExtValue());
709 if (!C)
710 return nullptr;
711 }
712
713 return C;
714}
715
716Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
717 const APInt &Offset,
718 const DataLayout &DL) {
719 if (Constant *AtOffset = getConstantAtOffset(Base: C, Offset, DL))
720 if (Constant *Result = ConstantFoldLoadThroughBitcast(C: AtOffset, DestTy: Ty, DL))
721 return Result;
722
723 // Explicitly check for out-of-bounds access, so we return poison even if the
724 // constant is a uniform value.
725 TypeSize Size = DL.getTypeAllocSize(Ty: C->getType());
726 if (!Size.isScalable() && Offset.sge(RHS: Size.getFixedValue()))
727 return PoisonValue::get(T: Ty);
728
729 // Try an offset-independent fold of a uniform value.
730 if (Constant *Result = ConstantFoldLoadFromUniformValue(C, Ty, DL))
731 return Result;
732
733 // Try hard to fold loads from bitcasted strange and non-type-safe things.
734 if (Offset.getSignificantBits() <= 64)
735 if (Constant *Result =
736 FoldReinterpretLoadFromConst(C, LoadTy: Ty, Offset: Offset.getSExtValue(), DL))
737 return Result;
738
739 return nullptr;
740}
741
742Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
743 const DataLayout &DL) {
744 return ConstantFoldLoadFromConst(C, Ty, Offset: APInt(64, 0), DL);
745}
746
747Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
748 APInt Offset,
749 const DataLayout &DL) {
750 // We can only fold loads from constant globals with a definitive initializer.
751 // Check this upfront, to skip expensive offset calculations.
752 auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: C));
753 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
754 return nullptr;
755
756 C = cast<Constant>(Val: C->stripAndAccumulateConstantOffsets(
757 DL, Offset, /* AllowNonInbounds */ true));
758
759 if (C == GV)
760 if (Constant *Result = ConstantFoldLoadFromConst(C: GV->getInitializer(), Ty,
761 Offset, DL))
762 return Result;
763
764 // If this load comes from anywhere in a uniform constant global, the value
765 // is always the same, regardless of the loaded offset.
766 return ConstantFoldLoadFromUniformValue(C: GV->getInitializer(), Ty, DL);
767}
768
769Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
770 const DataLayout &DL) {
771 APInt Offset(DL.getIndexTypeSizeInBits(Ty: C->getType()), 0);
772 return ConstantFoldLoadFromConstPtr(C, Ty, Offset: std::move(Offset), DL);
773}
774
775Constant *llvm::ConstantFoldLoadFromUniformValue(Constant *C, Type *Ty,
776 const DataLayout &DL) {
777 if (isa<PoisonValue>(Val: C))
778 return PoisonValue::get(T: Ty);
779 if (isa<UndefValue>(Val: C))
780 return UndefValue::get(T: Ty);
781 // If padding is needed when storing C to memory, then it isn't considered as
782 // uniform.
783 if (!DL.typeSizeEqualsStoreSize(Ty: C->getType()))
784 return nullptr;
785 if (C->isNullValue() && !Ty->isX86_AMXTy())
786 return Constant::getNullValue(Ty);
787 if (C->isAllOnesValue() &&
788 (Ty->isIntOrIntVectorTy() || Ty->isFPOrFPVectorTy()))
789 return Constant::getAllOnesValue(Ty);
790 return nullptr;
791}
792
793namespace {
794
795/// One of Op0/Op1 is a constant expression.
796/// Attempt to symbolically evaluate the result of a binary operator merging
797/// these together. If target data info is available, it is provided as DL,
798/// otherwise DL is null.
799Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
800 const DataLayout &DL) {
801 // SROA
802
803 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
804 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
805 // bits.
806
807 if (Opc == Instruction::And) {
808 KnownBits Known0 = computeKnownBits(V: Op0, DL);
809 KnownBits Known1 = computeKnownBits(V: Op1, DL);
810 if ((Known1.One | Known0.Zero).isAllOnes()) {
811 // All the bits of Op0 that the 'and' could be masking are already zero.
812 return Op0;
813 }
814 if ((Known0.One | Known1.Zero).isAllOnes()) {
815 // All the bits of Op1 that the 'and' could be masking are already zero.
816 return Op1;
817 }
818
819 Known0 &= Known1;
820 if (Known0.isConstant())
821 return ConstantInt::get(Ty: Op0->getType(), V: Known0.getConstant());
822 }
823
824 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
825 // constant. This happens frequently when iterating over a global array.
826 if (Opc == Instruction::Sub) {
827 GlobalValue *GV1, *GV2;
828 APInt Offs1, Offs2;
829
830 if (IsConstantOffsetFromGlobal(C: Op0, GV&: GV1, Offset&: Offs1, DL))
831 if (IsConstantOffsetFromGlobal(C: Op1, GV&: GV2, Offset&: Offs2, DL) && GV1 == GV2) {
832 unsigned OpSize = DL.getTypeSizeInBits(Ty: Op0->getType());
833
834 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
835 // PtrToInt may change the bitwidth so we have convert to the right size
836 // first.
837 return ConstantInt::get(Ty: Op0->getType(), V: Offs1.zextOrTrunc(width: OpSize) -
838 Offs2.zextOrTrunc(width: OpSize));
839 }
840 }
841
842 return nullptr;
843}
844
845/// If array indices are not pointer-sized integers, explicitly cast them so
846/// that they aren't implicitly casted by the getelementptr.
847Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
848 Type *ResultTy, GEPNoWrapFlags NW,
849 std::optional<ConstantRange> InRange,
850 const DataLayout &DL, const TargetLibraryInfo *TLI) {
851 Type *IntIdxTy = DL.getIndexType(PtrTy: ResultTy);
852 Type *IntIdxScalarTy = IntIdxTy->getScalarType();
853
854 bool Any = false;
855 SmallVector<Constant*, 32> NewIdxs;
856 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
857 if ((i == 1 ||
858 !isa<StructType>(Val: GetElementPtrInst::getIndexedType(
859 Ty: SrcElemTy, IdxList: Ops.slice(N: 1, M: i - 1)))) &&
860 Ops[i]->getType()->getScalarType() != IntIdxScalarTy) {
861 Any = true;
862 Type *NewType =
863 Ops[i]->getType()->isVectorTy() ? IntIdxTy : IntIdxScalarTy;
864 Constant *NewIdx = ConstantFoldCastOperand(
865 Opcode: CastInst::getCastOpcode(Val: Ops[i], SrcIsSigned: true, Ty: NewType, DstIsSigned: true), C: Ops[i], DestTy: NewType,
866 DL);
867 if (!NewIdx)
868 return nullptr;
869 NewIdxs.push_back(Elt: NewIdx);
870 } else
871 NewIdxs.push_back(Elt: Ops[i]);
872 }
873
874 if (!Any)
875 return nullptr;
876
877 Constant *C =
878 ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: NewIdxs, NW, InRange);
879 return ConstantFoldConstant(C, DL, TLI);
880}
881
882/// If we can symbolically evaluate the GEP constant expression, do so.
883Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
884 ArrayRef<Constant *> Ops,
885 const DataLayout &DL,
886 const TargetLibraryInfo *TLI) {
887 Type *SrcElemTy = GEP->getSourceElementType();
888 Type *ResTy = GEP->getType();
889 if (!SrcElemTy->isSized() || isa<ScalableVectorType>(Val: SrcElemTy))
890 return nullptr;
891
892 if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResultTy: ResTy, NW: GEP->getNoWrapFlags(),
893 InRange: GEP->getInRange(), DL, TLI))
894 return C;
895
896 Constant *Ptr = Ops[0];
897 if (!Ptr->getType()->isPointerTy())
898 return nullptr;
899
900 Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType());
901
902 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
903 if (!isa<ConstantInt>(Val: Ops[i]) || !Ops[i]->getType()->isIntegerTy())
904 return nullptr;
905
906 unsigned BitWidth = DL.getTypeSizeInBits(Ty: IntIdxTy);
907 APInt Offset = APInt(
908 BitWidth,
909 DL.getIndexedOffsetInType(
910 ElemTy: SrcElemTy, Indices: ArrayRef((Value *const *)Ops.data() + 1, Ops.size() - 1)),
911 /*isSigned=*/true, /*implicitTrunc=*/true);
912
913 std::optional<ConstantRange> InRange = GEP->getInRange();
914 if (InRange)
915 InRange = InRange->sextOrTrunc(BitWidth);
916
917 // If this is a GEP of a GEP, fold it all into a single GEP.
918 GEPNoWrapFlags NW = GEP->getNoWrapFlags();
919 bool Overflow = false;
920 while (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr)) {
921 NW &= GEP->getNoWrapFlags();
922
923 SmallVector<Value *, 4> NestedOps(llvm::drop_begin(RangeOrContainer: GEP->operands()));
924
925 // Do not try the incorporate the sub-GEP if some index is not a number.
926 bool AllConstantInt = true;
927 for (Value *NestedOp : NestedOps)
928 if (!isa<ConstantInt>(Val: NestedOp)) {
929 AllConstantInt = false;
930 break;
931 }
932 if (!AllConstantInt)
933 break;
934
935 // Adjust inrange offset and intersect inrange attributes
936 if (auto GEPRange = GEP->getInRange()) {
937 auto AdjustedGEPRange = GEPRange->sextOrTrunc(BitWidth).subtract(CI: Offset);
938 InRange =
939 InRange ? InRange->intersectWith(CR: AdjustedGEPRange) : AdjustedGEPRange;
940 }
941
942 Ptr = cast<Constant>(Val: GEP->getOperand(i_nocapture: 0));
943 SrcElemTy = GEP->getSourceElementType();
944 Offset = Offset.sadd_ov(
945 RHS: APInt(BitWidth, DL.getIndexedOffsetInType(ElemTy: SrcElemTy, Indices: NestedOps),
946 /*isSigned=*/true, /*implicitTrunc=*/true),
947 Overflow);
948 }
949
950 // Preserving nusw (without inbounds) also requires that the offset
951 // additions did not overflow.
952 if (NW.hasNoUnsignedSignedWrap() && !NW.isInBounds() && Overflow)
953 NW = NW.withoutNoUnsignedSignedWrap();
954
955 // If the base value for this address is a literal integer value, fold the
956 // getelementptr to the resulting integer value casted to the pointer type.
957 APInt BaseIntVal(DL.getPointerTypeSizeInBits(Ptr->getType()), 0);
958 if (auto *CE = dyn_cast<ConstantExpr>(Val: Ptr)) {
959 if (CE->getOpcode() == Instruction::IntToPtr) {
960 if (auto *Base = dyn_cast<ConstantInt>(Val: CE->getOperand(i_nocapture: 0)))
961 BaseIntVal = Base->getValue().zextOrTrunc(width: BaseIntVal.getBitWidth());
962 }
963 }
964
965 if ((Ptr->isNullValue() || BaseIntVal != 0) &&
966 !DL.mustNotIntroduceIntToPtr(Ty: Ptr->getType())) {
967
968 // If the index size is smaller than the pointer size, add to the low
969 // bits only.
970 BaseIntVal.insertBits(SubBits: BaseIntVal.trunc(width: BitWidth) + Offset, bitPosition: 0);
971 Constant *C = ConstantInt::get(Context&: Ptr->getContext(), V: BaseIntVal);
972 return ConstantExpr::getIntToPtr(C, Ty: ResTy);
973 }
974
975 // Try to infer inbounds for GEPs of globals.
976 if (!NW.isInBounds() && Offset.isNonNegative()) {
977 bool CanBeNull, CanBeFreed;
978 uint64_t DerefBytes =
979 Ptr->getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed);
980 if (DerefBytes != 0 && !CanBeNull && Offset.sle(RHS: DerefBytes))
981 NW |= GEPNoWrapFlags::inBounds();
982 }
983
984 // nusw + nneg -> nuw
985 if (NW.hasNoUnsignedSignedWrap() && Offset.isNonNegative())
986 NW |= GEPNoWrapFlags::noUnsignedWrap();
987
988 // Otherwise canonicalize this to a single ptradd.
989 LLVMContext &Ctx = Ptr->getContext();
990 return ConstantExpr::getPtrAdd(Ptr, Offset: ConstantInt::get(Context&: Ctx, V: Offset), NW,
991 InRange);
992}
993
994/// Attempt to constant fold an instruction with the
995/// specified opcode and operands. If successful, the constant result is
996/// returned, if not, null is returned. Note that this function can fail when
997/// attempting to fold instructions like loads and stores, which have no
998/// constant expression form.
999Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
1000 ArrayRef<Constant *> Ops,
1001 const DataLayout &DL,
1002 const TargetLibraryInfo *TLI,
1003 bool AllowNonDeterministic) {
1004 Type *DestTy = InstOrCE->getType();
1005
1006 if (Instruction::isUnaryOp(Opcode))
1007 return ConstantFoldUnaryOpOperand(Opcode, Op: Ops[0], DL);
1008
1009 if (Instruction::isBinaryOp(Opcode)) {
1010 switch (Opcode) {
1011 default:
1012 break;
1013 case Instruction::FAdd:
1014 case Instruction::FSub:
1015 case Instruction::FMul:
1016 case Instruction::FDiv:
1017 case Instruction::FRem:
1018 // Handle floating point instructions separately to account for denormals
1019 // TODO: If a constant expression is being folded rather than an
1020 // instruction, denormals will not be flushed/treated as zero
1021 if (const auto *I = dyn_cast<Instruction>(Val: InstOrCE)) {
1022 return ConstantFoldFPInstOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL, I,
1023 AllowNonDeterministic);
1024 }
1025 }
1026 return ConstantFoldBinaryOpOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL);
1027 }
1028
1029 if (Instruction::isCast(Opcode))
1030 return ConstantFoldCastOperand(Opcode, C: Ops[0], DestTy, DL);
1031
1032 if (auto *GEP = dyn_cast<GEPOperator>(Val: InstOrCE)) {
1033 Type *SrcElemTy = GEP->getSourceElementType();
1034 if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy))
1035 return nullptr;
1036
1037 if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1038 return C;
1039
1040 return ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: Ops.slice(N: 1),
1041 NW: GEP->getNoWrapFlags(),
1042 InRange: GEP->getInRange());
1043 }
1044
1045 if (auto *CE = dyn_cast<ConstantExpr>(Val: InstOrCE))
1046 return CE->getWithOperands(Ops);
1047
1048 switch (Opcode) {
1049 default: return nullptr;
1050 case Instruction::ICmp:
1051 case Instruction::FCmp: {
1052 auto *C = cast<CmpInst>(Val: InstOrCE);
1053 return ConstantFoldCompareInstOperands(Predicate: C->getPredicate(), LHS: Ops[0], RHS: Ops[1],
1054 DL, TLI, I: C);
1055 }
1056 case Instruction::Freeze:
1057 return isGuaranteedNotToBeUndefOrPoison(V: Ops[0]) ? Ops[0] : nullptr;
1058 case Instruction::Call:
1059 if (auto *F = dyn_cast<Function>(Val: Ops.back())) {
1060 const auto *Call = cast<CallBase>(Val: InstOrCE);
1061 if (canConstantFoldCallTo(Call, F))
1062 return ConstantFoldCall(Call, F, Operands: Ops.slice(N: 0, M: Ops.size() - 1), TLI,
1063 AllowNonDeterministic);
1064 }
1065 return nullptr;
1066 case Instruction::Select:
1067 return ConstantFoldSelectInstruction(Cond: Ops[0], V1: Ops[1], V2: Ops[2]);
1068 case Instruction::ExtractElement:
1069 return ConstantExpr::getExtractElement(Vec: Ops[0], Idx: Ops[1]);
1070 case Instruction::ExtractValue:
1071 return ConstantFoldExtractValueInstruction(
1072 Agg: Ops[0], Idxs: cast<ExtractValueInst>(Val: InstOrCE)->getIndices());
1073 case Instruction::InsertElement:
1074 return ConstantExpr::getInsertElement(Vec: Ops[0], Elt: Ops[1], Idx: Ops[2]);
1075 case Instruction::InsertValue:
1076 return ConstantFoldInsertValueInstruction(
1077 Agg: Ops[0], Val: Ops[1], Idxs: cast<InsertValueInst>(Val: InstOrCE)->getIndices());
1078 case Instruction::ShuffleVector:
1079 return ConstantExpr::getShuffleVector(
1080 V1: Ops[0], V2: Ops[1], Mask: cast<ShuffleVectorInst>(Val: InstOrCE)->getShuffleMask());
1081 case Instruction::Load: {
1082 const auto *LI = dyn_cast<LoadInst>(Val: InstOrCE);
1083 if (LI->isVolatile())
1084 return nullptr;
1085 return ConstantFoldLoadFromConstPtr(C: Ops[0], Ty: LI->getType(), DL);
1086 }
1087 }
1088}
1089
1090} // end anonymous namespace
1091
1092//===----------------------------------------------------------------------===//
1093// Constant Folding public APIs
1094//===----------------------------------------------------------------------===//
1095
1096namespace {
1097
1098Constant *
1099ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1100 const TargetLibraryInfo *TLI,
1101 SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1102 if (!isa<ConstantVector>(Val: C) && !isa<ConstantExpr>(Val: C))
1103 return const_cast<Constant *>(C);
1104
1105 SmallVector<Constant *, 8> Ops;
1106 for (const Use &OldU : C->operands()) {
1107 Constant *OldC = cast<Constant>(Val: &OldU);
1108 Constant *NewC = OldC;
1109 // Recursively fold the ConstantExpr's operands. If we have already folded
1110 // a ConstantExpr, we don't have to process it again.
1111 if (isa<ConstantVector>(Val: OldC) || isa<ConstantExpr>(Val: OldC)) {
1112 auto It = FoldedOps.find(Val: OldC);
1113 if (It == FoldedOps.end()) {
1114 NewC = ConstantFoldConstantImpl(C: OldC, DL, TLI, FoldedOps);
1115 FoldedOps.insert(KV: {OldC, NewC});
1116 } else {
1117 NewC = It->second;
1118 }
1119 }
1120 Ops.push_back(Elt: NewC);
1121 }
1122
1123 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
1124 if (Constant *Res = ConstantFoldInstOperandsImpl(
1125 InstOrCE: CE, Opcode: CE->getOpcode(), Ops, DL, TLI, /*AllowNonDeterministic=*/true))
1126 return Res;
1127 return const_cast<Constant *>(C);
1128 }
1129
1130 assert(isa<ConstantVector>(C));
1131 return ConstantVector::get(V: Ops);
1132}
1133
1134} // end anonymous namespace
1135
1136Constant *llvm::ConstantFoldInstruction(const Instruction *I,
1137 const DataLayout &DL,
1138 const TargetLibraryInfo *TLI) {
1139 // Handle PHI nodes quickly here...
1140 if (auto *PN = dyn_cast<PHINode>(Val: I)) {
1141 Constant *CommonValue = nullptr;
1142
1143 SmallDenseMap<Constant *, Constant *> FoldedOps;
1144 for (Value *Incoming : PN->incoming_values()) {
1145 // If the incoming value is undef then skip it. Note that while we could
1146 // skip the value if it is equal to the phi node itself we choose not to
1147 // because that would break the rule that constant folding only applies if
1148 // all operands are constants.
1149 if (isa<UndefValue>(Val: Incoming))
1150 continue;
1151 // If the incoming value is not a constant, then give up.
1152 auto *C = dyn_cast<Constant>(Val: Incoming);
1153 if (!C)
1154 return nullptr;
1155 // Fold the PHI's operands.
1156 C = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1157 // If the incoming value is a different constant to
1158 // the one we saw previously, then give up.
1159 if (CommonValue && C != CommonValue)
1160 return nullptr;
1161 CommonValue = C;
1162 }
1163
1164 // If we reach here, all incoming values are the same constant or undef.
1165 return CommonValue ? CommonValue : UndefValue::get(T: PN->getType());
1166 }
1167
1168 // Scan the operand list, checking to see if they are all constants, if so,
1169 // hand off to ConstantFoldInstOperandsImpl.
1170 if (!all_of(Range: I->operands(), P: [](const Use &U) { return isa<Constant>(Val: U); }))
1171 return nullptr;
1172
1173 SmallDenseMap<Constant *, Constant *> FoldedOps;
1174 SmallVector<Constant *, 8> Ops;
1175 for (const Use &OpU : I->operands()) {
1176 auto *Op = cast<Constant>(Val: &OpU);
1177 // Fold the Instruction's operands.
1178 Op = ConstantFoldConstantImpl(C: Op, DL, TLI, FoldedOps);
1179 Ops.push_back(Elt: Op);
1180 }
1181
1182 return ConstantFoldInstOperands(I, Ops, DL, TLI);
1183}
1184
1185Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1186 const TargetLibraryInfo *TLI) {
1187 SmallDenseMap<Constant *, Constant *> FoldedOps;
1188 return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1189}
1190
1191Constant *llvm::ConstantFoldInstOperands(const Instruction *I,
1192 ArrayRef<Constant *> Ops,
1193 const DataLayout &DL,
1194 const TargetLibraryInfo *TLI,
1195 bool AllowNonDeterministic) {
1196 return ConstantFoldInstOperandsImpl(InstOrCE: I, Opcode: I->getOpcode(), Ops, DL, TLI,
1197 AllowNonDeterministic);
1198}
1199
1200Constant *llvm::ConstantFoldCompareInstOperands(
1201 unsigned IntPredicate, Constant *Ops0, Constant *Ops1, const DataLayout &DL,
1202 const TargetLibraryInfo *TLI, const Instruction *I) {
1203 CmpInst::Predicate Predicate = (CmpInst::Predicate)IntPredicate;
1204 // fold: icmp (inttoptr x), null -> icmp x, 0
1205 // fold: icmp null, (inttoptr x) -> icmp 0, x
1206 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1207 // fold: icmp 0, (ptrtoint x) -> icmp null, x
1208 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1209 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1210 //
1211 // FIXME: The following comment is out of data and the DataLayout is here now.
1212 // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1213 // around to know if bit truncation is happening.
1214 if (auto *CE0 = dyn_cast<ConstantExpr>(Val: Ops0)) {
1215 if (Ops1->isNullValue()) {
1216 if (CE0->getOpcode() == Instruction::IntToPtr) {
1217 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1218 // Convert the integer value to the right size to ensure we get the
1219 // proper extension or truncation.
1220 if (Constant *C = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy,
1221 /*IsSigned*/ false, DL)) {
1222 Constant *Null = Constant::getNullValue(Ty: C->getType());
1223 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI);
1224 }
1225 }
1226
1227 // icmp only compares the address part of the pointer, so only do this
1228 // transform if the integer size matches the address size.
1229 if (CE0->getOpcode() == Instruction::PtrToInt ||
1230 CE0->getOpcode() == Instruction::PtrToAddr) {
1231 Type *AddrTy = DL.getAddressType(PtrTy: CE0->getOperand(i_nocapture: 0)->getType());
1232 if (CE0->getType() == AddrTy) {
1233 Constant *C = CE0->getOperand(i_nocapture: 0);
1234 Constant *Null = Constant::getNullValue(Ty: C->getType());
1235 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI);
1236 }
1237 }
1238 }
1239
1240 if (auto *CE1 = dyn_cast<ConstantExpr>(Val: Ops1)) {
1241 if (CE0->getOpcode() == CE1->getOpcode()) {
1242 if (CE0->getOpcode() == Instruction::IntToPtr) {
1243 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1244
1245 // Convert the integer value to the right size to ensure we get the
1246 // proper extension or truncation.
1247 Constant *C0 = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy,
1248 /*IsSigned*/ false, DL);
1249 Constant *C1 = ConstantFoldIntegerCast(C: CE1->getOperand(i_nocapture: 0), DestTy: IntPtrTy,
1250 /*IsSigned*/ false, DL);
1251 if (C0 && C1)
1252 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C0, Ops1: C1, DL, TLI);
1253 }
1254
1255 // icmp only compares the address part of the pointer, so only do this
1256 // transform if the integer size matches the address size.
1257 if (CE0->getOpcode() == Instruction::PtrToInt ||
1258 CE0->getOpcode() == Instruction::PtrToAddr) {
1259 Type *AddrTy = DL.getAddressType(PtrTy: CE0->getOperand(i_nocapture: 0)->getType());
1260 if (CE0->getType() == AddrTy &&
1261 CE0->getOperand(i_nocapture: 0)->getType() == CE1->getOperand(i_nocapture: 0)->getType()) {
1262 return ConstantFoldCompareInstOperands(
1263 IntPredicate: Predicate, Ops0: CE0->getOperand(i_nocapture: 0), Ops1: CE1->getOperand(i_nocapture: 0), DL, TLI);
1264 }
1265 }
1266 }
1267 }
1268
1269 // Convert pointer comparison (base+offset1) pred (base+offset2) into
1270 // offset1 pred offset2, for the case where the offset is inbounds. This
1271 // only works for equality and unsigned comparison, as inbounds permits
1272 // crossing the sign boundary. However, the offset comparison itself is
1273 // signed.
1274 if (Ops0->getType()->isPointerTy() && !ICmpInst::isSigned(Pred: Predicate)) {
1275 unsigned IndexWidth = DL.getIndexTypeSizeInBits(Ty: Ops0->getType());
1276 APInt Offset0(IndexWidth, 0);
1277 bool IsEqPred = ICmpInst::isEquality(P: Predicate);
1278 Value *Stripped0 = Ops0->stripAndAccumulateConstantOffsets(
1279 DL, Offset&: Offset0, /*AllowNonInbounds=*/IsEqPred,
1280 /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr,
1281 /*LookThroughIntToPtr=*/IsEqPred);
1282 APInt Offset1(IndexWidth, 0);
1283 Value *Stripped1 = Ops1->stripAndAccumulateConstantOffsets(
1284 DL, Offset&: Offset1, /*AllowNonInbounds=*/IsEqPred,
1285 /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr,
1286 /*LookThroughIntToPtr=*/IsEqPred);
1287 if (Stripped0 == Stripped1)
1288 return ConstantInt::getBool(
1289 Context&: Ops0->getContext(),
1290 V: ICmpInst::compare(LHS: Offset0, RHS: Offset1,
1291 Pred: ICmpInst::getSignedPredicate(Pred: Predicate)));
1292 }
1293 } else if (isa<ConstantExpr>(Val: Ops1)) {
1294 // If RHS is a constant expression, but the left side isn't, swap the
1295 // operands and try again.
1296 Predicate = ICmpInst::getSwappedPredicate(pred: Predicate);
1297 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: Ops1, Ops1: Ops0, DL, TLI);
1298 }
1299
1300 if (CmpInst::isFPPredicate(P: Predicate)) {
1301 // Flush any denormal constant float input according to denormal handling
1302 // mode.
1303 Ops0 = FlushFPConstant(Operand: Ops0, I, /*IsOutput=*/false);
1304 if (!Ops0)
1305 return nullptr;
1306 Ops1 = FlushFPConstant(Operand: Ops1, I, /*IsOutput=*/false);
1307 if (!Ops1)
1308 return nullptr;
1309 }
1310
1311 return ConstantFoldCompareInstruction(Predicate, C1: Ops0, C2: Ops1);
1312}
1313
1314Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op,
1315 const DataLayout &DL) {
1316 assert(Instruction::isUnaryOp(Opcode));
1317
1318 return ConstantFoldUnaryInstruction(Opcode, V: Op);
1319}
1320
1321Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1322 Constant *RHS,
1323 const DataLayout &DL) {
1324 assert(Instruction::isBinaryOp(Opcode));
1325 if (isa<ConstantExpr>(Val: LHS) || isa<ConstantExpr>(Val: RHS))
1326 if (Constant *C = SymbolicallyEvaluateBinop(Opc: Opcode, Op0: LHS, Op1: RHS, DL))
1327 return C;
1328
1329 if (ConstantExpr::isDesirableBinOp(Opcode))
1330 return ConstantExpr::get(Opcode, C1: LHS, C2: RHS);
1331 return ConstantFoldBinaryInstruction(Opcode, V1: LHS, V2: RHS);
1332}
1333
1334static ConstantFP *flushDenormalConstant(Type *Ty, const APFloat &APF,
1335 DenormalMode::DenormalModeKind Mode) {
1336 switch (Mode) {
1337 case DenormalMode::Dynamic:
1338 return nullptr;
1339 case DenormalMode::IEEE:
1340 return ConstantFP::get(Context&: Ty->getContext(), V: APF);
1341 case DenormalMode::PreserveSign:
1342 return ConstantFP::get(
1343 Context&: Ty->getContext(),
1344 V: APFloat::getZero(Sem: APF.getSemantics(), Negative: APF.isNegative()));
1345 case DenormalMode::PositiveZero:
1346 return ConstantFP::get(Context&: Ty->getContext(),
1347 V: APFloat::getZero(Sem: APF.getSemantics(), Negative: false));
1348 default:
1349 break;
1350 }
1351
1352 llvm_unreachable("unknown denormal mode");
1353}
1354
1355/// Return the denormal mode that can be assumed when executing a floating point
1356/// operation at \p CtxI.
1357static DenormalMode getInstrDenormalMode(const Instruction *CtxI, Type *Ty) {
1358 if (!CtxI || !CtxI->getParent() || !CtxI->getFunction())
1359 return DenormalMode::getDynamic();
1360 return CtxI->getFunction()->getDenormalMode(FPType: Ty->getFltSemantics());
1361}
1362
1363static ConstantFP *flushDenormalConstantFP(ConstantFP *CFP,
1364 const Instruction *Inst,
1365 bool IsOutput) {
1366 const APFloat &APF = CFP->getValueAPF();
1367 if (!APF.isDenormal())
1368 return CFP;
1369
1370 DenormalMode Mode = getInstrDenormalMode(CtxI: Inst, Ty: CFP->getType());
1371 return flushDenormalConstant(Ty: CFP->getType(), APF,
1372 Mode: IsOutput ? Mode.Output : Mode.Input);
1373}
1374
1375Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *Inst,
1376 bool IsOutput) {
1377 if (ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Operand))
1378 return flushDenormalConstantFP(CFP, Inst, IsOutput);
1379
1380 if (isa<ConstantAggregateZero, UndefValue>(Val: Operand))
1381 return Operand;
1382
1383 Type *Ty = Operand->getType();
1384 VectorType *VecTy = dyn_cast<VectorType>(Val: Ty);
1385 if (VecTy) {
1386 if (auto *Splat = dyn_cast_or_null<ConstantFP>(Val: Operand->getSplatValue())) {
1387 ConstantFP *Folded = flushDenormalConstantFP(CFP: Splat, Inst, IsOutput);
1388 if (!Folded)
1389 return nullptr;
1390 return ConstantVector::getSplat(EC: VecTy->getElementCount(), Elt: Folded);
1391 }
1392
1393 Ty = VecTy->getElementType();
1394 }
1395
1396 if (isa<ConstantExpr>(Val: Operand))
1397 return Operand;
1398
1399 if (const auto *CV = dyn_cast<ConstantVector>(Val: Operand)) {
1400 SmallVector<Constant *, 16> NewElts;
1401 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1402 Constant *Element = CV->getAggregateElement(Elt: i);
1403 if (isa<UndefValue>(Val: Element)) {
1404 NewElts.push_back(Elt: Element);
1405 continue;
1406 }
1407
1408 ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Element);
1409 if (!CFP)
1410 return nullptr;
1411
1412 ConstantFP *Folded = flushDenormalConstantFP(CFP, Inst, IsOutput);
1413 if (!Folded)
1414 return nullptr;
1415 NewElts.push_back(Elt: Folded);
1416 }
1417
1418 return ConstantVector::get(V: NewElts);
1419 }
1420
1421 if (const auto *CDV = dyn_cast<ConstantDataVector>(Val: Operand)) {
1422 SmallVector<Constant *, 16> NewElts;
1423 for (unsigned I = 0, E = CDV->getNumElements(); I < E; ++I) {
1424 const APFloat &Elt = CDV->getElementAsAPFloat(i: I);
1425 if (!Elt.isDenormal()) {
1426 NewElts.push_back(Elt: ConstantFP::get(Ty, V: Elt));
1427 } else {
1428 DenormalMode Mode = getInstrDenormalMode(CtxI: Inst, Ty);
1429 ConstantFP *Folded =
1430 flushDenormalConstant(Ty, APF: Elt, Mode: IsOutput ? Mode.Output : Mode.Input);
1431 if (!Folded)
1432 return nullptr;
1433 NewElts.push_back(Elt: Folded);
1434 }
1435 }
1436
1437 return ConstantVector::get(V: NewElts);
1438 }
1439
1440 return nullptr;
1441}
1442
1443Constant *llvm::ConstantFoldFPInstOperands(unsigned Opcode, Constant *LHS,
1444 Constant *RHS, const DataLayout &DL,
1445 const Instruction *I,
1446 bool AllowNonDeterministic) {
1447 if (Instruction::isBinaryOp(Opcode)) {
1448 // Flush denormal inputs if needed.
1449 Constant *Op0 = FlushFPConstant(Operand: LHS, Inst: I, /* IsOutput */ false);
1450 if (!Op0)
1451 return nullptr;
1452 Constant *Op1 = FlushFPConstant(Operand: RHS, Inst: I, /* IsOutput */ false);
1453 if (!Op1)
1454 return nullptr;
1455
1456 // If nsz or an algebraic FMF flag is set, the result of the FP operation
1457 // may change due to future optimization. Don't constant fold them if
1458 // non-deterministic results are not allowed.
1459 if (!AllowNonDeterministic)
1460 if (auto *FP = dyn_cast_or_null<FPMathOperator>(Val: I))
1461 if (FP->hasNoSignedZeros() || FP->hasAllowReassoc() ||
1462 FP->hasAllowContract() || FP->hasAllowReciprocal())
1463 return nullptr;
1464
1465 // Calculate constant result.
1466 Constant *C = ConstantFoldBinaryOpOperands(Opcode, LHS: Op0, RHS: Op1, DL);
1467 if (!C)
1468 return nullptr;
1469
1470 // Flush denormal output if needed.
1471 C = FlushFPConstant(Operand: C, Inst: I, /* IsOutput */ true);
1472 if (!C)
1473 return nullptr;
1474
1475 // The precise NaN value is non-deterministic.
1476 if (!AllowNonDeterministic && C->isNaN())
1477 return nullptr;
1478
1479 return C;
1480 }
1481 // If instruction lacks a parent/function and the denormal mode cannot be
1482 // determined, use the default (IEEE).
1483 return ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL);
1484}
1485
1486Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1487 Type *DestTy, const DataLayout &DL) {
1488 assert(Instruction::isCast(Opcode));
1489
1490 if (auto *CE = dyn_cast<ConstantExpr>(Val: C))
1491 if (CE->isCast())
1492 if (unsigned NewOp = CastInst::isEliminableCastPair(
1493 firstOpcode: Instruction::CastOps(CE->getOpcode()),
1494 secondOpcode: Instruction::CastOps(Opcode), SrcTy: CE->getOperand(i_nocapture: 0)->getType(),
1495 MidTy: C->getType(), DstTy: DestTy, DL: &DL))
1496 return ConstantFoldCastOperand(Opcode: NewOp, C: CE->getOperand(i_nocapture: 0), DestTy, DL);
1497
1498 switch (Opcode) {
1499 default:
1500 llvm_unreachable("Missing case");
1501 case Instruction::PtrToAddr:
1502 case Instruction::PtrToInt:
1503 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
1504 Constant *FoldedValue = nullptr;
1505 // If the input is an inttoptr, eliminate the pair. This requires knowing
1506 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1507 if (CE->getOpcode() == Instruction::IntToPtr) {
1508 // zext/trunc the inttoptr to pointer/address size.
1509 Type *MidTy = Opcode == Instruction::PtrToInt
1510 ? DL.getAddressType(PtrTy: CE->getType())
1511 : DL.getIntPtrType(CE->getType());
1512 FoldedValue = ConstantFoldIntegerCast(C: CE->getOperand(i_nocapture: 0), DestTy: MidTy,
1513 /*IsSigned=*/false, DL);
1514 } else if (auto *GEP = dyn_cast<GEPOperator>(Val: CE)) {
1515 // If we have GEP, we can perform the following folds:
1516 // (ptrtoint/ptrtoaddr (gep null, x)) -> x
1517 // (ptrtoint/ptrtoaddr (gep (gep null, x), y) -> x + y, etc.
1518 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType());
1519 APInt BaseOffset(BitWidth, 0);
1520 auto *Base = cast<Constant>(Val: GEP->stripAndAccumulateConstantOffsets(
1521 DL, Offset&: BaseOffset, /*AllowNonInbounds=*/true));
1522 if (Base->isNullValue()) {
1523 FoldedValue = ConstantInt::get(Context&: CE->getContext(), V: BaseOffset);
1524 } else {
1525 // ptrtoint/ptrtoaddr (gep i8, Ptr, (sub 0, V))
1526 // -> sub (ptrtoint/ptrtoaddr Ptr), V
1527 if (GEP->getNumIndices() == 1 &&
1528 GEP->getSourceElementType()->isIntegerTy(Bitwidth: 8)) {
1529 auto *Ptr = cast<Constant>(Val: GEP->getPointerOperand());
1530 auto *Sub = dyn_cast<ConstantExpr>(Val: GEP->getOperand(i_nocapture: 1));
1531 Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType());
1532 if (Sub && Sub->getType() == IntIdxTy &&
1533 Sub->getOpcode() == Instruction::Sub &&
1534 Sub->getOperand(i_nocapture: 0)->isNullValue())
1535 FoldedValue = ConstantExpr::getSub(
1536 C1: ConstantExpr::getCast(ops: Opcode, C: Ptr, Ty: IntIdxTy),
1537 C2: Sub->getOperand(i_nocapture: 1));
1538 }
1539 }
1540 }
1541 if (FoldedValue) {
1542 // Do a zext or trunc to get to the ptrtoint/ptrtoaddr dest size.
1543 return ConstantFoldIntegerCast(C: FoldedValue, DestTy, /*IsSigned=*/false,
1544 DL);
1545 }
1546 }
1547 break;
1548 case Instruction::IntToPtr:
1549 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1550 // the int size is >= the ptr size and the address spaces are the same.
1551 // This requires knowing the width of a pointer, so it can't be done in
1552 // ConstantExpr::getCast.
1553 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
1554 if (CE->getOpcode() == Instruction::PtrToInt) {
1555 Constant *SrcPtr = CE->getOperand(i_nocapture: 0);
1556 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1557 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1558
1559 if (MidIntSize >= SrcPtrSize) {
1560 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1561 if (SrcAS == DestTy->getPointerAddressSpace())
1562 return FoldBitCast(C: CE->getOperand(i_nocapture: 0), DestTy, DL);
1563 }
1564 }
1565 }
1566 break;
1567 case Instruction::Trunc:
1568 case Instruction::ZExt:
1569 case Instruction::SExt:
1570 case Instruction::FPTrunc:
1571 case Instruction::FPExt:
1572 case Instruction::UIToFP:
1573 case Instruction::SIToFP:
1574 case Instruction::FPToUI:
1575 case Instruction::FPToSI:
1576 case Instruction::AddrSpaceCast:
1577 break;
1578 case Instruction::BitCast:
1579 return FoldBitCast(C, DestTy, DL);
1580 }
1581
1582 if (ConstantExpr::isDesirableCastOp(Opcode))
1583 return ConstantExpr::getCast(ops: Opcode, C, Ty: DestTy);
1584 return ConstantFoldCastInstruction(opcode: Opcode, V: C, DestTy);
1585}
1586
1587Constant *llvm::ConstantFoldIntegerCast(Constant *C, Type *DestTy,
1588 bool IsSigned, const DataLayout &DL) {
1589 Type *SrcTy = C->getType();
1590 if (SrcTy == DestTy)
1591 return C;
1592 if (SrcTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits())
1593 return ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy, DL);
1594 if (IsSigned)
1595 return ConstantFoldCastOperand(Opcode: Instruction::SExt, C, DestTy, DL);
1596 return ConstantFoldCastOperand(Opcode: Instruction::ZExt, C, DestTy, DL);
1597}
1598
1599//===----------------------------------------------------------------------===//
1600// Constant Folding for Calls
1601//
1602
1603bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
1604 if (Call->isNoBuiltin())
1605 return false;
1606 if (Call->getFunctionType() != F->getFunctionType())
1607 return false;
1608
1609 // Allow FP calls (both libcalls and intrinsics) to avoid being folded.
1610 // This can be useful for GPU targets or in cross-compilation scenarios
1611 // when the exact target FP behaviour is required, and the host compiler's
1612 // behaviour may be slightly different from the device's run-time behaviour.
1613 if (DisableFPCallFolding && (F->getReturnType()->isFloatingPointTy() ||
1614 any_of(Range: F->args(), P: [](const Argument &Arg) {
1615 return Arg.getType()->isFloatingPointTy();
1616 })))
1617 return false;
1618
1619 switch (F->getIntrinsicID()) {
1620 // Operations that do not operate floating-point numbers and do not depend on
1621 // FP environment can be folded even in strictfp functions.
1622 case Intrinsic::bswap:
1623 case Intrinsic::ctpop:
1624 case Intrinsic::ctlz:
1625 case Intrinsic::cttz:
1626 case Intrinsic::fshl:
1627 case Intrinsic::fshr:
1628 case Intrinsic::launder_invariant_group:
1629 case Intrinsic::strip_invariant_group:
1630 case Intrinsic::masked_load:
1631 case Intrinsic::get_active_lane_mask:
1632 case Intrinsic::abs:
1633 case Intrinsic::smax:
1634 case Intrinsic::smin:
1635 case Intrinsic::umax:
1636 case Intrinsic::umin:
1637 case Intrinsic::scmp:
1638 case Intrinsic::ucmp:
1639 case Intrinsic::sadd_with_overflow:
1640 case Intrinsic::uadd_with_overflow:
1641 case Intrinsic::ssub_with_overflow:
1642 case Intrinsic::usub_with_overflow:
1643 case Intrinsic::smul_with_overflow:
1644 case Intrinsic::umul_with_overflow:
1645 case Intrinsic::sadd_sat:
1646 case Intrinsic::uadd_sat:
1647 case Intrinsic::ssub_sat:
1648 case Intrinsic::usub_sat:
1649 case Intrinsic::smul_fix:
1650 case Intrinsic::smul_fix_sat:
1651 case Intrinsic::bitreverse:
1652 case Intrinsic::is_constant:
1653 case Intrinsic::vector_reduce_add:
1654 case Intrinsic::vector_reduce_mul:
1655 case Intrinsic::vector_reduce_and:
1656 case Intrinsic::vector_reduce_or:
1657 case Intrinsic::vector_reduce_xor:
1658 case Intrinsic::vector_reduce_smin:
1659 case Intrinsic::vector_reduce_smax:
1660 case Intrinsic::vector_reduce_umin:
1661 case Intrinsic::vector_reduce_umax:
1662 case Intrinsic::vector_extract:
1663 case Intrinsic::vector_insert:
1664 case Intrinsic::vector_interleave2:
1665 case Intrinsic::vector_interleave3:
1666 case Intrinsic::vector_interleave4:
1667 case Intrinsic::vector_interleave5:
1668 case Intrinsic::vector_interleave6:
1669 case Intrinsic::vector_interleave7:
1670 case Intrinsic::vector_interleave8:
1671 case Intrinsic::vector_deinterleave2:
1672 case Intrinsic::vector_deinterleave3:
1673 case Intrinsic::vector_deinterleave4:
1674 case Intrinsic::vector_deinterleave5:
1675 case Intrinsic::vector_deinterleave6:
1676 case Intrinsic::vector_deinterleave7:
1677 case Intrinsic::vector_deinterleave8:
1678 // Target intrinsics
1679 case Intrinsic::amdgcn_perm:
1680 case Intrinsic::amdgcn_wave_reduce_umin:
1681 case Intrinsic::amdgcn_wave_reduce_umax:
1682 case Intrinsic::amdgcn_wave_reduce_max:
1683 case Intrinsic::amdgcn_wave_reduce_min:
1684 case Intrinsic::amdgcn_wave_reduce_add:
1685 case Intrinsic::amdgcn_wave_reduce_sub:
1686 case Intrinsic::amdgcn_wave_reduce_and:
1687 case Intrinsic::amdgcn_wave_reduce_or:
1688 case Intrinsic::amdgcn_wave_reduce_xor:
1689 case Intrinsic::amdgcn_s_wqm:
1690 case Intrinsic::amdgcn_s_quadmask:
1691 case Intrinsic::amdgcn_s_bitreplicate:
1692 case Intrinsic::arm_mve_vctp8:
1693 case Intrinsic::arm_mve_vctp16:
1694 case Intrinsic::arm_mve_vctp32:
1695 case Intrinsic::arm_mve_vctp64:
1696 case Intrinsic::aarch64_sve_convert_from_svbool:
1697 case Intrinsic::wasm_alltrue:
1698 case Intrinsic::wasm_anytrue:
1699 case Intrinsic::wasm_dot:
1700 // WebAssembly float semantics are always known
1701 case Intrinsic::wasm_trunc_signed:
1702 case Intrinsic::wasm_trunc_unsigned:
1703 return true;
1704
1705 // Floating point operations cannot be folded in strictfp functions in
1706 // general case. They can be folded if FP environment is known to compiler.
1707 case Intrinsic::minnum:
1708 case Intrinsic::maxnum:
1709 case Intrinsic::minimum:
1710 case Intrinsic::maximum:
1711 case Intrinsic::minimumnum:
1712 case Intrinsic::maximumnum:
1713 case Intrinsic::log:
1714 case Intrinsic::log2:
1715 case Intrinsic::log10:
1716 case Intrinsic::exp:
1717 case Intrinsic::exp2:
1718 case Intrinsic::exp10:
1719 case Intrinsic::sqrt:
1720 case Intrinsic::sin:
1721 case Intrinsic::cos:
1722 case Intrinsic::sincos:
1723 case Intrinsic::sinh:
1724 case Intrinsic::cosh:
1725 case Intrinsic::atan:
1726 case Intrinsic::pow:
1727 case Intrinsic::powi:
1728 case Intrinsic::ldexp:
1729 case Intrinsic::fma:
1730 case Intrinsic::fmuladd:
1731 case Intrinsic::frexp:
1732 case Intrinsic::fptoui_sat:
1733 case Intrinsic::fptosi_sat:
1734 case Intrinsic::amdgcn_cos:
1735 case Intrinsic::amdgcn_cubeid:
1736 case Intrinsic::amdgcn_cubema:
1737 case Intrinsic::amdgcn_cubesc:
1738 case Intrinsic::amdgcn_cubetc:
1739 case Intrinsic::amdgcn_fmul_legacy:
1740 case Intrinsic::amdgcn_fma_legacy:
1741 case Intrinsic::amdgcn_fract:
1742 case Intrinsic::amdgcn_sin:
1743 // The intrinsics below depend on rounding mode in MXCSR.
1744 case Intrinsic::x86_sse_cvtss2si:
1745 case Intrinsic::x86_sse_cvtss2si64:
1746 case Intrinsic::x86_sse_cvttss2si:
1747 case Intrinsic::x86_sse_cvttss2si64:
1748 case Intrinsic::x86_sse2_cvtsd2si:
1749 case Intrinsic::x86_sse2_cvtsd2si64:
1750 case Intrinsic::x86_sse2_cvttsd2si:
1751 case Intrinsic::x86_sse2_cvttsd2si64:
1752 case Intrinsic::x86_avx512_vcvtss2si32:
1753 case Intrinsic::x86_avx512_vcvtss2si64:
1754 case Intrinsic::x86_avx512_cvttss2si:
1755 case Intrinsic::x86_avx512_cvttss2si64:
1756 case Intrinsic::x86_avx512_vcvtsd2si32:
1757 case Intrinsic::x86_avx512_vcvtsd2si64:
1758 case Intrinsic::x86_avx512_cvttsd2si:
1759 case Intrinsic::x86_avx512_cvttsd2si64:
1760 case Intrinsic::x86_avx512_vcvtss2usi32:
1761 case Intrinsic::x86_avx512_vcvtss2usi64:
1762 case Intrinsic::x86_avx512_cvttss2usi:
1763 case Intrinsic::x86_avx512_cvttss2usi64:
1764 case Intrinsic::x86_avx512_vcvtsd2usi32:
1765 case Intrinsic::x86_avx512_vcvtsd2usi64:
1766 case Intrinsic::x86_avx512_cvttsd2usi:
1767 case Intrinsic::x86_avx512_cvttsd2usi64:
1768
1769 // NVVM FMax intrinsics
1770 case Intrinsic::nvvm_fmax_d:
1771 case Intrinsic::nvvm_fmax_f:
1772 case Intrinsic::nvvm_fmax_ftz_f:
1773 case Intrinsic::nvvm_fmax_ftz_nan_f:
1774 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
1775 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
1776 case Intrinsic::nvvm_fmax_nan_f:
1777 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
1778 case Intrinsic::nvvm_fmax_xorsign_abs_f:
1779
1780 // NVVM FMin intrinsics
1781 case Intrinsic::nvvm_fmin_d:
1782 case Intrinsic::nvvm_fmin_f:
1783 case Intrinsic::nvvm_fmin_ftz_f:
1784 case Intrinsic::nvvm_fmin_ftz_nan_f:
1785 case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
1786 case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
1787 case Intrinsic::nvvm_fmin_nan_f:
1788 case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
1789 case Intrinsic::nvvm_fmin_xorsign_abs_f:
1790
1791 // NVVM float/double to int32/uint32 conversion intrinsics
1792 case Intrinsic::nvvm_f2i_rm:
1793 case Intrinsic::nvvm_f2i_rn:
1794 case Intrinsic::nvvm_f2i_rp:
1795 case Intrinsic::nvvm_f2i_rz:
1796 case Intrinsic::nvvm_f2i_rm_ftz:
1797 case Intrinsic::nvvm_f2i_rn_ftz:
1798 case Intrinsic::nvvm_f2i_rp_ftz:
1799 case Intrinsic::nvvm_f2i_rz_ftz:
1800 case Intrinsic::nvvm_f2ui_rm:
1801 case Intrinsic::nvvm_f2ui_rn:
1802 case Intrinsic::nvvm_f2ui_rp:
1803 case Intrinsic::nvvm_f2ui_rz:
1804 case Intrinsic::nvvm_f2ui_rm_ftz:
1805 case Intrinsic::nvvm_f2ui_rn_ftz:
1806 case Intrinsic::nvvm_f2ui_rp_ftz:
1807 case Intrinsic::nvvm_f2ui_rz_ftz:
1808 case Intrinsic::nvvm_d2i_rm:
1809 case Intrinsic::nvvm_d2i_rn:
1810 case Intrinsic::nvvm_d2i_rp:
1811 case Intrinsic::nvvm_d2i_rz:
1812 case Intrinsic::nvvm_d2ui_rm:
1813 case Intrinsic::nvvm_d2ui_rn:
1814 case Intrinsic::nvvm_d2ui_rp:
1815 case Intrinsic::nvvm_d2ui_rz:
1816
1817 // NVVM float/double to int64/uint64 conversion intrinsics
1818 case Intrinsic::nvvm_f2ll_rm:
1819 case Intrinsic::nvvm_f2ll_rn:
1820 case Intrinsic::nvvm_f2ll_rp:
1821 case Intrinsic::nvvm_f2ll_rz:
1822 case Intrinsic::nvvm_f2ll_rm_ftz:
1823 case Intrinsic::nvvm_f2ll_rn_ftz:
1824 case Intrinsic::nvvm_f2ll_rp_ftz:
1825 case Intrinsic::nvvm_f2ll_rz_ftz:
1826 case Intrinsic::nvvm_f2ull_rm:
1827 case Intrinsic::nvvm_f2ull_rn:
1828 case Intrinsic::nvvm_f2ull_rp:
1829 case Intrinsic::nvvm_f2ull_rz:
1830 case Intrinsic::nvvm_f2ull_rm_ftz:
1831 case Intrinsic::nvvm_f2ull_rn_ftz:
1832 case Intrinsic::nvvm_f2ull_rp_ftz:
1833 case Intrinsic::nvvm_f2ull_rz_ftz:
1834 case Intrinsic::nvvm_d2ll_rm:
1835 case Intrinsic::nvvm_d2ll_rn:
1836 case Intrinsic::nvvm_d2ll_rp:
1837 case Intrinsic::nvvm_d2ll_rz:
1838 case Intrinsic::nvvm_d2ull_rm:
1839 case Intrinsic::nvvm_d2ull_rn:
1840 case Intrinsic::nvvm_d2ull_rp:
1841 case Intrinsic::nvvm_d2ull_rz:
1842
1843 // NVVM math intrinsics:
1844 case Intrinsic::nvvm_ceil_d:
1845 case Intrinsic::nvvm_ceil_f:
1846 case Intrinsic::nvvm_ceil_ftz_f:
1847
1848 case Intrinsic::nvvm_fabs:
1849 case Intrinsic::nvvm_fabs_ftz:
1850
1851 case Intrinsic::nvvm_floor_d:
1852 case Intrinsic::nvvm_floor_f:
1853 case Intrinsic::nvvm_floor_ftz_f:
1854
1855 case Intrinsic::nvvm_rcp_rm_d:
1856 case Intrinsic::nvvm_rcp_rm_f:
1857 case Intrinsic::nvvm_rcp_rm_ftz_f:
1858 case Intrinsic::nvvm_rcp_rn_d:
1859 case Intrinsic::nvvm_rcp_rn_f:
1860 case Intrinsic::nvvm_rcp_rn_ftz_f:
1861 case Intrinsic::nvvm_rcp_rp_d:
1862 case Intrinsic::nvvm_rcp_rp_f:
1863 case Intrinsic::nvvm_rcp_rp_ftz_f:
1864 case Intrinsic::nvvm_rcp_rz_d:
1865 case Intrinsic::nvvm_rcp_rz_f:
1866 case Intrinsic::nvvm_rcp_rz_ftz_f:
1867
1868 case Intrinsic::nvvm_round_d:
1869 case Intrinsic::nvvm_round_f:
1870 case Intrinsic::nvvm_round_ftz_f:
1871
1872 case Intrinsic::nvvm_saturate_d:
1873 case Intrinsic::nvvm_saturate_f:
1874 case Intrinsic::nvvm_saturate_ftz_f:
1875
1876 case Intrinsic::nvvm_sqrt_f:
1877 case Intrinsic::nvvm_sqrt_rn_d:
1878 case Intrinsic::nvvm_sqrt_rn_f:
1879 case Intrinsic::nvvm_sqrt_rn_ftz_f:
1880 return !Call->isStrictFP();
1881
1882 // NVVM add intrinsics with explicit rounding modes
1883 case Intrinsic::nvvm_add_rm_d:
1884 case Intrinsic::nvvm_add_rn_d:
1885 case Intrinsic::nvvm_add_rp_d:
1886 case Intrinsic::nvvm_add_rz_d:
1887 case Intrinsic::nvvm_add_rm_f:
1888 case Intrinsic::nvvm_add_rn_f:
1889 case Intrinsic::nvvm_add_rp_f:
1890 case Intrinsic::nvvm_add_rz_f:
1891 case Intrinsic::nvvm_add_rm_ftz_f:
1892 case Intrinsic::nvvm_add_rn_ftz_f:
1893 case Intrinsic::nvvm_add_rp_ftz_f:
1894 case Intrinsic::nvvm_add_rz_ftz_f:
1895
1896 // NVVM div intrinsics with explicit rounding modes
1897 case Intrinsic::nvvm_div_rm_d:
1898 case Intrinsic::nvvm_div_rn_d:
1899 case Intrinsic::nvvm_div_rp_d:
1900 case Intrinsic::nvvm_div_rz_d:
1901 case Intrinsic::nvvm_div_rm_f:
1902 case Intrinsic::nvvm_div_rn_f:
1903 case Intrinsic::nvvm_div_rp_f:
1904 case Intrinsic::nvvm_div_rz_f:
1905 case Intrinsic::nvvm_div_rm_ftz_f:
1906 case Intrinsic::nvvm_div_rn_ftz_f:
1907 case Intrinsic::nvvm_div_rp_ftz_f:
1908 case Intrinsic::nvvm_div_rz_ftz_f:
1909
1910 // NVVM mul intrinsics with explicit rounding modes
1911 case Intrinsic::nvvm_mul_rm_d:
1912 case Intrinsic::nvvm_mul_rn_d:
1913 case Intrinsic::nvvm_mul_rp_d:
1914 case Intrinsic::nvvm_mul_rz_d:
1915 case Intrinsic::nvvm_mul_rm_f:
1916 case Intrinsic::nvvm_mul_rn_f:
1917 case Intrinsic::nvvm_mul_rp_f:
1918 case Intrinsic::nvvm_mul_rz_f:
1919 case Intrinsic::nvvm_mul_rm_ftz_f:
1920 case Intrinsic::nvvm_mul_rn_ftz_f:
1921 case Intrinsic::nvvm_mul_rp_ftz_f:
1922 case Intrinsic::nvvm_mul_rz_ftz_f:
1923
1924 // NVVM fma intrinsics with explicit rounding modes
1925 case Intrinsic::nvvm_fma_rm_d:
1926 case Intrinsic::nvvm_fma_rn_d:
1927 case Intrinsic::nvvm_fma_rp_d:
1928 case Intrinsic::nvvm_fma_rz_d:
1929 case Intrinsic::nvvm_fma_rm_f:
1930 case Intrinsic::nvvm_fma_rn_f:
1931 case Intrinsic::nvvm_fma_rp_f:
1932 case Intrinsic::nvvm_fma_rz_f:
1933 case Intrinsic::nvvm_fma_rm_ftz_f:
1934 case Intrinsic::nvvm_fma_rn_ftz_f:
1935 case Intrinsic::nvvm_fma_rp_ftz_f:
1936 case Intrinsic::nvvm_fma_rz_ftz_f:
1937
1938 // Sign operations are actually bitwise operations, they do not raise
1939 // exceptions even for SNANs.
1940 case Intrinsic::fabs:
1941 case Intrinsic::copysign:
1942 case Intrinsic::is_fpclass:
1943 // Non-constrained variants of rounding operations means default FP
1944 // environment, they can be folded in any case.
1945 case Intrinsic::ceil:
1946 case Intrinsic::floor:
1947 case Intrinsic::round:
1948 case Intrinsic::roundeven:
1949 case Intrinsic::trunc:
1950 case Intrinsic::nearbyint:
1951 case Intrinsic::rint:
1952 case Intrinsic::canonicalize:
1953
1954 // Constrained intrinsics can be folded if FP environment is known
1955 // to compiler.
1956 case Intrinsic::experimental_constrained_fma:
1957 case Intrinsic::experimental_constrained_fmuladd:
1958 case Intrinsic::experimental_constrained_fadd:
1959 case Intrinsic::experimental_constrained_fsub:
1960 case Intrinsic::experimental_constrained_fmul:
1961 case Intrinsic::experimental_constrained_fdiv:
1962 case Intrinsic::experimental_constrained_frem:
1963 case Intrinsic::experimental_constrained_ceil:
1964 case Intrinsic::experimental_constrained_floor:
1965 case Intrinsic::experimental_constrained_round:
1966 case Intrinsic::experimental_constrained_roundeven:
1967 case Intrinsic::experimental_constrained_trunc:
1968 case Intrinsic::experimental_constrained_nearbyint:
1969 case Intrinsic::experimental_constrained_rint:
1970 case Intrinsic::experimental_constrained_fcmp:
1971 case Intrinsic::experimental_constrained_fcmps:
1972
1973 case Intrinsic::experimental_cttz_elts:
1974 return true;
1975 default:
1976 return false;
1977 case Intrinsic::not_intrinsic: break;
1978 }
1979
1980 if (!F->hasName() || Call->isStrictFP())
1981 return false;
1982
1983 // In these cases, the check of the length is required. We don't want to
1984 // return true for a name like "cos\0blah" which strcmp would return equal to
1985 // "cos", but has length 8.
1986 StringRef Name = F->getName();
1987 switch (Name[0]) {
1988 default:
1989 return false;
1990 // clang-format off
1991 case 'a':
1992 return Name == "acos" || Name == "acosf" ||
1993 Name == "asin" || Name == "asinf" ||
1994 Name == "atan" || Name == "atanf" ||
1995 Name == "atan2" || Name == "atan2f";
1996 case 'c':
1997 return Name == "ceil" || Name == "ceilf" ||
1998 Name == "cos" || Name == "cosf" ||
1999 Name == "cosh" || Name == "coshf";
2000 case 'e':
2001 return Name == "exp" || Name == "expf" || Name == "exp2" ||
2002 Name == "exp2f" || Name == "erf" || Name == "erff";
2003 case 'f':
2004 return Name == "fabs" || Name == "fabsf" ||
2005 Name == "floor" || Name == "floorf" ||
2006 Name == "fmod" || Name == "fmodf";
2007 case 'i':
2008 return Name == "ilogb" || Name == "ilogbf";
2009 case 'l':
2010 return Name == "log" || Name == "logf" || Name == "logl" ||
2011 Name == "log2" || Name == "log2f" || Name == "log10" ||
2012 Name == "log10f" || Name == "logb" || Name == "logbf" ||
2013 Name == "log1p" || Name == "log1pf";
2014 case 'n':
2015 return Name == "nearbyint" || Name == "nearbyintf";
2016 case 'p':
2017 return Name == "pow" || Name == "powf";
2018 case 'r':
2019 return Name == "remainder" || Name == "remainderf" ||
2020 Name == "rint" || Name == "rintf" ||
2021 Name == "round" || Name == "roundf" ||
2022 Name == "roundeven" || Name == "roundevenf";
2023 case 's':
2024 return Name == "sin" || Name == "sinf" ||
2025 Name == "sinh" || Name == "sinhf" ||
2026 Name == "sqrt" || Name == "sqrtf";
2027 case 't':
2028 return Name == "tan" || Name == "tanf" ||
2029 Name == "tanh" || Name == "tanhf" ||
2030 Name == "trunc" || Name == "truncf";
2031 case '_':
2032 // Check for various function names that get used for the math functions
2033 // when the header files are preprocessed with the macro
2034 // __FINITE_MATH_ONLY__ enabled.
2035 // The '12' here is the length of the shortest name that can match.
2036 // We need to check the size before looking at Name[1] and Name[2]
2037 // so we may as well check a limit that will eliminate mismatches.
2038 if (Name.size() < 12 || Name[1] != '_')
2039 return false;
2040 switch (Name[2]) {
2041 default:
2042 return false;
2043 case 'a':
2044 return Name == "__acos_finite" || Name == "__acosf_finite" ||
2045 Name == "__asin_finite" || Name == "__asinf_finite" ||
2046 Name == "__atan2_finite" || Name == "__atan2f_finite";
2047 case 'c':
2048 return Name == "__cosh_finite" || Name == "__coshf_finite";
2049 case 'e':
2050 return Name == "__exp_finite" || Name == "__expf_finite" ||
2051 Name == "__exp2_finite" || Name == "__exp2f_finite";
2052 case 'l':
2053 return Name == "__log_finite" || Name == "__logf_finite" ||
2054 Name == "__log10_finite" || Name == "__log10f_finite";
2055 case 'p':
2056 return Name == "__pow_finite" || Name == "__powf_finite";
2057 case 's':
2058 return Name == "__sinh_finite" || Name == "__sinhf_finite";
2059 }
2060 // clang-format on
2061 }
2062}
2063
2064namespace {
2065
2066Constant *GetConstantFoldFPValue(double V, Type *Ty) {
2067 if (Ty->isHalfTy() || Ty->isFloatTy()) {
2068 APFloat APF(V);
2069 bool unused;
2070 APF.convert(ToSemantics: Ty->getFltSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused);
2071 return ConstantFP::get(Context&: Ty->getContext(), V: APF);
2072 }
2073 if (Ty->isDoubleTy())
2074 return ConstantFP::get(Context&: Ty->getContext(), V: APFloat(V));
2075 llvm_unreachable("Can only constant fold half/float/double");
2076}
2077
2078#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
2079Constant *GetConstantFoldFPValue128(float128 V, Type *Ty) {
2080 if (Ty->isFP128Ty())
2081 return ConstantFP::get(Ty, V);
2082 llvm_unreachable("Can only constant fold fp128");
2083}
2084#endif
2085
2086/// Clear the floating-point exception state.
2087inline void llvm_fenv_clearexcept() {
2088#if HAVE_DECL_FE_ALL_EXCEPT
2089 feclearexcept(FE_ALL_EXCEPT);
2090#endif
2091 errno = 0;
2092}
2093
2094/// Test if a floating-point exception was raised.
2095inline bool llvm_fenv_testexcept() {
2096 int errno_val = errno;
2097 if (errno_val == ERANGE || errno_val == EDOM)
2098 return true;
2099#if HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
2100 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
2101 return true;
2102#endif
2103 return false;
2104}
2105
2106static APFloat FTZPreserveSign(const APFloat &V) {
2107 if (V.isDenormal())
2108 return APFloat::getZero(Sem: V.getSemantics(), Negative: V.isNegative());
2109 return V;
2110}
2111
2112static APFloat FlushToPositiveZero(const APFloat &V) {
2113 if (V.isDenormal())
2114 return APFloat::getZero(Sem: V.getSemantics(), Negative: false);
2115 return V;
2116}
2117
2118static APFloat FlushWithDenormKind(const APFloat &V,
2119 DenormalMode::DenormalModeKind DenormKind) {
2120 assert(DenormKind != DenormalMode::DenormalModeKind::Invalid &&
2121 DenormKind != DenormalMode::DenormalModeKind::Dynamic);
2122 switch (DenormKind) {
2123 case DenormalMode::DenormalModeKind::IEEE:
2124 return V;
2125 case DenormalMode::DenormalModeKind::PreserveSign:
2126 return FTZPreserveSign(V);
2127 case DenormalMode::DenormalModeKind::PositiveZero:
2128 return FlushToPositiveZero(V);
2129 default:
2130 llvm_unreachable("Invalid denormal mode!");
2131 }
2132}
2133
2134Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V, Type *Ty,
2135 DenormalMode DenormMode = DenormalMode::getIEEE()) {
2136 if (!DenormMode.isValid() ||
2137 DenormMode.Input == DenormalMode::DenormalModeKind::Dynamic ||
2138 DenormMode.Output == DenormalMode::DenormalModeKind::Dynamic)
2139 return nullptr;
2140
2141 llvm_fenv_clearexcept();
2142 auto Input = FlushWithDenormKind(V, DenormKind: DenormMode.Input);
2143 double Result = NativeFP(Input.convertToDouble());
2144 if (llvm_fenv_testexcept()) {
2145 llvm_fenv_clearexcept();
2146 return nullptr;
2147 }
2148
2149 Constant *Output = GetConstantFoldFPValue(V: Result, Ty);
2150 if (DenormMode.Output == DenormalMode::DenormalModeKind::IEEE)
2151 return Output;
2152 const auto *CFP = static_cast<ConstantFP *>(Output);
2153 const auto Res = FlushWithDenormKind(V: CFP->getValueAPF(), DenormKind: DenormMode.Output);
2154 return ConstantFP::get(Context&: Ty->getContext(), V: Res);
2155}
2156
2157#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
2158Constant *ConstantFoldFP128(float128 (*NativeFP)(float128), const APFloat &V,
2159 Type *Ty) {
2160 llvm_fenv_clearexcept();
2161 float128 Result = NativeFP(V.convertToQuad());
2162 if (llvm_fenv_testexcept()) {
2163 llvm_fenv_clearexcept();
2164 return nullptr;
2165 }
2166
2167 return GetConstantFoldFPValue128(V: Result, Ty);
2168}
2169#endif
2170
2171Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
2172 const APFloat &V, const APFloat &W, Type *Ty) {
2173 llvm_fenv_clearexcept();
2174 double Result = NativeFP(V.convertToDouble(), W.convertToDouble());
2175 if (llvm_fenv_testexcept()) {
2176 llvm_fenv_clearexcept();
2177 return nullptr;
2178 }
2179
2180 return GetConstantFoldFPValue(V: Result, Ty);
2181}
2182
2183Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) {
2184 auto *OpVT = cast<VectorType>(Val: Op->getType());
2185
2186 // This is the same as the underlying binops - poison propagates.
2187 if (Op->containsPoisonElement())
2188 return PoisonValue::get(T: OpVT->getElementType());
2189
2190 // Shortcut non-accumulating reductions.
2191 if (Constant *SplatVal = Op->getSplatValue()) {
2192 switch (IID) {
2193 case Intrinsic::vector_reduce_and:
2194 case Intrinsic::vector_reduce_or:
2195 case Intrinsic::vector_reduce_smin:
2196 case Intrinsic::vector_reduce_smax:
2197 case Intrinsic::vector_reduce_umin:
2198 case Intrinsic::vector_reduce_umax:
2199 return SplatVal;
2200 case Intrinsic::vector_reduce_add:
2201 if (SplatVal->isNullValue())
2202 return SplatVal;
2203 break;
2204 case Intrinsic::vector_reduce_mul:
2205 if (SplatVal->isNullValue() || SplatVal->isOneValue())
2206 return SplatVal;
2207 break;
2208 case Intrinsic::vector_reduce_xor:
2209 if (SplatVal->isNullValue())
2210 return SplatVal;
2211 if (OpVT->getElementCount().isKnownMultipleOf(RHS: 2))
2212 return Constant::getNullValue(Ty: OpVT->getElementType());
2213 break;
2214 }
2215 }
2216
2217 FixedVectorType *VT = dyn_cast<FixedVectorType>(Val: OpVT);
2218 if (!VT)
2219 return nullptr;
2220
2221 // TODO: Handle undef.
2222 auto *EltC = dyn_cast_or_null<ConstantInt>(Val: Op->getAggregateElement(Elt: 0U));
2223 if (!EltC)
2224 return nullptr;
2225
2226 APInt Acc = EltC->getValue();
2227 for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) {
2228 if (!(EltC = dyn_cast_or_null<ConstantInt>(Val: Op->getAggregateElement(Elt: I))))
2229 return nullptr;
2230 const APInt &X = EltC->getValue();
2231 switch (IID) {
2232 case Intrinsic::vector_reduce_add:
2233 Acc = Acc + X;
2234 break;
2235 case Intrinsic::vector_reduce_mul:
2236 Acc = Acc * X;
2237 break;
2238 case Intrinsic::vector_reduce_and:
2239 Acc = Acc & X;
2240 break;
2241 case Intrinsic::vector_reduce_or:
2242 Acc = Acc | X;
2243 break;
2244 case Intrinsic::vector_reduce_xor:
2245 Acc = Acc ^ X;
2246 break;
2247 case Intrinsic::vector_reduce_smin:
2248 Acc = APIntOps::smin(A: Acc, B: X);
2249 break;
2250 case Intrinsic::vector_reduce_smax:
2251 Acc = APIntOps::smax(A: Acc, B: X);
2252 break;
2253 case Intrinsic::vector_reduce_umin:
2254 Acc = APIntOps::umin(A: Acc, B: X);
2255 break;
2256 case Intrinsic::vector_reduce_umax:
2257 Acc = APIntOps::umax(A: Acc, B: X);
2258 break;
2259 }
2260 }
2261
2262 return ConstantInt::get(Context&: Op->getContext(), V: Acc);
2263}
2264
2265/// Attempt to fold an SSE floating point to integer conversion of a constant
2266/// floating point. If roundTowardZero is false, the default IEEE rounding is
2267/// used (toward nearest, ties to even). This matches the behavior of the
2268/// non-truncating SSE instructions in the default rounding mode. The desired
2269/// integer type Ty is used to select how many bits are available for the
2270/// result. Returns null if the conversion cannot be performed, otherwise
2271/// returns the Constant value resulting from the conversion.
2272Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
2273 Type *Ty, bool IsSigned) {
2274 // All of these conversion intrinsics form an integer of at most 64bits.
2275 unsigned ResultWidth = Ty->getIntegerBitWidth();
2276 assert(ResultWidth <= 64 &&
2277 "Can only constant fold conversions to 64 and 32 bit ints");
2278
2279 uint64_t UIntVal;
2280 bool isExact = false;
2281 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
2282 : APFloat::rmNearestTiesToEven;
2283 APFloat::opStatus status =
2284 Val.convertToInteger(Input: MutableArrayRef(UIntVal), Width: ResultWidth,
2285 IsSigned, RM: mode, IsExact: &isExact);
2286 if (status != APFloat::opOK &&
2287 (!roundTowardZero || status != APFloat::opInexact))
2288 return nullptr;
2289 return ConstantInt::get(Ty, V: UIntVal, IsSigned);
2290}
2291
2292double getValueAsDouble(ConstantFP *Op) {
2293 Type *Ty = Op->getType();
2294
2295 if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())
2296 return Op->getValueAPF().convertToDouble();
2297
2298 bool unused;
2299 APFloat APF = Op->getValueAPF();
2300 APF.convert(ToSemantics: APFloat::IEEEdouble(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused);
2301 return APF.convertToDouble();
2302}
2303
2304static bool getConstIntOrUndef(Value *Op, const APInt *&C) {
2305 if (auto *CI = dyn_cast<ConstantInt>(Val: Op)) {
2306 C = &CI->getValue();
2307 return true;
2308 }
2309 if (isa<UndefValue>(Val: Op)) {
2310 C = nullptr;
2311 return true;
2312 }
2313 return false;
2314}
2315
2316/// Checks if the given intrinsic call, which evaluates to constant, is allowed
2317/// to be folded.
2318///
2319/// \param CI Constrained intrinsic call.
2320/// \param St Exception flags raised during constant evaluation.
2321static bool mayFoldConstrained(ConstrainedFPIntrinsic *CI,
2322 APFloat::opStatus St) {
2323 std::optional<RoundingMode> ORM = CI->getRoundingMode();
2324 std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2325
2326 // If the operation does not change exception status flags, it is safe
2327 // to fold.
2328 if (St == APFloat::opStatus::opOK)
2329 return true;
2330
2331 // If evaluation raised FP exception, the result can depend on rounding
2332 // mode. If the latter is unknown, folding is not possible.
2333 if (ORM == RoundingMode::Dynamic)
2334 return false;
2335
2336 // If FP exceptions are ignored, fold the call, even if such exception is
2337 // raised.
2338 if (EB && *EB != fp::ExceptionBehavior::ebStrict)
2339 return true;
2340
2341 // Leave the calculation for runtime so that exception flags be correctly set
2342 // in hardware.
2343 return false;
2344}
2345
2346/// Returns the rounding mode that should be used for constant evaluation.
2347static RoundingMode
2348getEvaluationRoundingMode(const ConstrainedFPIntrinsic *CI) {
2349 std::optional<RoundingMode> ORM = CI->getRoundingMode();
2350 if (!ORM || *ORM == RoundingMode::Dynamic)
2351 // Even if the rounding mode is unknown, try evaluating the operation.
2352 // If it does not raise inexact exception, rounding was not applied,
2353 // so the result is exact and does not depend on rounding mode. Whether
2354 // other FP exceptions are raised, it does not depend on rounding mode.
2355 return RoundingMode::NearestTiesToEven;
2356 return *ORM;
2357}
2358
2359/// Try to constant fold llvm.canonicalize for the given caller and value.
2360static Constant *constantFoldCanonicalize(const Type *Ty, const CallBase *CI,
2361 const APFloat &Src) {
2362 // Zero, positive and negative, is always OK to fold.
2363 if (Src.isZero()) {
2364 // Get a fresh 0, since ppc_fp128 does have non-canonical zeros.
2365 return ConstantFP::get(
2366 Context&: CI->getContext(),
2367 V: APFloat::getZero(Sem: Src.getSemantics(), Negative: Src.isNegative()));
2368 }
2369
2370 if (!Ty->isIEEELikeFPTy())
2371 return nullptr;
2372
2373 // Zero is always canonical and the sign must be preserved.
2374 //
2375 // Denorms and nans may have special encodings, but it should be OK to fold a
2376 // totally average number.
2377 if (Src.isNormal() || Src.isInfinity())
2378 return ConstantFP::get(Context&: CI->getContext(), V: Src);
2379
2380 if (Src.isDenormal() && CI->getParent() && CI->getFunction()) {
2381 DenormalMode DenormMode =
2382 CI->getFunction()->getDenormalMode(FPType: Src.getSemantics());
2383
2384 if (DenormMode == DenormalMode::getIEEE())
2385 return ConstantFP::get(Context&: CI->getContext(), V: Src);
2386
2387 if (DenormMode.Input == DenormalMode::Dynamic)
2388 return nullptr;
2389
2390 // If we know if either input or output is flushed, we can fold.
2391 if ((DenormMode.Input == DenormalMode::Dynamic &&
2392 DenormMode.Output == DenormalMode::IEEE) ||
2393 (DenormMode.Input == DenormalMode::IEEE &&
2394 DenormMode.Output == DenormalMode::Dynamic))
2395 return nullptr;
2396
2397 bool IsPositive =
2398 (!Src.isNegative() || DenormMode.Input == DenormalMode::PositiveZero ||
2399 (DenormMode.Output == DenormalMode::PositiveZero &&
2400 DenormMode.Input == DenormalMode::IEEE));
2401
2402 return ConstantFP::get(Context&: CI->getContext(),
2403 V: APFloat::getZero(Sem: Src.getSemantics(), Negative: !IsPositive));
2404 }
2405
2406 return nullptr;
2407}
2408
2409static Constant *ConstantFoldScalarCall1(StringRef Name,
2410 Intrinsic::ID IntrinsicID,
2411 Type *Ty,
2412 ArrayRef<Constant *> Operands,
2413 const TargetLibraryInfo *TLI,
2414 const CallBase *Call) {
2415 assert(Operands.size() == 1 && "Wrong number of operands.");
2416
2417 if (IntrinsicID == Intrinsic::is_constant) {
2418 // We know we have a "Constant" argument. But we want to only
2419 // return true for manifest constants, not those that depend on
2420 // constants with unknowable values, e.g. GlobalValue or BlockAddress.
2421 if (Operands[0]->isManifestConstant())
2422 return ConstantInt::getTrue(Context&: Ty->getContext());
2423 return nullptr;
2424 }
2425
2426 if (isa<UndefValue>(Val: Operands[0])) {
2427 // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN.
2428 // ctpop() is between 0 and bitwidth, pick 0 for undef.
2429 // fptoui.sat and fptosi.sat can always fold to zero (for a zero input).
2430 if (IntrinsicID == Intrinsic::cos ||
2431 IntrinsicID == Intrinsic::ctpop ||
2432 IntrinsicID == Intrinsic::fptoui_sat ||
2433 IntrinsicID == Intrinsic::fptosi_sat ||
2434 IntrinsicID == Intrinsic::canonicalize)
2435 return Constant::getNullValue(Ty);
2436 if (IntrinsicID == Intrinsic::bswap ||
2437 IntrinsicID == Intrinsic::bitreverse ||
2438 IntrinsicID == Intrinsic::launder_invariant_group ||
2439 IntrinsicID == Intrinsic::strip_invariant_group)
2440 return Operands[0];
2441 }
2442
2443 if (isa<ConstantPointerNull>(Val: Operands[0])) {
2444 // launder(null) == null == strip(null) iff in addrspace 0
2445 if (IntrinsicID == Intrinsic::launder_invariant_group ||
2446 IntrinsicID == Intrinsic::strip_invariant_group) {
2447 // If instruction is not yet put in a basic block (e.g. when cloning
2448 // a function during inlining), Call's caller may not be available.
2449 // So check Call's BB first before querying Call->getCaller.
2450 const Function *Caller =
2451 Call->getParent() ? Call->getCaller() : nullptr;
2452 if (Caller &&
2453 !NullPointerIsDefined(
2454 F: Caller, AS: Operands[0]->getType()->getPointerAddressSpace())) {
2455 return Operands[0];
2456 }
2457 return nullptr;
2458 }
2459 }
2460
2461 if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[0])) {
2462 APFloat U = Op->getValueAPF();
2463
2464 if (IntrinsicID == Intrinsic::wasm_trunc_signed ||
2465 IntrinsicID == Intrinsic::wasm_trunc_unsigned) {
2466 bool Signed = IntrinsicID == Intrinsic::wasm_trunc_signed;
2467
2468 if (U.isNaN())
2469 return nullptr;
2470
2471 unsigned Width = Ty->getIntegerBitWidth();
2472 APSInt Int(Width, !Signed);
2473 bool IsExact = false;
2474 APFloat::opStatus Status =
2475 U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact);
2476
2477 if (Status == APFloat::opOK || Status == APFloat::opInexact)
2478 return ConstantInt::get(Ty, V: Int);
2479
2480 return nullptr;
2481 }
2482
2483 if (IntrinsicID == Intrinsic::fptoui_sat ||
2484 IntrinsicID == Intrinsic::fptosi_sat) {
2485 // convertToInteger() already has the desired saturation semantics.
2486 APSInt Int(Ty->getIntegerBitWidth(),
2487 IntrinsicID == Intrinsic::fptoui_sat);
2488 bool IsExact;
2489 U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact);
2490 return ConstantInt::get(Ty, V: Int);
2491 }
2492
2493 if (IntrinsicID == Intrinsic::canonicalize)
2494 return constantFoldCanonicalize(Ty, CI: Call, Src: U);
2495
2496#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
2497 if (Ty->isFP128Ty()) {
2498 if (IntrinsicID == Intrinsic::log) {
2499 float128 Result = logf128(Op->getValueAPF().convertToQuad());
2500 return GetConstantFoldFPValue128(V: Result, Ty);
2501 }
2502
2503 LibFunc Fp128Func = NotLibFunc;
2504 if (TLI && TLI->getLibFunc(funcName: Name, F&: Fp128Func) && TLI->has(F: Fp128Func) &&
2505 Fp128Func == LibFunc_logl)
2506 return ConstantFoldFP128(logf128, Op->getValueAPF(), Ty);
2507 }
2508#endif
2509
2510 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy() &&
2511 !Ty->isIntegerTy())
2512 return nullptr;
2513
2514 // Use internal versions of these intrinsics.
2515
2516 if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint ||
2517 IntrinsicID == Intrinsic::roundeven) {
2518 U.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
2519 return ConstantFP::get(Ty, V: U);
2520 }
2521
2522 if (IntrinsicID == Intrinsic::round) {
2523 U.roundToIntegral(RM: APFloat::rmNearestTiesToAway);
2524 return ConstantFP::get(Ty, V: U);
2525 }
2526
2527 if (IntrinsicID == Intrinsic::roundeven) {
2528 U.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
2529 return ConstantFP::get(Ty, V: U);
2530 }
2531
2532 if (IntrinsicID == Intrinsic::ceil) {
2533 U.roundToIntegral(RM: APFloat::rmTowardPositive);
2534 return ConstantFP::get(Ty, V: U);
2535 }
2536
2537 if (IntrinsicID == Intrinsic::floor) {
2538 U.roundToIntegral(RM: APFloat::rmTowardNegative);
2539 return ConstantFP::get(Ty, V: U);
2540 }
2541
2542 if (IntrinsicID == Intrinsic::trunc) {
2543 U.roundToIntegral(RM: APFloat::rmTowardZero);
2544 return ConstantFP::get(Ty, V: U);
2545 }
2546
2547 if (IntrinsicID == Intrinsic::fabs) {
2548 U.clearSign();
2549 return ConstantFP::get(Ty, V: U);
2550 }
2551
2552 if (IntrinsicID == Intrinsic::amdgcn_fract) {
2553 // The v_fract instruction behaves like the OpenCL spec, which defines
2554 // fract(x) as fmin(x - floor(x), 0x1.fffffep-1f): "The min() operator is
2555 // there to prevent fract(-small) from returning 1.0. It returns the
2556 // largest positive floating-point number less than 1.0."
2557 APFloat FloorU(U);
2558 FloorU.roundToIntegral(RM: APFloat::rmTowardNegative);
2559 APFloat FractU(U - FloorU);
2560 APFloat AlmostOne(U.getSemantics(), 1);
2561 AlmostOne.next(/*nextDown*/ true);
2562 return ConstantFP::get(Ty, V: minimum(A: FractU, B: AlmostOne));
2563 }
2564
2565 // Rounding operations (floor, trunc, ceil, round and nearbyint) do not
2566 // raise FP exceptions, unless the argument is signaling NaN.
2567
2568 std::optional<APFloat::roundingMode> RM;
2569 switch (IntrinsicID) {
2570 default:
2571 break;
2572 case Intrinsic::experimental_constrained_nearbyint:
2573 case Intrinsic::experimental_constrained_rint: {
2574 auto CI = cast<ConstrainedFPIntrinsic>(Val: Call);
2575 RM = CI->getRoundingMode();
2576 if (!RM || *RM == RoundingMode::Dynamic)
2577 return nullptr;
2578 break;
2579 }
2580 case Intrinsic::experimental_constrained_round:
2581 RM = APFloat::rmNearestTiesToAway;
2582 break;
2583 case Intrinsic::experimental_constrained_ceil:
2584 RM = APFloat::rmTowardPositive;
2585 break;
2586 case Intrinsic::experimental_constrained_floor:
2587 RM = APFloat::rmTowardNegative;
2588 break;
2589 case Intrinsic::experimental_constrained_trunc:
2590 RM = APFloat::rmTowardZero;
2591 break;
2592 }
2593 if (RM) {
2594 auto CI = cast<ConstrainedFPIntrinsic>(Val: Call);
2595 if (U.isFinite()) {
2596 APFloat::opStatus St = U.roundToIntegral(RM: *RM);
2597 if (IntrinsicID == Intrinsic::experimental_constrained_rint &&
2598 St == APFloat::opInexact) {
2599 std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2600 if (EB == fp::ebStrict)
2601 return nullptr;
2602 }
2603 } else if (U.isSignaling()) {
2604 std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2605 if (EB && *EB != fp::ebIgnore)
2606 return nullptr;
2607 U = APFloat::getQNaN(Sem: U.getSemantics());
2608 }
2609 return ConstantFP::get(Ty, V: U);
2610 }
2611
2612 // NVVM float/double to signed/unsigned int32/int64 conversions:
2613 switch (IntrinsicID) {
2614 // f2i
2615 case Intrinsic::nvvm_f2i_rm:
2616 case Intrinsic::nvvm_f2i_rn:
2617 case Intrinsic::nvvm_f2i_rp:
2618 case Intrinsic::nvvm_f2i_rz:
2619 case Intrinsic::nvvm_f2i_rm_ftz:
2620 case Intrinsic::nvvm_f2i_rn_ftz:
2621 case Intrinsic::nvvm_f2i_rp_ftz:
2622 case Intrinsic::nvvm_f2i_rz_ftz:
2623 // f2ui
2624 case Intrinsic::nvvm_f2ui_rm:
2625 case Intrinsic::nvvm_f2ui_rn:
2626 case Intrinsic::nvvm_f2ui_rp:
2627 case Intrinsic::nvvm_f2ui_rz:
2628 case Intrinsic::nvvm_f2ui_rm_ftz:
2629 case Intrinsic::nvvm_f2ui_rn_ftz:
2630 case Intrinsic::nvvm_f2ui_rp_ftz:
2631 case Intrinsic::nvvm_f2ui_rz_ftz:
2632 // d2i
2633 case Intrinsic::nvvm_d2i_rm:
2634 case Intrinsic::nvvm_d2i_rn:
2635 case Intrinsic::nvvm_d2i_rp:
2636 case Intrinsic::nvvm_d2i_rz:
2637 // d2ui
2638 case Intrinsic::nvvm_d2ui_rm:
2639 case Intrinsic::nvvm_d2ui_rn:
2640 case Intrinsic::nvvm_d2ui_rp:
2641 case Intrinsic::nvvm_d2ui_rz:
2642 // f2ll
2643 case Intrinsic::nvvm_f2ll_rm:
2644 case Intrinsic::nvvm_f2ll_rn:
2645 case Intrinsic::nvvm_f2ll_rp:
2646 case Intrinsic::nvvm_f2ll_rz:
2647 case Intrinsic::nvvm_f2ll_rm_ftz:
2648 case Intrinsic::nvvm_f2ll_rn_ftz:
2649 case Intrinsic::nvvm_f2ll_rp_ftz:
2650 case Intrinsic::nvvm_f2ll_rz_ftz:
2651 // f2ull
2652 case Intrinsic::nvvm_f2ull_rm:
2653 case Intrinsic::nvvm_f2ull_rn:
2654 case Intrinsic::nvvm_f2ull_rp:
2655 case Intrinsic::nvvm_f2ull_rz:
2656 case Intrinsic::nvvm_f2ull_rm_ftz:
2657 case Intrinsic::nvvm_f2ull_rn_ftz:
2658 case Intrinsic::nvvm_f2ull_rp_ftz:
2659 case Intrinsic::nvvm_f2ull_rz_ftz:
2660 // d2ll
2661 case Intrinsic::nvvm_d2ll_rm:
2662 case Intrinsic::nvvm_d2ll_rn:
2663 case Intrinsic::nvvm_d2ll_rp:
2664 case Intrinsic::nvvm_d2ll_rz:
2665 // d2ull
2666 case Intrinsic::nvvm_d2ull_rm:
2667 case Intrinsic::nvvm_d2ull_rn:
2668 case Intrinsic::nvvm_d2ull_rp:
2669 case Intrinsic::nvvm_d2ull_rz: {
2670 // In float-to-integer conversion, NaN inputs are converted to 0.
2671 if (U.isNaN()) {
2672 // In float-to-integer conversion, NaN inputs are converted to 0
2673 // when the source and destination bitwidths are both less than 64.
2674 if (nvvm::FPToIntegerIntrinsicNaNZero(IntrinsicID))
2675 return ConstantInt::get(Ty, V: 0);
2676
2677 // Otherwise, the most significant bit is set.
2678 unsigned BitWidth = Ty->getIntegerBitWidth();
2679 uint64_t Val = 1ULL << (BitWidth - 1);
2680 return ConstantInt::get(Ty, V: APInt(BitWidth, Val, /*IsSigned=*/false));
2681 }
2682
2683 APFloat::roundingMode RMode =
2684 nvvm::GetFPToIntegerRoundingMode(IntrinsicID);
2685 bool IsFTZ = nvvm::FPToIntegerIntrinsicShouldFTZ(IntrinsicID);
2686 bool IsSigned = nvvm::FPToIntegerIntrinsicResultIsSigned(IntrinsicID);
2687
2688 APSInt ResInt(Ty->getIntegerBitWidth(), !IsSigned);
2689 auto FloatToRound = IsFTZ ? FTZPreserveSign(V: U) : U;
2690
2691 // Return max/min value for integers if the result is +/-inf or
2692 // is too large to fit in the result's integer bitwidth.
2693 bool IsExact = false;
2694 FloatToRound.convertToInteger(Result&: ResInt, RM: RMode, IsExact: &IsExact);
2695 return ConstantInt::get(Ty, V: ResInt);
2696 }
2697 }
2698
2699 /// We only fold functions with finite arguments. Folding NaN and inf is
2700 /// likely to be aborted with an exception anyway, and some host libms
2701 /// have known errors raising exceptions.
2702 if (!U.isFinite())
2703 return nullptr;
2704
2705 /// Currently APFloat versions of these functions do not exist, so we use
2706 /// the host native double versions. Float versions are not called
2707 /// directly but for all these it is true (float)(f((double)arg)) ==
2708 /// f(arg). Long double not supported yet.
2709 const APFloat &APF = Op->getValueAPF();
2710
2711 switch (IntrinsicID) {
2712 default: break;
2713 case Intrinsic::log:
2714 if (U.isZero())
2715 return ConstantFP::getInfinity(Ty, Negative: true);
2716 if (U.isNegative())
2717 return ConstantFP::getNaN(Ty);
2718 if (U.isExactlyValue(V: 1.0))
2719 return ConstantFP::getZero(Ty);
2720 return ConstantFoldFP(NativeFP: log, V: APF, Ty);
2721 case Intrinsic::log2:
2722 if (U.isZero())
2723 return ConstantFP::getInfinity(Ty, Negative: true);
2724 if (U.isNegative())
2725 return ConstantFP::getNaN(Ty);
2726 if (U.isExactlyValue(V: 1.0))
2727 return ConstantFP::getZero(Ty);
2728 // TODO: What about hosts that lack a C99 library?
2729 return ConstantFoldFP(NativeFP: log2, V: APF, Ty);
2730 case Intrinsic::log10:
2731 if (U.isZero())
2732 return ConstantFP::getInfinity(Ty, Negative: true);
2733 if (U.isNegative())
2734 return ConstantFP::getNaN(Ty);
2735 if (U.isExactlyValue(V: 1.0))
2736 return ConstantFP::getZero(Ty);
2737 // TODO: What about hosts that lack a C99 library?
2738 return ConstantFoldFP(NativeFP: log10, V: APF, Ty);
2739 case Intrinsic::exp:
2740 return ConstantFoldFP(NativeFP: exp, V: APF, Ty);
2741 case Intrinsic::exp2:
2742 // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
2743 return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty);
2744 case Intrinsic::exp10:
2745 // Fold exp10(x) as pow(10, x), in case the host lacks a C99 library.
2746 return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(10.0), W: APF, Ty);
2747 case Intrinsic::sin:
2748 return ConstantFoldFP(NativeFP: sin, V: APF, Ty);
2749 case Intrinsic::cos:
2750 return ConstantFoldFP(NativeFP: cos, V: APF, Ty);
2751 case Intrinsic::sinh:
2752 return ConstantFoldFP(NativeFP: sinh, V: APF, Ty);
2753 case Intrinsic::cosh:
2754 return ConstantFoldFP(NativeFP: cosh, V: APF, Ty);
2755 case Intrinsic::atan:
2756 // Implement optional behavior from C's Annex F for +/-0.0.
2757 if (U.isZero())
2758 return ConstantFP::get(Ty, V: U);
2759 return ConstantFoldFP(NativeFP: atan, V: APF, Ty);
2760 case Intrinsic::sqrt:
2761 return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty);
2762
2763 // NVVM Intrinsics:
2764 case Intrinsic::nvvm_ceil_ftz_f:
2765 case Intrinsic::nvvm_ceil_f:
2766 case Intrinsic::nvvm_ceil_d:
2767 return ConstantFoldFP(
2768 NativeFP: ceil, V: APF, Ty,
2769 DenormMode: nvvm::GetNVVMDenormMode(
2770 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2771
2772 case Intrinsic::nvvm_fabs_ftz:
2773 case Intrinsic::nvvm_fabs:
2774 return ConstantFoldFP(
2775 NativeFP: fabs, V: APF, Ty,
2776 DenormMode: nvvm::GetNVVMDenormMode(
2777 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2778
2779 case Intrinsic::nvvm_floor_ftz_f:
2780 case Intrinsic::nvvm_floor_f:
2781 case Intrinsic::nvvm_floor_d:
2782 return ConstantFoldFP(
2783 NativeFP: floor, V: APF, Ty,
2784 DenormMode: nvvm::GetNVVMDenormMode(
2785 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2786
2787 case Intrinsic::nvvm_rcp_rm_ftz_f:
2788 case Intrinsic::nvvm_rcp_rn_ftz_f:
2789 case Intrinsic::nvvm_rcp_rp_ftz_f:
2790 case Intrinsic::nvvm_rcp_rz_ftz_f:
2791 case Intrinsic::nvvm_rcp_rm_d:
2792 case Intrinsic::nvvm_rcp_rm_f:
2793 case Intrinsic::nvvm_rcp_rn_d:
2794 case Intrinsic::nvvm_rcp_rn_f:
2795 case Intrinsic::nvvm_rcp_rp_d:
2796 case Intrinsic::nvvm_rcp_rp_f:
2797 case Intrinsic::nvvm_rcp_rz_d:
2798 case Intrinsic::nvvm_rcp_rz_f: {
2799 APFloat::roundingMode RoundMode = nvvm::GetRCPRoundingMode(IntrinsicID);
2800 bool IsFTZ = nvvm::RCPShouldFTZ(IntrinsicID);
2801
2802 auto Denominator = IsFTZ ? FTZPreserveSign(V: APF) : APF;
2803 APFloat Res = APFloat::getOne(Sem: APF.getSemantics());
2804 APFloat::opStatus Status = Res.divide(RHS: Denominator, RM: RoundMode);
2805
2806 if (Status == APFloat::opOK || Status == APFloat::opInexact) {
2807 if (IsFTZ)
2808 Res = FTZPreserveSign(V: Res);
2809 return ConstantFP::get(Ty, V: Res);
2810 }
2811 return nullptr;
2812 }
2813
2814 case Intrinsic::nvvm_round_ftz_f:
2815 case Intrinsic::nvvm_round_f:
2816 case Intrinsic::nvvm_round_d: {
2817 // nvvm_round is lowered to PTX cvt.rni, which will round to nearest
2818 // integer, choosing even integer if source is equidistant between two
2819 // integers, so the semantics are closer to "rint" rather than "round".
2820 bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);
2821 auto V = IsFTZ ? FTZPreserveSign(V: APF) : APF;
2822 V.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
2823 return ConstantFP::get(Ty, V);
2824 }
2825
2826 case Intrinsic::nvvm_saturate_ftz_f:
2827 case Intrinsic::nvvm_saturate_d:
2828 case Intrinsic::nvvm_saturate_f: {
2829 bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);
2830 auto V = IsFTZ ? FTZPreserveSign(V: APF) : APF;
2831 if (V.isNegative() || V.isZero() || V.isNaN())
2832 return ConstantFP::getZero(Ty);
2833 APFloat One = APFloat::getOne(Sem: APF.getSemantics());
2834 if (V > One)
2835 return ConstantFP::get(Ty, V: One);
2836 return ConstantFP::get(Ty, V: APF);
2837 }
2838
2839 case Intrinsic::nvvm_sqrt_rn_ftz_f:
2840 case Intrinsic::nvvm_sqrt_f:
2841 case Intrinsic::nvvm_sqrt_rn_d:
2842 case Intrinsic::nvvm_sqrt_rn_f:
2843 if (APF.isNegative())
2844 return nullptr;
2845 return ConstantFoldFP(
2846 NativeFP: sqrt, V: APF, Ty,
2847 DenormMode: nvvm::GetNVVMDenormMode(
2848 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2849
2850 // AMDGCN Intrinsics:
2851 case Intrinsic::amdgcn_cos:
2852 case Intrinsic::amdgcn_sin: {
2853 double V = getValueAsDouble(Op);
2854 if (V < -256.0 || V > 256.0)
2855 // The gfx8 and gfx9 architectures handle arguments outside the range
2856 // [-256, 256] differently. This should be a rare case so bail out
2857 // rather than trying to handle the difference.
2858 return nullptr;
2859 bool IsCos = IntrinsicID == Intrinsic::amdgcn_cos;
2860 double V4 = V * 4.0;
2861 if (V4 == floor(x: V4)) {
2862 // Force exact results for quarter-integer inputs.
2863 const double SinVals[4] = { 0.0, 1.0, 0.0, -1.0 };
2864 V = SinVals[((int)V4 + (IsCos ? 1 : 0)) & 3];
2865 } else {
2866 if (IsCos)
2867 V = cos(x: V * 2.0 * numbers::pi);
2868 else
2869 V = sin(x: V * 2.0 * numbers::pi);
2870 }
2871 return GetConstantFoldFPValue(V, Ty);
2872 }
2873 }
2874
2875 if (!TLI)
2876 return nullptr;
2877
2878 LibFunc Func = NotLibFunc;
2879 if (!TLI->getLibFunc(funcName: Name, F&: Func))
2880 return nullptr;
2881
2882 switch (Func) {
2883 default:
2884 break;
2885 case LibFunc_acos:
2886 case LibFunc_acosf:
2887 case LibFunc_acos_finite:
2888 case LibFunc_acosf_finite:
2889 if (TLI->has(F: Func))
2890 return ConstantFoldFP(NativeFP: acos, V: APF, Ty);
2891 break;
2892 case LibFunc_asin:
2893 case LibFunc_asinf:
2894 case LibFunc_asin_finite:
2895 case LibFunc_asinf_finite:
2896 if (TLI->has(F: Func))
2897 return ConstantFoldFP(NativeFP: asin, V: APF, Ty);
2898 break;
2899 case LibFunc_atan:
2900 case LibFunc_atanf:
2901 // Implement optional behavior from C's Annex F for +/-0.0.
2902 if (U.isZero())
2903 return ConstantFP::get(Ty, V: U);
2904 if (TLI->has(F: Func))
2905 return ConstantFoldFP(NativeFP: atan, V: APF, Ty);
2906 break;
2907 case LibFunc_ceil:
2908 case LibFunc_ceilf:
2909 if (TLI->has(F: Func)) {
2910 U.roundToIntegral(RM: APFloat::rmTowardPositive);
2911 return ConstantFP::get(Ty, V: U);
2912 }
2913 break;
2914 case LibFunc_cos:
2915 case LibFunc_cosf:
2916 if (TLI->has(F: Func))
2917 return ConstantFoldFP(NativeFP: cos, V: APF, Ty);
2918 break;
2919 case LibFunc_cosh:
2920 case LibFunc_coshf:
2921 case LibFunc_cosh_finite:
2922 case LibFunc_coshf_finite:
2923 if (TLI->has(F: Func))
2924 return ConstantFoldFP(NativeFP: cosh, V: APF, Ty);
2925 break;
2926 case LibFunc_exp:
2927 case LibFunc_expf:
2928 case LibFunc_exp_finite:
2929 case LibFunc_expf_finite:
2930 if (TLI->has(F: Func))
2931 return ConstantFoldFP(NativeFP: exp, V: APF, Ty);
2932 break;
2933 case LibFunc_exp2:
2934 case LibFunc_exp2f:
2935 case LibFunc_exp2_finite:
2936 case LibFunc_exp2f_finite:
2937 if (TLI->has(F: Func))
2938 // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
2939 return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty);
2940 break;
2941 case LibFunc_fabs:
2942 case LibFunc_fabsf:
2943 if (TLI->has(F: Func)) {
2944 U.clearSign();
2945 return ConstantFP::get(Ty, V: U);
2946 }
2947 break;
2948 case LibFunc_floor:
2949 case LibFunc_floorf:
2950 if (TLI->has(F: Func)) {
2951 U.roundToIntegral(RM: APFloat::rmTowardNegative);
2952 return ConstantFP::get(Ty, V: U);
2953 }
2954 break;
2955 case LibFunc_log:
2956 case LibFunc_logf:
2957 case LibFunc_log_finite:
2958 case LibFunc_logf_finite:
2959 if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func))
2960 return ConstantFoldFP(NativeFP: log, V: APF, Ty);
2961 break;
2962 case LibFunc_log2:
2963 case LibFunc_log2f:
2964 case LibFunc_log2_finite:
2965 case LibFunc_log2f_finite:
2966 if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func))
2967 // TODO: What about hosts that lack a C99 library?
2968 return ConstantFoldFP(NativeFP: log2, V: APF, Ty);
2969 break;
2970 case LibFunc_log10:
2971 case LibFunc_log10f:
2972 case LibFunc_log10_finite:
2973 case LibFunc_log10f_finite:
2974 if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func))
2975 // TODO: What about hosts that lack a C99 library?
2976 return ConstantFoldFP(NativeFP: log10, V: APF, Ty);
2977 break;
2978 case LibFunc_ilogb:
2979 case LibFunc_ilogbf:
2980 if (!APF.isZero() && TLI->has(F: Func))
2981 return ConstantInt::get(Ty, V: ilogb(Arg: APF), IsSigned: true);
2982 break;
2983 case LibFunc_logb:
2984 case LibFunc_logbf:
2985 if (!APF.isZero() && TLI->has(F: Func))
2986 return ConstantFoldFP(NativeFP: logb, V: APF, Ty);
2987 break;
2988 case LibFunc_log1p:
2989 case LibFunc_log1pf:
2990 // Implement optional behavior from C's Annex F for +/-0.0.
2991 if (U.isZero())
2992 return ConstantFP::get(Ty, V: U);
2993 if (APF > APFloat::getOne(Sem: APF.getSemantics(), Negative: true) && TLI->has(F: Func))
2994 return ConstantFoldFP(NativeFP: log1p, V: APF, Ty);
2995 break;
2996 case LibFunc_logl:
2997 return nullptr;
2998 case LibFunc_erf:
2999 case LibFunc_erff:
3000 if (TLI->has(F: Func))
3001 return ConstantFoldFP(NativeFP: erf, V: APF, Ty);
3002 break;
3003 case LibFunc_nearbyint:
3004 case LibFunc_nearbyintf:
3005 case LibFunc_rint:
3006 case LibFunc_rintf:
3007 case LibFunc_roundeven:
3008 case LibFunc_roundevenf:
3009 if (TLI->has(F: Func)) {
3010 U.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
3011 return ConstantFP::get(Ty, V: U);
3012 }
3013 break;
3014 case LibFunc_round:
3015 case LibFunc_roundf:
3016 if (TLI->has(F: Func)) {
3017 U.roundToIntegral(RM: APFloat::rmNearestTiesToAway);
3018 return ConstantFP::get(Ty, V: U);
3019 }
3020 break;
3021 case LibFunc_sin:
3022 case LibFunc_sinf:
3023 if (TLI->has(F: Func))
3024 return ConstantFoldFP(NativeFP: sin, V: APF, Ty);
3025 break;
3026 case LibFunc_sinh:
3027 case LibFunc_sinhf:
3028 case LibFunc_sinh_finite:
3029 case LibFunc_sinhf_finite:
3030 if (TLI->has(F: Func))
3031 return ConstantFoldFP(NativeFP: sinh, V: APF, Ty);
3032 break;
3033 case LibFunc_sqrt:
3034 case LibFunc_sqrtf:
3035 if (!APF.isNegative() && TLI->has(F: Func))
3036 return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty);
3037 break;
3038 case LibFunc_tan:
3039 case LibFunc_tanf:
3040 if (TLI->has(F: Func))
3041 return ConstantFoldFP(NativeFP: tan, V: APF, Ty);
3042 break;
3043 case LibFunc_tanh:
3044 case LibFunc_tanhf:
3045 if (TLI->has(F: Func))
3046 return ConstantFoldFP(NativeFP: tanh, V: APF, Ty);
3047 break;
3048 case LibFunc_trunc:
3049 case LibFunc_truncf:
3050 if (TLI->has(F: Func)) {
3051 U.roundToIntegral(RM: APFloat::rmTowardZero);
3052 return ConstantFP::get(Ty, V: U);
3053 }
3054 break;
3055 }
3056 return nullptr;
3057 }
3058
3059 if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) {
3060 switch (IntrinsicID) {
3061 case Intrinsic::bswap:
3062 return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().byteSwap());
3063 case Intrinsic::ctpop:
3064 return ConstantInt::get(Ty, V: Op->getValue().popcount());
3065 case Intrinsic::bitreverse:
3066 return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().reverseBits());
3067 case Intrinsic::amdgcn_s_wqm: {
3068 uint64_t Val = Op->getZExtValue();
3069 Val |= (Val & 0x5555555555555555ULL) << 1 |
3070 ((Val >> 1) & 0x5555555555555555ULL);
3071 Val |= (Val & 0x3333333333333333ULL) << 2 |
3072 ((Val >> 2) & 0x3333333333333333ULL);
3073 return ConstantInt::get(Ty, V: Val);
3074 }
3075
3076 case Intrinsic::amdgcn_s_quadmask: {
3077 uint64_t Val = Op->getZExtValue();
3078 uint64_t QuadMask = 0;
3079 for (unsigned I = 0; I < Op->getBitWidth() / 4; ++I, Val >>= 4) {
3080 if (!(Val & 0xF))
3081 continue;
3082
3083 QuadMask |= (1ULL << I);
3084 }
3085 return ConstantInt::get(Ty, V: QuadMask);
3086 }
3087
3088 case Intrinsic::amdgcn_s_bitreplicate: {
3089 uint64_t Val = Op->getZExtValue();
3090 Val = (Val & 0x000000000000FFFFULL) | (Val & 0x00000000FFFF0000ULL) << 16;
3091 Val = (Val & 0x000000FF000000FFULL) | (Val & 0x0000FF000000FF00ULL) << 8;
3092 Val = (Val & 0x000F000F000F000FULL) | (Val & 0x00F000F000F000F0ULL) << 4;
3093 Val = (Val & 0x0303030303030303ULL) | (Val & 0x0C0C0C0C0C0C0C0CULL) << 2;
3094 Val = (Val & 0x1111111111111111ULL) | (Val & 0x2222222222222222ULL) << 1;
3095 Val = Val | Val << 1;
3096 return ConstantInt::get(Ty, V: Val);
3097 }
3098 }
3099 }
3100
3101 if (Operands[0]->getType()->isVectorTy()) {
3102 auto *Op = cast<Constant>(Val: Operands[0]);
3103 switch (IntrinsicID) {
3104 default: break;
3105 case Intrinsic::vector_reduce_add:
3106 case Intrinsic::vector_reduce_mul:
3107 case Intrinsic::vector_reduce_and:
3108 case Intrinsic::vector_reduce_or:
3109 case Intrinsic::vector_reduce_xor:
3110 case Intrinsic::vector_reduce_smin:
3111 case Intrinsic::vector_reduce_smax:
3112 case Intrinsic::vector_reduce_umin:
3113 case Intrinsic::vector_reduce_umax:
3114 if (Constant *C = constantFoldVectorReduce(IID: IntrinsicID, Op: Operands[0]))
3115 return C;
3116 break;
3117 case Intrinsic::x86_sse_cvtss2si:
3118 case Intrinsic::x86_sse_cvtss2si64:
3119 case Intrinsic::x86_sse2_cvtsd2si:
3120 case Intrinsic::x86_sse2_cvtsd2si64:
3121 if (ConstantFP *FPOp =
3122 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3123 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3124 /*roundTowardZero=*/false, Ty,
3125 /*IsSigned*/true);
3126 break;
3127 case Intrinsic::x86_sse_cvttss2si:
3128 case Intrinsic::x86_sse_cvttss2si64:
3129 case Intrinsic::x86_sse2_cvttsd2si:
3130 case Intrinsic::x86_sse2_cvttsd2si64:
3131 if (ConstantFP *FPOp =
3132 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3133 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3134 /*roundTowardZero=*/true, Ty,
3135 /*IsSigned*/true);
3136 break;
3137
3138 case Intrinsic::wasm_anytrue:
3139 return Op->isNullValue() ? ConstantInt::get(Ty, V: 0)
3140 : ConstantInt::get(Ty, V: 1);
3141
3142 case Intrinsic::wasm_alltrue:
3143 // Check each element individually
3144 unsigned E = cast<FixedVectorType>(Val: Op->getType())->getNumElements();
3145 for (unsigned I = 0; I != E; ++I) {
3146 Constant *Elt = Op->getAggregateElement(Elt: I);
3147 // Return false as soon as we find a non-true element.
3148 if (Elt && Elt->isNullValue())
3149 return ConstantInt::get(Ty, V: 0);
3150 // Bail as soon as we find an element we cannot prove to be true.
3151 if (!Elt || !isa<ConstantInt>(Val: Elt))
3152 return nullptr;
3153 }
3154
3155 return ConstantInt::get(Ty, V: 1);
3156 }
3157 }
3158
3159 return nullptr;
3160}
3161
3162static Constant *evaluateCompare(const APFloat &Op1, const APFloat &Op2,
3163 const ConstrainedFPIntrinsic *Call) {
3164 APFloat::opStatus St = APFloat::opOK;
3165 auto *FCmp = cast<ConstrainedFPCmpIntrinsic>(Val: Call);
3166 FCmpInst::Predicate Cond = FCmp->getPredicate();
3167 if (FCmp->isSignaling()) {
3168 if (Op1.isNaN() || Op2.isNaN())
3169 St = APFloat::opInvalidOp;
3170 } else {
3171 if (Op1.isSignaling() || Op2.isSignaling())
3172 St = APFloat::opInvalidOp;
3173 }
3174 bool Result = FCmpInst::compare(LHS: Op1, RHS: Op2, Pred: Cond);
3175 if (mayFoldConstrained(CI: const_cast<ConstrainedFPCmpIntrinsic *>(FCmp), St))
3176 return ConstantInt::get(Ty: Call->getType()->getScalarType(), V: Result);
3177 return nullptr;
3178}
3179
3180static Constant *ConstantFoldLibCall2(StringRef Name, Type *Ty,
3181 ArrayRef<Constant *> Operands,
3182 const TargetLibraryInfo *TLI) {
3183 if (!TLI)
3184 return nullptr;
3185
3186 LibFunc Func = NotLibFunc;
3187 if (!TLI->getLibFunc(funcName: Name, F&: Func))
3188 return nullptr;
3189
3190 const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0]);
3191 if (!Op1)
3192 return nullptr;
3193
3194 const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1]);
3195 if (!Op2)
3196 return nullptr;
3197
3198 const APFloat &Op1V = Op1->getValueAPF();
3199 const APFloat &Op2V = Op2->getValueAPF();
3200
3201 switch (Func) {
3202 default:
3203 break;
3204 case LibFunc_pow:
3205 case LibFunc_powf:
3206 case LibFunc_pow_finite:
3207 case LibFunc_powf_finite:
3208 if (TLI->has(F: Func))
3209 return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty);
3210 break;
3211 case LibFunc_fmod:
3212 case LibFunc_fmodf:
3213 if (TLI->has(F: Func)) {
3214 APFloat V = Op1->getValueAPF();
3215 if (APFloat::opStatus::opOK == V.mod(RHS: Op2->getValueAPF()))
3216 return ConstantFP::get(Ty, V);
3217 }
3218 break;
3219 case LibFunc_remainder:
3220 case LibFunc_remainderf:
3221 if (TLI->has(F: Func)) {
3222 APFloat V = Op1->getValueAPF();
3223 if (APFloat::opStatus::opOK == V.remainder(RHS: Op2->getValueAPF()))
3224 return ConstantFP::get(Ty, V);
3225 }
3226 break;
3227 case LibFunc_atan2:
3228 case LibFunc_atan2f:
3229 // atan2(+/-0.0, +/-0.0) is known to raise an exception on some libm
3230 // (Solaris), so we do not assume a known result for that.
3231 if (Op1V.isZero() && Op2V.isZero())
3232 return nullptr;
3233 [[fallthrough]];
3234 case LibFunc_atan2_finite:
3235 case LibFunc_atan2f_finite:
3236 if (TLI->has(F: Func))
3237 return ConstantFoldBinaryFP(NativeFP: atan2, V: Op1V, W: Op2V, Ty);
3238 break;
3239 }
3240
3241 return nullptr;
3242}
3243
3244static Constant *ConstantFoldIntrinsicCall2(Intrinsic::ID IntrinsicID, Type *Ty,
3245 ArrayRef<Constant *> Operands,
3246 const CallBase *Call) {
3247 assert(Operands.size() == 2 && "Wrong number of operands.");
3248
3249 if (Ty->isFloatingPointTy()) {
3250 // TODO: We should have undef handling for all of the FP intrinsics that
3251 // are attempted to be folded in this function.
3252 bool IsOp0Undef = isa<UndefValue>(Val: Operands[0]);
3253 bool IsOp1Undef = isa<UndefValue>(Val: Operands[1]);
3254 switch (IntrinsicID) {
3255 case Intrinsic::maxnum:
3256 case Intrinsic::minnum:
3257 case Intrinsic::maximum:
3258 case Intrinsic::minimum:
3259 case Intrinsic::maximumnum:
3260 case Intrinsic::minimumnum:
3261 case Intrinsic::nvvm_fmax_d:
3262 case Intrinsic::nvvm_fmin_d:
3263 // If one argument is undef, return the other argument.
3264 if (IsOp0Undef)
3265 return Operands[1];
3266 if (IsOp1Undef)
3267 return Operands[0];
3268 break;
3269
3270 case Intrinsic::nvvm_fmax_f:
3271 case Intrinsic::nvvm_fmax_ftz_f:
3272 case Intrinsic::nvvm_fmax_ftz_nan_f:
3273 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
3274 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
3275 case Intrinsic::nvvm_fmax_nan_f:
3276 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
3277 case Intrinsic::nvvm_fmax_xorsign_abs_f:
3278
3279 case Intrinsic::nvvm_fmin_f:
3280 case Intrinsic::nvvm_fmin_ftz_f:
3281 case Intrinsic::nvvm_fmin_ftz_nan_f:
3282 case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
3283 case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
3284 case Intrinsic::nvvm_fmin_nan_f:
3285 case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
3286 case Intrinsic::nvvm_fmin_xorsign_abs_f:
3287 // If one arg is undef, the other arg can be returned only if it is
3288 // constant, as we may need to flush it to sign-preserving zero or
3289 // canonicalize the NaN.
3290 if (!IsOp0Undef && !IsOp1Undef)
3291 break;
3292 if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[IsOp0Undef ? 1 : 0])) {
3293 if (Op->isNaN()) {
3294 APInt NVCanonicalNaN(32, 0x7fffffff);
3295 return ConstantFP::get(
3296 Ty, V: APFloat(Ty->getFltSemantics(), NVCanonicalNaN));
3297 }
3298 if (nvvm::FMinFMaxShouldFTZ(IntrinsicID))
3299 return ConstantFP::get(Ty, V: FTZPreserveSign(V: Op->getValueAPF()));
3300 else
3301 return Op;
3302 }
3303 break;
3304 }
3305 }
3306
3307 if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) {
3308 const APFloat &Op1V = Op1->getValueAPF();
3309
3310 if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) {
3311 if (Op2->getType() != Op1->getType())
3312 return nullptr;
3313 const APFloat &Op2V = Op2->getValueAPF();
3314
3315 if (const auto *ConstrIntr =
3316 dyn_cast_if_present<ConstrainedFPIntrinsic>(Val: Call)) {
3317 RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr);
3318 APFloat Res = Op1V;
3319 APFloat::opStatus St;
3320 switch (IntrinsicID) {
3321 default:
3322 return nullptr;
3323 case Intrinsic::experimental_constrained_fadd:
3324 St = Res.add(RHS: Op2V, RM);
3325 break;
3326 case Intrinsic::experimental_constrained_fsub:
3327 St = Res.subtract(RHS: Op2V, RM);
3328 break;
3329 case Intrinsic::experimental_constrained_fmul:
3330 St = Res.multiply(RHS: Op2V, RM);
3331 break;
3332 case Intrinsic::experimental_constrained_fdiv:
3333 St = Res.divide(RHS: Op2V, RM);
3334 break;
3335 case Intrinsic::experimental_constrained_frem:
3336 St = Res.mod(RHS: Op2V);
3337 break;
3338 case Intrinsic::experimental_constrained_fcmp:
3339 case Intrinsic::experimental_constrained_fcmps:
3340 return evaluateCompare(Op1: Op1V, Op2: Op2V, Call: ConstrIntr);
3341 }
3342 if (mayFoldConstrained(CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr),
3343 St))
3344 return ConstantFP::get(Ty, V: Res);
3345 return nullptr;
3346 }
3347
3348 switch (IntrinsicID) {
3349 default:
3350 break;
3351 case Intrinsic::copysign:
3352 return ConstantFP::get(Ty, V: APFloat::copySign(Value: Op1V, Sign: Op2V));
3353 case Intrinsic::minnum:
3354 return ConstantFP::get(Ty, V: minnum(A: Op1V, B: Op2V));
3355 case Intrinsic::maxnum:
3356 return ConstantFP::get(Ty, V: maxnum(A: Op1V, B: Op2V));
3357 case Intrinsic::minimum:
3358 return ConstantFP::get(Ty, V: minimum(A: Op1V, B: Op2V));
3359 case Intrinsic::maximum:
3360 return ConstantFP::get(Ty, V: maximum(A: Op1V, B: Op2V));
3361 case Intrinsic::minimumnum:
3362 return ConstantFP::get(Ty, V: minimumnum(A: Op1V, B: Op2V));
3363 case Intrinsic::maximumnum:
3364 return ConstantFP::get(Ty, V: maximumnum(A: Op1V, B: Op2V));
3365
3366 case Intrinsic::nvvm_fmax_d:
3367 case Intrinsic::nvvm_fmax_f:
3368 case Intrinsic::nvvm_fmax_ftz_f:
3369 case Intrinsic::nvvm_fmax_ftz_nan_f:
3370 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
3371 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
3372 case Intrinsic::nvvm_fmax_nan_f:
3373 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
3374 case Intrinsic::nvvm_fmax_xorsign_abs_f:
3375
3376 case Intrinsic::nvvm_fmin_d:
3377 case Intrinsic::nvvm_fmin_f:
3378 case Intrinsic::nvvm_fmin_ftz_f:
3379 case Intrinsic::nvvm_fmin_ftz_nan_f:
3380 case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
3381 case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
3382 case Intrinsic::nvvm_fmin_nan_f:
3383 case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
3384 case Intrinsic::nvvm_fmin_xorsign_abs_f: {
3385
3386 bool ShouldCanonicalizeNaNs = !(IntrinsicID == Intrinsic::nvvm_fmax_d ||
3387 IntrinsicID == Intrinsic::nvvm_fmin_d);
3388 bool IsFTZ = nvvm::FMinFMaxShouldFTZ(IntrinsicID);
3389 bool IsNaNPropagating = nvvm::FMinFMaxPropagatesNaNs(IntrinsicID);
3390 bool IsXorSignAbs = nvvm::FMinFMaxIsXorSignAbs(IntrinsicID);
3391
3392 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3393 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3394
3395 bool XorSign = false;
3396 if (IsXorSignAbs) {
3397 XorSign = A.isNegative() ^ B.isNegative();
3398 A = abs(X: A);
3399 B = abs(X: B);
3400 }
3401
3402 bool IsFMax = false;
3403 switch (IntrinsicID) {
3404 case Intrinsic::nvvm_fmax_d:
3405 case Intrinsic::nvvm_fmax_f:
3406 case Intrinsic::nvvm_fmax_ftz_f:
3407 case Intrinsic::nvvm_fmax_ftz_nan_f:
3408 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
3409 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
3410 case Intrinsic::nvvm_fmax_nan_f:
3411 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
3412 case Intrinsic::nvvm_fmax_xorsign_abs_f:
3413 IsFMax = true;
3414 break;
3415 }
3416 APFloat Res = IsFMax ? maximum(A, B) : minimum(A, B);
3417
3418 if (ShouldCanonicalizeNaNs) {
3419 APFloat NVCanonicalNaN(Res.getSemantics(), APInt(32, 0x7fffffff));
3420 if (A.isNaN() && B.isNaN())
3421 return ConstantFP::get(Ty, V: NVCanonicalNaN);
3422 else if (IsNaNPropagating && (A.isNaN() || B.isNaN()))
3423 return ConstantFP::get(Ty, V: NVCanonicalNaN);
3424 }
3425
3426 if (A.isNaN() && B.isNaN())
3427 return Operands[1];
3428 else if (A.isNaN())
3429 Res = B;
3430 else if (B.isNaN())
3431 Res = A;
3432
3433 if (IsXorSignAbs && XorSign != Res.isNegative())
3434 Res.changeSign();
3435
3436 return ConstantFP::get(Ty, V: Res);
3437 }
3438
3439 case Intrinsic::nvvm_add_rm_f:
3440 case Intrinsic::nvvm_add_rn_f:
3441 case Intrinsic::nvvm_add_rp_f:
3442 case Intrinsic::nvvm_add_rz_f:
3443 case Intrinsic::nvvm_add_rm_d:
3444 case Intrinsic::nvvm_add_rn_d:
3445 case Intrinsic::nvvm_add_rp_d:
3446 case Intrinsic::nvvm_add_rz_d:
3447 case Intrinsic::nvvm_add_rm_ftz_f:
3448 case Intrinsic::nvvm_add_rn_ftz_f:
3449 case Intrinsic::nvvm_add_rp_ftz_f:
3450 case Intrinsic::nvvm_add_rz_ftz_f: {
3451
3452 bool IsFTZ = nvvm::FAddShouldFTZ(IntrinsicID);
3453 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3454 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3455
3456 APFloat::roundingMode RoundMode =
3457 nvvm::GetFAddRoundingMode(IntrinsicID);
3458
3459 APFloat Res = A;
3460 APFloat::opStatus Status = Res.add(RHS: B, RM: RoundMode);
3461
3462 if (!Res.isNaN() &&
3463 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3464 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3465 return ConstantFP::get(Ty, V: Res);
3466 }
3467 return nullptr;
3468 }
3469
3470 case Intrinsic::nvvm_mul_rm_f:
3471 case Intrinsic::nvvm_mul_rn_f:
3472 case Intrinsic::nvvm_mul_rp_f:
3473 case Intrinsic::nvvm_mul_rz_f:
3474 case Intrinsic::nvvm_mul_rm_d:
3475 case Intrinsic::nvvm_mul_rn_d:
3476 case Intrinsic::nvvm_mul_rp_d:
3477 case Intrinsic::nvvm_mul_rz_d:
3478 case Intrinsic::nvvm_mul_rm_ftz_f:
3479 case Intrinsic::nvvm_mul_rn_ftz_f:
3480 case Intrinsic::nvvm_mul_rp_ftz_f:
3481 case Intrinsic::nvvm_mul_rz_ftz_f: {
3482
3483 bool IsFTZ = nvvm::FMulShouldFTZ(IntrinsicID);
3484 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3485 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3486
3487 APFloat::roundingMode RoundMode =
3488 nvvm::GetFMulRoundingMode(IntrinsicID);
3489
3490 APFloat Res = A;
3491 APFloat::opStatus Status = Res.multiply(RHS: B, RM: RoundMode);
3492
3493 if (!Res.isNaN() &&
3494 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3495 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3496 return ConstantFP::get(Ty, V: Res);
3497 }
3498 return nullptr;
3499 }
3500
3501 case Intrinsic::nvvm_div_rm_f:
3502 case Intrinsic::nvvm_div_rn_f:
3503 case Intrinsic::nvvm_div_rp_f:
3504 case Intrinsic::nvvm_div_rz_f:
3505 case Intrinsic::nvvm_div_rm_d:
3506 case Intrinsic::nvvm_div_rn_d:
3507 case Intrinsic::nvvm_div_rp_d:
3508 case Intrinsic::nvvm_div_rz_d:
3509 case Intrinsic::nvvm_div_rm_ftz_f:
3510 case Intrinsic::nvvm_div_rn_ftz_f:
3511 case Intrinsic::nvvm_div_rp_ftz_f:
3512 case Intrinsic::nvvm_div_rz_ftz_f: {
3513 bool IsFTZ = nvvm::FDivShouldFTZ(IntrinsicID);
3514 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3515 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3516 APFloat::roundingMode RoundMode =
3517 nvvm::GetFDivRoundingMode(IntrinsicID);
3518
3519 APFloat Res = A;
3520 APFloat::opStatus Status = Res.divide(RHS: B, RM: RoundMode);
3521 if (!Res.isNaN() &&
3522 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3523 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3524 return ConstantFP::get(Ty, V: Res);
3525 }
3526 return nullptr;
3527 }
3528 }
3529
3530 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
3531 return nullptr;
3532
3533 switch (IntrinsicID) {
3534 default:
3535 break;
3536 case Intrinsic::pow:
3537 return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty);
3538 case Intrinsic::amdgcn_fmul_legacy:
3539 // The legacy behaviour is that multiplying +/- 0.0 by anything, even
3540 // NaN or infinity, gives +0.0.
3541 if (Op1V.isZero() || Op2V.isZero())
3542 return ConstantFP::getZero(Ty);
3543 return ConstantFP::get(Ty, V: Op1V * Op2V);
3544 }
3545
3546 } else if (auto *Op2C = dyn_cast<ConstantInt>(Val: Operands[1])) {
3547 switch (IntrinsicID) {
3548 case Intrinsic::ldexp: {
3549 return ConstantFP::get(
3550 Context&: Ty->getContext(),
3551 V: scalbn(X: Op1V, Exp: Op2C->getSExtValue(), RM: APFloat::rmNearestTiesToEven));
3552 }
3553 case Intrinsic::is_fpclass: {
3554 FPClassTest Mask = static_cast<FPClassTest>(Op2C->getZExtValue());
3555 bool Result =
3556 ((Mask & fcSNan) && Op1V.isNaN() && Op1V.isSignaling()) ||
3557 ((Mask & fcQNan) && Op1V.isNaN() && !Op1V.isSignaling()) ||
3558 ((Mask & fcNegInf) && Op1V.isNegInfinity()) ||
3559 ((Mask & fcNegNormal) && Op1V.isNormal() && Op1V.isNegative()) ||
3560 ((Mask & fcNegSubnormal) && Op1V.isDenormal() && Op1V.isNegative()) ||
3561 ((Mask & fcNegZero) && Op1V.isZero() && Op1V.isNegative()) ||
3562 ((Mask & fcPosZero) && Op1V.isZero() && !Op1V.isNegative()) ||
3563 ((Mask & fcPosSubnormal) && Op1V.isDenormal() && !Op1V.isNegative()) ||
3564 ((Mask & fcPosNormal) && Op1V.isNormal() && !Op1V.isNegative()) ||
3565 ((Mask & fcPosInf) && Op1V.isPosInfinity());
3566 return ConstantInt::get(Ty, V: Result);
3567 }
3568 case Intrinsic::powi: {
3569 int Exp = static_cast<int>(Op2C->getSExtValue());
3570 switch (Ty->getTypeID()) {
3571 case Type::HalfTyID:
3572 case Type::FloatTyID: {
3573 APFloat Res(static_cast<float>(std::pow(x: Op1V.convertToFloat(), y: Exp)));
3574 if (Ty->isHalfTy()) {
3575 bool Unused;
3576 Res.convert(ToSemantics: APFloat::IEEEhalf(), RM: APFloat::rmNearestTiesToEven,
3577 losesInfo: &Unused);
3578 }
3579 return ConstantFP::get(Ty, V: Res);
3580 }
3581 case Type::DoubleTyID:
3582 return ConstantFP::get(Ty, V: std::pow(x: Op1V.convertToDouble(), y: Exp));
3583 default:
3584 return nullptr;
3585 }
3586 }
3587 default:
3588 break;
3589 }
3590 }
3591 return nullptr;
3592 }
3593
3594 if (Operands[0]->getType()->isIntegerTy() &&
3595 Operands[1]->getType()->isIntegerTy()) {
3596 const APInt *C0, *C1;
3597 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
3598 !getConstIntOrUndef(Op: Operands[1], C&: C1))
3599 return nullptr;
3600
3601 switch (IntrinsicID) {
3602 default: break;
3603 case Intrinsic::smax:
3604 case Intrinsic::smin:
3605 case Intrinsic::umax:
3606 case Intrinsic::umin:
3607 if (!C0 && !C1)
3608 return UndefValue::get(T: Ty);
3609 if (!C0 || !C1)
3610 return MinMaxIntrinsic::getSaturationPoint(ID: IntrinsicID, Ty);
3611 return ConstantInt::get(
3612 Ty, V: ICmpInst::compare(LHS: *C0, RHS: *C1,
3613 Pred: MinMaxIntrinsic::getPredicate(ID: IntrinsicID))
3614 ? *C0
3615 : *C1);
3616
3617 case Intrinsic::scmp:
3618 case Intrinsic::ucmp:
3619 if (!C0 || !C1)
3620 return ConstantInt::get(Ty, V: 0);
3621
3622 int Res;
3623 if (IntrinsicID == Intrinsic::scmp)
3624 Res = C0->sgt(RHS: *C1) ? 1 : C0->slt(RHS: *C1) ? -1 : 0;
3625 else
3626 Res = C0->ugt(RHS: *C1) ? 1 : C0->ult(RHS: *C1) ? -1 : 0;
3627 return ConstantInt::get(Ty, V: Res, /*IsSigned=*/true);
3628
3629 case Intrinsic::usub_with_overflow:
3630 case Intrinsic::ssub_with_overflow:
3631 // X - undef -> { 0, false }
3632 // undef - X -> { 0, false }
3633 if (!C0 || !C1)
3634 return Constant::getNullValue(Ty);
3635 [[fallthrough]];
3636 case Intrinsic::uadd_with_overflow:
3637 case Intrinsic::sadd_with_overflow:
3638 // X + undef -> { -1, false }
3639 // undef + x -> { -1, false }
3640 if (!C0 || !C1) {
3641 return ConstantStruct::get(
3642 T: cast<StructType>(Val: Ty),
3643 V: {Constant::getAllOnesValue(Ty: Ty->getStructElementType(N: 0)),
3644 Constant::getNullValue(Ty: Ty->getStructElementType(N: 1))});
3645 }
3646 [[fallthrough]];
3647 case Intrinsic::smul_with_overflow:
3648 case Intrinsic::umul_with_overflow: {
3649 // undef * X -> { 0, false }
3650 // X * undef -> { 0, false }
3651 if (!C0 || !C1)
3652 return Constant::getNullValue(Ty);
3653
3654 APInt Res;
3655 bool Overflow;
3656 switch (IntrinsicID) {
3657 default: llvm_unreachable("Invalid case");
3658 case Intrinsic::sadd_with_overflow:
3659 Res = C0->sadd_ov(RHS: *C1, Overflow);
3660 break;
3661 case Intrinsic::uadd_with_overflow:
3662 Res = C0->uadd_ov(RHS: *C1, Overflow);
3663 break;
3664 case Intrinsic::ssub_with_overflow:
3665 Res = C0->ssub_ov(RHS: *C1, Overflow);
3666 break;
3667 case Intrinsic::usub_with_overflow:
3668 Res = C0->usub_ov(RHS: *C1, Overflow);
3669 break;
3670 case Intrinsic::smul_with_overflow:
3671 Res = C0->smul_ov(RHS: *C1, Overflow);
3672 break;
3673 case Intrinsic::umul_with_overflow:
3674 Res = C0->umul_ov(RHS: *C1, Overflow);
3675 break;
3676 }
3677 Constant *Ops[] = {
3678 ConstantInt::get(Context&: Ty->getContext(), V: Res),
3679 ConstantInt::get(Ty: Type::getInt1Ty(C&: Ty->getContext()), V: Overflow)
3680 };
3681 return ConstantStruct::get(T: cast<StructType>(Val: Ty), V: Ops);
3682 }
3683 case Intrinsic::uadd_sat:
3684 case Intrinsic::sadd_sat:
3685 if (!C0 && !C1)
3686 return UndefValue::get(T: Ty);
3687 if (!C0 || !C1)
3688 return Constant::getAllOnesValue(Ty);
3689 if (IntrinsicID == Intrinsic::uadd_sat)
3690 return ConstantInt::get(Ty, V: C0->uadd_sat(RHS: *C1));
3691 else
3692 return ConstantInt::get(Ty, V: C0->sadd_sat(RHS: *C1));
3693 case Intrinsic::usub_sat:
3694 case Intrinsic::ssub_sat:
3695 if (!C0 && !C1)
3696 return UndefValue::get(T: Ty);
3697 if (!C0 || !C1)
3698 return Constant::getNullValue(Ty);
3699 if (IntrinsicID == Intrinsic::usub_sat)
3700 return ConstantInt::get(Ty, V: C0->usub_sat(RHS: *C1));
3701 else
3702 return ConstantInt::get(Ty, V: C0->ssub_sat(RHS: *C1));
3703 case Intrinsic::cttz:
3704 case Intrinsic::ctlz:
3705 assert(C1 && "Must be constant int");
3706
3707 // cttz(0, 1) and ctlz(0, 1) are poison.
3708 if (C1->isOne() && (!C0 || C0->isZero()))
3709 return PoisonValue::get(T: Ty);
3710 if (!C0)
3711 return Constant::getNullValue(Ty);
3712 if (IntrinsicID == Intrinsic::cttz)
3713 return ConstantInt::get(Ty, V: C0->countr_zero());
3714 else
3715 return ConstantInt::get(Ty, V: C0->countl_zero());
3716
3717 case Intrinsic::abs:
3718 assert(C1 && "Must be constant int");
3719 assert((C1->isOne() || C1->isZero()) && "Must be 0 or 1");
3720
3721 // Undef or minimum val operand with poison min --> poison
3722 if (C1->isOne() && (!C0 || C0->isMinSignedValue()))
3723 return PoisonValue::get(T: Ty);
3724
3725 // Undef operand with no poison min --> 0 (sign bit must be clear)
3726 if (!C0)
3727 return Constant::getNullValue(Ty);
3728
3729 return ConstantInt::get(Ty, V: C0->abs());
3730 case Intrinsic::amdgcn_wave_reduce_umin:
3731 case Intrinsic::amdgcn_wave_reduce_umax:
3732 case Intrinsic::amdgcn_wave_reduce_max:
3733 case Intrinsic::amdgcn_wave_reduce_min:
3734 case Intrinsic::amdgcn_wave_reduce_add:
3735 case Intrinsic::amdgcn_wave_reduce_sub:
3736 case Intrinsic::amdgcn_wave_reduce_and:
3737 case Intrinsic::amdgcn_wave_reduce_or:
3738 case Intrinsic::amdgcn_wave_reduce_xor:
3739 return dyn_cast<Constant>(Val: Operands[0]);
3740 }
3741
3742 return nullptr;
3743 }
3744
3745 // Support ConstantVector in case we have an Undef in the top.
3746 if ((isa<ConstantVector>(Val: Operands[0]) ||
3747 isa<ConstantDataVector>(Val: Operands[0])) &&
3748 // Check for default rounding mode.
3749 // FIXME: Support other rounding modes?
3750 isa<ConstantInt>(Val: Operands[1]) &&
3751 cast<ConstantInt>(Val: Operands[1])->getValue() == 4) {
3752 auto *Op = cast<Constant>(Val: Operands[0]);
3753 switch (IntrinsicID) {
3754 default: break;
3755 case Intrinsic::x86_avx512_vcvtss2si32:
3756 case Intrinsic::x86_avx512_vcvtss2si64:
3757 case Intrinsic::x86_avx512_vcvtsd2si32:
3758 case Intrinsic::x86_avx512_vcvtsd2si64:
3759 if (ConstantFP *FPOp =
3760 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3761 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3762 /*roundTowardZero=*/false, Ty,
3763 /*IsSigned*/true);
3764 break;
3765 case Intrinsic::x86_avx512_vcvtss2usi32:
3766 case Intrinsic::x86_avx512_vcvtss2usi64:
3767 case Intrinsic::x86_avx512_vcvtsd2usi32:
3768 case Intrinsic::x86_avx512_vcvtsd2usi64:
3769 if (ConstantFP *FPOp =
3770 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3771 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3772 /*roundTowardZero=*/false, Ty,
3773 /*IsSigned*/false);
3774 break;
3775 case Intrinsic::x86_avx512_cvttss2si:
3776 case Intrinsic::x86_avx512_cvttss2si64:
3777 case Intrinsic::x86_avx512_cvttsd2si:
3778 case Intrinsic::x86_avx512_cvttsd2si64:
3779 if (ConstantFP *FPOp =
3780 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3781 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3782 /*roundTowardZero=*/true, Ty,
3783 /*IsSigned*/true);
3784 break;
3785 case Intrinsic::x86_avx512_cvttss2usi:
3786 case Intrinsic::x86_avx512_cvttss2usi64:
3787 case Intrinsic::x86_avx512_cvttsd2usi:
3788 case Intrinsic::x86_avx512_cvttsd2usi64:
3789 if (ConstantFP *FPOp =
3790 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3791 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3792 /*roundTowardZero=*/true, Ty,
3793 /*IsSigned*/false);
3794 break;
3795 }
3796 }
3797
3798 if (IntrinsicID == Intrinsic::experimental_cttz_elts) {
3799 auto *FVTy = dyn_cast<FixedVectorType>(Val: Operands[0]->getType());
3800 bool ZeroIsPoison = cast<ConstantInt>(Val: Operands[1])->isOne();
3801 if (!FVTy)
3802 return nullptr;
3803 unsigned Width = Ty->getIntegerBitWidth();
3804 if (APInt::getMaxValue(numBits: Width).ult(RHS: FVTy->getNumElements()))
3805 return PoisonValue::get(T: Ty);
3806 for (unsigned I = 0; I < FVTy->getNumElements(); ++I) {
3807 Constant *Elt = Operands[0]->getAggregateElement(Elt: I);
3808 if (!Elt)
3809 return nullptr;
3810 if (isa<UndefValue>(Val: Elt) || Elt->isNullValue())
3811 continue;
3812 return ConstantInt::get(Ty, V: I);
3813 }
3814 if (ZeroIsPoison)
3815 return PoisonValue::get(T: Ty);
3816 return ConstantInt::get(Ty, V: FVTy->getNumElements());
3817 }
3818 return nullptr;
3819}
3820
3821static APFloat ConstantFoldAMDGCNCubeIntrinsic(Intrinsic::ID IntrinsicID,
3822 const APFloat &S0,
3823 const APFloat &S1,
3824 const APFloat &S2) {
3825 unsigned ID;
3826 const fltSemantics &Sem = S0.getSemantics();
3827 APFloat MA(Sem), SC(Sem), TC(Sem);
3828 if (abs(X: S2) >= abs(X: S0) && abs(X: S2) >= abs(X: S1)) {
3829 if (S2.isNegative() && S2.isNonZero() && !S2.isNaN()) {
3830 // S2 < 0
3831 ID = 5;
3832 SC = -S0;
3833 } else {
3834 ID = 4;
3835 SC = S0;
3836 }
3837 MA = S2;
3838 TC = -S1;
3839 } else if (abs(X: S1) >= abs(X: S0)) {
3840 if (S1.isNegative() && S1.isNonZero() && !S1.isNaN()) {
3841 // S1 < 0
3842 ID = 3;
3843 TC = -S2;
3844 } else {
3845 ID = 2;
3846 TC = S2;
3847 }
3848 MA = S1;
3849 SC = S0;
3850 } else {
3851 if (S0.isNegative() && S0.isNonZero() && !S0.isNaN()) {
3852 // S0 < 0
3853 ID = 1;
3854 SC = S2;
3855 } else {
3856 ID = 0;
3857 SC = -S2;
3858 }
3859 MA = S0;
3860 TC = -S1;
3861 }
3862 switch (IntrinsicID) {
3863 default:
3864 llvm_unreachable("unhandled amdgcn cube intrinsic");
3865 case Intrinsic::amdgcn_cubeid:
3866 return APFloat(Sem, ID);
3867 case Intrinsic::amdgcn_cubema:
3868 return MA + MA;
3869 case Intrinsic::amdgcn_cubesc:
3870 return SC;
3871 case Intrinsic::amdgcn_cubetc:
3872 return TC;
3873 }
3874}
3875
3876static Constant *ConstantFoldAMDGCNPermIntrinsic(ArrayRef<Constant *> Operands,
3877 Type *Ty) {
3878 const APInt *C0, *C1, *C2;
3879 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
3880 !getConstIntOrUndef(Op: Operands[1], C&: C1) ||
3881 !getConstIntOrUndef(Op: Operands[2], C&: C2))
3882 return nullptr;
3883
3884 if (!C2)
3885 return UndefValue::get(T: Ty);
3886
3887 APInt Val(32, 0);
3888 unsigned NumUndefBytes = 0;
3889 for (unsigned I = 0; I < 32; I += 8) {
3890 unsigned Sel = C2->extractBitsAsZExtValue(numBits: 8, bitPosition: I);
3891 unsigned B = 0;
3892
3893 if (Sel >= 13)
3894 B = 0xff;
3895 else if (Sel == 12)
3896 B = 0x00;
3897 else {
3898 const APInt *Src = ((Sel & 10) == 10 || (Sel & 12) == 4) ? C0 : C1;
3899 if (!Src)
3900 ++NumUndefBytes;
3901 else if (Sel < 8)
3902 B = Src->extractBitsAsZExtValue(numBits: 8, bitPosition: (Sel & 3) * 8);
3903 else
3904 B = Src->extractBitsAsZExtValue(numBits: 1, bitPosition: (Sel & 1) ? 31 : 15) * 0xff;
3905 }
3906
3907 Val.insertBits(SubBits: B, bitPosition: I, numBits: 8);
3908 }
3909
3910 if (NumUndefBytes == 4)
3911 return UndefValue::get(T: Ty);
3912
3913 return ConstantInt::get(Ty, V: Val);
3914}
3915
3916static Constant *ConstantFoldScalarCall3(StringRef Name,
3917 Intrinsic::ID IntrinsicID,
3918 Type *Ty,
3919 ArrayRef<Constant *> Operands,
3920 const TargetLibraryInfo *TLI,
3921 const CallBase *Call) {
3922 assert(Operands.size() == 3 && "Wrong number of operands.");
3923
3924 if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) {
3925 if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) {
3926 if (const auto *Op3 = dyn_cast<ConstantFP>(Val: Operands[2])) {
3927 const APFloat &C1 = Op1->getValueAPF();
3928 const APFloat &C2 = Op2->getValueAPF();
3929 const APFloat &C3 = Op3->getValueAPF();
3930
3931 if (const auto *ConstrIntr = dyn_cast<ConstrainedFPIntrinsic>(Val: Call)) {
3932 RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr);
3933 APFloat Res = C1;
3934 APFloat::opStatus St;
3935 switch (IntrinsicID) {
3936 default:
3937 return nullptr;
3938 case Intrinsic::experimental_constrained_fma:
3939 case Intrinsic::experimental_constrained_fmuladd:
3940 St = Res.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM);
3941 break;
3942 }
3943 if (mayFoldConstrained(
3944 CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), St))
3945 return ConstantFP::get(Ty, V: Res);
3946 return nullptr;
3947 }
3948
3949 switch (IntrinsicID) {
3950 default: break;
3951 case Intrinsic::amdgcn_fma_legacy: {
3952 // The legacy behaviour is that multiplying +/- 0.0 by anything, even
3953 // NaN or infinity, gives +0.0.
3954 if (C1.isZero() || C2.isZero()) {
3955 // It's tempting to just return C3 here, but that would give the
3956 // wrong result if C3 was -0.0.
3957 return ConstantFP::get(Ty, V: APFloat(0.0f) + C3);
3958 }
3959 [[fallthrough]];
3960 }
3961 case Intrinsic::fma:
3962 case Intrinsic::fmuladd: {
3963 APFloat V = C1;
3964 V.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM: APFloat::rmNearestTiesToEven);
3965 return ConstantFP::get(Ty, V);
3966 }
3967
3968 case Intrinsic::nvvm_fma_rm_f:
3969 case Intrinsic::nvvm_fma_rn_f:
3970 case Intrinsic::nvvm_fma_rp_f:
3971 case Intrinsic::nvvm_fma_rz_f:
3972 case Intrinsic::nvvm_fma_rm_d:
3973 case Intrinsic::nvvm_fma_rn_d:
3974 case Intrinsic::nvvm_fma_rp_d:
3975 case Intrinsic::nvvm_fma_rz_d:
3976 case Intrinsic::nvvm_fma_rm_ftz_f:
3977 case Intrinsic::nvvm_fma_rn_ftz_f:
3978 case Intrinsic::nvvm_fma_rp_ftz_f:
3979 case Intrinsic::nvvm_fma_rz_ftz_f: {
3980 bool IsFTZ = nvvm::FMAShouldFTZ(IntrinsicID);
3981 APFloat A = IsFTZ ? FTZPreserveSign(V: C1) : C1;
3982 APFloat B = IsFTZ ? FTZPreserveSign(V: C2) : C2;
3983 APFloat C = IsFTZ ? FTZPreserveSign(V: C3) : C3;
3984
3985 APFloat::roundingMode RoundMode =
3986 nvvm::GetFMARoundingMode(IntrinsicID);
3987
3988 APFloat Res = A;
3989 APFloat::opStatus Status = Res.fusedMultiplyAdd(Multiplicand: B, Addend: C, RM: RoundMode);
3990
3991 if (!Res.isNaN() &&
3992 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3993 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3994 return ConstantFP::get(Ty, V: Res);
3995 }
3996 return nullptr;
3997 }
3998
3999 case Intrinsic::amdgcn_cubeid:
4000 case Intrinsic::amdgcn_cubema:
4001 case Intrinsic::amdgcn_cubesc:
4002 case Intrinsic::amdgcn_cubetc: {
4003 APFloat V = ConstantFoldAMDGCNCubeIntrinsic(IntrinsicID, S0: C1, S1: C2, S2: C3);
4004 return ConstantFP::get(Ty, V);
4005 }
4006 }
4007 }
4008 }
4009 }
4010
4011 if (IntrinsicID == Intrinsic::smul_fix ||
4012 IntrinsicID == Intrinsic::smul_fix_sat) {
4013 const APInt *C0, *C1;
4014 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
4015 !getConstIntOrUndef(Op: Operands[1], C&: C1))
4016 return nullptr;
4017
4018 // undef * C -> 0
4019 // C * undef -> 0
4020 if (!C0 || !C1)
4021 return Constant::getNullValue(Ty);
4022
4023 // This code performs rounding towards negative infinity in case the result
4024 // cannot be represented exactly for the given scale. Targets that do care
4025 // about rounding should use a target hook for specifying how rounding
4026 // should be done, and provide their own folding to be consistent with
4027 // rounding. This is the same approach as used by
4028 // DAGTypeLegalizer::ExpandIntRes_MULFIX.
4029 unsigned Scale = cast<ConstantInt>(Val: Operands[2])->getZExtValue();
4030 unsigned Width = C0->getBitWidth();
4031 assert(Scale < Width && "Illegal scale.");
4032 unsigned ExtendedWidth = Width * 2;
4033 APInt Product =
4034 (C0->sext(width: ExtendedWidth) * C1->sext(width: ExtendedWidth)).ashr(ShiftAmt: Scale);
4035 if (IntrinsicID == Intrinsic::smul_fix_sat) {
4036 APInt Max = APInt::getSignedMaxValue(numBits: Width).sext(width: ExtendedWidth);
4037 APInt Min = APInt::getSignedMinValue(numBits: Width).sext(width: ExtendedWidth);
4038 Product = APIntOps::smin(A: Product, B: Max);
4039 Product = APIntOps::smax(A: Product, B: Min);
4040 }
4041 return ConstantInt::get(Context&: Ty->getContext(), V: Product.sextOrTrunc(width: Width));
4042 }
4043
4044 if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {
4045 const APInt *C0, *C1, *C2;
4046 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
4047 !getConstIntOrUndef(Op: Operands[1], C&: C1) ||
4048 !getConstIntOrUndef(Op: Operands[2], C&: C2))
4049 return nullptr;
4050
4051 bool IsRight = IntrinsicID == Intrinsic::fshr;
4052 if (!C2)
4053 return Operands[IsRight ? 1 : 0];
4054 if (!C0 && !C1)
4055 return UndefValue::get(T: Ty);
4056
4057 // The shift amount is interpreted as modulo the bitwidth. If the shift
4058 // amount is effectively 0, avoid UB due to oversized inverse shift below.
4059 unsigned BitWidth = C2->getBitWidth();
4060 unsigned ShAmt = C2->urem(RHS: BitWidth);
4061 if (!ShAmt)
4062 return Operands[IsRight ? 1 : 0];
4063
4064 // (C0 << ShlAmt) | (C1 >> LshrAmt)
4065 unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt;
4066 unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt;
4067 if (!C0)
4068 return ConstantInt::get(Ty, V: C1->lshr(shiftAmt: LshrAmt));
4069 if (!C1)
4070 return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt));
4071 return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt) | C1->lshr(shiftAmt: LshrAmt));
4072 }
4073
4074 if (IntrinsicID == Intrinsic::amdgcn_perm)
4075 return ConstantFoldAMDGCNPermIntrinsic(Operands, Ty);
4076
4077 return nullptr;
4078}
4079
4080static Constant *ConstantFoldScalarCall(StringRef Name,
4081 Intrinsic::ID IntrinsicID,
4082 Type *Ty,
4083 ArrayRef<Constant *> Operands,
4084 const TargetLibraryInfo *TLI,
4085 const CallBase *Call) {
4086 if (IntrinsicID != Intrinsic::not_intrinsic &&
4087 any_of(Range&: Operands, P: IsaPred<PoisonValue>) &&
4088 intrinsicPropagatesPoison(IID: IntrinsicID))
4089 return PoisonValue::get(T: Ty);
4090
4091 if (Operands.size() == 1)
4092 return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call);
4093
4094 if (Operands.size() == 2) {
4095 if (Constant *FoldedLibCall =
4096 ConstantFoldLibCall2(Name, Ty, Operands, TLI)) {
4097 return FoldedLibCall;
4098 }
4099 return ConstantFoldIntrinsicCall2(IntrinsicID, Ty, Operands, Call);
4100 }
4101
4102 if (Operands.size() == 3)
4103 return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call);
4104
4105 return nullptr;
4106}
4107
4108static Constant *ConstantFoldFixedVectorCall(
4109 StringRef Name, Intrinsic::ID IntrinsicID, FixedVectorType *FVTy,
4110 ArrayRef<Constant *> Operands, const DataLayout &DL,
4111 const TargetLibraryInfo *TLI, const CallBase *Call) {
4112 SmallVector<Constant *, 4> Result(FVTy->getNumElements());
4113 SmallVector<Constant *, 4> Lane(Operands.size());
4114 Type *Ty = FVTy->getElementType();
4115
4116 switch (IntrinsicID) {
4117 case Intrinsic::masked_load: {
4118 auto *SrcPtr = Operands[0];
4119 auto *Mask = Operands[1];
4120 auto *Passthru = Operands[2];
4121
4122 Constant *VecData = ConstantFoldLoadFromConstPtr(C: SrcPtr, Ty: FVTy, DL);
4123
4124 SmallVector<Constant *, 32> NewElements;
4125 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
4126 auto *MaskElt = Mask->getAggregateElement(Elt: I);
4127 if (!MaskElt)
4128 break;
4129 auto *PassthruElt = Passthru->getAggregateElement(Elt: I);
4130 auto *VecElt = VecData ? VecData->getAggregateElement(Elt: I) : nullptr;
4131 if (isa<UndefValue>(Val: MaskElt)) {
4132 if (PassthruElt)
4133 NewElements.push_back(Elt: PassthruElt);
4134 else if (VecElt)
4135 NewElements.push_back(Elt: VecElt);
4136 else
4137 return nullptr;
4138 }
4139 if (MaskElt->isNullValue()) {
4140 if (!PassthruElt)
4141 return nullptr;
4142 NewElements.push_back(Elt: PassthruElt);
4143 } else if (MaskElt->isOneValue()) {
4144 if (!VecElt)
4145 return nullptr;
4146 NewElements.push_back(Elt: VecElt);
4147 } else {
4148 return nullptr;
4149 }
4150 }
4151 if (NewElements.size() != FVTy->getNumElements())
4152 return nullptr;
4153 return ConstantVector::get(V: NewElements);
4154 }
4155 case Intrinsic::arm_mve_vctp8:
4156 case Intrinsic::arm_mve_vctp16:
4157 case Intrinsic::arm_mve_vctp32:
4158 case Intrinsic::arm_mve_vctp64: {
4159 if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) {
4160 unsigned Lanes = FVTy->getNumElements();
4161 uint64_t Limit = Op->getZExtValue();
4162
4163 SmallVector<Constant *, 16> NCs;
4164 for (unsigned i = 0; i < Lanes; i++) {
4165 if (i < Limit)
4166 NCs.push_back(Elt: ConstantInt::getTrue(Ty));
4167 else
4168 NCs.push_back(Elt: ConstantInt::getFalse(Ty));
4169 }
4170 return ConstantVector::get(V: NCs);
4171 }
4172 return nullptr;
4173 }
4174 case Intrinsic::get_active_lane_mask: {
4175 auto *Op0 = dyn_cast<ConstantInt>(Val: Operands[0]);
4176 auto *Op1 = dyn_cast<ConstantInt>(Val: Operands[1]);
4177 if (Op0 && Op1) {
4178 unsigned Lanes = FVTy->getNumElements();
4179 uint64_t Base = Op0->getZExtValue();
4180 uint64_t Limit = Op1->getZExtValue();
4181
4182 SmallVector<Constant *, 16> NCs;
4183 for (unsigned i = 0; i < Lanes; i++) {
4184 if (Base + i < Limit)
4185 NCs.push_back(Elt: ConstantInt::getTrue(Ty));
4186 else
4187 NCs.push_back(Elt: ConstantInt::getFalse(Ty));
4188 }
4189 return ConstantVector::get(V: NCs);
4190 }
4191 return nullptr;
4192 }
4193 case Intrinsic::vector_extract: {
4194 auto *Idx = dyn_cast<ConstantInt>(Val: Operands[1]);
4195 Constant *Vec = Operands[0];
4196 if (!Idx || !isa<FixedVectorType>(Val: Vec->getType()))
4197 return nullptr;
4198
4199 unsigned NumElements = FVTy->getNumElements();
4200 unsigned VecNumElements =
4201 cast<FixedVectorType>(Val: Vec->getType())->getNumElements();
4202 unsigned StartingIndex = Idx->getZExtValue();
4203
4204 // Extracting entire vector is nop
4205 if (NumElements == VecNumElements && StartingIndex == 0)
4206 return Vec;
4207
4208 for (unsigned I = StartingIndex, E = StartingIndex + NumElements; I < E;
4209 ++I) {
4210 Constant *Elt = Vec->getAggregateElement(Elt: I);
4211 if (!Elt)
4212 return nullptr;
4213 Result[I - StartingIndex] = Elt;
4214 }
4215
4216 return ConstantVector::get(V: Result);
4217 }
4218 case Intrinsic::vector_insert: {
4219 Constant *Vec = Operands[0];
4220 Constant *SubVec = Operands[1];
4221 auto *Idx = dyn_cast<ConstantInt>(Val: Operands[2]);
4222 if (!Idx || !isa<FixedVectorType>(Val: Vec->getType()))
4223 return nullptr;
4224
4225 unsigned SubVecNumElements =
4226 cast<FixedVectorType>(Val: SubVec->getType())->getNumElements();
4227 unsigned VecNumElements =
4228 cast<FixedVectorType>(Val: Vec->getType())->getNumElements();
4229 unsigned IdxN = Idx->getZExtValue();
4230 // Replacing entire vector with a subvec is nop
4231 if (SubVecNumElements == VecNumElements && IdxN == 0)
4232 return SubVec;
4233
4234 for (unsigned I = 0; I < VecNumElements; ++I) {
4235 Constant *Elt;
4236 if (I < IdxN + SubVecNumElements)
4237 Elt = SubVec->getAggregateElement(Elt: I - IdxN);
4238 else
4239 Elt = Vec->getAggregateElement(Elt: I);
4240 if (!Elt)
4241 return nullptr;
4242 Result[I] = Elt;
4243 }
4244 return ConstantVector::get(V: Result);
4245 }
4246 case Intrinsic::vector_interleave2:
4247 case Intrinsic::vector_interleave3:
4248 case Intrinsic::vector_interleave4:
4249 case Intrinsic::vector_interleave5:
4250 case Intrinsic::vector_interleave6:
4251 case Intrinsic::vector_interleave7:
4252 case Intrinsic::vector_interleave8: {
4253 unsigned NumElements =
4254 cast<FixedVectorType>(Val: Operands[0]->getType())->getNumElements();
4255 unsigned NumOperands = Operands.size();
4256 for (unsigned I = 0; I < NumElements; ++I) {
4257 for (unsigned J = 0; J < NumOperands; ++J) {
4258 Constant *Elt = Operands[J]->getAggregateElement(Elt: I);
4259 if (!Elt)
4260 return nullptr;
4261 Result[NumOperands * I + J] = Elt;
4262 }
4263 }
4264 return ConstantVector::get(V: Result);
4265 }
4266 case Intrinsic::wasm_dot: {
4267 unsigned NumElements =
4268 cast<FixedVectorType>(Val: Operands[0]->getType())->getNumElements();
4269
4270 assert(NumElements == 8 && Result.size() == 4 &&
4271 "wasm dot takes i16x8 and produces i32x4");
4272 assert(Ty->isIntegerTy());
4273 int32_t MulVector[8];
4274
4275 for (unsigned I = 0; I < NumElements; ++I) {
4276 ConstantInt *Elt0 =
4277 cast<ConstantInt>(Val: Operands[0]->getAggregateElement(Elt: I));
4278 ConstantInt *Elt1 =
4279 cast<ConstantInt>(Val: Operands[1]->getAggregateElement(Elt: I));
4280
4281 MulVector[I] = Elt0->getSExtValue() * Elt1->getSExtValue();
4282 }
4283 for (unsigned I = 0; I < Result.size(); I++) {
4284 int64_t IAdd = (int64_t)MulVector[I * 2] + (int64_t)MulVector[I * 2 + 1];
4285 Result[I] = ConstantInt::getSigned(Ty, V: IAdd, /*ImplicitTrunc=*/true);
4286 }
4287
4288 return ConstantVector::get(V: Result);
4289 }
4290 default:
4291 break;
4292 }
4293
4294 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
4295 // Gather a column of constants.
4296 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
4297 // Some intrinsics use a scalar type for certain arguments.
4298 if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: J, /*TTI=*/nullptr)) {
4299 Lane[J] = Operands[J];
4300 continue;
4301 }
4302
4303 Constant *Agg = Operands[J]->getAggregateElement(Elt: I);
4304 if (!Agg)
4305 return nullptr;
4306
4307 Lane[J] = Agg;
4308 }
4309
4310 // Use the regular scalar folding to simplify this column.
4311 Constant *Folded =
4312 ConstantFoldScalarCall(Name, IntrinsicID, Ty, Operands: Lane, TLI, Call);
4313 if (!Folded)
4314 return nullptr;
4315 Result[I] = Folded;
4316 }
4317
4318 return ConstantVector::get(V: Result);
4319}
4320
4321static Constant *ConstantFoldScalableVectorCall(
4322 StringRef Name, Intrinsic::ID IntrinsicID, ScalableVectorType *SVTy,
4323 ArrayRef<Constant *> Operands, const DataLayout &DL,
4324 const TargetLibraryInfo *TLI, const CallBase *Call) {
4325 switch (IntrinsicID) {
4326 case Intrinsic::aarch64_sve_convert_from_svbool: {
4327 auto *Src = dyn_cast<Constant>(Val: Operands[0]);
4328 if (!Src || !Src->isNullValue())
4329 break;
4330
4331 return ConstantInt::getFalse(Ty: SVTy);
4332 }
4333 case Intrinsic::get_active_lane_mask: {
4334 auto *Op0 = dyn_cast<ConstantInt>(Val: Operands[0]);
4335 auto *Op1 = dyn_cast<ConstantInt>(Val: Operands[1]);
4336 if (Op0 && Op1 && Op0->getValue().uge(RHS: Op1->getValue()))
4337 return ConstantVector::getNullValue(Ty: SVTy);
4338 break;
4339 }
4340 case Intrinsic::vector_interleave2:
4341 case Intrinsic::vector_interleave3:
4342 case Intrinsic::vector_interleave4:
4343 case Intrinsic::vector_interleave5:
4344 case Intrinsic::vector_interleave6:
4345 case Intrinsic::vector_interleave7:
4346 case Intrinsic::vector_interleave8: {
4347 Constant *SplatVal = Operands[0]->getSplatValue();
4348 if (!SplatVal)
4349 return nullptr;
4350
4351 if (!llvm::all_equal(Range&: Operands))
4352 return nullptr;
4353
4354 return ConstantVector::getSplat(EC: SVTy->getElementCount(), Elt: SplatVal);
4355 }
4356 default:
4357 break;
4358 }
4359
4360 // If trivially vectorizable, try folding it via the scalar call if all
4361 // operands are splats.
4362
4363 // TODO: ConstantFoldFixedVectorCall should probably check this too?
4364 if (!isTriviallyVectorizable(ID: IntrinsicID))
4365 return nullptr;
4366
4367 SmallVector<Constant *, 4> SplatOps;
4368 for (auto [I, Op] : enumerate(First&: Operands)) {
4369 if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: I, /*TTI=*/nullptr)) {
4370 SplatOps.push_back(Elt: Op);
4371 continue;
4372 }
4373 Constant *Splat = Op->getSplatValue();
4374 if (!Splat)
4375 return nullptr;
4376 SplatOps.push_back(Elt: Splat);
4377 }
4378 Constant *Folded = ConstantFoldScalarCall(
4379 Name, IntrinsicID, Ty: SVTy->getElementType(), Operands: SplatOps, TLI, Call);
4380 if (!Folded)
4381 return nullptr;
4382 return ConstantVector::getSplat(EC: SVTy->getElementCount(), Elt: Folded);
4383}
4384
4385static std::pair<Constant *, Constant *>
4386ConstantFoldScalarFrexpCall(Constant *Op, Type *IntTy) {
4387 if (isa<PoisonValue>(Val: Op))
4388 return {Op, PoisonValue::get(T: IntTy)};
4389
4390 auto *ConstFP = dyn_cast<ConstantFP>(Val: Op);
4391 if (!ConstFP)
4392 return {};
4393
4394 const APFloat &U = ConstFP->getValueAPF();
4395 int FrexpExp;
4396 APFloat FrexpMant = frexp(X: U, Exp&: FrexpExp, RM: APFloat::rmNearestTiesToEven);
4397 Constant *Result0 = ConstantFP::get(Ty: ConstFP->getType(), V: FrexpMant);
4398
4399 // The exponent is an "unspecified value" for inf/nan. We use zero to avoid
4400 // using undef.
4401 Constant *Result1 = FrexpMant.isFinite()
4402 ? ConstantInt::getSigned(Ty: IntTy, V: FrexpExp)
4403 : ConstantInt::getNullValue(Ty: IntTy);
4404 return {Result0, Result1};
4405}
4406
4407/// Handle intrinsics that return tuples, which may be tuples of vectors.
4408static Constant *
4409ConstantFoldStructCall(StringRef Name, Intrinsic::ID IntrinsicID,
4410 StructType *StTy, ArrayRef<Constant *> Operands,
4411 const DataLayout &DL, const TargetLibraryInfo *TLI,
4412 const CallBase *Call) {
4413
4414 switch (IntrinsicID) {
4415 case Intrinsic::frexp: {
4416 Type *Ty0 = StTy->getContainedType(i: 0);
4417 Type *Ty1 = StTy->getContainedType(i: 1)->getScalarType();
4418
4419 if (auto *FVTy0 = dyn_cast<FixedVectorType>(Val: Ty0)) {
4420 SmallVector<Constant *, 4> Results0(FVTy0->getNumElements());
4421 SmallVector<Constant *, 4> Results1(FVTy0->getNumElements());
4422
4423 for (unsigned I = 0, E = FVTy0->getNumElements(); I != E; ++I) {
4424 Constant *Lane = Operands[0]->getAggregateElement(Elt: I);
4425 std::tie(args&: Results0[I], args&: Results1[I]) =
4426 ConstantFoldScalarFrexpCall(Op: Lane, IntTy: Ty1);
4427 if (!Results0[I])
4428 return nullptr;
4429 }
4430
4431 return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: Results0),
4432 Vs: ConstantVector::get(V: Results1));
4433 }
4434
4435 auto [Result0, Result1] = ConstantFoldScalarFrexpCall(Op: Operands[0], IntTy: Ty1);
4436 if (!Result0)
4437 return nullptr;
4438 return ConstantStruct::get(T: StTy, Vs: Result0, Vs: Result1);
4439 }
4440 case Intrinsic::sincos: {
4441 Type *Ty = StTy->getContainedType(i: 0);
4442 Type *TyScalar = Ty->getScalarType();
4443
4444 auto ConstantFoldScalarSincosCall =
4445 [&](Constant *Op) -> std::pair<Constant *, Constant *> {
4446 Constant *SinResult =
4447 ConstantFoldScalarCall(Name, IntrinsicID: Intrinsic::sin, Ty: TyScalar, Operands: Op, TLI, Call);
4448 Constant *CosResult =
4449 ConstantFoldScalarCall(Name, IntrinsicID: Intrinsic::cos, Ty: TyScalar, Operands: Op, TLI, Call);
4450 return std::make_pair(x&: SinResult, y&: CosResult);
4451 };
4452
4453 if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) {
4454 SmallVector<Constant *> SinResults(FVTy->getNumElements());
4455 SmallVector<Constant *> CosResults(FVTy->getNumElements());
4456
4457 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
4458 Constant *Lane = Operands[0]->getAggregateElement(Elt: I);
4459 std::tie(args&: SinResults[I], args&: CosResults[I]) =
4460 ConstantFoldScalarSincosCall(Lane);
4461 if (!SinResults[I] || !CosResults[I])
4462 return nullptr;
4463 }
4464
4465 return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: SinResults),
4466 Vs: ConstantVector::get(V: CosResults));
4467 }
4468
4469 auto [SinResult, CosResult] = ConstantFoldScalarSincosCall(Operands[0]);
4470 if (!SinResult || !CosResult)
4471 return nullptr;
4472 return ConstantStruct::get(T: StTy, Vs: SinResult, Vs: CosResult);
4473 }
4474 case Intrinsic::vector_deinterleave2:
4475 case Intrinsic::vector_deinterleave3:
4476 case Intrinsic::vector_deinterleave4:
4477 case Intrinsic::vector_deinterleave5:
4478 case Intrinsic::vector_deinterleave6:
4479 case Intrinsic::vector_deinterleave7:
4480 case Intrinsic::vector_deinterleave8: {
4481 unsigned NumResults = StTy->getNumElements();
4482 auto *Vec = Operands[0];
4483 auto *VecTy = cast<VectorType>(Val: Vec->getType());
4484
4485 ElementCount ResultEC =
4486 VecTy->getElementCount().divideCoefficientBy(RHS: NumResults);
4487
4488 if (auto *EltC = Vec->getSplatValue()) {
4489 auto *ResultVec = ConstantVector::getSplat(EC: ResultEC, Elt: EltC);
4490 SmallVector<Constant *, 8> Results(NumResults, ResultVec);
4491 return ConstantStruct::get(T: StTy, V: Results);
4492 }
4493
4494 if (!ResultEC.isFixed())
4495 return nullptr;
4496
4497 unsigned NumElements = ResultEC.getFixedValue();
4498 SmallVector<Constant *, 8> Results(NumResults);
4499 SmallVector<Constant *> Elements(NumElements);
4500 for (unsigned I = 0; I != NumResults; ++I) {
4501 for (unsigned J = 0; J != NumElements; ++J) {
4502 Constant *Elt = Vec->getAggregateElement(Elt: J * NumResults + I);
4503 if (!Elt)
4504 return nullptr;
4505 Elements[J] = Elt;
4506 }
4507 Results[I] = ConstantVector::get(V: Elements);
4508 }
4509 return ConstantStruct::get(T: StTy, V: Results);
4510 }
4511 default:
4512 // TODO: Constant folding of vector intrinsics that fall through here does
4513 // not work (e.g. overflow intrinsics)
4514 return ConstantFoldScalarCall(Name, IntrinsicID, Ty: StTy, Operands, TLI, Call);
4515 }
4516
4517 return nullptr;
4518}
4519
4520} // end anonymous namespace
4521
4522Constant *llvm::ConstantFoldBinaryIntrinsic(Intrinsic::ID ID, Constant *LHS,
4523 Constant *RHS, Type *Ty,
4524 Instruction *FMFSource) {
4525 auto *Call = dyn_cast_if_present<CallBase>(Val: FMFSource);
4526 // Ensure we check flags like StrictFP that might prevent this from getting
4527 // folded before generating a result.
4528 if (Call && !canConstantFoldCallTo(Call, F: Call->getCalledFunction()))
4529 return nullptr;
4530 return ConstantFoldIntrinsicCall2(IntrinsicID: ID, Ty, Operands: {LHS, RHS}, Call);
4531}
4532
4533Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F,
4534 ArrayRef<Constant *> Operands,
4535 const TargetLibraryInfo *TLI,
4536 bool AllowNonDeterministic) {
4537 if (Call->isNoBuiltin())
4538 return nullptr;
4539 if (!F->hasName())
4540 return nullptr;
4541
4542 // If this is not an intrinsic and not recognized as a library call, bail out.
4543 Intrinsic::ID IID = F->getIntrinsicID();
4544 if (IID == Intrinsic::not_intrinsic) {
4545 if (!TLI)
4546 return nullptr;
4547 LibFunc LibF;
4548 if (!TLI->getLibFunc(FDecl: *F, F&: LibF))
4549 return nullptr;
4550 }
4551
4552 // Conservatively assume that floating-point libcalls may be
4553 // non-deterministic.
4554 Type *Ty = F->getReturnType();
4555 if (!AllowNonDeterministic && Ty->isFPOrFPVectorTy())
4556 return nullptr;
4557
4558 StringRef Name = F->getName();
4559 if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty))
4560 return ConstantFoldFixedVectorCall(
4561 Name, IntrinsicID: IID, FVTy, Operands, DL: F->getDataLayout(), TLI, Call);
4562
4563 if (auto *SVTy = dyn_cast<ScalableVectorType>(Val: Ty))
4564 return ConstantFoldScalableVectorCall(
4565 Name, IntrinsicID: IID, SVTy, Operands, DL: F->getDataLayout(), TLI, Call);
4566
4567 if (auto *StTy = dyn_cast<StructType>(Val: Ty))
4568 return ConstantFoldStructCall(Name, IntrinsicID: IID, StTy, Operands,
4569 DL: F->getDataLayout(), TLI, Call);
4570
4571 // TODO: If this is a library function, we already discovered that above,
4572 // so we should pass the LibFunc, not the name (and it might be better
4573 // still to separate intrinsic handling from libcalls).
4574 return ConstantFoldScalarCall(Name, IntrinsicID: IID, Ty, Operands, TLI, Call);
4575}
4576
4577bool llvm::isMathLibCallNoop(const CallBase *Call,
4578 const TargetLibraryInfo *TLI) {
4579 // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
4580 // (and to some extent ConstantFoldScalarCall).
4581 if (Call->isNoBuiltin() || Call->isStrictFP())
4582 return false;
4583 Function *F = Call->getCalledFunction();
4584 if (!F)
4585 return false;
4586
4587 LibFunc Func;
4588 if (!TLI || !TLI->getLibFunc(FDecl: *F, F&: Func))
4589 return false;
4590
4591 if (Call->arg_size() == 1) {
4592 if (ConstantFP *OpC = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0))) {
4593 const APFloat &Op = OpC->getValueAPF();
4594 switch (Func) {
4595 case LibFunc_logl:
4596 case LibFunc_log:
4597 case LibFunc_logf:
4598 case LibFunc_log2l:
4599 case LibFunc_log2:
4600 case LibFunc_log2f:
4601 case LibFunc_log10l:
4602 case LibFunc_log10:
4603 case LibFunc_log10f:
4604 return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
4605
4606 case LibFunc_ilogb:
4607 return !Op.isNaN() && !Op.isZero() && !Op.isInfinity();
4608
4609 case LibFunc_expl:
4610 case LibFunc_exp:
4611 case LibFunc_expf:
4612 // FIXME: These boundaries are slightly conservative.
4613 if (OpC->getType()->isDoubleTy())
4614 return !(Op < APFloat(-745.0) || Op > APFloat(709.0));
4615 if (OpC->getType()->isFloatTy())
4616 return !(Op < APFloat(-103.0f) || Op > APFloat(88.0f));
4617 break;
4618
4619 case LibFunc_exp2l:
4620 case LibFunc_exp2:
4621 case LibFunc_exp2f:
4622 // FIXME: These boundaries are slightly conservative.
4623 if (OpC->getType()->isDoubleTy())
4624 return !(Op < APFloat(-1074.0) || Op > APFloat(1023.0));
4625 if (OpC->getType()->isFloatTy())
4626 return !(Op < APFloat(-149.0f) || Op > APFloat(127.0f));
4627 break;
4628
4629 case LibFunc_sinl:
4630 case LibFunc_sin:
4631 case LibFunc_sinf:
4632 case LibFunc_cosl:
4633 case LibFunc_cos:
4634 case LibFunc_cosf:
4635 return !Op.isInfinity();
4636
4637 case LibFunc_tanl:
4638 case LibFunc_tan:
4639 case LibFunc_tanf: {
4640 // FIXME: Stop using the host math library.
4641 // FIXME: The computation isn't done in the right precision.
4642 Type *Ty = OpC->getType();
4643 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy())
4644 return ConstantFoldFP(NativeFP: tan, V: OpC->getValueAPF(), Ty) != nullptr;
4645 break;
4646 }
4647
4648 case LibFunc_atan:
4649 case LibFunc_atanf:
4650 case LibFunc_atanl:
4651 // Per POSIX, this MAY fail if Op is denormal. We choose not failing.
4652 return true;
4653
4654 case LibFunc_asinl:
4655 case LibFunc_asin:
4656 case LibFunc_asinf:
4657 case LibFunc_acosl:
4658 case LibFunc_acos:
4659 case LibFunc_acosf:
4660 return !(Op < APFloat::getOne(Sem: Op.getSemantics(), Negative: true) ||
4661 Op > APFloat::getOne(Sem: Op.getSemantics()));
4662
4663 case LibFunc_sinh:
4664 case LibFunc_cosh:
4665 case LibFunc_sinhf:
4666 case LibFunc_coshf:
4667 case LibFunc_sinhl:
4668 case LibFunc_coshl:
4669 // FIXME: These boundaries are slightly conservative.
4670 if (OpC->getType()->isDoubleTy())
4671 return !(Op < APFloat(-710.0) || Op > APFloat(710.0));
4672 if (OpC->getType()->isFloatTy())
4673 return !(Op < APFloat(-89.0f) || Op > APFloat(89.0f));
4674 break;
4675
4676 case LibFunc_sqrtl:
4677 case LibFunc_sqrt:
4678 case LibFunc_sqrtf:
4679 return Op.isNaN() || Op.isZero() || !Op.isNegative();
4680
4681 // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
4682 // maybe others?
4683 default:
4684 break;
4685 }
4686 }
4687 }
4688
4689 if (Call->arg_size() == 2) {
4690 ConstantFP *Op0C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0));
4691 ConstantFP *Op1C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 1));
4692 if (Op0C && Op1C) {
4693 const APFloat &Op0 = Op0C->getValueAPF();
4694 const APFloat &Op1 = Op1C->getValueAPF();
4695
4696 switch (Func) {
4697 case LibFunc_powl:
4698 case LibFunc_pow:
4699 case LibFunc_powf: {
4700 // FIXME: Stop using the host math library.
4701 // FIXME: The computation isn't done in the right precision.
4702 Type *Ty = Op0C->getType();
4703 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
4704 if (Ty == Op1C->getType())
4705 return ConstantFoldBinaryFP(NativeFP: pow, V: Op0, W: Op1, Ty) != nullptr;
4706 }
4707 break;
4708 }
4709
4710 case LibFunc_fmodl:
4711 case LibFunc_fmod:
4712 case LibFunc_fmodf:
4713 case LibFunc_remainderl:
4714 case LibFunc_remainder:
4715 case LibFunc_remainderf:
4716 return Op0.isNaN() || Op1.isNaN() ||
4717 (!Op0.isInfinity() && !Op1.isZero());
4718
4719 case LibFunc_atan2:
4720 case LibFunc_atan2f:
4721 case LibFunc_atan2l:
4722 // Although IEEE-754 says atan2(+/-0.0, +/-0.0) are well-defined, and
4723 // GLIBC and MSVC do not appear to raise an error on those, we
4724 // cannot rely on that behavior. POSIX and C11 say that a domain error
4725 // may occur, so allow for that possibility.
4726 return !Op0.isZero() || !Op1.isZero();
4727
4728 default:
4729 break;
4730 }
4731 }
4732 }
4733
4734 return false;
4735}
4736
4737Constant *llvm::getLosslessInvCast(Constant *C, Type *InvCastTo,
4738 unsigned CastOp, const DataLayout &DL,
4739 PreservedCastFlags *Flags) {
4740 switch (CastOp) {
4741 case Instruction::BitCast:
4742 // Bitcast is always lossless.
4743 return ConstantFoldCastOperand(Opcode: Instruction::BitCast, C, DestTy: InvCastTo, DL);
4744 case Instruction::Trunc: {
4745 auto *ZExtC = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C, DestTy: InvCastTo, DL);
4746 if (Flags) {
4747 // Truncation back on ZExt value is always NUW.
4748 Flags->NUW = true;
4749 // Test positivity of C.
4750 auto *SExtC =
4751 ConstantFoldCastOperand(Opcode: Instruction::SExt, C, DestTy: InvCastTo, DL);
4752 Flags->NSW = ZExtC == SExtC;
4753 }
4754 return ZExtC;
4755 }
4756 case Instruction::SExt:
4757 case Instruction::ZExt: {
4758 auto *InvC = ConstantExpr::getTrunc(C, Ty: InvCastTo);
4759 auto *CastInvC = ConstantFoldCastOperand(Opcode: CastOp, C: InvC, DestTy: C->getType(), DL);
4760 // Must satisfy CastOp(InvC) == C.
4761 if (!CastInvC || CastInvC != C)
4762 return nullptr;
4763 if (Flags && CastOp == Instruction::ZExt) {
4764 auto *SExtInvC =
4765 ConstantFoldCastOperand(Opcode: Instruction::SExt, C: InvC, DestTy: C->getType(), DL);
4766 // Test positivity of InvC.
4767 Flags->NNeg = CastInvC == SExtInvC;
4768 }
4769 return InvC;
4770 }
4771 case Instruction::FPExt: {
4772 Constant *InvC =
4773 ConstantFoldCastOperand(Opcode: Instruction::FPTrunc, C, DestTy: InvCastTo, DL);
4774 if (InvC) {
4775 Constant *CastInvC =
4776 ConstantFoldCastOperand(Opcode: CastOp, C: InvC, DestTy: C->getType(), DL);
4777 if (CastInvC == C)
4778 return InvC;
4779 }
4780 return nullptr;
4781 }
4782 default:
4783 return nullptr;
4784 }
4785}
4786
4787Constant *llvm::getLosslessUnsignedTrunc(Constant *C, Type *DestTy,
4788 const DataLayout &DL,
4789 PreservedCastFlags *Flags) {
4790 return getLosslessInvCast(C, InvCastTo: DestTy, CastOp: Instruction::ZExt, DL, Flags);
4791}
4792
4793Constant *llvm::getLosslessSignedTrunc(Constant *C, Type *DestTy,
4794 const DataLayout &DL,
4795 PreservedCastFlags *Flags) {
4796 return getLosslessInvCast(C, InvCastTo: DestTy, CastOp: Instruction::SExt, DL, Flags);
4797}
4798
4799void TargetFolder::anchor() {}
4800