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/SmallBitVector.h"
26#include "llvm/ADT/SmallVector.h"
27#include "llvm/ADT/StringRef.h"
28#include "llvm/Analysis/TargetFolder.h"
29#include "llvm/Analysis/TargetLibraryInfo.h"
30#include "llvm/Analysis/ValueTracking.h"
31#include "llvm/Analysis/VectorUtils.h"
32#include "llvm/Config/config.h"
33#include "llvm/IR/Constant.h"
34#include "llvm/IR/ConstantFold.h"
35#include "llvm/IR/Constants.h"
36#include "llvm/IR/DataLayout.h"
37#include "llvm/IR/DerivedTypes.h"
38#include "llvm/IR/Function.h"
39#include "llvm/IR/GlobalValue.h"
40#include "llvm/IR/GlobalVariable.h"
41#include "llvm/IR/InstrTypes.h"
42#include "llvm/IR/Instruction.h"
43#include "llvm/IR/Instructions.h"
44#include "llvm/IR/IntrinsicInst.h"
45#include "llvm/IR/Intrinsics.h"
46#include "llvm/IR/IntrinsicsAArch64.h"
47#include "llvm/IR/IntrinsicsAMDGPU.h"
48#include "llvm/IR/IntrinsicsARM.h"
49#include "llvm/IR/IntrinsicsNVPTX.h"
50#include "llvm/IR/IntrinsicsWebAssembly.h"
51#include "llvm/IR/IntrinsicsX86.h"
52#include "llvm/IR/NVVMIntrinsicUtils.h"
53#include "llvm/IR/Operator.h"
54#include "llvm/IR/Type.h"
55#include "llvm/IR/Value.h"
56#include "llvm/Support/Casting.h"
57#include "llvm/Support/ErrorHandling.h"
58#include "llvm/Support/KnownBits.h"
59#include "llvm/Support/MathExtras.h"
60#include <cassert>
61#include <cerrno>
62#include <cfenv>
63#include <cmath>
64#include <cstdint>
65
66using namespace llvm;
67
68static cl::opt<bool> DisableFPCallFolding(
69 "disable-fp-call-folding",
70 cl::desc("Disable constant-folding of FP intrinsics and libcalls."),
71 cl::init(Val: false), cl::Hidden);
72
73namespace {
74
75//===----------------------------------------------------------------------===//
76// Constant Folding internal helper functions
77//===----------------------------------------------------------------------===//
78
79static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
80 Constant *C, Type *SrcEltTy,
81 unsigned NumSrcElts,
82 const DataLayout &DL) {
83 // Now that we know that the input value is a vector of integers, just shift
84 // and insert them into our result.
85 unsigned BitShift = DL.getTypeSizeInBits(Ty: SrcEltTy);
86 for (unsigned i = 0; i != NumSrcElts; ++i) {
87 Constant *Element;
88 if (DL.isLittleEndian())
89 Element = C->getAggregateElement(Elt: NumSrcElts - i - 1);
90 else
91 Element = C->getAggregateElement(Elt: i);
92
93 if (isa_and_nonnull<UndefValue>(Val: Element)) {
94 Result <<= BitShift;
95 continue;
96 }
97
98 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Val: Element);
99 if (!ElementCI)
100 return ConstantExpr::getBitCast(C, Ty: DestTy);
101
102 Result <<= BitShift;
103 Result |= ElementCI->getValue().zext(width: Result.getBitWidth());
104 }
105
106 return nullptr;
107}
108
109/// Check whether folding this bitcast into a byte vector would mix poison and
110/// non-poison bits in the same output lane. While integer types track poison on
111/// a per-value basis, byte types track it on a per-bit basis. However,
112/// `ConstantByte` cannot represent values with both poison and non-poison bits.
113///
114/// Source elements are grouped by the output lane they map to. Returns true if
115/// any group contains both poison and non-poison elements.
116static bool foldMixesPoisonBits(Constant *C, unsigned NumSrcElt,
117 unsigned NumDstElt) {
118 // If element counts don't divide evenly, bail out if a poison source element
119 // might span multiple destination lanes.
120 if (NumSrcElt % NumDstElt != 0)
121 return C->containsPoisonElement();
122 unsigned Ratio = NumSrcElt / NumDstElt;
123 for (unsigned i = 0; i != NumSrcElt; i += Ratio) {
124 bool HasPoison = false;
125 bool HasNonPoison = false;
126 for (unsigned j = 0; j != Ratio; ++j) {
127 Constant *Src = C->getAggregateElement(Elt: i + j);
128 // Conservatively bail out.
129 if (!Src)
130 return true;
131 if (isa<PoisonValue>(Val: Src))
132 HasPoison = true;
133 else
134 HasNonPoison = true;
135 }
136 if (HasPoison && HasNonPoison)
137 return true;
138 }
139 return false;
140}
141
142/// Track which destination lanes of a bitcast are produced from poison bytes.
143/// A destination lane is marked if any source element mapped to it is poison.
144/// Returns false if an aggregate element cannot be inspected. The caller should
145/// bail out of folding.
146static bool computePoisonDstLanes(Constant *C, unsigned NumSrcElt,
147 unsigned NumDstElt,
148 SmallBitVector &PoisonDstElts) {
149 // If element counts don't divide evenly, bail out if a poison source element
150 // might span multiple destination lanes.
151 if ((NumDstElt < NumSrcElt ? NumSrcElt % NumDstElt : NumDstElt % NumSrcElt))
152 return !C->containsPoisonElement();
153 if (NumDstElt < NumSrcElt) {
154 unsigned Ratio = NumSrcElt / NumDstElt;
155 for (unsigned i = 0; i != NumDstElt; ++i) {
156 for (unsigned j = 0; j != Ratio; ++j) {
157 Constant *Src = C->getAggregateElement(Elt: i * Ratio + j);
158 if (!Src)
159 return false;
160 if (isa<PoisonValue>(Val: Src)) {
161 PoisonDstElts[i] = true;
162 break;
163 }
164 }
165 }
166 } else {
167 unsigned Ratio = NumDstElt / NumSrcElt;
168 for (unsigned i = 0; i != NumSrcElt; ++i) {
169 Constant *Src = C->getAggregateElement(Elt: i);
170 if (!Src)
171 return false;
172 if (isa<PoisonValue>(Val: Src))
173 PoisonDstElts.set(I: i * Ratio, E: (i + 1) * Ratio);
174 }
175 }
176 return true;
177}
178
179/// Constant fold bitcast, symbolically evaluating it with DataLayout.
180/// This always returns a non-null constant, but it may be a
181/// ConstantExpr if unfoldable.
182Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
183 assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) &&
184 "Invalid constantexpr bitcast!");
185
186 // Catch the obvious splat cases.
187 if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL))
188 return Res;
189
190 if (auto *VTy = dyn_cast<VectorType>(Val: C->getType())) {
191 // Handle a vector->scalar integer/fp cast.
192 if (isa<IntegerType>(Val: DestTy) || DestTy->isFloatingPointTy()) {
193 unsigned NumSrcElts = cast<FixedVectorType>(Val: VTy)->getNumElements();
194 Type *SrcEltTy = VTy->getElementType();
195
196 // Bitcasting a byte containing any poison bit to an integer or fp type
197 // yields poison.
198 if (SrcEltTy->isByteTy() && C->containsPoisonElement())
199 return PoisonValue::get(T: DestTy);
200
201 // If the vector is a vector of floating point or bytes, convert it to a
202 // vector of int to simplify things.
203 if (SrcEltTy->isFloatingPointTy() || SrcEltTy->isByteTy()) {
204 unsigned Width = SrcEltTy->getPrimitiveSizeInBits();
205 auto *SrcIVTy = FixedVectorType::get(
206 ElementType: IntegerType::get(C&: C->getContext(), NumBits: Width), NumElts: NumSrcElts);
207 // Ask IR to do the conversion now that #elts line up.
208 C = ConstantExpr::getBitCast(C, Ty: SrcIVTy);
209 }
210
211 APInt Result(DL.getTypeSizeInBits(Ty: DestTy), 0);
212 if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
213 SrcEltTy, NumSrcElts, DL))
214 return CE;
215
216 if (isa<IntegerType>(Val: DestTy))
217 return ConstantInt::get(Ty: DestTy, V: Result);
218
219 APFloat FP(DestTy->getFltSemantics(), Result);
220 return ConstantFP::get(Context&: DestTy->getContext(), V: FP);
221 }
222 }
223
224 // The code below only handles casts to vectors currently.
225 auto *DestVTy = dyn_cast<VectorType>(Val: DestTy);
226 if (!DestVTy)
227 return ConstantExpr::getBitCast(C, Ty: DestTy);
228
229 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
230 // vector so the code below can handle it uniformly.
231 if (!isa<VectorType>(Val: C->getType()) &&
232 (isa<ConstantFP>(Val: C) || isa<ConstantInt>(Val: C) || isa<ConstantByte>(Val: C))) {
233 Constant *Ops = C; // don't take the address of C!
234 return FoldBitCast(C: ConstantVector::get(V: Ops), DestTy, DL);
235 }
236
237 // Some of what follows may extend to cover scalable vectors but the current
238 // implementation is fixed length specific.
239 if (!isa<FixedVectorType>(Val: C->getType()))
240 return ConstantExpr::getBitCast(C, Ty: DestTy);
241
242 // If this is a bitcast from constant vector -> vector, fold it.
243 if (!isa<ConstantDataVector>(Val: C) && !isa<ConstantVector>(Val: C) &&
244 !isa<ConstantInt>(Val: C) && !isa<ConstantFP>(Val: C) && !isa<ConstantByte>(Val: C))
245 return ConstantExpr::getBitCast(C, Ty: DestTy);
246
247 // If the element types match, IR can fold it.
248 unsigned NumDstElt = cast<FixedVectorType>(Val: DestVTy)->getNumElements();
249 unsigned NumSrcElt = cast<FixedVectorType>(Val: C->getType())->getNumElements();
250 if (NumDstElt == NumSrcElt)
251 return ConstantExpr::getBitCast(C, Ty: DestTy);
252
253 Type *SrcEltTy = cast<VectorType>(Val: C->getType())->getElementType();
254 Type *DstEltTy = DestVTy->getElementType();
255
256 // Otherwise, we're changing the number of elements in a vector, which
257 // requires endianness information to do the right thing. For example,
258 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
259 // folds to (little endian):
260 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
261 // and to (big endian):
262 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
263
264 // First thing is first. We only want to think about integer here, so if
265 // we have something in FP form, recast it as integer.
266 if (DstEltTy->isFloatingPointTy()) {
267 // Fold to an vector of integers with same size as our FP type.
268 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
269 auto *DestIVTy = FixedVectorType::get(
270 ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumDstElt);
271 // Recursively handle this integer conversion, if possible.
272 C = FoldBitCast(C, DestTy: DestIVTy, DL);
273
274 // Finally, IR can handle this now that #elts line up.
275 return ConstantExpr::getBitCast(C, Ty: DestTy);
276 }
277
278 // Handle byte destination type by folding through integers.
279 if (DstEltTy->isByteTy()) {
280 // When combining elements into larger byte values, bail out if the fold
281 // mixes poison and non-poison bits in the same destination element. Byte
282 // types track poison per bit, and no constant value can represent that.
283 if (NumDstElt < NumSrcElt && foldMixesPoisonBits(C, NumSrcElt, NumDstElt))
284 return ConstantExpr::getBitCast(C, Ty: DestTy);
285
286 // Fold to a vector of integers with same size as the byte type.
287 unsigned ByteWidth = DstEltTy->getPrimitiveSizeInBits();
288 auto *DestIVTy = FixedVectorType::get(
289 ElementType: IntegerType::get(C&: C->getContext(), NumBits: ByteWidth), NumElts: NumDstElt);
290 C = FoldBitCast(C, DestTy: DestIVTy, DL);
291 return ConstantExpr::getBitCast(C, Ty: DestTy);
292 }
293
294 // Okay, we know the destination is integer, if the input is FP, convert
295 // it to integer first.
296 if (SrcEltTy->isFloatingPointTy()) {
297 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
298 auto *SrcIVTy = FixedVectorType::get(
299 ElementType: IntegerType::get(C&: C->getContext(), NumBits: FPWidth), NumElts: NumSrcElt);
300 // Ask IR to do the conversion now that #elts line up.
301 C = ConstantExpr::getBitCast(C, Ty: SrcIVTy);
302 assert((isa<ConstantVector>(C) || // FIXME: Remove ConstantVector.
303 isa<ConstantDataVector>(C) || isa<ConstantInt>(C)) &&
304 "Constant folding cannot fail for plain fp->int bitcast!");
305 }
306
307 // Handle byte source type by folding through integers. Byte types track
308 // poison per bit, so any poison bit makes the destination lane poison.
309 // Record which destination lanes contain poison bits, before the generic
310 // fold below refines them to undef/zero, so they can be restored.
311 SmallBitVector PoisonDstElts(NumDstElt);
312 if (SrcEltTy->isByteTy()) {
313 if (!computePoisonDstLanes(C, NumSrcElt, NumDstElt, PoisonDstElts))
314 return ConstantExpr::getBitCast(C, Ty: DestTy);
315
316 unsigned ByteWidth = SrcEltTy->getPrimitiveSizeInBits();
317 auto *SrcIVTy = FixedVectorType::get(
318 ElementType: IntegerType::get(C&: C->getContext(), NumBits: ByteWidth), NumElts: NumSrcElt);
319 // Ask IR to do the conversion now that #elts line up.
320 C = ConstantExpr::getBitCast(C, Ty: SrcIVTy);
321 assert((isa<ConstantVector>(C) || // FIXME: Remove ConstantVector.
322 isa<ConstantDataVector>(C) || isa<ConstantInt>(C)) &&
323 "Constant folding cannot fail for plain byte->int bitcast!");
324 }
325
326 // Now we know that the input and output vectors are both integer vectors
327 // of the same size, and that their #elements is not the same.
328 // Use data buffer for easy non-integer element ratio vectors handling,
329 // For example: <4 x i24> to <3 x i32>.
330 bool isLittleEndian = DL.isLittleEndian();
331 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
332 unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
333 SmallVector<Constant*, 32> Result;
334 unsigned SrcElt = 0;
335
336 APInt Buffer(2 * std::max(a: SrcBitSize, b: DstBitSize), 0);
337 APInt UndefMask(Buffer.getBitWidth(), 0);
338 APInt PoisonMask(Buffer.getBitWidth(), 0);
339 unsigned BufferBitSize = 0;
340
341 while (Result.size() != NumDstElt) {
342 // Load SrcElts into Buffer.
343 while (BufferBitSize < DstBitSize) {
344 Constant *Element = C->getAggregateElement(Elt: SrcElt++);
345 if (!Element) // Reject constantexpr elements
346 return ConstantExpr::getBitCast(C, Ty: DestTy);
347
348 // Shift Buffer & Masks to fit next SrcElt.
349 if (!isLittleEndian) {
350 Buffer <<= SrcBitSize;
351 UndefMask <<= SrcBitSize;
352 PoisonMask <<= SrcBitSize;
353 }
354
355 APInt SrcValue;
356 unsigned BitPosition = isLittleEndian ? BufferBitSize : 0;
357 if (isa<UndefValue>(Val: Element)) {
358 // Set masks fragments bits.
359 UndefMask.setBits(loBit: BitPosition, hiBit: BitPosition + SrcBitSize);
360 if (isa<PoisonValue>(Val: Element))
361 PoisonMask.setBits(loBit: BitPosition, hiBit: BitPosition + SrcBitSize);
362 SrcValue = APInt::getZero(numBits: SrcBitSize);
363 } else {
364 auto *Src = dyn_cast<ConstantInt>(Val: Element);
365 if (!Src)
366 return ConstantExpr::getBitCast(C, Ty: DestTy);
367 SrcValue = Src->getValue();
368 }
369
370 // Insert src element bits into Buffer on correct position.
371 Buffer.insertBits(SubBits: SrcValue, bitPosition: BitPosition);
372 BufferBitSize += SrcBitSize;
373 }
374
375 // Create DstElts from Buffer.
376 while (BufferBitSize >= DstBitSize) {
377 unsigned ShiftAmt = isLittleEndian ? 0 : BufferBitSize - DstBitSize;
378 // Emit undef/poison, if all undef mask fragment bits are set.
379 if (UndefMask.extractBits(numBits: DstBitSize, bitPosition: ShiftAmt).isAllOnes()) {
380 // Push poison, if any bit in poison mask fragment is set.
381 if (!PoisonMask.extractBits(numBits: DstBitSize, bitPosition: ShiftAmt).isZero()) {
382 Result.push_back(Elt: PoisonValue::get(T: DstEltTy));
383 } else {
384 Result.push_back(Elt: UndefValue::get(T: DstEltTy));
385 }
386 } else {
387 // Create and push DstElt.
388 APInt Elt = Buffer.extractBits(numBits: DstBitSize, bitPosition: ShiftAmt);
389 Result.push_back(Elt: ConstantInt::get(Ty: DstEltTy, V: Elt));
390 }
391
392 // Shift unused Buffer fragment to lower bits.
393 if (isLittleEndian) {
394 Buffer.lshrInPlace(ShiftAmt: DstBitSize);
395 UndefMask.lshrInPlace(ShiftAmt: DstBitSize);
396 PoisonMask.lshrInPlace(ShiftAmt: DstBitSize);
397 }
398 BufferBitSize -= DstBitSize;
399 }
400 }
401
402 // Restore destination lanes whose source bytes contained poison bits.
403 for (unsigned I : PoisonDstElts.set_bits())
404 Result[I] = PoisonValue::get(T: DstEltTy);
405
406 return ConstantVector::get(V: Result);
407}
408
409} // end anonymous namespace
410
411/// If this constant is a constant offset from a global, return the global and
412/// the constant. Because of constantexprs, this function is recursive.
413bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
414 APInt &Offset, const DataLayout &DL,
415 DSOLocalEquivalent **DSOEquiv) {
416 if (DSOEquiv)
417 *DSOEquiv = nullptr;
418
419 // Trivial case, constant is the global.
420 if ((GV = dyn_cast<GlobalValue>(Val: C))) {
421 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType());
422 Offset = APInt(BitWidth, 0);
423 return true;
424 }
425
426 if (auto *FoundDSOEquiv = dyn_cast<DSOLocalEquivalent>(Val: C)) {
427 if (DSOEquiv)
428 *DSOEquiv = FoundDSOEquiv;
429 GV = FoundDSOEquiv->getGlobalValue();
430 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GV->getType());
431 Offset = APInt(BitWidth, 0);
432 return true;
433 }
434
435 // Otherwise, if this isn't a constant expr, bail out.
436 auto *CE = dyn_cast<ConstantExpr>(Val: C);
437 if (!CE) return false;
438
439 // Look through ptr->int and ptr->ptr casts.
440 if (CE->getOpcode() == Instruction::PtrToInt ||
441 CE->getOpcode() == Instruction::PtrToAddr)
442 return IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset, DL,
443 DSOEquiv);
444
445 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
446 auto *GEP = dyn_cast<GEPOperator>(Val: CE);
447 if (!GEP)
448 return false;
449
450 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType());
451 APInt TmpOffset(BitWidth, 0);
452
453 // If the base isn't a global+constant, we aren't either.
454 if (!IsConstantOffsetFromGlobal(C: CE->getOperand(i_nocapture: 0), GV, Offset&: TmpOffset, DL,
455 DSOEquiv))
456 return false;
457
458 // Otherwise, add any offset that our operands provide.
459 if (!GEP->accumulateConstantOffset(DL, Offset&: TmpOffset))
460 return false;
461
462 Offset = TmpOffset;
463 return true;
464}
465
466Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
467 const DataLayout &DL) {
468 do {
469 Type *SrcTy = C->getType();
470 if (SrcTy == DestTy)
471 return C;
472
473 TypeSize DestSize = DL.getTypeSizeInBits(Ty: DestTy);
474 TypeSize SrcSize = DL.getTypeSizeInBits(Ty: SrcTy);
475 if (!TypeSize::isKnownGE(LHS: SrcSize, RHS: DestSize))
476 return nullptr;
477
478 // Catch the obvious splat cases (since all-zeros can coerce non-integral
479 // pointers legally).
480 if (Constant *Res = ConstantFoldLoadFromUniformValue(C, Ty: DestTy, DL))
481 return Res;
482
483 // If the type sizes are the same and a cast is legal, just directly
484 // cast the constant.
485 // But be careful not to coerce non-integral pointers illegally.
486 if (SrcSize == DestSize &&
487 DL.isNonIntegralPointerType(Ty: SrcTy->getScalarType()) ==
488 DL.isNonIntegralPointerType(Ty: DestTy->getScalarType())) {
489 Instruction::CastOps Cast = Instruction::BitCast;
490 // If we are going from a pointer to int or vice versa, we spell the cast
491 // differently.
492 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
493 Cast = Instruction::IntToPtr;
494 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
495 Cast = Instruction::PtrToInt;
496
497 if (CastInst::castIsValid(op: Cast, S: C, DstTy: DestTy))
498 return ConstantFoldCastOperand(Opcode: Cast, C, DestTy, DL);
499 }
500
501 // If this isn't an aggregate type, there is nothing we can do to drill down
502 // and find a bitcastable constant.
503 if (!SrcTy->isAggregateType() && !SrcTy->isVectorTy())
504 return nullptr;
505
506 // We're simulating a load through a pointer that was bitcast to point to
507 // a different type, so we can try to walk down through the initial
508 // elements of an aggregate to see if some part of the aggregate is
509 // castable to implement the "load" semantic model.
510 if (SrcTy->isStructTy()) {
511 // Struct types might have leading zero-length elements like [0 x i32],
512 // which are certainly not what we are looking for, so skip them.
513 unsigned Elem = 0;
514 Constant *ElemC;
515 do {
516 ElemC = C->getAggregateElement(Elt: Elem++);
517 } while (ElemC && DL.getTypeSizeInBits(Ty: ElemC->getType()).isZero());
518 C = ElemC;
519 } else {
520 // For non-byte-sized vector elements, the first element is not
521 // necessarily located at the vector base address.
522 if (auto *VT = dyn_cast<VectorType>(Val: SrcTy))
523 if (!DL.typeSizeEqualsStoreSize(Ty: VT->getElementType()))
524 return nullptr;
525
526 C = C->getAggregateElement(Elt: 0u);
527 }
528 } while (C);
529
530 return nullptr;
531}
532
533namespace {
534
535/// Recursive helper to read bits out of global. C is the constant being copied
536/// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
537/// results into and BytesLeft is the number of bytes left in
538/// the CurPtr buffer. DL is the DataLayout. When IsByteLoad is true, do not
539/// unwrap inttoptr constant expressions. The caller would reconstruct those
540/// bits as a ConstantByte, dropping the pointer's provenance.
541bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
542 unsigned BytesLeft, const DataLayout &DL,
543 bool IsByteLoad = false) {
544 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&
545 "Out of range access");
546
547 // Reading type padding, return zero.
548 if (ByteOffset >= DL.getTypeStoreSize(Ty: C->getType()))
549 return true;
550
551 // If this element is zero or undefined, we can just return since *CurPtr is
552 // zero initialized.
553 if (isa<ConstantAggregateZero>(Val: C) || isa<UndefValue>(Val: C))
554 return true;
555
556 auto *CI = dyn_cast<ConstantInt>(Val: C);
557 if (CI && CI->getType()->isIntegerTy()) {
558 if ((CI->getBitWidth() & 7) != 0)
559 return false;
560 const APInt &Val = CI->getValue();
561 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
562
563 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
564 unsigned n = ByteOffset;
565 if (!DL.isLittleEndian())
566 n = IntBytes - n - 1;
567 CurPtr[i] = Val.extractBits(numBits: 8, bitPosition: n * 8).getZExtValue();
568 ++ByteOffset;
569 }
570 return true;
571 }
572
573 auto *CFP = dyn_cast<ConstantFP>(Val: C);
574 if (CFP && CFP->getType()->isFloatingPointTy()) {
575 if (CFP->getType()->isDoubleTy()) {
576 C = FoldBitCast(C, DestTy: Type::getInt64Ty(C&: C->getContext()), DL);
577 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL,
578 IsByteLoad);
579 }
580 if (CFP->getType()->isFloatTy()){
581 C = FoldBitCast(C, DestTy: Type::getInt32Ty(C&: C->getContext()), DL);
582 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL,
583 IsByteLoad);
584 }
585 if (CFP->getType()->isHalfTy()){
586 C = FoldBitCast(C, DestTy: Type::getInt16Ty(C&: C->getContext()), DL);
587 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL,
588 IsByteLoad);
589 }
590 return false;
591 }
592
593 if (auto *CS = dyn_cast<ConstantStruct>(Val: C)) {
594 const StructLayout *SL = DL.getStructLayout(Ty: CS->getType());
595 unsigned Index = SL->getElementContainingOffset(FixedOffset: ByteOffset);
596 uint64_t CurEltOffset = SL->getElementOffset(Idx: Index);
597 ByteOffset -= CurEltOffset;
598
599 while (true) {
600 // If the element access is to the element itself and not to tail padding,
601 // read the bytes from the element.
602 uint64_t EltSize = DL.getTypeAllocSize(Ty: CS->getOperand(i_nocapture: Index)->getType());
603
604 if (ByteOffset < EltSize &&
605 !ReadDataFromGlobal(C: CS->getOperand(i_nocapture: Index), ByteOffset, CurPtr,
606 BytesLeft, DL, IsByteLoad))
607 return false;
608
609 ++Index;
610
611 // Check to see if we read from the last struct element, if so we're done.
612 if (Index == CS->getType()->getNumElements())
613 return true;
614
615 // If we read all of the bytes we needed from this element we're done.
616 uint64_t NextEltOffset = SL->getElementOffset(Idx: Index);
617
618 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
619 return true;
620
621 // Move to the next element of the struct.
622 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
623 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
624 ByteOffset = 0;
625 CurEltOffset = NextEltOffset;
626 }
627 // not reached.
628 }
629
630 if (isa<ConstantArray>(Val: C) || isa<ConstantVector>(Val: C) ||
631 isa<ConstantDataSequential>(Val: C) || isa<ConstantInt>(Val: C) ||
632 isa<ConstantFP>(Val: C)) {
633 uint64_t NumElts, EltSize;
634 Type *EltTy;
635 if (auto *AT = dyn_cast<ArrayType>(Val: C->getType())) {
636 NumElts = AT->getNumElements();
637 EltTy = AT->getElementType();
638 EltSize = DL.getTypeAllocSize(Ty: EltTy);
639 } else {
640 NumElts = cast<FixedVectorType>(Val: C->getType())->getNumElements();
641 EltTy = cast<FixedVectorType>(Val: C->getType())->getElementType();
642 // TODO: For non-byte-sized vectors, current implementation assumes there is
643 // padding to the next byte boundary between elements.
644 if (!DL.typeSizeEqualsStoreSize(Ty: EltTy))
645 return false;
646
647 EltSize = DL.getTypeStoreSize(Ty: EltTy);
648 }
649 uint64_t Index = ByteOffset / EltSize;
650 uint64_t Offset = ByteOffset - Index * EltSize;
651
652 for (; Index != NumElts; ++Index) {
653 if (!ReadDataFromGlobal(C: C->getAggregateElement(Elt: Index), ByteOffset: Offset, CurPtr,
654 BytesLeft, DL, IsByteLoad))
655 return false;
656
657 uint64_t BytesWritten = EltSize - Offset;
658 assert(BytesWritten <= EltSize && "Not indexing into this element?");
659 if (BytesWritten >= BytesLeft)
660 return true;
661
662 Offset = 0;
663 BytesLeft -= BytesWritten;
664 CurPtr += BytesWritten;
665 }
666 return true;
667 }
668
669 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
670 if (CE->getOpcode() == Instruction::IntToPtr &&
671 CE->getOperand(i_nocapture: 0)->getType() == DL.getIntPtrType(CE->getType())) {
672 // Folding byte loads through the integer operand would rebuild the result
673 // as a `ConstantByte`, dropping the pointer's provenance.
674 if (IsByteLoad)
675 return false;
676 return ReadDataFromGlobal(C: CE->getOperand(i_nocapture: 0), ByteOffset, CurPtr,
677 BytesLeft, DL, IsByteLoad);
678 }
679 }
680
681 // Otherwise, unknown initializer type.
682 return false;
683}
684
685/// OrigLoadTy is the original type being loaded, while LoadTy is the type
686/// currently being folded (which may be integer type mapped from OrigLoadTy).
687Constant *FoldReinterpretLoadFromConst(Constant *C, Type *LoadTy,
688 Type *OrigLoadTy, int64_t Offset,
689 const DataLayout &DL) {
690 // Bail out early. Not expect to load from scalable global variable.
691 if (isa<ScalableVectorType>(Val: LoadTy))
692 return nullptr;
693
694 auto *IntType = dyn_cast<IntegerType>(Val: LoadTy);
695
696 // If this isn't an integer load we can't fold it directly.
697 if (!IntType) {
698 // If this is a non-integer load, we can try folding it as an int load and
699 // then bitcast the result. This can be useful for union cases. Note
700 // that address spaces don't matter here since we're not going to result in
701 // an actual new load.
702 if (!LoadTy->isFloatingPointTy() && !LoadTy->isPointerTy() &&
703 !LoadTy->isByteTy() && !LoadTy->isVectorTy())
704 return nullptr;
705
706 Type *MapTy = Type::getIntNTy(C&: C->getContext(),
707 N: DL.getTypeSizeInBits(Ty: LoadTy).getFixedValue());
708 if (Constant *Res =
709 FoldReinterpretLoadFromConst(C, LoadTy: MapTy, OrigLoadTy, Offset, DL)) {
710 if (Res->isNullValue() && !LoadTy->isX86_AMXTy())
711 // Materializing a zero can be done trivially without a bitcast
712 return Constant::getNullValue(Ty: LoadTy);
713 Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy;
714 Res = FoldBitCast(C: Res, DestTy: CastTy, DL);
715 if (LoadTy->isPtrOrPtrVectorTy()) {
716 // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr
717 if (Res->isNullValue() && !LoadTy->isX86_AMXTy())
718 return Constant::getNullValue(Ty: LoadTy);
719 if (DL.isNonIntegralPointerType(Ty: LoadTy->getScalarType()))
720 // Be careful not to replace a load of an addrspace value with an inttoptr here
721 return nullptr;
722 Res = ConstantExpr::getIntToPtr(C: Res, Ty: LoadTy);
723 }
724 return Res;
725 }
726 return nullptr;
727 }
728
729 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
730 // Allow folding of large type loads (e.g. <16 x double>).
731 if (BytesLoaded > 128 || BytesLoaded == 0)
732 return nullptr;
733
734 // For scalar integer load, use smaller limit to avoid regression during
735 // memcmp expansion. Codegen may generate inefficient string operations.
736 if (BytesLoaded > 32 && OrigLoadTy->isIntegerTy())
737 return nullptr;
738
739 // If we're not accessing anything in this constant, the result is undefined.
740 if (Offset <= -1 * static_cast<int64_t>(BytesLoaded))
741 return PoisonValue::get(T: IntType);
742
743 // TODO: We should be able to support scalable types.
744 TypeSize InitializerSize = DL.getTypeAllocSize(Ty: C->getType());
745 if (InitializerSize.isScalable())
746 return nullptr;
747
748 // If we're not accessing anything in this constant, the result is undefined.
749 if (Offset >= (int64_t)InitializerSize.getFixedValue())
750 return PoisonValue::get(T: IntType);
751
752 SmallVector<unsigned char, 64> RawBytes(BytesLoaded);
753 unsigned char *CurPtr = RawBytes.data();
754 unsigned BytesLeft = BytesLoaded;
755
756 // If we're loading off the beginning of the global, some bytes may be valid.
757 if (Offset < 0) {
758 CurPtr += -Offset;
759 BytesLeft += Offset;
760 Offset = 0;
761 }
762
763 if (!ReadDataFromGlobal(C, ByteOffset: Offset, CurPtr, BytesLeft, DL,
764 /*IsByteLoad=*/OrigLoadTy->isByteOrByteVectorTy()))
765 return nullptr;
766
767 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
768 if (DL.isLittleEndian()) {
769 ResultVal = RawBytes[BytesLoaded - 1];
770 for (unsigned i = 1; i != BytesLoaded; ++i) {
771 ResultVal <<= 8;
772 ResultVal |= RawBytes[BytesLoaded - 1 - i];
773 }
774 } else {
775 ResultVal = RawBytes[0];
776 for (unsigned i = 1; i != BytesLoaded; ++i) {
777 ResultVal <<= 8;
778 ResultVal |= RawBytes[i];
779 }
780 }
781
782 return ConstantInt::get(Context&: IntType->getContext(), V: ResultVal);
783}
784
785} // anonymous namespace
786
787// If GV is a constant with an initializer read its representation starting
788// at Offset and return it as a constant array of unsigned char. Otherwise
789// return null.
790Constant *llvm::ReadByteArrayFromGlobal(const GlobalVariable *GV,
791 uint64_t Offset) {
792 if (!GV->isConstant() || !GV->hasDefinitiveInitializer())
793 return nullptr;
794
795 const DataLayout &DL = GV->getDataLayout();
796 Constant *Init = const_cast<Constant *>(GV->getInitializer());
797 TypeSize InitSize = DL.getTypeAllocSize(Ty: Init->getType());
798 if (InitSize < Offset)
799 return nullptr;
800
801 uint64_t NBytes = InitSize - Offset;
802 if (NBytes > UINT16_MAX)
803 // Bail for large initializers in excess of 64K to avoid allocating
804 // too much memory.
805 // Offset is assumed to be less than or equal than InitSize (this
806 // is enforced in ReadDataFromGlobal).
807 return nullptr;
808
809 SmallVector<unsigned char, 256> RawBytes(static_cast<size_t>(NBytes));
810 unsigned char *CurPtr = RawBytes.data();
811
812 if (!ReadDataFromGlobal(C: Init, ByteOffset: Offset, CurPtr, BytesLeft: NBytes, DL))
813 return nullptr;
814
815 return ConstantDataArray::get(Context&: GV->getContext(), Elts&: RawBytes);
816}
817
818/// If this Offset points exactly to the start of an aggregate element, return
819/// that element, otherwise return nullptr.
820Constant *getConstantAtOffset(Constant *Base, APInt Offset,
821 const DataLayout &DL) {
822 if (Offset.isZero())
823 return Base;
824
825 if (!isa<ConstantAggregate>(Val: Base) && !isa<ConstantDataSequential>(Val: Base))
826 return nullptr;
827
828 Type *ElemTy = Base->getType();
829 SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset);
830 if (!Offset.isZero() || !Indices[0].isZero())
831 return nullptr;
832
833 Constant *C = Base;
834 for (const APInt &Index : drop_begin(RangeOrContainer&: Indices)) {
835 if (Index.isNegative() || Index.getActiveBits() >= 32)
836 return nullptr;
837
838 C = C->getAggregateElement(Elt: Index.getZExtValue());
839 if (!C)
840 return nullptr;
841 }
842
843 return C;
844}
845
846Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
847 const APInt &Offset,
848 const DataLayout &DL) {
849 if (Constant *AtOffset = getConstantAtOffset(Base: C, Offset, DL))
850 if (Constant *Result = ConstantFoldLoadThroughBitcast(C: AtOffset, DestTy: Ty, DL))
851 return Result;
852
853 // Explicitly check for out-of-bounds access, so we return poison even if the
854 // constant is a uniform value.
855 TypeSize Size = DL.getTypeAllocSize(Ty: C->getType());
856 if (!Size.isScalable() && Offset.sge(RHS: Size.getFixedValue()))
857 return PoisonValue::get(T: Ty);
858
859 // Try an offset-independent fold of a uniform value.
860 if (Constant *Result = ConstantFoldLoadFromUniformValue(C, Ty, DL))
861 return Result;
862
863 // Try hard to fold loads from bitcasted strange and non-type-safe things.
864 if (Offset.getSignificantBits() <= 64)
865 if (Constant *Result =
866 FoldReinterpretLoadFromConst(C, LoadTy: Ty, OrigLoadTy: Ty, Offset: Offset.getSExtValue(), DL))
867 return Result;
868
869 return nullptr;
870}
871
872Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
873 const DataLayout &DL) {
874 return ConstantFoldLoadFromConst(C, Ty, Offset: APInt(64, 0), DL);
875}
876
877Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
878 APInt Offset,
879 const DataLayout &DL) {
880 // We can only fold loads from constant globals with a definitive initializer.
881 // Check this upfront, to skip expensive offset calculations.
882 auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: C));
883 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
884 return nullptr;
885
886 C = cast<Constant>(Val: C->stripAndAccumulateConstantOffsets(
887 DL, Offset, /* AllowNonInbounds */ true));
888
889 if (C == GV)
890 if (Constant *Result = ConstantFoldLoadFromConst(C: GV->getInitializer(), Ty,
891 Offset, DL))
892 return Result;
893
894 // If this load comes from anywhere in a uniform constant global, the value
895 // is always the same, regardless of the loaded offset.
896 return ConstantFoldLoadFromUniformValue(C: GV->getInitializer(), Ty, DL);
897}
898
899Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
900 const DataLayout &DL) {
901 APInt Offset(DL.getIndexTypeSizeInBits(Ty: C->getType()), 0);
902 return ConstantFoldLoadFromConstPtr(C, Ty, Offset: std::move(Offset), DL);
903}
904
905Constant *llvm::ConstantFoldLoadFromUniformValue(Constant *C, Type *Ty,
906 const DataLayout &DL) {
907 if (isa<PoisonValue>(Val: C))
908 return PoisonValue::get(T: Ty);
909 if (isa<UndefValue>(Val: C))
910 return UndefValue::get(T: Ty);
911 // If padding is needed when storing C to memory, then it isn't considered as
912 // uniform.
913 if (!DL.typeSizeEqualsStoreSize(Ty: C->getType()))
914 return nullptr;
915 if (C->isNullValue() && !Ty->isX86_AMXTy())
916 return Constant::getNullValue(Ty);
917 if (C->isAllOnesValue() &&
918 (Ty->isIntOrIntVectorTy() || Ty->isByteOrByteVectorTy() ||
919 Ty->isFPOrFPVectorTy()))
920 return Constant::getAllOnesValue(Ty);
921 return nullptr;
922}
923
924namespace {
925
926/// One of Op0/Op1 is a constant expression.
927/// Attempt to symbolically evaluate the result of a binary operator merging
928/// these together. If target data info is available, it is provided as DL,
929/// otherwise DL is null.
930Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
931 const DataLayout &DL) {
932 // SROA
933
934 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
935 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
936 // bits.
937
938 if (Opc == Instruction::And) {
939 KnownBits Known0 = computeKnownBits(V: Op0, DL);
940 KnownBits Known1 = computeKnownBits(V: Op1, DL);
941 if ((Known1.One | Known0.Zero).isAllOnes()) {
942 // All the bits of Op0 that the 'and' could be masking are already zero.
943 return Op0;
944 }
945 if ((Known0.One | Known1.Zero).isAllOnes()) {
946 // All the bits of Op1 that the 'and' could be masking are already zero.
947 return Op1;
948 }
949
950 Known0 &= Known1;
951 if (Known0.isConstant())
952 return ConstantInt::get(Ty: Op0->getType(), V: Known0.getConstant());
953 }
954
955 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
956 // constant. This happens frequently when iterating over a global array.
957 if (Opc == Instruction::Sub) {
958 GlobalValue *GV1, *GV2;
959 APInt Offs1, Offs2;
960
961 if (IsConstantOffsetFromGlobal(C: Op0, GV&: GV1, Offset&: Offs1, DL))
962 if (IsConstantOffsetFromGlobal(C: Op1, GV&: GV2, Offset&: Offs2, DL) && GV1 == GV2) {
963 unsigned OpSize = DL.getTypeSizeInBits(Ty: Op0->getType());
964
965 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
966 // PtrToInt may change the bitwidth so we have convert to the right size
967 // first.
968 return ConstantInt::get(Ty: Op0->getType(), V: Offs1.zextOrTrunc(width: OpSize) -
969 Offs2.zextOrTrunc(width: OpSize));
970 }
971 }
972
973 return nullptr;
974}
975
976/// If array indices are not pointer-sized integers, explicitly cast them so
977/// that they aren't implicitly casted by the getelementptr.
978Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
979 Type *ResultTy, GEPNoWrapFlags NW,
980 std::optional<ConstantRange> InRange,
981 const DataLayout &DL, const TargetLibraryInfo *TLI) {
982 Type *IntIdxTy = DL.getIndexType(PtrTy: ResultTy);
983 Type *IntIdxScalarTy = IntIdxTy->getScalarType();
984
985 bool Any = false;
986 SmallVector<Constant*, 32> NewIdxs;
987 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
988 if ((i == 1 ||
989 !isa<StructType>(Val: GetElementPtrInst::getIndexedType(
990 Ty: SrcElemTy, IdxList: Ops.slice(N: 1, M: i - 1)))) &&
991 Ops[i]->getType()->getScalarType() != IntIdxScalarTy) {
992 Any = true;
993 Type *NewType =
994 Ops[i]->getType()->isVectorTy() ? IntIdxTy : IntIdxScalarTy;
995 Constant *NewIdx = ConstantFoldCastOperand(
996 Opcode: CastInst::getCastOpcode(Val: Ops[i], SrcIsSigned: true, Ty: NewType, DstIsSigned: true), C: Ops[i], DestTy: NewType,
997 DL);
998 if (!NewIdx)
999 return nullptr;
1000 NewIdxs.push_back(Elt: NewIdx);
1001 } else
1002 NewIdxs.push_back(Elt: Ops[i]);
1003 }
1004
1005 if (!Any)
1006 return nullptr;
1007
1008 Constant *C =
1009 ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: NewIdxs, NW, InRange);
1010 return ConstantFoldConstant(C, DL, TLI);
1011}
1012
1013/// If we can symbolically evaluate the GEP constant expression, do so.
1014Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
1015 ArrayRef<Constant *> Ops,
1016 const DataLayout &DL,
1017 const TargetLibraryInfo *TLI) {
1018 Type *SrcElemTy = GEP->getSourceElementType();
1019 Type *ResTy = GEP->getType();
1020 if (!SrcElemTy->isSized() || isa<ScalableVectorType>(Val: SrcElemTy))
1021 return nullptr;
1022
1023 if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResultTy: ResTy, NW: GEP->getNoWrapFlags(),
1024 InRange: GEP->getInRange(), DL, TLI))
1025 return C;
1026
1027 Constant *Ptr = Ops[0];
1028 if (!Ptr->getType()->isPointerTy())
1029 return nullptr;
1030
1031 Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType());
1032
1033 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1034 if (!isa<ConstantInt>(Val: Ops[i]) || !Ops[i]->getType()->isIntegerTy())
1035 return nullptr;
1036
1037 unsigned BitWidth = DL.getTypeSizeInBits(Ty: IntIdxTy);
1038 APInt Offset = APInt(
1039 BitWidth,
1040 DL.getIndexedOffsetInType(
1041 ElemTy: SrcElemTy, Indices: ArrayRef((Value *const *)Ops.data() + 1, Ops.size() - 1)),
1042 /*isSigned=*/true, /*implicitTrunc=*/true);
1043
1044 std::optional<ConstantRange> InRange = GEP->getInRange();
1045 if (InRange)
1046 InRange = InRange->sextOrTrunc(BitWidth);
1047
1048 // If this is a GEP of a GEP, fold it all into a single GEP.
1049 GEPNoWrapFlags NW = GEP->getNoWrapFlags();
1050 bool Overflow = false;
1051 while (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr)) {
1052 NW &= GEP->getNoWrapFlags();
1053
1054 SmallVector<Value *, 4> NestedOps(llvm::drop_begin(RangeOrContainer: GEP->operands()));
1055
1056 // Do not try the incorporate the sub-GEP if some index is not a number.
1057 bool AllConstantInt = true;
1058 for (Value *NestedOp : NestedOps)
1059 if (!isa<ConstantInt>(Val: NestedOp)) {
1060 AllConstantInt = false;
1061 break;
1062 }
1063 if (!AllConstantInt)
1064 break;
1065
1066 // Adjust inrange offset and intersect inrange attributes
1067 if (auto GEPRange = GEP->getInRange()) {
1068 auto AdjustedGEPRange = GEPRange->sextOrTrunc(BitWidth).subtract(CI: Offset);
1069 InRange =
1070 InRange ? InRange->intersectWith(CR: AdjustedGEPRange) : AdjustedGEPRange;
1071 }
1072
1073 Ptr = cast<Constant>(Val: GEP->getOperand(i_nocapture: 0));
1074 SrcElemTy = GEP->getSourceElementType();
1075 Offset = Offset.sadd_ov(
1076 RHS: APInt(BitWidth, DL.getIndexedOffsetInType(ElemTy: SrcElemTy, Indices: NestedOps),
1077 /*isSigned=*/true, /*implicitTrunc=*/true),
1078 Overflow);
1079 }
1080
1081 // Preserving nusw (without inbounds) also requires that the offset
1082 // additions did not overflow.
1083 if (NW.hasNoUnsignedSignedWrap() && !NW.isInBounds() && Overflow)
1084 NW = NW.withoutNoUnsignedSignedWrap();
1085
1086 // If the base value for this address is a literal integer value, fold the
1087 // getelementptr to the resulting integer value casted to the pointer type.
1088 APInt BaseIntVal(DL.getPointerTypeSizeInBits(Ptr->getType()), 0);
1089 if (auto *CE = dyn_cast<ConstantExpr>(Val: Ptr)) {
1090 if (CE->getOpcode() == Instruction::IntToPtr) {
1091 if (auto *Base = dyn_cast<ConstantInt>(Val: CE->getOperand(i_nocapture: 0)))
1092 BaseIntVal = Base->getValue().zextOrTrunc(width: BaseIntVal.getBitWidth());
1093 }
1094 }
1095
1096 if ((Ptr->isNullValue() || BaseIntVal != 0) &&
1097 !DL.mustNotIntroduceIntToPtr(Ty: Ptr->getType())) {
1098
1099 // If the index size is smaller than the pointer size, add to the low
1100 // bits only.
1101 BaseIntVal.insertBits(SubBits: BaseIntVal.trunc(width: BitWidth) + Offset, bitPosition: 0);
1102 Constant *C = ConstantInt::get(Context&: Ptr->getContext(), V: BaseIntVal);
1103 return ConstantExpr::getIntToPtr(C, Ty: ResTy);
1104 }
1105
1106 // Try to infer inbounds for GEPs of globals.
1107 if (!NW.isInBounds() && Offset.isNonNegative()) {
1108 bool CanBeNull;
1109 uint64_t DerefBytes = Ptr->getPointerDereferenceableBytes(
1110 DL, CanBeNull, /*CanBeFreed=*/nullptr);
1111 if (DerefBytes != 0 && !CanBeNull && Offset.sle(RHS: DerefBytes))
1112 NW |= GEPNoWrapFlags::inBounds();
1113 }
1114
1115 // nusw + nneg -> nuw
1116 if (NW.hasNoUnsignedSignedWrap() && Offset.isNonNegative())
1117 NW |= GEPNoWrapFlags::noUnsignedWrap();
1118
1119 // Otherwise canonicalize this to a single ptradd.
1120 LLVMContext &Ctx = Ptr->getContext();
1121 return ConstantExpr::getPtrAdd(Ptr, Offset: ConstantInt::get(Context&: Ctx, V: Offset), NW,
1122 InRange);
1123}
1124
1125/// Attempt to constant fold an instruction with the
1126/// specified opcode and operands. If successful, the constant result is
1127/// returned, if not, null is returned. Note that this function can fail when
1128/// attempting to fold instructions like loads and stores, which have no
1129/// constant expression form.
1130Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
1131 ArrayRef<Constant *> Ops,
1132 const DataLayout &DL,
1133 const TargetLibraryInfo *TLI,
1134 bool AllowNonDeterministic) {
1135 Type *DestTy = InstOrCE->getType();
1136
1137 if (Instruction::isUnaryOp(Opcode))
1138 return ConstantFoldUnaryOpOperand(Opcode, Op: Ops[0], DL);
1139
1140 if (Instruction::isBinaryOp(Opcode)) {
1141 switch (Opcode) {
1142 default:
1143 break;
1144 case Instruction::FAdd:
1145 case Instruction::FSub:
1146 case Instruction::FMul:
1147 case Instruction::FDiv:
1148 case Instruction::FRem:
1149 // Handle floating point instructions separately to account for denormals
1150 // TODO: If a constant expression is being folded rather than an
1151 // instruction, denormals will not be flushed/treated as zero
1152 if (const auto *I = dyn_cast<Instruction>(Val: InstOrCE)) {
1153 return ConstantFoldFPInstOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL, I,
1154 AllowNonDeterministic);
1155 }
1156 }
1157 return ConstantFoldBinaryOpOperands(Opcode, LHS: Ops[0], RHS: Ops[1], DL);
1158 }
1159
1160 if (Instruction::isCast(Opcode))
1161 return ConstantFoldCastOperand(Opcode, C: Ops[0], DestTy, DL);
1162
1163 if (auto *GEP = dyn_cast<GEPOperator>(Val: InstOrCE)) {
1164 Type *SrcElemTy = GEP->getSourceElementType();
1165 if (!ConstantExpr::isSupportedGetElementPtr(SrcElemTy))
1166 return nullptr;
1167
1168 if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1169 return C;
1170
1171 return ConstantExpr::getGetElementPtr(Ty: SrcElemTy, C: Ops[0], IdxList: Ops.slice(N: 1),
1172 NW: GEP->getNoWrapFlags(),
1173 InRange: GEP->getInRange());
1174 }
1175
1176 if (auto *CE = dyn_cast<ConstantExpr>(Val: InstOrCE))
1177 return CE->getWithOperands(Ops);
1178
1179 switch (Opcode) {
1180 default: return nullptr;
1181 case Instruction::ICmp:
1182 case Instruction::FCmp: {
1183 auto *C = cast<CmpInst>(Val: InstOrCE);
1184 return ConstantFoldCompareInstOperands(Predicate: C->getPredicate(), LHS: Ops[0], RHS: Ops[1],
1185 DL, TLI, I: C);
1186 }
1187 case Instruction::Freeze:
1188 return isGuaranteedNotToBeUndefOrPoison(V: Ops[0]) ? Ops[0] : nullptr;
1189 case Instruction::Call:
1190 if (auto *F = dyn_cast<Function>(Val: Ops.back())) {
1191 const auto *Call = cast<CallBase>(Val: InstOrCE);
1192 if (canConstantFoldCallTo(Call, F))
1193 return ConstantFoldCall(Call, F, Operands: Ops.slice(N: 0, M: Ops.size() - 1), TLI,
1194 AllowNonDeterministic);
1195 }
1196 return nullptr;
1197 case Instruction::Select:
1198 return ConstantFoldSelectInstruction(Cond: Ops[0], V1: Ops[1], V2: Ops[2]);
1199 case Instruction::ExtractElement:
1200 return ConstantExpr::getExtractElement(Vec: Ops[0], Idx: Ops[1]);
1201 case Instruction::ExtractValue:
1202 return ConstantFoldExtractValueInstruction(
1203 Agg: Ops[0], Idxs: cast<ExtractValueInst>(Val: InstOrCE)->getIndices());
1204 case Instruction::InsertElement:
1205 return ConstantExpr::getInsertElement(Vec: Ops[0], Elt: Ops[1], Idx: Ops[2]);
1206 case Instruction::InsertValue:
1207 return ConstantFoldInsertValueInstruction(
1208 Agg: Ops[0], Val: Ops[1], Idxs: cast<InsertValueInst>(Val: InstOrCE)->getIndices());
1209 case Instruction::ShuffleVector:
1210 return ConstantExpr::getShuffleVector(
1211 V1: Ops[0], V2: Ops[1], Mask: cast<ShuffleVectorInst>(Val: InstOrCE)->getShuffleMask());
1212 case Instruction::Load: {
1213 const auto *LI = dyn_cast<LoadInst>(Val: InstOrCE);
1214 if (LI->isVolatile())
1215 return nullptr;
1216 return ConstantFoldLoadFromConstPtr(C: Ops[0], Ty: LI->getType(), DL);
1217 }
1218 }
1219}
1220
1221} // end anonymous namespace
1222
1223//===----------------------------------------------------------------------===//
1224// Constant Folding public APIs
1225//===----------------------------------------------------------------------===//
1226
1227namespace {
1228
1229Constant *
1230ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1231 const TargetLibraryInfo *TLI,
1232 SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1233 if (!isa<ConstantVector>(Val: C) && !isa<ConstantExpr>(Val: C))
1234 return const_cast<Constant *>(C);
1235
1236 SmallVector<Constant *, 8> Ops;
1237 for (const Use &OldU : C->operands()) {
1238 Constant *OldC = cast<Constant>(Val: &OldU);
1239 Constant *NewC = OldC;
1240 // Recursively fold the ConstantExpr's operands. If we have already folded
1241 // a ConstantExpr, we don't have to process it again.
1242 if (isa<ConstantVector>(Val: OldC) || isa<ConstantExpr>(Val: OldC)) {
1243 auto It = FoldedOps.find(Val: OldC);
1244 if (It == FoldedOps.end()) {
1245 NewC = ConstantFoldConstantImpl(C: OldC, DL, TLI, FoldedOps);
1246 FoldedOps.insert(KV: {OldC, NewC});
1247 } else {
1248 NewC = It->second;
1249 }
1250 }
1251 Ops.push_back(Elt: NewC);
1252 }
1253
1254 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
1255 if (Constant *Res = ConstantFoldInstOperandsImpl(
1256 InstOrCE: CE, Opcode: CE->getOpcode(), Ops, DL, TLI, /*AllowNonDeterministic=*/true))
1257 return Res;
1258 return const_cast<Constant *>(C);
1259 }
1260
1261 assert(isa<ConstantVector>(C));
1262 return ConstantVector::get(V: Ops);
1263}
1264
1265} // end anonymous namespace
1266
1267Constant *llvm::ConstantFoldInstruction(const Instruction *I,
1268 const DataLayout &DL,
1269 const TargetLibraryInfo *TLI) {
1270 // Handle PHI nodes quickly here...
1271 if (auto *PN = dyn_cast<PHINode>(Val: I)) {
1272 Constant *CommonValue = nullptr;
1273
1274 SmallDenseMap<Constant *, Constant *> FoldedOps;
1275 for (Value *Incoming : PN->incoming_values()) {
1276 // If the incoming value is undef then skip it. Note that while we could
1277 // skip the value if it is equal to the phi node itself we choose not to
1278 // because that would break the rule that constant folding only applies if
1279 // all operands are constants.
1280 if (isa<UndefValue>(Val: Incoming))
1281 continue;
1282 // If the incoming value is not a constant, then give up.
1283 auto *C = dyn_cast<Constant>(Val: Incoming);
1284 if (!C)
1285 return nullptr;
1286 // Fold the PHI's operands.
1287 C = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1288 // If the incoming value is a different constant to
1289 // the one we saw previously, then give up.
1290 if (CommonValue && C != CommonValue)
1291 return nullptr;
1292 CommonValue = C;
1293 }
1294
1295 // If we reach here, all incoming values are the same constant or undef.
1296 return CommonValue ? CommonValue : UndefValue::get(T: PN->getType());
1297 }
1298
1299 // Scan the operand list, checking to see if they are all constants, if so,
1300 // hand off to ConstantFoldInstOperandsImpl.
1301 if (!all_of(Range: I->operands(), P: [](const Use &U) { return isa<Constant>(Val: U); }))
1302 return nullptr;
1303
1304 SmallDenseMap<Constant *, Constant *> FoldedOps;
1305 SmallVector<Constant *, 8> Ops;
1306 for (const Use &OpU : I->operands()) {
1307 auto *Op = cast<Constant>(Val: &OpU);
1308 // Fold the Instruction's operands.
1309 Op = ConstantFoldConstantImpl(C: Op, DL, TLI, FoldedOps);
1310 Ops.push_back(Elt: Op);
1311 }
1312
1313 return ConstantFoldInstOperands(I, Ops, DL, TLI);
1314}
1315
1316Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1317 const TargetLibraryInfo *TLI) {
1318 SmallDenseMap<Constant *, Constant *> FoldedOps;
1319 return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1320}
1321
1322Constant *llvm::ConstantFoldInstOperands(const Instruction *I,
1323 ArrayRef<Constant *> Ops,
1324 const DataLayout &DL,
1325 const TargetLibraryInfo *TLI,
1326 bool AllowNonDeterministic) {
1327 return ConstantFoldInstOperandsImpl(InstOrCE: I, Opcode: I->getOpcode(), Ops, DL, TLI,
1328 AllowNonDeterministic);
1329}
1330
1331Constant *llvm::ConstantFoldCompareInstOperands(
1332 unsigned IntPredicate, Constant *Ops0, Constant *Ops1, const DataLayout &DL,
1333 const TargetLibraryInfo *TLI, const Instruction *I) {
1334 CmpInst::Predicate Predicate = (CmpInst::Predicate)IntPredicate;
1335 // fold: icmp (inttoptr x), null -> icmp x, 0
1336 // fold: icmp null, (inttoptr x) -> icmp 0, x
1337 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1338 // fold: icmp 0, (ptrtoint x) -> icmp null, x
1339 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1340 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1341 //
1342 // FIXME: The following comment is out of data and the DataLayout is here now.
1343 // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1344 // around to know if bit truncation is happening.
1345 if (auto *CE0 = dyn_cast<ConstantExpr>(Val: Ops0)) {
1346 if (Ops1->isNullValue()) {
1347 if (CE0->getOpcode() == Instruction::IntToPtr) {
1348 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1349 // Convert the integer value to the right size to ensure we get the
1350 // proper extension or truncation.
1351 if (Constant *C = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy,
1352 /*IsSigned*/ false, DL)) {
1353 Constant *Null = Constant::getNullValue(Ty: C->getType());
1354 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI);
1355 }
1356 }
1357
1358 // icmp only compares the address part of the pointer, so only do this
1359 // transform if the integer size matches the address size.
1360 if (CE0->getOpcode() == Instruction::PtrToInt ||
1361 CE0->getOpcode() == Instruction::PtrToAddr) {
1362 Type *AddrTy = DL.getAddressType(PtrTy: CE0->getOperand(i_nocapture: 0)->getType());
1363 if (CE0->getType() == AddrTy) {
1364 Constant *C = CE0->getOperand(i_nocapture: 0);
1365 Constant *Null = Constant::getNullValue(Ty: C->getType());
1366 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C, Ops1: Null, DL, TLI);
1367 }
1368 }
1369 }
1370
1371 if (auto *CE1 = dyn_cast<ConstantExpr>(Val: Ops1)) {
1372 if (CE0->getOpcode() == CE1->getOpcode()) {
1373 if (CE0->getOpcode() == Instruction::IntToPtr) {
1374 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1375
1376 // Convert the integer value to the right size to ensure we get the
1377 // proper extension or truncation.
1378 Constant *C0 = ConstantFoldIntegerCast(C: CE0->getOperand(i_nocapture: 0), DestTy: IntPtrTy,
1379 /*IsSigned*/ false, DL);
1380 Constant *C1 = ConstantFoldIntegerCast(C: CE1->getOperand(i_nocapture: 0), DestTy: IntPtrTy,
1381 /*IsSigned*/ false, DL);
1382 if (C0 && C1)
1383 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: C0, Ops1: C1, DL, TLI);
1384 }
1385
1386 // icmp only compares the address part of the pointer, so only do this
1387 // transform if the integer size matches the address size.
1388 if (CE0->getOpcode() == Instruction::PtrToInt ||
1389 CE0->getOpcode() == Instruction::PtrToAddr) {
1390 Type *AddrTy = DL.getAddressType(PtrTy: CE0->getOperand(i_nocapture: 0)->getType());
1391 if (CE0->getType() == AddrTy &&
1392 CE0->getOperand(i_nocapture: 0)->getType() == CE1->getOperand(i_nocapture: 0)->getType()) {
1393 return ConstantFoldCompareInstOperands(
1394 IntPredicate: Predicate, Ops0: CE0->getOperand(i_nocapture: 0), Ops1: CE1->getOperand(i_nocapture: 0), DL, TLI);
1395 }
1396 }
1397 }
1398 }
1399
1400 // Convert pointer comparison (base+offset1) pred (base+offset2) into
1401 // offset1 pred offset2, for the case where the offset is inbounds. This
1402 // only works for equality and unsigned comparison, as inbounds permits
1403 // crossing the sign boundary. However, the offset comparison itself is
1404 // signed.
1405 if (Ops0->getType()->isPointerTy() && !ICmpInst::isSigned(Pred: Predicate)) {
1406 unsigned IndexWidth = DL.getIndexTypeSizeInBits(Ty: Ops0->getType());
1407 APInt Offset0(IndexWidth, 0);
1408 bool IsEqPred = ICmpInst::isEquality(P: Predicate);
1409 Value *Stripped0 = Ops0->stripAndAccumulateConstantOffsets(
1410 DL, Offset&: Offset0, /*AllowNonInbounds=*/IsEqPred,
1411 /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr,
1412 /*LookThroughIntToPtr=*/IsEqPred);
1413 APInt Offset1(IndexWidth, 0);
1414 Value *Stripped1 = Ops1->stripAndAccumulateConstantOffsets(
1415 DL, Offset&: Offset1, /*AllowNonInbounds=*/IsEqPred,
1416 /*AllowInvariantGroup=*/false, /*ExternalAnalysis=*/nullptr,
1417 /*LookThroughIntToPtr=*/IsEqPred);
1418 if (Stripped0 == Stripped1)
1419 return ConstantInt::getBool(
1420 Context&: Ops0->getContext(),
1421 V: ICmpInst::compare(LHS: Offset0, RHS: Offset1,
1422 Pred: ICmpInst::getSignedPredicate(Pred: Predicate)));
1423 }
1424 } else if (isa<ConstantExpr>(Val: Ops1)) {
1425 // If RHS is a constant expression, but the left side isn't, swap the
1426 // operands and try again.
1427 Predicate = ICmpInst::getSwappedPredicate(pred: Predicate);
1428 return ConstantFoldCompareInstOperands(IntPredicate: Predicate, Ops0: Ops1, Ops1: Ops0, DL, TLI);
1429 }
1430
1431 if (CmpInst::isFPPredicate(P: Predicate)) {
1432 // Flush any denormal constant float input according to denormal handling
1433 // mode.
1434 Ops0 = FlushFPConstant(Operand: Ops0, I, /*IsOutput=*/false);
1435 if (!Ops0)
1436 return nullptr;
1437 Ops1 = FlushFPConstant(Operand: Ops1, I, /*IsOutput=*/false);
1438 if (!Ops1)
1439 return nullptr;
1440 }
1441
1442 return ConstantFoldCompareInstruction(Predicate, C1: Ops0, C2: Ops1);
1443}
1444
1445Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op,
1446 const DataLayout &DL) {
1447 assert(Instruction::isUnaryOp(Opcode));
1448
1449 return ConstantFoldUnaryInstruction(Opcode, V: Op);
1450}
1451
1452Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1453 Constant *RHS,
1454 const DataLayout &DL) {
1455 assert(Instruction::isBinaryOp(Opcode));
1456 if (isa<ConstantExpr>(Val: LHS) || isa<ConstantExpr>(Val: RHS))
1457 if (Constant *C = SymbolicallyEvaluateBinop(Opc: Opcode, Op0: LHS, Op1: RHS, DL))
1458 return C;
1459
1460 if (ConstantExpr::isDesirableBinOp(Opcode))
1461 return ConstantExpr::get(Opcode, C1: LHS, C2: RHS);
1462 return ConstantFoldBinaryInstruction(Opcode, V1: LHS, V2: RHS);
1463}
1464
1465static ConstantFP *flushDenormalConstant(Type *Ty, const APFloat &APF,
1466 DenormalMode::DenormalModeKind Mode) {
1467 switch (Mode) {
1468 case DenormalMode::Dynamic:
1469 return nullptr;
1470 case DenormalMode::IEEE:
1471 return ConstantFP::get(Ty, V: APF);
1472 case DenormalMode::PreserveSign:
1473 return ConstantFP::get(
1474 Ty, V: APFloat::getZero(Sem: APF.getSemantics(), Negative: APF.isNegative()));
1475 case DenormalMode::PositiveZero:
1476 return ConstantFP::get(Ty, V: APFloat::getZero(Sem: APF.getSemantics(), Negative: false));
1477 default:
1478 break;
1479 }
1480
1481 llvm_unreachable("unknown denormal mode");
1482}
1483
1484/// Return the denormal mode that can be assumed when executing a floating point
1485/// operation at \p CtxI.
1486static DenormalMode getInstrDenormalMode(const Instruction *CtxI, Type *Ty) {
1487 if (!CtxI || !CtxI->getParent() || !CtxI->getFunction())
1488 return DenormalMode::getDynamic();
1489 return CtxI->getFunction()->getDenormalMode(
1490 FPType: Ty->getScalarType()->getFltSemantics());
1491}
1492
1493static ConstantFP *flushDenormalConstantFP(ConstantFP *CFP,
1494 const Instruction *Inst,
1495 bool IsOutput) {
1496 const APFloat &APF = CFP->getValueAPF();
1497 if (!APF.isDenormal())
1498 return CFP;
1499
1500 DenormalMode Mode = getInstrDenormalMode(CtxI: Inst, Ty: CFP->getType());
1501 return flushDenormalConstant(Ty: CFP->getType(), APF,
1502 Mode: IsOutput ? Mode.Output : Mode.Input);
1503}
1504
1505Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *Inst,
1506 bool IsOutput) {
1507 if (ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Operand))
1508 return flushDenormalConstantFP(CFP, Inst, IsOutput);
1509
1510 if (isa<ConstantAggregateZero, UndefValue>(Val: Operand))
1511 return Operand;
1512
1513 Type *Ty = Operand->getType();
1514 VectorType *VecTy = dyn_cast<VectorType>(Val: Ty);
1515 if (VecTy) {
1516 if (auto *Splat = dyn_cast_or_null<ConstantFP>(Val: Operand->getSplatValue())) {
1517 ConstantFP *Folded = flushDenormalConstantFP(CFP: Splat, Inst, IsOutput);
1518 if (!Folded)
1519 return nullptr;
1520 return ConstantVector::getSplat(EC: VecTy->getElementCount(), Elt: Folded);
1521 }
1522
1523 Ty = VecTy->getElementType();
1524 }
1525
1526 if (isa<ConstantExpr>(Val: Operand))
1527 return Operand;
1528
1529 if (const auto *CV = dyn_cast<ConstantVector>(Val: Operand)) {
1530 SmallVector<Constant *, 16> NewElts;
1531 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1532 Constant *Element = CV->getAggregateElement(Elt: i);
1533 if (isa<UndefValue>(Val: Element)) {
1534 NewElts.push_back(Elt: Element);
1535 continue;
1536 }
1537
1538 ConstantFP *CFP = dyn_cast<ConstantFP>(Val: Element);
1539 if (!CFP)
1540 return nullptr;
1541
1542 ConstantFP *Folded = flushDenormalConstantFP(CFP, Inst, IsOutput);
1543 if (!Folded)
1544 return nullptr;
1545 NewElts.push_back(Elt: Folded);
1546 }
1547
1548 return ConstantVector::get(V: NewElts);
1549 }
1550
1551 if (const auto *CDV = dyn_cast<ConstantDataVector>(Val: Operand)) {
1552 SmallVector<Constant *, 16> NewElts;
1553 for (unsigned I = 0, E = CDV->getNumElements(); I < E; ++I) {
1554 const APFloat &Elt = CDV->getElementAsAPFloat(i: I);
1555 if (!Elt.isDenormal()) {
1556 NewElts.push_back(Elt: ConstantFP::get(Ty, V: Elt));
1557 } else {
1558 DenormalMode Mode = getInstrDenormalMode(CtxI: Inst, Ty);
1559 ConstantFP *Folded =
1560 flushDenormalConstant(Ty, APF: Elt, Mode: IsOutput ? Mode.Output : Mode.Input);
1561 if (!Folded)
1562 return nullptr;
1563 NewElts.push_back(Elt: Folded);
1564 }
1565 }
1566
1567 return ConstantVector::get(V: NewElts);
1568 }
1569
1570 return nullptr;
1571}
1572
1573Constant *llvm::ConstantFoldFPInstOperands(unsigned Opcode, Constant *LHS,
1574 Constant *RHS, const DataLayout &DL,
1575 const Instruction *I,
1576 bool AllowNonDeterministic) {
1577 if (Instruction::isBinaryOp(Opcode)) {
1578 // Flush denormal inputs if needed.
1579 Constant *Op0 = FlushFPConstant(Operand: LHS, Inst: I, /* IsOutput */ false);
1580 if (!Op0)
1581 return nullptr;
1582 Constant *Op1 = FlushFPConstant(Operand: RHS, Inst: I, /* IsOutput */ false);
1583 if (!Op1)
1584 return nullptr;
1585
1586 // If nsz or an algebraic FMF flag is set, the result of the FP operation
1587 // may change due to future optimization. Don't constant fold them if
1588 // non-deterministic results are not allowed.
1589 if (!AllowNonDeterministic)
1590 if (auto *FP = dyn_cast_or_null<FPMathOperator>(Val: I))
1591 if (FP->hasNoSignedZeros() || FP->hasAllowReassoc() ||
1592 FP->hasAllowContract() || FP->hasAllowReciprocal())
1593 return nullptr;
1594
1595 // Calculate constant result.
1596 Constant *C = ConstantFoldBinaryOpOperands(Opcode, LHS: Op0, RHS: Op1, DL);
1597 if (!C)
1598 return nullptr;
1599
1600 // Flush denormal output if needed.
1601 C = FlushFPConstant(Operand: C, Inst: I, /* IsOutput */ true);
1602 if (!C)
1603 return nullptr;
1604
1605 // The precise NaN value is non-deterministic.
1606 if (!AllowNonDeterministic && C->isNaN())
1607 return nullptr;
1608
1609 return C;
1610 }
1611 // If instruction lacks a parent/function and the denormal mode cannot be
1612 // determined, use the default (IEEE).
1613 return ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL);
1614}
1615
1616Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1617 Type *DestTy, const DataLayout &DL) {
1618 assert(Instruction::isCast(Opcode));
1619
1620 if (auto *CE = dyn_cast<ConstantExpr>(Val: C))
1621 if (CE->isCast())
1622 if (unsigned NewOp = CastInst::isEliminableCastPair(
1623 firstOpcode: Instruction::CastOps(CE->getOpcode()),
1624 secondOpcode: Instruction::CastOps(Opcode), SrcTy: CE->getOperand(i_nocapture: 0)->getType(),
1625 MidTy: C->getType(), DstTy: DestTy, DL: &DL))
1626 return ConstantFoldCastOperand(Opcode: NewOp, C: CE->getOperand(i_nocapture: 0), DestTy, DL);
1627
1628 switch (Opcode) {
1629 default:
1630 llvm_unreachable("Missing case");
1631 case Instruction::PtrToAddr:
1632 case Instruction::PtrToInt:
1633 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
1634 Constant *FoldedValue = nullptr;
1635 // If the input is an inttoptr, eliminate the pair. This requires knowing
1636 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1637 if (CE->getOpcode() == Instruction::IntToPtr) {
1638 // zext/trunc the inttoptr to pointer/address size.
1639 Type *MidTy = Opcode == Instruction::PtrToInt
1640 ? DL.getAddressType(PtrTy: CE->getType())
1641 : DL.getIntPtrType(CE->getType());
1642 FoldedValue = ConstantFoldIntegerCast(C: CE->getOperand(i_nocapture: 0), DestTy: MidTy,
1643 /*IsSigned=*/false, DL);
1644 } else if (auto *GEP = dyn_cast<GEPOperator>(Val: CE)) {
1645 // If we have GEP, we can perform the following folds:
1646 // (ptrtoint/ptrtoaddr (gep null, x)) -> x
1647 // (ptrtoint/ptrtoaddr (gep (gep null, x), y) -> x + y, etc.
1648 unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: GEP->getType());
1649 APInt BaseOffset(BitWidth, 0);
1650 auto *Base = cast<Constant>(Val: GEP->stripAndAccumulateConstantOffsets(
1651 DL, Offset&: BaseOffset, /*AllowNonInbounds=*/true));
1652 if (Base->isNullValue()) {
1653 FoldedValue = ConstantInt::get(Context&: CE->getContext(), V: BaseOffset);
1654 } else {
1655 // ptrtoint/ptrtoaddr (gep i8, Ptr, (sub 0, V))
1656 // -> sub (ptrtoint/ptrtoaddr Ptr), V
1657 if (GEP->getNumIndices() == 1 &&
1658 GEP->getSourceElementType()->isIntegerTy(BitWidth: 8)) {
1659 auto *Ptr = cast<Constant>(Val: GEP->getPointerOperand());
1660 auto *Sub = dyn_cast<ConstantExpr>(Val: GEP->getOperand(i_nocapture: 1));
1661 Type *IntIdxTy = DL.getIndexType(PtrTy: Ptr->getType());
1662 if (Sub && Sub->getType() == IntIdxTy &&
1663 Sub->getOpcode() == Instruction::Sub &&
1664 Sub->getOperand(i_nocapture: 0)->isNullValue())
1665 FoldedValue = ConstantExpr::getSub(
1666 C1: ConstantExpr::getCast(ops: Opcode, C: Ptr, Ty: IntIdxTy),
1667 C2: Sub->getOperand(i_nocapture: 1));
1668 }
1669 }
1670 }
1671 if (FoldedValue) {
1672 // Do a zext or trunc to get to the ptrtoint/ptrtoaddr dest size.
1673 return ConstantFoldIntegerCast(C: FoldedValue, DestTy, /*IsSigned=*/false,
1674 DL);
1675 }
1676 }
1677 break;
1678 case Instruction::IntToPtr:
1679 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1680 // the int size is >= the ptr size and the address spaces are the same.
1681 // This requires knowing the width of a pointer, so it can't be done in
1682 // ConstantExpr::getCast.
1683 if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) {
1684 if (CE->getOpcode() == Instruction::PtrToInt) {
1685 Constant *SrcPtr = CE->getOperand(i_nocapture: 0);
1686 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1687 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1688
1689 if (MidIntSize >= SrcPtrSize) {
1690 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1691 if (SrcAS == DestTy->getPointerAddressSpace())
1692 return FoldBitCast(C: CE->getOperand(i_nocapture: 0), DestTy, DL);
1693 }
1694 }
1695 }
1696 break;
1697 case Instruction::Trunc:
1698 case Instruction::ZExt:
1699 case Instruction::SExt:
1700 case Instruction::FPTrunc:
1701 case Instruction::FPExt:
1702 case Instruction::UIToFP:
1703 case Instruction::SIToFP:
1704 case Instruction::FPToUI:
1705 case Instruction::FPToSI:
1706 case Instruction::AddrSpaceCast:
1707 break;
1708 case Instruction::BitCast:
1709 return FoldBitCast(C, DestTy, DL);
1710 }
1711
1712 if (ConstantExpr::isDesirableCastOp(Opcode))
1713 return ConstantExpr::getCast(ops: Opcode, C, Ty: DestTy);
1714 return ConstantFoldCastInstruction(opcode: Opcode, V: C, DestTy);
1715}
1716
1717Constant *llvm::ConstantFoldIntegerCast(Constant *C, Type *DestTy,
1718 bool IsSigned, const DataLayout &DL) {
1719 Type *SrcTy = C->getType();
1720 if (SrcTy == DestTy)
1721 return C;
1722 if (SrcTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits())
1723 return ConstantFoldCastOperand(Opcode: Instruction::Trunc, C, DestTy, DL);
1724 if (IsSigned)
1725 return ConstantFoldCastOperand(Opcode: Instruction::SExt, C, DestTy, DL);
1726 return ConstantFoldCastOperand(Opcode: Instruction::ZExt, C, DestTy, DL);
1727}
1728
1729//===----------------------------------------------------------------------===//
1730// Constant Folding for Calls
1731//
1732
1733bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
1734 if (Call->isNoBuiltin())
1735 return false;
1736 if (Call->getFunctionType() != F->getFunctionType())
1737 return false;
1738
1739 // Allow FP calls (both libcalls and intrinsics) to avoid being folded.
1740 // This can be useful for GPU targets or in cross-compilation scenarios
1741 // when the exact target FP behaviour is required, and the host compiler's
1742 // behaviour may be slightly different from the device's run-time behaviour.
1743 if (DisableFPCallFolding && (F->getReturnType()->isFloatingPointTy() ||
1744 any_of(Range: F->args(), P: [](const Argument &Arg) {
1745 return Arg.getType()->isFloatingPointTy();
1746 })))
1747 return false;
1748
1749 switch (F->getIntrinsicID()) {
1750 // Operations that do not operate floating-point numbers and do not depend on
1751 // FP environment can be folded even in strictfp functions.
1752 case Intrinsic::bswap:
1753 case Intrinsic::ctpop:
1754 case Intrinsic::ctlz:
1755 case Intrinsic::cttz:
1756 case Intrinsic::fshl:
1757 case Intrinsic::fshr:
1758 case Intrinsic::clmul:
1759 case Intrinsic::pdep:
1760 case Intrinsic::pext:
1761 case Intrinsic::launder_invariant_group:
1762 case Intrinsic::strip_invariant_group:
1763 case Intrinsic::masked_load:
1764 case Intrinsic::get_active_lane_mask:
1765 case Intrinsic::abs:
1766 case Intrinsic::smax:
1767 case Intrinsic::smin:
1768 case Intrinsic::umax:
1769 case Intrinsic::umin:
1770 case Intrinsic::scmp:
1771 case Intrinsic::ucmp:
1772 case Intrinsic::sadd_with_overflow:
1773 case Intrinsic::uadd_with_overflow:
1774 case Intrinsic::ssub_with_overflow:
1775 case Intrinsic::usub_with_overflow:
1776 case Intrinsic::smul_with_overflow:
1777 case Intrinsic::umul_with_overflow:
1778 case Intrinsic::sadd_sat:
1779 case Intrinsic::uadd_sat:
1780 case Intrinsic::ssub_sat:
1781 case Intrinsic::usub_sat:
1782 case Intrinsic::smul_fix:
1783 case Intrinsic::smul_fix_sat:
1784 case Intrinsic::bitreverse:
1785 case Intrinsic::is_constant:
1786 case Intrinsic::vector_reduce_add:
1787 case Intrinsic::vector_reduce_mul:
1788 case Intrinsic::vector_reduce_and:
1789 case Intrinsic::vector_reduce_or:
1790 case Intrinsic::vector_reduce_xor:
1791 case Intrinsic::vector_reduce_smin:
1792 case Intrinsic::vector_reduce_smax:
1793 case Intrinsic::vector_reduce_umin:
1794 case Intrinsic::vector_reduce_umax:
1795 case Intrinsic::vector_extract:
1796 case Intrinsic::vector_insert:
1797 case Intrinsic::vector_interleave2:
1798 case Intrinsic::vector_interleave3:
1799 case Intrinsic::vector_interleave4:
1800 case Intrinsic::vector_interleave5:
1801 case Intrinsic::vector_interleave6:
1802 case Intrinsic::vector_interleave7:
1803 case Intrinsic::vector_interleave8:
1804 case Intrinsic::vector_deinterleave2:
1805 case Intrinsic::vector_deinterleave3:
1806 case Intrinsic::vector_deinterleave4:
1807 case Intrinsic::vector_deinterleave5:
1808 case Intrinsic::vector_deinterleave6:
1809 case Intrinsic::vector_deinterleave7:
1810 case Intrinsic::vector_deinterleave8:
1811 // Target intrinsics
1812 case Intrinsic::amdgcn_perm:
1813 case Intrinsic::amdgcn_wave_reduce_umin:
1814 case Intrinsic::amdgcn_wave_reduce_umax:
1815 case Intrinsic::amdgcn_wave_reduce_max:
1816 case Intrinsic::amdgcn_wave_reduce_min:
1817 case Intrinsic::amdgcn_wave_reduce_and:
1818 case Intrinsic::amdgcn_wave_reduce_or:
1819 case Intrinsic::amdgcn_s_wqm:
1820 case Intrinsic::amdgcn_s_quadmask:
1821 case Intrinsic::amdgcn_s_bitreplicate:
1822 case Intrinsic::arm_mve_vctp8:
1823 case Intrinsic::arm_mve_vctp16:
1824 case Intrinsic::arm_mve_vctp32:
1825 case Intrinsic::arm_mve_vctp64:
1826 case Intrinsic::aarch64_sve_convert_from_svbool:
1827 case Intrinsic::wasm_alltrue:
1828 case Intrinsic::wasm_anytrue:
1829 case Intrinsic::wasm_dot:
1830 // WebAssembly float semantics are always known
1831 case Intrinsic::wasm_trunc_signed:
1832 case Intrinsic::wasm_trunc_unsigned:
1833 return true;
1834
1835 // Floating point operations cannot be folded in strictfp functions in
1836 // general case. They can be folded if FP environment is known to compiler.
1837 case Intrinsic::minnum:
1838 case Intrinsic::maxnum:
1839 case Intrinsic::minimum:
1840 case Intrinsic::maximum:
1841 case Intrinsic::minimumnum:
1842 case Intrinsic::maximumnum:
1843 case Intrinsic::log:
1844 case Intrinsic::log2:
1845 case Intrinsic::log10:
1846 case Intrinsic::exp:
1847 case Intrinsic::exp2:
1848 case Intrinsic::exp10:
1849 case Intrinsic::sqrt:
1850 case Intrinsic::sin:
1851 case Intrinsic::cos:
1852 case Intrinsic::sincos:
1853 case Intrinsic::sinh:
1854 case Intrinsic::cosh:
1855 case Intrinsic::atan:
1856 case Intrinsic::pow:
1857 case Intrinsic::powi:
1858 case Intrinsic::ldexp:
1859 case Intrinsic::fma:
1860 case Intrinsic::fmuladd:
1861 case Intrinsic::frexp:
1862 case Intrinsic::fptoui_sat:
1863 case Intrinsic::fptosi_sat:
1864 case Intrinsic::amdgcn_cos:
1865 case Intrinsic::amdgcn_cubeid:
1866 case Intrinsic::amdgcn_cubema:
1867 case Intrinsic::amdgcn_cubesc:
1868 case Intrinsic::amdgcn_cubetc:
1869 case Intrinsic::amdgcn_fmul_legacy:
1870 case Intrinsic::amdgcn_fma_legacy:
1871 case Intrinsic::amdgcn_fract:
1872 case Intrinsic::amdgcn_sin:
1873 // The intrinsics below depend on rounding mode in MXCSR.
1874 case Intrinsic::x86_sse_cvtss2si:
1875 case Intrinsic::x86_sse_cvtss2si64:
1876 case Intrinsic::x86_sse_cvttss2si:
1877 case Intrinsic::x86_sse_cvttss2si64:
1878 case Intrinsic::x86_sse2_cvtsd2si:
1879 case Intrinsic::x86_sse2_cvtsd2si64:
1880 case Intrinsic::x86_sse2_cvttsd2si:
1881 case Intrinsic::x86_sse2_cvttsd2si64:
1882 case Intrinsic::x86_avx512_vcvtss2si32:
1883 case Intrinsic::x86_avx512_vcvtss2si64:
1884 case Intrinsic::x86_avx512_cvttss2si:
1885 case Intrinsic::x86_avx512_cvttss2si64:
1886 case Intrinsic::x86_avx512_vcvtsd2si32:
1887 case Intrinsic::x86_avx512_vcvtsd2si64:
1888 case Intrinsic::x86_avx512_cvttsd2si:
1889 case Intrinsic::x86_avx512_cvttsd2si64:
1890 case Intrinsic::x86_avx512_vcvtss2usi32:
1891 case Intrinsic::x86_avx512_vcvtss2usi64:
1892 case Intrinsic::x86_avx512_cvttss2usi:
1893 case Intrinsic::x86_avx512_cvttss2usi64:
1894 case Intrinsic::x86_avx512_vcvtsd2usi32:
1895 case Intrinsic::x86_avx512_vcvtsd2usi64:
1896 case Intrinsic::x86_avx512_cvttsd2usi:
1897 case Intrinsic::x86_avx512_cvttsd2usi64:
1898
1899 // NVVM FMax intrinsics
1900 case Intrinsic::nvvm_fmax_d:
1901 case Intrinsic::nvvm_fmax_f:
1902 case Intrinsic::nvvm_fmax_ftz_f:
1903 case Intrinsic::nvvm_fmax_ftz_nan_f:
1904 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
1905 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
1906 case Intrinsic::nvvm_fmax_nan_f:
1907 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
1908 case Intrinsic::nvvm_fmax_xorsign_abs_f:
1909
1910 // NVVM FMin intrinsics
1911 case Intrinsic::nvvm_fmin_d:
1912 case Intrinsic::nvvm_fmin_f:
1913 case Intrinsic::nvvm_fmin_ftz_f:
1914 case Intrinsic::nvvm_fmin_ftz_nan_f:
1915 case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
1916 case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
1917 case Intrinsic::nvvm_fmin_nan_f:
1918 case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
1919 case Intrinsic::nvvm_fmin_xorsign_abs_f:
1920
1921 // NVVM float/double to int32/uint32 conversion intrinsics
1922 case Intrinsic::nvvm_f2i_rm:
1923 case Intrinsic::nvvm_f2i_rn:
1924 case Intrinsic::nvvm_f2i_rp:
1925 case Intrinsic::nvvm_f2i_rz:
1926 case Intrinsic::nvvm_f2i_rm_ftz:
1927 case Intrinsic::nvvm_f2i_rn_ftz:
1928 case Intrinsic::nvvm_f2i_rp_ftz:
1929 case Intrinsic::nvvm_f2i_rz_ftz:
1930 case Intrinsic::nvvm_f2ui_rm:
1931 case Intrinsic::nvvm_f2ui_rn:
1932 case Intrinsic::nvvm_f2ui_rp:
1933 case Intrinsic::nvvm_f2ui_rz:
1934 case Intrinsic::nvvm_f2ui_rm_ftz:
1935 case Intrinsic::nvvm_f2ui_rn_ftz:
1936 case Intrinsic::nvvm_f2ui_rp_ftz:
1937 case Intrinsic::nvvm_f2ui_rz_ftz:
1938 case Intrinsic::nvvm_d2i_rm:
1939 case Intrinsic::nvvm_d2i_rn:
1940 case Intrinsic::nvvm_d2i_rp:
1941 case Intrinsic::nvvm_d2i_rz:
1942 case Intrinsic::nvvm_d2ui_rm:
1943 case Intrinsic::nvvm_d2ui_rn:
1944 case Intrinsic::nvvm_d2ui_rp:
1945 case Intrinsic::nvvm_d2ui_rz:
1946
1947 // NVVM float/double to int64/uint64 conversion intrinsics
1948 case Intrinsic::nvvm_f2ll_rm:
1949 case Intrinsic::nvvm_f2ll_rn:
1950 case Intrinsic::nvvm_f2ll_rp:
1951 case Intrinsic::nvvm_f2ll_rz:
1952 case Intrinsic::nvvm_f2ll_rm_ftz:
1953 case Intrinsic::nvvm_f2ll_rn_ftz:
1954 case Intrinsic::nvvm_f2ll_rp_ftz:
1955 case Intrinsic::nvvm_f2ll_rz_ftz:
1956 case Intrinsic::nvvm_f2ull_rm:
1957 case Intrinsic::nvvm_f2ull_rn:
1958 case Intrinsic::nvvm_f2ull_rp:
1959 case Intrinsic::nvvm_f2ull_rz:
1960 case Intrinsic::nvvm_f2ull_rm_ftz:
1961 case Intrinsic::nvvm_f2ull_rn_ftz:
1962 case Intrinsic::nvvm_f2ull_rp_ftz:
1963 case Intrinsic::nvvm_f2ull_rz_ftz:
1964 case Intrinsic::nvvm_d2ll_rm:
1965 case Intrinsic::nvvm_d2ll_rn:
1966 case Intrinsic::nvvm_d2ll_rp:
1967 case Intrinsic::nvvm_d2ll_rz:
1968 case Intrinsic::nvvm_d2ull_rm:
1969 case Intrinsic::nvvm_d2ull_rn:
1970 case Intrinsic::nvvm_d2ull_rp:
1971 case Intrinsic::nvvm_d2ull_rz:
1972
1973 // NVVM math intrinsics:
1974 case Intrinsic::nvvm_ceil_d:
1975 case Intrinsic::nvvm_ceil_f:
1976 case Intrinsic::nvvm_ceil_ftz_f:
1977
1978 case Intrinsic::nvvm_fabs:
1979 case Intrinsic::nvvm_fabs_ftz:
1980
1981 case Intrinsic::nvvm_floor_d:
1982 case Intrinsic::nvvm_floor_f:
1983 case Intrinsic::nvvm_floor_ftz_f:
1984
1985 case Intrinsic::nvvm_rcp_rm_d:
1986 case Intrinsic::nvvm_rcp_rm_f:
1987 case Intrinsic::nvvm_rcp_rm_ftz_f:
1988 case Intrinsic::nvvm_rcp_rn_d:
1989 case Intrinsic::nvvm_rcp_rn_f:
1990 case Intrinsic::nvvm_rcp_rn_ftz_f:
1991 case Intrinsic::nvvm_rcp_rp_d:
1992 case Intrinsic::nvvm_rcp_rp_f:
1993 case Intrinsic::nvvm_rcp_rp_ftz_f:
1994 case Intrinsic::nvvm_rcp_rz_d:
1995 case Intrinsic::nvvm_rcp_rz_f:
1996 case Intrinsic::nvvm_rcp_rz_ftz_f:
1997
1998 case Intrinsic::nvvm_round_d:
1999 case Intrinsic::nvvm_round_f:
2000 case Intrinsic::nvvm_round_ftz_f:
2001
2002 case Intrinsic::nvvm_saturate_d:
2003 case Intrinsic::nvvm_saturate_f:
2004 case Intrinsic::nvvm_saturate_ftz_f:
2005
2006 case Intrinsic::nvvm_sqrt_f:
2007 case Intrinsic::nvvm_sqrt_rn_d:
2008 case Intrinsic::nvvm_sqrt_rn_f:
2009 case Intrinsic::nvvm_sqrt_rn_ftz_f:
2010 return !Call->isStrictFP();
2011
2012 // NVVM add intrinsics with explicit rounding modes
2013 case Intrinsic::nvvm_add_rm_d:
2014 case Intrinsic::nvvm_add_rn_d:
2015 case Intrinsic::nvvm_add_rp_d:
2016 case Intrinsic::nvvm_add_rz_d:
2017 case Intrinsic::nvvm_add_rm_f:
2018 case Intrinsic::nvvm_add_rn_f:
2019 case Intrinsic::nvvm_add_rp_f:
2020 case Intrinsic::nvvm_add_rz_f:
2021 case Intrinsic::nvvm_add_rm_ftz_f:
2022 case Intrinsic::nvvm_add_rn_ftz_f:
2023 case Intrinsic::nvvm_add_rp_ftz_f:
2024 case Intrinsic::nvvm_add_rz_ftz_f:
2025
2026 // NVVM div intrinsics with explicit rounding modes
2027 case Intrinsic::nvvm_div_rm_d:
2028 case Intrinsic::nvvm_div_rn_d:
2029 case Intrinsic::nvvm_div_rp_d:
2030 case Intrinsic::nvvm_div_rz_d:
2031 case Intrinsic::nvvm_div_rm_f:
2032 case Intrinsic::nvvm_div_rn_f:
2033 case Intrinsic::nvvm_div_rp_f:
2034 case Intrinsic::nvvm_div_rz_f:
2035 case Intrinsic::nvvm_div_rm_ftz_f:
2036 case Intrinsic::nvvm_div_rn_ftz_f:
2037 case Intrinsic::nvvm_div_rp_ftz_f:
2038 case Intrinsic::nvvm_div_rz_ftz_f:
2039
2040 // NVVM mul intrinsics with explicit rounding modes
2041 case Intrinsic::nvvm_mul_rm_d:
2042 case Intrinsic::nvvm_mul_rn_d:
2043 case Intrinsic::nvvm_mul_rp_d:
2044 case Intrinsic::nvvm_mul_rz_d:
2045 case Intrinsic::nvvm_mul_rm_f:
2046 case Intrinsic::nvvm_mul_rn_f:
2047 case Intrinsic::nvvm_mul_rp_f:
2048 case Intrinsic::nvvm_mul_rz_f:
2049 case Intrinsic::nvvm_mul_rm_ftz_f:
2050 case Intrinsic::nvvm_mul_rn_ftz_f:
2051 case Intrinsic::nvvm_mul_rp_ftz_f:
2052 case Intrinsic::nvvm_mul_rz_ftz_f:
2053
2054 // NVVM fma intrinsics with explicit rounding modes
2055 case Intrinsic::nvvm_fma_rm_d:
2056 case Intrinsic::nvvm_fma_rn_d:
2057 case Intrinsic::nvvm_fma_rp_d:
2058 case Intrinsic::nvvm_fma_rz_d:
2059 case Intrinsic::nvvm_fma_rm_f:
2060 case Intrinsic::nvvm_fma_rn_f:
2061 case Intrinsic::nvvm_fma_rp_f:
2062 case Intrinsic::nvvm_fma_rz_f:
2063 case Intrinsic::nvvm_fma_rm_ftz_f:
2064 case Intrinsic::nvvm_fma_rn_ftz_f:
2065 case Intrinsic::nvvm_fma_rp_ftz_f:
2066 case Intrinsic::nvvm_fma_rz_ftz_f:
2067
2068 // Sign operations are actually bitwise operations, they do not raise
2069 // exceptions even for SNANs.
2070 case Intrinsic::fabs:
2071 case Intrinsic::copysign:
2072 case Intrinsic::is_fpclass:
2073 // Non-constrained variants of rounding operations means default FP
2074 // environment, they can be folded in any case.
2075 case Intrinsic::ceil:
2076 case Intrinsic::floor:
2077 case Intrinsic::round:
2078 case Intrinsic::roundeven:
2079 case Intrinsic::trunc:
2080 case Intrinsic::nearbyint:
2081 case Intrinsic::rint:
2082 case Intrinsic::canonicalize:
2083
2084 // Constrained intrinsics can be folded if FP environment is known
2085 // to compiler.
2086 case Intrinsic::experimental_constrained_fma:
2087 case Intrinsic::experimental_constrained_fmuladd:
2088 case Intrinsic::experimental_constrained_fadd:
2089 case Intrinsic::experimental_constrained_fsub:
2090 case Intrinsic::experimental_constrained_fmul:
2091 case Intrinsic::experimental_constrained_fdiv:
2092 case Intrinsic::experimental_constrained_frem:
2093 case Intrinsic::experimental_constrained_ceil:
2094 case Intrinsic::experimental_constrained_floor:
2095 case Intrinsic::experimental_constrained_round:
2096 case Intrinsic::experimental_constrained_roundeven:
2097 case Intrinsic::experimental_constrained_trunc:
2098 case Intrinsic::experimental_constrained_nearbyint:
2099 case Intrinsic::experimental_constrained_rint:
2100 case Intrinsic::experimental_constrained_fcmp:
2101 case Intrinsic::experimental_constrained_fcmps:
2102
2103 case Intrinsic::experimental_cttz_elts:
2104 return true;
2105 default:
2106 return false;
2107 case Intrinsic::not_intrinsic: break;
2108 }
2109
2110 if (!F->hasName() || Call->isStrictFP())
2111 return false;
2112
2113 // In these cases, the check of the length is required. We don't want to
2114 // return true for a name like "cos\0blah" which strcmp would return equal to
2115 // "cos", but has length 8.
2116 StringRef Name = F->getName();
2117 switch (Name[0]) {
2118 default:
2119 return false;
2120 // clang-format off
2121 case 'a':
2122 return Name == "acos" || Name == "acosf" ||
2123 Name == "asin" || Name == "asinf" ||
2124 Name == "atan" || Name == "atanf" ||
2125 Name == "atan2" || Name == "atan2f";
2126 case 'c':
2127 return Name == "ceil" || Name == "ceilf" ||
2128 Name == "cos" || Name == "cosf" ||
2129 Name == "cosh" || Name == "coshf";
2130 case 'e':
2131 return Name == "exp" || Name == "expf" || Name == "exp2" ||
2132 Name == "exp2f" || Name == "erf" || Name == "erff";
2133 case 'f':
2134 return Name == "fabs" || Name == "fabsf" ||
2135 Name == "floor" || Name == "floorf" ||
2136 Name == "fmod" || Name == "fmodf";
2137 case 'i':
2138 return Name == "ilogb" || Name == "ilogbf";
2139 case 'l':
2140 return Name == "log" || Name == "logf" || Name == "logl" ||
2141 Name == "log2" || Name == "log2f" || Name == "log10" ||
2142 Name == "log10f" || Name == "logb" || Name == "logbf" ||
2143 Name == "log1p" || Name == "log1pf";
2144 case 'n':
2145 return Name == "nearbyint" || Name == "nearbyintf" || Name == "nextafter" ||
2146 Name == "nextafterf" || Name == "nexttoward" ||
2147 Name == "nexttowardf";
2148 case 'p':
2149 return Name == "pow" || Name == "powf";
2150 case 'r':
2151 return Name == "remainder" || Name == "remainderf" ||
2152 Name == "rint" || Name == "rintf" ||
2153 Name == "round" || Name == "roundf" ||
2154 Name == "roundeven" || Name == "roundevenf";
2155 case 's':
2156 return Name == "sin" || Name == "sinf" ||
2157 Name == "sinh" || Name == "sinhf" ||
2158 Name == "sqrt" || Name == "sqrtf";
2159 case 't':
2160 return Name == "tan" || Name == "tanf" ||
2161 Name == "tanh" || Name == "tanhf" ||
2162 Name == "trunc" || Name == "truncf";
2163 case '_':
2164 // Check for various function names that get used for the math functions
2165 // when the header files are preprocessed with the macro
2166 // __FINITE_MATH_ONLY__ enabled.
2167 // The '12' here is the length of the shortest name that can match.
2168 // We need to check the size before looking at Name[1] and Name[2]
2169 // so we may as well check a limit that will eliminate mismatches.
2170 if (Name.size() < 12 || Name[1] != '_')
2171 return false;
2172 switch (Name[2]) {
2173 default:
2174 return false;
2175 case 'a':
2176 return Name == "__acos_finite" || Name == "__acosf_finite" ||
2177 Name == "__asin_finite" || Name == "__asinf_finite" ||
2178 Name == "__atan2_finite" || Name == "__atan2f_finite";
2179 case 'c':
2180 return Name == "__cosh_finite" || Name == "__coshf_finite";
2181 case 'e':
2182 return Name == "__exp_finite" || Name == "__expf_finite" ||
2183 Name == "__exp2_finite" || Name == "__exp2f_finite";
2184 case 'l':
2185 return Name == "__log_finite" || Name == "__logf_finite" ||
2186 Name == "__log10_finite" || Name == "__log10f_finite";
2187 case 'p':
2188 return Name == "__pow_finite" || Name == "__powf_finite";
2189 case 's':
2190 return Name == "__sinh_finite" || Name == "__sinhf_finite";
2191 }
2192 // clang-format on
2193 }
2194}
2195
2196namespace {
2197
2198Constant *GetConstantFoldFPValue(double V, Type *Ty) {
2199 if (Ty->isHalfTy() || Ty->isFloatTy()) {
2200 APFloat APF(V);
2201 bool unused;
2202 APF.convert(ToSemantics: Ty->getFltSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused);
2203 return ConstantFP::get(Context&: Ty->getContext(), V: APF);
2204 }
2205 if (Ty->isDoubleTy())
2206 return ConstantFP::get(Context&: Ty->getContext(), V: APFloat(V));
2207 llvm_unreachable("Can only constant fold half/float/double");
2208}
2209
2210#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
2211Constant *GetConstantFoldFPValue128(float128 V, Type *Ty) {
2212 if (Ty->isFP128Ty())
2213 return ConstantFP::get(Ty, V);
2214 llvm_unreachable("Can only constant fold fp128");
2215}
2216#endif
2217
2218/// Clear the floating-point exception state.
2219inline void llvm_fenv_clearexcept() {
2220#if HAVE_DECL_FE_ALL_EXCEPT
2221 feclearexcept(FE_ALL_EXCEPT);
2222#endif
2223 errno = 0;
2224}
2225
2226/// Test if a floating-point exception was raised.
2227inline bool llvm_fenv_testexcept() {
2228 int errno_val = errno;
2229 if (errno_val == ERANGE || errno_val == EDOM)
2230 return true;
2231#if HAVE_DECL_FE_ALL_EXCEPT && HAVE_DECL_FE_INEXACT
2232 if (fetestexcept(FE_ALL_EXCEPT & ~FE_INEXACT))
2233 return true;
2234#endif
2235 return false;
2236}
2237
2238static APFloat FTZPreserveSign(const APFloat &V) {
2239 if (V.isDenormal())
2240 return APFloat::getZero(Sem: V.getSemantics(), Negative: V.isNegative());
2241 return V;
2242}
2243
2244static APFloat FlushToPositiveZero(const APFloat &V) {
2245 if (V.isDenormal())
2246 return APFloat::getZero(Sem: V.getSemantics(), Negative: false);
2247 return V;
2248}
2249
2250static APFloat FlushWithDenormKind(const APFloat &V,
2251 DenormalMode::DenormalModeKind DenormKind) {
2252 assert(DenormKind != DenormalMode::DenormalModeKind::Invalid &&
2253 DenormKind != DenormalMode::DenormalModeKind::Dynamic);
2254 switch (DenormKind) {
2255 case DenormalMode::DenormalModeKind::IEEE:
2256 return V;
2257 case DenormalMode::DenormalModeKind::PreserveSign:
2258 return FTZPreserveSign(V);
2259 case DenormalMode::DenormalModeKind::PositiveZero:
2260 return FlushToPositiveZero(V);
2261 default:
2262 llvm_unreachable("Invalid denormal mode!");
2263 }
2264}
2265
2266Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V, Type *Ty,
2267 DenormalMode DenormMode = DenormalMode::getIEEE()) {
2268 if (!DenormMode.isValid() ||
2269 DenormMode.Input == DenormalMode::DenormalModeKind::Dynamic ||
2270 DenormMode.Output == DenormalMode::DenormalModeKind::Dynamic)
2271 return nullptr;
2272
2273 llvm_fenv_clearexcept();
2274 auto Input = FlushWithDenormKind(V, DenormKind: DenormMode.Input);
2275 double Result = NativeFP(Input.convertToDouble());
2276 if (llvm_fenv_testexcept()) {
2277 llvm_fenv_clearexcept();
2278 return nullptr;
2279 }
2280
2281 Constant *Output = GetConstantFoldFPValue(V: Result, Ty);
2282 if (DenormMode.Output == DenormalMode::DenormalModeKind::IEEE)
2283 return Output;
2284 const auto *CFP = static_cast<ConstantFP *>(Output);
2285 const auto Res = FlushWithDenormKind(V: CFP->getValueAPF(), DenormKind: DenormMode.Output);
2286 return ConstantFP::get(Context&: Ty->getContext(), V: Res);
2287}
2288
2289#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
2290Constant *ConstantFoldFP128(float128 (*NativeFP)(float128), const APFloat &V,
2291 Type *Ty) {
2292 llvm_fenv_clearexcept();
2293 float128 Result = NativeFP(V.convertToQuad());
2294 if (llvm_fenv_testexcept()) {
2295 llvm_fenv_clearexcept();
2296 return nullptr;
2297 }
2298
2299 return GetConstantFoldFPValue128(V: Result, Ty);
2300}
2301#endif
2302
2303Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
2304 const APFloat &V, const APFloat &W, Type *Ty) {
2305 llvm_fenv_clearexcept();
2306 double Result = NativeFP(V.convertToDouble(), W.convertToDouble());
2307 if (llvm_fenv_testexcept()) {
2308 llvm_fenv_clearexcept();
2309 return nullptr;
2310 }
2311
2312 return GetConstantFoldFPValue(V: Result, Ty);
2313}
2314
2315Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) {
2316 auto *OpVT = cast<VectorType>(Val: Op->getType());
2317
2318 // This is the same as the underlying binops - poison propagates.
2319 if (Op->containsPoisonElement())
2320 return PoisonValue::get(T: OpVT->getElementType());
2321
2322 // Shortcut non-accumulating reductions.
2323 if (Constant *SplatVal = Op->getSplatValue()) {
2324 switch (IID) {
2325 case Intrinsic::vector_reduce_and:
2326 case Intrinsic::vector_reduce_or:
2327 case Intrinsic::vector_reduce_smin:
2328 case Intrinsic::vector_reduce_smax:
2329 case Intrinsic::vector_reduce_umin:
2330 case Intrinsic::vector_reduce_umax:
2331 return SplatVal;
2332 case Intrinsic::vector_reduce_add:
2333 if (SplatVal->isNullValue())
2334 return SplatVal;
2335 break;
2336 case Intrinsic::vector_reduce_mul:
2337 if (SplatVal->isNullValue() || SplatVal->isOneValue())
2338 return SplatVal;
2339 break;
2340 case Intrinsic::vector_reduce_xor:
2341 if (SplatVal->isNullValue())
2342 return SplatVal;
2343 if (OpVT->getElementCount().isKnownMultipleOf(RHS: 2))
2344 return Constant::getNullValue(Ty: OpVT->getElementType());
2345 break;
2346 }
2347 }
2348
2349 FixedVectorType *VT = dyn_cast<FixedVectorType>(Val: OpVT);
2350 if (!VT)
2351 return nullptr;
2352
2353 auto *EltC = dyn_cast_or_null<ConstantInt>(Val: Op->getAggregateElement(Elt: 0U));
2354 if (!EltC)
2355 return nullptr;
2356
2357 APInt Acc = EltC->getValue();
2358 for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) {
2359 if (!(EltC = dyn_cast_or_null<ConstantInt>(Val: Op->getAggregateElement(Elt: I))))
2360 return nullptr;
2361 const APInt &X = EltC->getValue();
2362 switch (IID) {
2363 case Intrinsic::vector_reduce_add:
2364 Acc = Acc + X;
2365 break;
2366 case Intrinsic::vector_reduce_mul:
2367 Acc = Acc * X;
2368 break;
2369 case Intrinsic::vector_reduce_and:
2370 Acc = Acc & X;
2371 break;
2372 case Intrinsic::vector_reduce_or:
2373 Acc = Acc | X;
2374 break;
2375 case Intrinsic::vector_reduce_xor:
2376 Acc = Acc ^ X;
2377 break;
2378 case Intrinsic::vector_reduce_smin:
2379 Acc = APIntOps::smin(A: Acc, B: X);
2380 break;
2381 case Intrinsic::vector_reduce_smax:
2382 Acc = APIntOps::smax(A: Acc, B: X);
2383 break;
2384 case Intrinsic::vector_reduce_umin:
2385 Acc = APIntOps::umin(A: Acc, B: X);
2386 break;
2387 case Intrinsic::vector_reduce_umax:
2388 Acc = APIntOps::umax(A: Acc, B: X);
2389 break;
2390 }
2391 }
2392
2393 return ConstantInt::get(Context&: Op->getContext(), V: Acc);
2394}
2395
2396/// Attempt to fold an SSE floating point to integer conversion of a constant
2397/// floating point. If roundTowardZero is false, the default IEEE rounding is
2398/// used (toward nearest, ties to even). This matches the behavior of the
2399/// non-truncating SSE instructions in the default rounding mode. The desired
2400/// integer type Ty is used to select how many bits are available for the
2401/// result. Returns null if the conversion cannot be performed, otherwise
2402/// returns the Constant value resulting from the conversion.
2403Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
2404 Type *Ty, bool IsSigned) {
2405 // All of these conversion intrinsics form an integer of at most 64bits.
2406 unsigned ResultWidth = Ty->getIntegerBitWidth();
2407 assert(ResultWidth <= 64 &&
2408 "Can only constant fold conversions to 64 and 32 bit ints");
2409
2410 uint64_t UIntVal;
2411 bool isExact = false;
2412 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
2413 : APFloat::rmNearestTiesToEven;
2414 APFloat::opStatus status =
2415 Val.convertToInteger(Input: MutableArrayRef(UIntVal), Width: ResultWidth,
2416 IsSigned, RM: mode, IsExact: &isExact);
2417 if (status != APFloat::opOK &&
2418 (!roundTowardZero || status != APFloat::opInexact))
2419 return nullptr;
2420 return ConstantInt::get(Ty, V: UIntVal, IsSigned);
2421}
2422
2423double getValueAsDouble(ConstantFP *Op) {
2424 Type *Ty = Op->getType();
2425
2426 if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())
2427 return Op->getValueAPF().convertToDouble();
2428
2429 bool unused;
2430 APFloat APF = Op->getValueAPF();
2431 APF.convert(ToSemantics: APFloat::IEEEdouble(), RM: APFloat::rmNearestTiesToEven, losesInfo: &unused);
2432 return APF.convertToDouble();
2433}
2434
2435static bool getConstIntOrUndef(Value *Op, const APInt *&C) {
2436 if (auto *CI = dyn_cast<ConstantInt>(Val: Op)) {
2437 C = &CI->getValue();
2438 return true;
2439 }
2440 if (isa<UndefValue>(Val: Op)) {
2441 C = nullptr;
2442 return true;
2443 }
2444 return false;
2445}
2446
2447/// Checks if the given intrinsic call, which evaluates to constant, is allowed
2448/// to be folded.
2449///
2450/// \param CI Constrained intrinsic call.
2451/// \param St Exception flags raised during constant evaluation.
2452static bool mayFoldConstrained(ConstrainedFPIntrinsic *CI,
2453 APFloat::opStatus St) {
2454 std::optional<RoundingMode> ORM = CI->getRoundingMode();
2455 std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2456
2457 // If the operation does not change exception status flags, it is safe
2458 // to fold.
2459 if (St == APFloat::opStatus::opOK)
2460 return true;
2461
2462 // If evaluation raised FP exception, the result can depend on rounding
2463 // mode. If the latter is unknown, folding is not possible.
2464 if (ORM == RoundingMode::Dynamic)
2465 return false;
2466
2467 // If FP exceptions are ignored, fold the call, even if such exception is
2468 // raised.
2469 if (EB && *EB != fp::ExceptionBehavior::ebStrict)
2470 return true;
2471
2472 // Leave the calculation for runtime so that exception flags be correctly set
2473 // in hardware.
2474 return false;
2475}
2476
2477/// Returns the rounding mode that should be used for constant evaluation.
2478static RoundingMode
2479getEvaluationRoundingMode(const ConstrainedFPIntrinsic *CI) {
2480 std::optional<RoundingMode> ORM = CI->getRoundingMode();
2481 if (!ORM || *ORM == RoundingMode::Dynamic)
2482 // Even if the rounding mode is unknown, try evaluating the operation.
2483 // If it does not raise inexact exception, rounding was not applied,
2484 // so the result is exact and does not depend on rounding mode. Whether
2485 // other FP exceptions are raised, it does not depend on rounding mode.
2486 return RoundingMode::NearestTiesToEven;
2487 return *ORM;
2488}
2489
2490/// Try to constant fold llvm.canonicalize for the given caller and value.
2491static Constant *constantFoldCanonicalize(const Type *Ty, const APFloat &Src,
2492 const Function *CtxF = nullptr) {
2493 // Zero, positive and negative, is always OK to fold.
2494 if (Src.isZero()) {
2495 // Get a fresh 0, since ppc_fp128 does have non-canonical zeros.
2496 return ConstantFP::get(
2497 Context&: Ty->getContext(),
2498 V: APFloat::getZero(Sem: Src.getSemantics(), Negative: Src.isNegative()));
2499 }
2500
2501 if (!Ty->isIEEELikeFPTy())
2502 return nullptr;
2503
2504 // Zero is always canonical and the sign must be preserved.
2505 //
2506 // Denorms and nans may have special encodings, but it should be OK to fold a
2507 // totally average number.
2508 if (Src.isNormal() || Src.isInfinity())
2509 return ConstantFP::get(Context&: Ty->getContext(), V: Src);
2510
2511 if (Src.isDenormal() && CtxF) {
2512 DenormalMode DenormMode = CtxF->getDenormalMode(FPType: Src.getSemantics());
2513
2514 if (DenormMode == DenormalMode::getIEEE())
2515 return ConstantFP::get(Context&: Ty->getContext(), V: Src);
2516
2517 if (DenormMode.Input == DenormalMode::Dynamic)
2518 return nullptr;
2519
2520 // If we know if either input or output is flushed, we can fold.
2521 if ((DenormMode.Input == DenormalMode::Dynamic &&
2522 DenormMode.Output == DenormalMode::IEEE) ||
2523 (DenormMode.Input == DenormalMode::IEEE &&
2524 DenormMode.Output == DenormalMode::Dynamic))
2525 return nullptr;
2526
2527 bool IsPositive =
2528 (!Src.isNegative() || DenormMode.Input == DenormalMode::PositiveZero ||
2529 (DenormMode.Output == DenormalMode::PositiveZero &&
2530 DenormMode.Input == DenormalMode::IEEE));
2531
2532 return ConstantFP::get(Context&: Ty->getContext(),
2533 V: APFloat::getZero(Sem: Src.getSemantics(), Negative: !IsPositive));
2534 }
2535
2536 return nullptr;
2537}
2538
2539static Constant *ConstantFoldScalarCall1(StringRef Name,
2540 Intrinsic::ID IntrinsicID, Type *Ty,
2541 ArrayRef<Constant *> Operands,
2542 const TargetLibraryInfo *TLI = nullptr,
2543 const CallBase *Call = nullptr) {
2544 assert(Operands.size() == 1 && "Wrong number of operands.");
2545
2546 if (IntrinsicID == Intrinsic::is_constant) {
2547 // We know we have a "Constant" argument. But we want to only
2548 // return true for manifest constants, not those that depend on
2549 // constants with unknowable values, e.g. GlobalValue or BlockAddress.
2550 if (Operands[0]->isManifestConstant())
2551 return ConstantInt::getTrue(Context&: Ty->getContext());
2552 return nullptr;
2553 }
2554
2555 if (isa<UndefValue>(Val: Operands[0])) {
2556 // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN.
2557 // ctpop() is between 0 and bitwidth, pick 0 for undef.
2558 // fptoui.sat and fptosi.sat can always fold to zero (for a zero input).
2559 if (IntrinsicID == Intrinsic::cos ||
2560 IntrinsicID == Intrinsic::ctpop ||
2561 IntrinsicID == Intrinsic::fptoui_sat ||
2562 IntrinsicID == Intrinsic::fptosi_sat ||
2563 IntrinsicID == Intrinsic::canonicalize)
2564 return Constant::getNullValue(Ty);
2565 if (IntrinsicID == Intrinsic::bswap ||
2566 IntrinsicID == Intrinsic::bitreverse ||
2567 IntrinsicID == Intrinsic::launder_invariant_group ||
2568 IntrinsicID == Intrinsic::strip_invariant_group)
2569 return Operands[0];
2570 }
2571
2572 if (isa<ConstantPointerNull>(Val: Operands[0])) {
2573 // launder(null) == null == strip(null) iff in addrspace 0
2574 if (IntrinsicID == Intrinsic::launder_invariant_group ||
2575 IntrinsicID == Intrinsic::strip_invariant_group) {
2576 // If instruction is not yet put in a basic block (e.g. when cloning
2577 // a function during inlining), Call's caller may not be available.
2578 // So check Call's BB first before querying Call->getCaller.
2579 const Function *Caller =
2580 Call && Call->getParent() ? Call->getCaller() : nullptr;
2581 if (Caller &&
2582 !NullPointerIsDefined(
2583 F: Caller, AS: Operands[0]->getType()->getPointerAddressSpace())) {
2584 return Operands[0];
2585 }
2586 return nullptr;
2587 }
2588 }
2589
2590 if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[0])) {
2591 APFloat U = Op->getValueAPF();
2592
2593 if (IntrinsicID == Intrinsic::wasm_trunc_signed ||
2594 IntrinsicID == Intrinsic::wasm_trunc_unsigned) {
2595 bool Signed = IntrinsicID == Intrinsic::wasm_trunc_signed;
2596
2597 if (U.isNaN())
2598 return nullptr;
2599
2600 unsigned Width = Ty->getIntegerBitWidth();
2601 APSInt Int(Width, !Signed);
2602 bool IsExact = false;
2603 APFloat::opStatus Status =
2604 U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact);
2605
2606 if (Status == APFloat::opOK || Status == APFloat::opInexact)
2607 return ConstantInt::get(Ty, V: Int);
2608
2609 return nullptr;
2610 }
2611
2612 if (IntrinsicID == Intrinsic::fptoui_sat ||
2613 IntrinsicID == Intrinsic::fptosi_sat) {
2614 // convertToInteger() already has the desired saturation semantics.
2615 APSInt Int(Ty->getIntegerBitWidth(),
2616 IntrinsicID == Intrinsic::fptoui_sat);
2617 bool IsExact;
2618 U.convertToInteger(Result&: Int, RM: APFloat::rmTowardZero, IsExact: &IsExact);
2619 return ConstantInt::get(Ty, V: Int);
2620 }
2621
2622 if (IntrinsicID == Intrinsic::canonicalize) {
2623 const Function *CtxF =
2624 Call && Call->getParent() ? Call->getFunction() : nullptr;
2625 return constantFoldCanonicalize(Ty, Src: U, CtxF);
2626 }
2627
2628#if defined(HAS_IEE754_FLOAT128) && defined(HAS_LOGF128)
2629 if (Ty->isFP128Ty()) {
2630 if (IntrinsicID == Intrinsic::log) {
2631 float128 Result = logf128(x: Op->getValueAPF().convertToQuad());
2632 return GetConstantFoldFPValue128(V: Result, Ty);
2633 }
2634
2635 LibFunc Fp128Func = NotLibFunc;
2636 if (TLI && TLI->getLibFunc(funcName: Name, F&: Fp128Func) && TLI->has(F: Fp128Func) &&
2637 Fp128Func == LibFunc_logl)
2638 return ConstantFoldFP128(NativeFP: logf128, V: Op->getValueAPF(), Ty);
2639 }
2640#endif
2641
2642 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy() &&
2643 !Ty->isIntegerTy())
2644 return nullptr;
2645
2646 // Use internal versions of these intrinsics.
2647
2648 if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint ||
2649 IntrinsicID == Intrinsic::roundeven) {
2650 U.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
2651 return ConstantFP::get(Ty, V: U);
2652 }
2653
2654 if (IntrinsicID == Intrinsic::round) {
2655 U.roundToIntegral(RM: APFloat::rmNearestTiesToAway);
2656 return ConstantFP::get(Ty, V: U);
2657 }
2658
2659 if (IntrinsicID == Intrinsic::roundeven) {
2660 U.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
2661 return ConstantFP::get(Ty, V: U);
2662 }
2663
2664 if (IntrinsicID == Intrinsic::ceil) {
2665 U.roundToIntegral(RM: APFloat::rmTowardPositive);
2666 return ConstantFP::get(Ty, V: U);
2667 }
2668
2669 if (IntrinsicID == Intrinsic::floor) {
2670 U.roundToIntegral(RM: APFloat::rmTowardNegative);
2671 return ConstantFP::get(Ty, V: U);
2672 }
2673
2674 if (IntrinsicID == Intrinsic::trunc) {
2675 U.roundToIntegral(RM: APFloat::rmTowardZero);
2676 return ConstantFP::get(Ty, V: U);
2677 }
2678
2679 if (IntrinsicID == Intrinsic::fabs) {
2680 U.clearSign();
2681 return ConstantFP::get(Ty, V: U);
2682 }
2683
2684 if (IntrinsicID == Intrinsic::amdgcn_fract) {
2685 // The v_fract instruction behaves like the OpenCL spec, which defines
2686 // fract(x) as fmin(x - floor(x), 0x1.fffffep-1f): "The min() operator is
2687 // there to prevent fract(-small) from returning 1.0. It returns the
2688 // largest positive floating-point number less than 1.0."
2689 APFloat FloorU(U);
2690 FloorU.roundToIntegral(RM: APFloat::rmTowardNegative);
2691 APFloat FractU(U - FloorU);
2692 APFloat AlmostOne(U.getSemantics(), 1);
2693 AlmostOne.next(/*nextDown*/ true);
2694 return ConstantFP::get(Ty, V: minimum(A: FractU, B: AlmostOne));
2695 }
2696
2697 // Rounding operations (floor, trunc, ceil, round and nearbyint) do not
2698 // raise FP exceptions, unless the argument is signaling NaN.
2699
2700 if (auto *CI = dyn_cast_or_null<ConstrainedFPIntrinsic>(Val: Call)) {
2701 std::optional<APFloat::roundingMode> RM;
2702 switch (IntrinsicID) {
2703 default:
2704 break;
2705 case Intrinsic::experimental_constrained_nearbyint:
2706 case Intrinsic::experimental_constrained_rint: {
2707 RM = CI->getRoundingMode();
2708 if (!RM || *RM == RoundingMode::Dynamic)
2709 return nullptr;
2710 break;
2711 }
2712 case Intrinsic::experimental_constrained_round:
2713 RM = APFloat::rmNearestTiesToAway;
2714 break;
2715 case Intrinsic::experimental_constrained_ceil:
2716 RM = APFloat::rmTowardPositive;
2717 break;
2718 case Intrinsic::experimental_constrained_floor:
2719 RM = APFloat::rmTowardNegative;
2720 break;
2721 case Intrinsic::experimental_constrained_trunc:
2722 RM = APFloat::rmTowardZero;
2723 break;
2724 }
2725 if (RM) {
2726 if (U.isFinite()) {
2727 APFloat::opStatus St = U.roundToIntegral(RM: *RM);
2728 if (IntrinsicID == Intrinsic::experimental_constrained_rint &&
2729 St == APFloat::opInexact) {
2730 std::optional<fp::ExceptionBehavior> EB =
2731 CI->getExceptionBehavior();
2732 if (EB == fp::ebStrict)
2733 return nullptr;
2734 }
2735 } else if (U.isSignaling()) {
2736 std::optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2737 if (EB && *EB != fp::ebIgnore)
2738 return nullptr;
2739 U = APFloat::getQNaN(Sem: U.getSemantics());
2740 }
2741 return ConstantFP::get(Ty, V: U);
2742 }
2743 }
2744
2745 // NVVM float/double to signed/unsigned int32/int64 conversions:
2746 switch (IntrinsicID) {
2747 // f2i
2748 case Intrinsic::nvvm_f2i_rm:
2749 case Intrinsic::nvvm_f2i_rn:
2750 case Intrinsic::nvvm_f2i_rp:
2751 case Intrinsic::nvvm_f2i_rz:
2752 case Intrinsic::nvvm_f2i_rm_ftz:
2753 case Intrinsic::nvvm_f2i_rn_ftz:
2754 case Intrinsic::nvvm_f2i_rp_ftz:
2755 case Intrinsic::nvvm_f2i_rz_ftz:
2756 // f2ui
2757 case Intrinsic::nvvm_f2ui_rm:
2758 case Intrinsic::nvvm_f2ui_rn:
2759 case Intrinsic::nvvm_f2ui_rp:
2760 case Intrinsic::nvvm_f2ui_rz:
2761 case Intrinsic::nvvm_f2ui_rm_ftz:
2762 case Intrinsic::nvvm_f2ui_rn_ftz:
2763 case Intrinsic::nvvm_f2ui_rp_ftz:
2764 case Intrinsic::nvvm_f2ui_rz_ftz:
2765 // d2i
2766 case Intrinsic::nvvm_d2i_rm:
2767 case Intrinsic::nvvm_d2i_rn:
2768 case Intrinsic::nvvm_d2i_rp:
2769 case Intrinsic::nvvm_d2i_rz:
2770 // d2ui
2771 case Intrinsic::nvvm_d2ui_rm:
2772 case Intrinsic::nvvm_d2ui_rn:
2773 case Intrinsic::nvvm_d2ui_rp:
2774 case Intrinsic::nvvm_d2ui_rz:
2775 // f2ll
2776 case Intrinsic::nvvm_f2ll_rm:
2777 case Intrinsic::nvvm_f2ll_rn:
2778 case Intrinsic::nvvm_f2ll_rp:
2779 case Intrinsic::nvvm_f2ll_rz:
2780 case Intrinsic::nvvm_f2ll_rm_ftz:
2781 case Intrinsic::nvvm_f2ll_rn_ftz:
2782 case Intrinsic::nvvm_f2ll_rp_ftz:
2783 case Intrinsic::nvvm_f2ll_rz_ftz:
2784 // f2ull
2785 case Intrinsic::nvvm_f2ull_rm:
2786 case Intrinsic::nvvm_f2ull_rn:
2787 case Intrinsic::nvvm_f2ull_rp:
2788 case Intrinsic::nvvm_f2ull_rz:
2789 case Intrinsic::nvvm_f2ull_rm_ftz:
2790 case Intrinsic::nvvm_f2ull_rn_ftz:
2791 case Intrinsic::nvvm_f2ull_rp_ftz:
2792 case Intrinsic::nvvm_f2ull_rz_ftz:
2793 // d2ll
2794 case Intrinsic::nvvm_d2ll_rm:
2795 case Intrinsic::nvvm_d2ll_rn:
2796 case Intrinsic::nvvm_d2ll_rp:
2797 case Intrinsic::nvvm_d2ll_rz:
2798 // d2ull
2799 case Intrinsic::nvvm_d2ull_rm:
2800 case Intrinsic::nvvm_d2ull_rn:
2801 case Intrinsic::nvvm_d2ull_rp:
2802 case Intrinsic::nvvm_d2ull_rz: {
2803 // In float-to-integer conversion, NaN inputs are converted to 0.
2804 if (U.isNaN()) {
2805 // In float-to-integer conversion, NaN inputs are converted to 0
2806 // when the source and destination bitwidths are both less than 64.
2807 if (nvvm::FPToIntegerIntrinsicNaNZero(IntrinsicID))
2808 return ConstantInt::get(Ty, V: 0);
2809
2810 // Otherwise, the most significant bit is set.
2811 unsigned BitWidth = Ty->getIntegerBitWidth();
2812 uint64_t Val = 1ULL << (BitWidth - 1);
2813 return ConstantInt::get(Ty, V: APInt(BitWidth, Val, /*IsSigned=*/false));
2814 }
2815
2816 APFloat::roundingMode RMode =
2817 nvvm::GetFPToIntegerRoundingMode(IntrinsicID);
2818 bool IsFTZ = nvvm::FPToIntegerIntrinsicShouldFTZ(IntrinsicID);
2819 bool IsSigned = nvvm::FPToIntegerIntrinsicResultIsSigned(IntrinsicID);
2820
2821 APSInt ResInt(Ty->getIntegerBitWidth(), !IsSigned);
2822 auto FloatToRound = IsFTZ ? FTZPreserveSign(V: U) : U;
2823
2824 // Return max/min value for integers if the result is +/-inf or
2825 // is too large to fit in the result's integer bitwidth.
2826 bool IsExact = false;
2827 FloatToRound.convertToInteger(Result&: ResInt, RM: RMode, IsExact: &IsExact);
2828 return ConstantInt::get(Ty, V: ResInt);
2829 }
2830 }
2831
2832 /// We only fold functions with finite arguments. Folding NaN and inf is
2833 /// likely to be aborted with an exception anyway, and some host libms
2834 /// have known errors raising exceptions.
2835 if (!U.isFinite())
2836 return nullptr;
2837
2838 /// Currently APFloat versions of these functions do not exist, so we use
2839 /// the host native double versions. Float versions are not called
2840 /// directly but for all these it is true (float)(f((double)arg)) ==
2841 /// f(arg). Long double not supported yet.
2842 const APFloat &APF = Op->getValueAPF();
2843
2844 switch (IntrinsicID) {
2845 default: break;
2846 case Intrinsic::log:
2847 if (U.isZero())
2848 return ConstantFP::getInfinity(Ty, Negative: true);
2849 if (U.isNegative())
2850 return ConstantFP::getNaN(Ty);
2851 if (U.isOne())
2852 return ConstantFP::getZero(Ty);
2853 return ConstantFoldFP(NativeFP: log, V: APF, Ty);
2854 case Intrinsic::log2:
2855 if (U.isZero())
2856 return ConstantFP::getInfinity(Ty, Negative: true);
2857 if (U.isNegative())
2858 return ConstantFP::getNaN(Ty);
2859 if (U.isOne())
2860 return ConstantFP::getZero(Ty);
2861 // TODO: What about hosts that lack a C99 library?
2862 return ConstantFoldFP(NativeFP: log2, V: APF, Ty);
2863 case Intrinsic::log10:
2864 if (U.isZero())
2865 return ConstantFP::getInfinity(Ty, Negative: true);
2866 if (U.isNegative())
2867 return ConstantFP::getNaN(Ty);
2868 if (U.isOne())
2869 return ConstantFP::getZero(Ty);
2870 // TODO: What about hosts that lack a C99 library?
2871 return ConstantFoldFP(NativeFP: log10, V: APF, Ty);
2872 case Intrinsic::exp:
2873 return ConstantFoldFP(NativeFP: exp, V: APF, Ty);
2874 case Intrinsic::exp2:
2875 // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
2876 return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty);
2877 case Intrinsic::exp10:
2878 // Fold exp10(x) as pow(10, x), in case the host lacks a C99 library.
2879 return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(10.0), W: APF, Ty);
2880 case Intrinsic::sin:
2881 return ConstantFoldFP(NativeFP: sin, V: APF, Ty);
2882 case Intrinsic::cos:
2883 return ConstantFoldFP(NativeFP: cos, V: APF, Ty);
2884 case Intrinsic::sinh:
2885 return ConstantFoldFP(NativeFP: sinh, V: APF, Ty);
2886 case Intrinsic::cosh:
2887 return ConstantFoldFP(NativeFP: cosh, V: APF, Ty);
2888 case Intrinsic::atan:
2889 // Implement optional behavior from C's Annex F for +/-0.0.
2890 if (U.isZero())
2891 return ConstantFP::get(Ty, V: U);
2892 return ConstantFoldFP(NativeFP: atan, V: APF, Ty);
2893 case Intrinsic::sqrt:
2894 return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty);
2895
2896 // NVVM Intrinsics:
2897 case Intrinsic::nvvm_ceil_ftz_f:
2898 case Intrinsic::nvvm_ceil_f:
2899 case Intrinsic::nvvm_ceil_d:
2900 return ConstantFoldFP(
2901 NativeFP: ceil, V: APF, Ty,
2902 DenormMode: nvvm::GetNVVMDenormMode(
2903 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2904
2905 case Intrinsic::nvvm_fabs_ftz:
2906 case Intrinsic::nvvm_fabs:
2907 return ConstantFoldFP(
2908 NativeFP: fabs, V: APF, Ty,
2909 DenormMode: nvvm::GetNVVMDenormMode(
2910 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2911
2912 case Intrinsic::nvvm_floor_ftz_f:
2913 case Intrinsic::nvvm_floor_f:
2914 case Intrinsic::nvvm_floor_d:
2915 return ConstantFoldFP(
2916 NativeFP: floor, V: APF, Ty,
2917 DenormMode: nvvm::GetNVVMDenormMode(
2918 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2919
2920 case Intrinsic::nvvm_rcp_rm_ftz_f:
2921 case Intrinsic::nvvm_rcp_rn_ftz_f:
2922 case Intrinsic::nvvm_rcp_rp_ftz_f:
2923 case Intrinsic::nvvm_rcp_rz_ftz_f:
2924 case Intrinsic::nvvm_rcp_rm_d:
2925 case Intrinsic::nvvm_rcp_rm_f:
2926 case Intrinsic::nvvm_rcp_rn_d:
2927 case Intrinsic::nvvm_rcp_rn_f:
2928 case Intrinsic::nvvm_rcp_rp_d:
2929 case Intrinsic::nvvm_rcp_rp_f:
2930 case Intrinsic::nvvm_rcp_rz_d:
2931 case Intrinsic::nvvm_rcp_rz_f: {
2932 APFloat::roundingMode RoundMode = nvvm::GetRCPRoundingMode(IntrinsicID);
2933 bool IsFTZ = nvvm::RCPShouldFTZ(IntrinsicID);
2934
2935 auto Denominator = IsFTZ ? FTZPreserveSign(V: APF) : APF;
2936 APFloat Res = APFloat::getOne(Sem: APF.getSemantics());
2937 APFloat::opStatus Status = Res.divide(RHS: Denominator, RM: RoundMode);
2938
2939 if (Status == APFloat::opOK || Status == APFloat::opInexact) {
2940 if (IsFTZ)
2941 Res = FTZPreserveSign(V: Res);
2942 return ConstantFP::get(Ty, V: Res);
2943 }
2944 return nullptr;
2945 }
2946
2947 case Intrinsic::nvvm_round_ftz_f:
2948 case Intrinsic::nvvm_round_f:
2949 case Intrinsic::nvvm_round_d: {
2950 // nvvm_round is lowered to PTX cvt.rni, which will round to nearest
2951 // integer, choosing even integer if source is equidistant between two
2952 // integers, so the semantics are closer to "rint" rather than "round".
2953 bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);
2954 auto V = IsFTZ ? FTZPreserveSign(V: APF) : APF;
2955 V.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
2956 return ConstantFP::get(Ty, V);
2957 }
2958
2959 case Intrinsic::nvvm_saturate_ftz_f:
2960 case Intrinsic::nvvm_saturate_d:
2961 case Intrinsic::nvvm_saturate_f: {
2962 bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);
2963 auto V = IsFTZ ? FTZPreserveSign(V: APF) : APF;
2964 if (V.isNegative() || V.isZero() || V.isNaN())
2965 return ConstantFP::getZero(Ty);
2966 APFloat One = APFloat::getOne(Sem: APF.getSemantics());
2967 if (V > One)
2968 return ConstantFP::get(Ty, V: One);
2969 return ConstantFP::get(Ty, V: APF);
2970 }
2971
2972 case Intrinsic::nvvm_sqrt_rn_ftz_f:
2973 case Intrinsic::nvvm_sqrt_f:
2974 case Intrinsic::nvvm_sqrt_rn_d:
2975 case Intrinsic::nvvm_sqrt_rn_f:
2976 if (APF.isNegative())
2977 return nullptr;
2978 return ConstantFoldFP(
2979 NativeFP: sqrt, V: APF, Ty,
2980 DenormMode: nvvm::GetNVVMDenormMode(
2981 ShouldFTZ: nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID)));
2982
2983 // AMDGCN Intrinsics:
2984 case Intrinsic::amdgcn_cos:
2985 case Intrinsic::amdgcn_sin: {
2986 double V = getValueAsDouble(Op);
2987 if (V < -256.0 || V > 256.0)
2988 // The gfx8 and gfx9 architectures handle arguments outside the range
2989 // [-256, 256] differently. This should be a rare case so bail out
2990 // rather than trying to handle the difference.
2991 return nullptr;
2992 bool IsCos = IntrinsicID == Intrinsic::amdgcn_cos;
2993 double V4 = V * 4.0;
2994 if (V4 == floor(x: V4)) {
2995 // Force exact results for quarter-integer inputs.
2996 const double SinVals[4] = { 0.0, 1.0, 0.0, -1.0 };
2997 V = SinVals[((int)V4 + (IsCos ? 1 : 0)) & 3];
2998 } else {
2999 if (IsCos)
3000 V = cos(x: V * 2.0 * numbers::pi);
3001 else
3002 V = sin(x: V * 2.0 * numbers::pi);
3003 }
3004 return GetConstantFoldFPValue(V, Ty);
3005 }
3006 }
3007
3008 if (!TLI)
3009 return nullptr;
3010
3011 LibFunc Func = NotLibFunc;
3012 if (!TLI->getLibFunc(funcName: Name, F&: Func))
3013 return nullptr;
3014
3015 switch (Func) {
3016 default:
3017 break;
3018 case LibFunc_acos:
3019 case LibFunc_acosf:
3020 case LibFunc_acos_finite:
3021 case LibFunc_acosf_finite:
3022 if (TLI->has(F: Func))
3023 return ConstantFoldFP(NativeFP: acos, V: APF, Ty);
3024 break;
3025 case LibFunc_asin:
3026 case LibFunc_asinf:
3027 case LibFunc_asin_finite:
3028 case LibFunc_asinf_finite:
3029 if (TLI->has(F: Func))
3030 return ConstantFoldFP(NativeFP: asin, V: APF, Ty);
3031 break;
3032 case LibFunc_atan:
3033 case LibFunc_atanf:
3034 // Implement optional behavior from C's Annex F for +/-0.0.
3035 if (U.isZero())
3036 return ConstantFP::get(Ty, V: U);
3037 if (TLI->has(F: Func))
3038 return ConstantFoldFP(NativeFP: atan, V: APF, Ty);
3039 break;
3040 case LibFunc_ceil:
3041 case LibFunc_ceilf:
3042 if (TLI->has(F: Func)) {
3043 U.roundToIntegral(RM: APFloat::rmTowardPositive);
3044 return ConstantFP::get(Ty, V: U);
3045 }
3046 break;
3047 case LibFunc_cos:
3048 case LibFunc_cosf:
3049 if (TLI->has(F: Func))
3050 return ConstantFoldFP(NativeFP: cos, V: APF, Ty);
3051 break;
3052 case LibFunc_cosh:
3053 case LibFunc_coshf:
3054 case LibFunc_cosh_finite:
3055 case LibFunc_coshf_finite:
3056 if (TLI->has(F: Func))
3057 return ConstantFoldFP(NativeFP: cosh, V: APF, Ty);
3058 break;
3059 case LibFunc_exp:
3060 case LibFunc_expf:
3061 case LibFunc_exp_finite:
3062 case LibFunc_expf_finite:
3063 if (TLI->has(F: Func))
3064 return ConstantFoldFP(NativeFP: exp, V: APF, Ty);
3065 break;
3066 case LibFunc_exp2:
3067 case LibFunc_exp2f:
3068 case LibFunc_exp2_finite:
3069 case LibFunc_exp2f_finite:
3070 if (TLI->has(F: Func))
3071 // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
3072 return ConstantFoldBinaryFP(NativeFP: pow, V: APFloat(2.0), W: APF, Ty);
3073 break;
3074 case LibFunc_fabs:
3075 case LibFunc_fabsf:
3076 if (TLI->has(F: Func)) {
3077 U.clearSign();
3078 return ConstantFP::get(Ty, V: U);
3079 }
3080 break;
3081 case LibFunc_floor:
3082 case LibFunc_floorf:
3083 if (TLI->has(F: Func)) {
3084 U.roundToIntegral(RM: APFloat::rmTowardNegative);
3085 return ConstantFP::get(Ty, V: U);
3086 }
3087 break;
3088 case LibFunc_log:
3089 case LibFunc_logf:
3090 case LibFunc_log_finite:
3091 case LibFunc_logf_finite:
3092 if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func))
3093 return ConstantFoldFP(NativeFP: log, V: APF, Ty);
3094 break;
3095 case LibFunc_log2:
3096 case LibFunc_log2f:
3097 case LibFunc_log2_finite:
3098 case LibFunc_log2f_finite:
3099 if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func))
3100 // TODO: What about hosts that lack a C99 library?
3101 return ConstantFoldFP(NativeFP: log2, V: APF, Ty);
3102 break;
3103 case LibFunc_log10:
3104 case LibFunc_log10f:
3105 case LibFunc_log10_finite:
3106 case LibFunc_log10f_finite:
3107 if (!APF.isNegative() && !APF.isZero() && TLI->has(F: Func))
3108 // TODO: What about hosts that lack a C99 library?
3109 return ConstantFoldFP(NativeFP: log10, V: APF, Ty);
3110 break;
3111 case LibFunc_ilogb:
3112 case LibFunc_ilogbf:
3113 if (!APF.isZero() && TLI->has(F: Func))
3114 return ConstantInt::get(Ty, V: ilogb(Arg: APF), IsSigned: true);
3115 break;
3116 case LibFunc_logb:
3117 case LibFunc_logbf:
3118 if (!APF.isZero() && TLI->has(F: Func))
3119 return ConstantFoldFP(NativeFP: logb, V: APF, Ty);
3120 break;
3121 case LibFunc_log1p:
3122 case LibFunc_log1pf:
3123 // Implement optional behavior from C's Annex F for +/-0.0.
3124 if (U.isZero())
3125 return ConstantFP::get(Ty, V: U);
3126 if (APF > APFloat::getOne(Sem: APF.getSemantics(), Negative: true) && TLI->has(F: Func))
3127 return ConstantFoldFP(NativeFP: log1p, V: APF, Ty);
3128 break;
3129 case LibFunc_logl:
3130 return nullptr;
3131 case LibFunc_erf:
3132 case LibFunc_erff:
3133 if (TLI->has(F: Func))
3134 return ConstantFoldFP(NativeFP: erf, V: APF, Ty);
3135 break;
3136 case LibFunc_nearbyint:
3137 case LibFunc_nearbyintf:
3138 case LibFunc_rint:
3139 case LibFunc_rintf:
3140 case LibFunc_roundeven:
3141 case LibFunc_roundevenf:
3142 if (TLI->has(F: Func)) {
3143 U.roundToIntegral(RM: APFloat::rmNearestTiesToEven);
3144 return ConstantFP::get(Ty, V: U);
3145 }
3146 break;
3147 case LibFunc_round:
3148 case LibFunc_roundf:
3149 if (TLI->has(F: Func)) {
3150 U.roundToIntegral(RM: APFloat::rmNearestTiesToAway);
3151 return ConstantFP::get(Ty, V: U);
3152 }
3153 break;
3154 case LibFunc_sin:
3155 case LibFunc_sinf:
3156 if (TLI->has(F: Func))
3157 return ConstantFoldFP(NativeFP: sin, V: APF, Ty);
3158 break;
3159 case LibFunc_sinh:
3160 case LibFunc_sinhf:
3161 case LibFunc_sinh_finite:
3162 case LibFunc_sinhf_finite:
3163 if (TLI->has(F: Func))
3164 return ConstantFoldFP(NativeFP: sinh, V: APF, Ty);
3165 break;
3166 case LibFunc_sqrt:
3167 case LibFunc_sqrtf:
3168 if (!APF.isNegative() && TLI->has(F: Func))
3169 return ConstantFoldFP(NativeFP: sqrt, V: APF, Ty);
3170 break;
3171 case LibFunc_tan:
3172 case LibFunc_tanf:
3173 if (TLI->has(F: Func))
3174 return ConstantFoldFP(NativeFP: tan, V: APF, Ty);
3175 break;
3176 case LibFunc_tanh:
3177 case LibFunc_tanhf:
3178 if (TLI->has(F: Func))
3179 return ConstantFoldFP(NativeFP: tanh, V: APF, Ty);
3180 break;
3181 case LibFunc_trunc:
3182 case LibFunc_truncf:
3183 if (TLI->has(F: Func)) {
3184 U.roundToIntegral(RM: APFloat::rmTowardZero);
3185 return ConstantFP::get(Ty, V: U);
3186 }
3187 break;
3188 }
3189 return nullptr;
3190 }
3191
3192 if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) {
3193 switch (IntrinsicID) {
3194 case Intrinsic::bswap:
3195 return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().byteSwap());
3196 case Intrinsic::ctpop:
3197 return ConstantInt::get(Ty, V: Op->getValue().popcount());
3198 case Intrinsic::bitreverse:
3199 return ConstantInt::get(Context&: Ty->getContext(), V: Op->getValue().reverseBits());
3200 case Intrinsic::amdgcn_s_wqm: {
3201 uint64_t Val = Op->getZExtValue();
3202 Val |= (Val & 0x5555555555555555ULL) << 1 |
3203 ((Val >> 1) & 0x5555555555555555ULL);
3204 Val |= (Val & 0x3333333333333333ULL) << 2 |
3205 ((Val >> 2) & 0x3333333333333333ULL);
3206 return ConstantInt::get(Ty, V: Val);
3207 }
3208
3209 case Intrinsic::amdgcn_s_quadmask: {
3210 uint64_t Val = Op->getZExtValue();
3211 uint64_t QuadMask = 0;
3212 for (unsigned I = 0; I < Op->getBitWidth() / 4; ++I, Val >>= 4) {
3213 if (!(Val & 0xF))
3214 continue;
3215
3216 QuadMask |= (1ULL << I);
3217 }
3218 return ConstantInt::get(Ty, V: QuadMask);
3219 }
3220
3221 case Intrinsic::amdgcn_s_bitreplicate: {
3222 uint64_t Val = Op->getZExtValue();
3223 Val = (Val & 0x000000000000FFFFULL) | (Val & 0x00000000FFFF0000ULL) << 16;
3224 Val = (Val & 0x000000FF000000FFULL) | (Val & 0x0000FF000000FF00ULL) << 8;
3225 Val = (Val & 0x000F000F000F000FULL) | (Val & 0x00F000F000F000F0ULL) << 4;
3226 Val = (Val & 0x0303030303030303ULL) | (Val & 0x0C0C0C0C0C0C0C0CULL) << 2;
3227 Val = (Val & 0x1111111111111111ULL) | (Val & 0x2222222222222222ULL) << 1;
3228 Val = Val | Val << 1;
3229 return ConstantInt::get(Ty, V: Val);
3230 }
3231 }
3232 }
3233
3234 if (Operands[0]->getType()->isVectorTy()) {
3235 auto *Op = cast<Constant>(Val: Operands[0]);
3236 switch (IntrinsicID) {
3237 default: break;
3238 case Intrinsic::vector_reduce_add:
3239 case Intrinsic::vector_reduce_mul:
3240 case Intrinsic::vector_reduce_and:
3241 case Intrinsic::vector_reduce_or:
3242 case Intrinsic::vector_reduce_xor:
3243 case Intrinsic::vector_reduce_smin:
3244 case Intrinsic::vector_reduce_smax:
3245 case Intrinsic::vector_reduce_umin:
3246 case Intrinsic::vector_reduce_umax:
3247 if (Constant *C = constantFoldVectorReduce(IID: IntrinsicID, Op: Operands[0]))
3248 return C;
3249 break;
3250 case Intrinsic::x86_sse_cvtss2si:
3251 case Intrinsic::x86_sse_cvtss2si64:
3252 case Intrinsic::x86_sse2_cvtsd2si:
3253 case Intrinsic::x86_sse2_cvtsd2si64:
3254 if (ConstantFP *FPOp =
3255 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3256 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3257 /*roundTowardZero=*/false, Ty,
3258 /*IsSigned*/true);
3259 break;
3260 case Intrinsic::x86_sse_cvttss2si:
3261 case Intrinsic::x86_sse_cvttss2si64:
3262 case Intrinsic::x86_sse2_cvttsd2si:
3263 case Intrinsic::x86_sse2_cvttsd2si64:
3264 if (ConstantFP *FPOp =
3265 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3266 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3267 /*roundTowardZero=*/true, Ty,
3268 /*IsSigned*/true);
3269 break;
3270
3271 case Intrinsic::wasm_anytrue:
3272 return Op->isNullValue() ? ConstantInt::get(Ty, V: 0)
3273 : ConstantInt::get(Ty, V: 1);
3274
3275 case Intrinsic::wasm_alltrue:
3276 // Check each element individually
3277 unsigned E = cast<FixedVectorType>(Val: Op->getType())->getNumElements();
3278 for (unsigned I = 0; I != E; ++I) {
3279 Constant *Elt = Op->getAggregateElement(Elt: I);
3280 // Return false as soon as we find a non-true element.
3281 if (Elt && Elt->isNullValue())
3282 return ConstantInt::get(Ty, V: 0);
3283 // Bail as soon as we find an element we cannot prove to be true.
3284 if (!Elt || !isa<ConstantInt>(Val: Elt))
3285 return nullptr;
3286 }
3287
3288 return ConstantInt::get(Ty, V: 1);
3289 }
3290 }
3291
3292 return nullptr;
3293}
3294
3295static Constant *evaluateCompare(const APFloat &Op1, const APFloat &Op2,
3296 const ConstrainedFPIntrinsic *Call) {
3297 APFloat::opStatus St = APFloat::opOK;
3298 auto *FCmp = cast<ConstrainedFPCmpIntrinsic>(Val: Call);
3299 FCmpInst::Predicate Cond = FCmp->getPredicate();
3300 if (FCmp->isSignaling()) {
3301 if (Op1.isNaN() || Op2.isNaN())
3302 St = APFloat::opInvalidOp;
3303 } else {
3304 if (Op1.isSignaling() || Op2.isSignaling())
3305 St = APFloat::opInvalidOp;
3306 }
3307 bool Result = FCmpInst::compare(LHS: Op1, RHS: Op2, Pred: Cond);
3308 if (mayFoldConstrained(CI: const_cast<ConstrainedFPCmpIntrinsic *>(FCmp), St))
3309 return ConstantInt::get(Ty: Call->getType()->getScalarType(), V: Result);
3310 return nullptr;
3311}
3312
3313static Constant *ConstantFoldNextToward(const APFloat &Op0, const APFloat &Op1,
3314 const Type *RetTy) {
3315 assert(RetTy != nullptr);
3316 bool LosesInfo;
3317
3318 if (Op1.isSignaling())
3319 return nullptr;
3320 if (Op1.isNaN()) {
3321 APFloat Ret(Op1);
3322 Ret.convert(ToSemantics: RetTy->getFltSemantics(), RM: detail::rmNearestTiesToEven,
3323 losesInfo: &LosesInfo);
3324 return ConstantFP::get(Context&: RetTy->getContext(), V: Ret);
3325 }
3326
3327 // Recall that the second argument of nexttoward is always a long double,
3328 // so we may need to promote the first argument for comparisons to be valid.
3329 APFloat PromotedOp0(Op0);
3330 PromotedOp0.convert(ToSemantics: Op1.getSemantics(), RM: detail::rmNearestTiesToEven,
3331 losesInfo: &LosesInfo);
3332 assert(!LosesInfo && "Unexpected lossy promotion");
3333 const APFloat::cmpResult Result = PromotedOp0.compare(RHS: Op1);
3334
3335 // When equal, the standard says we must return the second argument.
3336 // This allows nice behavior such as nexttoward(0.0, -0.0) = -0.0 and
3337 // nexttoward(-0.0, 0.0) = 0.0
3338 if (Result == detail::cmpEqual) {
3339 APFloat Ret(Op1);
3340 Ret.convert(ToSemantics: RetTy->getFltSemantics(), RM: detail::rmNearestTiesToEven,
3341 losesInfo: &LosesInfo);
3342 return ConstantFP::get(Context&: RetTy->getContext(), V: Ret);
3343 }
3344
3345 APFloat Next(Op0);
3346 Next.next(/*nextDown=*/Result == APFloat::cmpGreaterThan);
3347 if (Next.isZero() || Next.isDenormal() || Next.isSignaling())
3348 return nullptr;
3349 return ConstantFP::get(Context&: RetTy->getContext(), V: Next);
3350}
3351
3352static Constant *ConstantFoldLibCall2(StringRef Name, Type *Ty,
3353 ArrayRef<Constant *> Operands,
3354 const TargetLibraryInfo *TLI = nullptr) {
3355 if (!TLI)
3356 return nullptr;
3357
3358 LibFunc Func = NotLibFunc;
3359 if (!TLI->getLibFunc(funcName: Name, F&: Func))
3360 return nullptr;
3361
3362 const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0]);
3363 if (!Op1)
3364 return nullptr;
3365
3366 const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1]);
3367 if (!Op2)
3368 return nullptr;
3369
3370 const APFloat &Op1V = Op1->getValueAPF();
3371 const APFloat &Op2V = Op2->getValueAPF();
3372
3373 switch (Func) {
3374 default:
3375 break;
3376 case LibFunc_pow:
3377 case LibFunc_powf:
3378 case LibFunc_pow_finite:
3379 case LibFunc_powf_finite:
3380 if (TLI->has(F: Func))
3381 return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty);
3382 break;
3383 case LibFunc_fmod:
3384 case LibFunc_fmodf:
3385 if (TLI->has(F: Func)) {
3386 APFloat V = Op1->getValueAPF();
3387 if (APFloat::opStatus::opOK == V.mod(RHS: Op2->getValueAPF()))
3388 return ConstantFP::get(Ty, V);
3389 }
3390 break;
3391 case LibFunc_remainder:
3392 case LibFunc_remainderf:
3393 if (TLI->has(F: Func)) {
3394 APFloat V = Op1->getValueAPF();
3395 if (APFloat::opStatus::opOK == V.remainder(RHS: Op2->getValueAPF()))
3396 return ConstantFP::get(Ty, V);
3397 }
3398 break;
3399 case LibFunc_atan2:
3400 case LibFunc_atan2f:
3401 // atan2(+/-0.0, +/-0.0) is known to raise an exception on some libm
3402 // (Solaris), so we do not assume a known result for that.
3403 if (Op1V.isZero() && Op2V.isZero())
3404 return nullptr;
3405 [[fallthrough]];
3406 case LibFunc_atan2_finite:
3407 case LibFunc_atan2f_finite:
3408 if (TLI->has(F: Func))
3409 return ConstantFoldBinaryFP(NativeFP: atan2, V: Op1V, W: Op2V, Ty);
3410 break;
3411 case LibFunc_nextafter:
3412 case LibFunc_nextafterf:
3413 case LibFunc_nexttoward:
3414 case LibFunc_nexttowardf:
3415 if (TLI->has(F: Func))
3416 return ConstantFoldNextToward(Op0: Op1V, Op1: Op2V, RetTy: Ty);
3417 break;
3418 }
3419
3420 return nullptr;
3421}
3422
3423static Constant *ConstantFoldIntrinsicCall2(Intrinsic::ID IntrinsicID, Type *Ty,
3424 ArrayRef<Constant *> Operands,
3425 const CallBase *Call = nullptr) {
3426 assert(Operands.size() == 2 && "Wrong number of operands.");
3427
3428 if (Ty->isFloatingPointTy()) {
3429 // TODO: We should have undef handling for all of the FP intrinsics that
3430 // are attempted to be folded in this function.
3431 bool IsOp0Undef = isa<UndefValue>(Val: Operands[0]);
3432 bool IsOp1Undef = isa<UndefValue>(Val: Operands[1]);
3433 switch (IntrinsicID) {
3434 case Intrinsic::maxnum:
3435 case Intrinsic::minnum:
3436 case Intrinsic::maximum:
3437 case Intrinsic::minimum:
3438 case Intrinsic::maximumnum:
3439 case Intrinsic::minimumnum:
3440 case Intrinsic::nvvm_fmax_d:
3441 case Intrinsic::nvvm_fmin_d:
3442 // If one argument is undef, return the other argument.
3443 if (IsOp0Undef)
3444 return Operands[1];
3445 if (IsOp1Undef)
3446 return Operands[0];
3447 break;
3448
3449 case Intrinsic::nvvm_fmax_f:
3450 case Intrinsic::nvvm_fmax_ftz_f:
3451 case Intrinsic::nvvm_fmax_ftz_nan_f:
3452 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
3453 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
3454 case Intrinsic::nvvm_fmax_nan_f:
3455 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
3456 case Intrinsic::nvvm_fmax_xorsign_abs_f:
3457
3458 case Intrinsic::nvvm_fmin_f:
3459 case Intrinsic::nvvm_fmin_ftz_f:
3460 case Intrinsic::nvvm_fmin_ftz_nan_f:
3461 case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
3462 case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
3463 case Intrinsic::nvvm_fmin_nan_f:
3464 case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
3465 case Intrinsic::nvvm_fmin_xorsign_abs_f:
3466 // If one arg is undef, the other arg can be returned only if it is
3467 // constant, as we may need to flush it to sign-preserving zero or
3468 // canonicalize the NaN.
3469 if (!IsOp0Undef && !IsOp1Undef)
3470 break;
3471 if (auto *Op = dyn_cast<ConstantFP>(Val: Operands[IsOp0Undef ? 1 : 0])) {
3472 if (Op->isNaN()) {
3473 APInt NVCanonicalNaN(32, 0x7fffffff);
3474 return ConstantFP::get(
3475 Ty, V: APFloat(Ty->getFltSemantics(), NVCanonicalNaN));
3476 }
3477 if (nvvm::FMinFMaxShouldFTZ(IntrinsicID))
3478 return ConstantFP::get(Ty, V: FTZPreserveSign(V: Op->getValueAPF()));
3479 else
3480 return Op;
3481 }
3482 break;
3483 }
3484 }
3485
3486 if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) {
3487 const APFloat &Op1V = Op1->getValueAPF();
3488
3489 if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) {
3490 if (Op2->getType() != Op1->getType())
3491 return nullptr;
3492 const APFloat &Op2V = Op2->getValueAPF();
3493
3494 if (const auto *ConstrIntr =
3495 dyn_cast_if_present<ConstrainedFPIntrinsic>(Val: Call)) {
3496 RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr);
3497 APFloat Res = Op1V;
3498 APFloat::opStatus St;
3499 switch (IntrinsicID) {
3500 default:
3501 return nullptr;
3502 case Intrinsic::experimental_constrained_fadd:
3503 St = Res.add(RHS: Op2V, RM);
3504 break;
3505 case Intrinsic::experimental_constrained_fsub:
3506 St = Res.subtract(RHS: Op2V, RM);
3507 break;
3508 case Intrinsic::experimental_constrained_fmul:
3509 St = Res.multiply(RHS: Op2V, RM);
3510 break;
3511 case Intrinsic::experimental_constrained_fdiv:
3512 St = Res.divide(RHS: Op2V, RM);
3513 break;
3514 case Intrinsic::experimental_constrained_frem:
3515 St = Res.mod(RHS: Op2V);
3516 break;
3517 case Intrinsic::experimental_constrained_fcmp:
3518 case Intrinsic::experimental_constrained_fcmps:
3519 return evaluateCompare(Op1: Op1V, Op2: Op2V, Call: ConstrIntr);
3520 }
3521 if (mayFoldConstrained(CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr),
3522 St))
3523 return ConstantFP::get(Ty, V: Res);
3524 return nullptr;
3525 }
3526
3527 switch (IntrinsicID) {
3528 default:
3529 break;
3530 case Intrinsic::copysign:
3531 return ConstantFP::get(Ty, V: APFloat::copySign(Value: Op1V, Sign: Op2V));
3532 case Intrinsic::minnum:
3533 return ConstantFP::get(Ty, V: minnum(A: Op1V, B: Op2V));
3534 case Intrinsic::maxnum:
3535 return ConstantFP::get(Ty, V: maxnum(A: Op1V, B: Op2V));
3536 case Intrinsic::minimum:
3537 return ConstantFP::get(Ty, V: minimum(A: Op1V, B: Op2V));
3538 case Intrinsic::maximum:
3539 return ConstantFP::get(Ty, V: maximum(A: Op1V, B: Op2V));
3540 case Intrinsic::minimumnum:
3541 return ConstantFP::get(Ty, V: minimumnum(A: Op1V, B: Op2V));
3542 case Intrinsic::maximumnum:
3543 return ConstantFP::get(Ty, V: maximumnum(A: Op1V, B: Op2V));
3544
3545 case Intrinsic::nvvm_fmax_d:
3546 case Intrinsic::nvvm_fmax_f:
3547 case Intrinsic::nvvm_fmax_ftz_f:
3548 case Intrinsic::nvvm_fmax_ftz_nan_f:
3549 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
3550 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
3551 case Intrinsic::nvvm_fmax_nan_f:
3552 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
3553 case Intrinsic::nvvm_fmax_xorsign_abs_f:
3554
3555 case Intrinsic::nvvm_fmin_d:
3556 case Intrinsic::nvvm_fmin_f:
3557 case Intrinsic::nvvm_fmin_ftz_f:
3558 case Intrinsic::nvvm_fmin_ftz_nan_f:
3559 case Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f:
3560 case Intrinsic::nvvm_fmin_ftz_xorsign_abs_f:
3561 case Intrinsic::nvvm_fmin_nan_f:
3562 case Intrinsic::nvvm_fmin_nan_xorsign_abs_f:
3563 case Intrinsic::nvvm_fmin_xorsign_abs_f: {
3564
3565 bool ShouldCanonicalizeNaNs = !(IntrinsicID == Intrinsic::nvvm_fmax_d ||
3566 IntrinsicID == Intrinsic::nvvm_fmin_d);
3567 bool IsFTZ = nvvm::FMinFMaxShouldFTZ(IntrinsicID);
3568 bool IsNaNPropagating = nvvm::FMinFMaxPropagatesNaNs(IntrinsicID);
3569 bool IsXorSignAbs = nvvm::FMinFMaxIsXorSignAbs(IntrinsicID);
3570
3571 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3572 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3573
3574 bool XorSign = false;
3575 if (IsXorSignAbs) {
3576 XorSign = A.isNegative() ^ B.isNegative();
3577 A = abs(X: A);
3578 B = abs(X: B);
3579 }
3580
3581 bool IsFMax = false;
3582 switch (IntrinsicID) {
3583 case Intrinsic::nvvm_fmax_d:
3584 case Intrinsic::nvvm_fmax_f:
3585 case Intrinsic::nvvm_fmax_ftz_f:
3586 case Intrinsic::nvvm_fmax_ftz_nan_f:
3587 case Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f:
3588 case Intrinsic::nvvm_fmax_ftz_xorsign_abs_f:
3589 case Intrinsic::nvvm_fmax_nan_f:
3590 case Intrinsic::nvvm_fmax_nan_xorsign_abs_f:
3591 case Intrinsic::nvvm_fmax_xorsign_abs_f:
3592 IsFMax = true;
3593 break;
3594 }
3595 APFloat Res =
3596 IsFMax ? (IsNaNPropagating ? maximum(A, B) : maximumnum(A, B))
3597 : (IsNaNPropagating ? minimum(A, B) : minimumnum(A, B));
3598
3599 if (ShouldCanonicalizeNaNs && Res.isNaN()) {
3600 APFloat NVCanonicalNaN(Res.getSemantics(), APInt(32, 0x7fffffff));
3601 return ConstantFP::get(Ty, V: NVCanonicalNaN);
3602 }
3603
3604 if (IsXorSignAbs && XorSign != Res.isNegative())
3605 Res.changeSign();
3606
3607 return ConstantFP::get(Ty, V: Res);
3608 }
3609
3610 case Intrinsic::nvvm_add_rm_f:
3611 case Intrinsic::nvvm_add_rn_f:
3612 case Intrinsic::nvvm_add_rp_f:
3613 case Intrinsic::nvvm_add_rz_f:
3614 case Intrinsic::nvvm_add_rm_d:
3615 case Intrinsic::nvvm_add_rn_d:
3616 case Intrinsic::nvvm_add_rp_d:
3617 case Intrinsic::nvvm_add_rz_d:
3618 case Intrinsic::nvvm_add_rm_ftz_f:
3619 case Intrinsic::nvvm_add_rn_ftz_f:
3620 case Intrinsic::nvvm_add_rp_ftz_f:
3621 case Intrinsic::nvvm_add_rz_ftz_f: {
3622
3623 bool IsFTZ = nvvm::FAddShouldFTZ(IntrinsicID);
3624 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3625 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3626
3627 APFloat::roundingMode RoundMode =
3628 nvvm::GetFAddRoundingMode(IntrinsicID);
3629
3630 APFloat Res = A;
3631 APFloat::opStatus Status = Res.add(RHS: B, RM: RoundMode);
3632
3633 if (!Res.isNaN() &&
3634 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3635 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3636 return ConstantFP::get(Ty, V: Res);
3637 }
3638 return nullptr;
3639 }
3640
3641 case Intrinsic::nvvm_mul_rm_f:
3642 case Intrinsic::nvvm_mul_rn_f:
3643 case Intrinsic::nvvm_mul_rp_f:
3644 case Intrinsic::nvvm_mul_rz_f:
3645 case Intrinsic::nvvm_mul_rm_d:
3646 case Intrinsic::nvvm_mul_rn_d:
3647 case Intrinsic::nvvm_mul_rp_d:
3648 case Intrinsic::nvvm_mul_rz_d:
3649 case Intrinsic::nvvm_mul_rm_ftz_f:
3650 case Intrinsic::nvvm_mul_rn_ftz_f:
3651 case Intrinsic::nvvm_mul_rp_ftz_f:
3652 case Intrinsic::nvvm_mul_rz_ftz_f: {
3653
3654 bool IsFTZ = nvvm::FMulShouldFTZ(IntrinsicID);
3655 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3656 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3657
3658 APFloat::roundingMode RoundMode =
3659 nvvm::GetFMulRoundingMode(IntrinsicID);
3660
3661 APFloat Res = A;
3662 APFloat::opStatus Status = Res.multiply(RHS: B, RM: RoundMode);
3663
3664 if (!Res.isNaN() &&
3665 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3666 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3667 return ConstantFP::get(Ty, V: Res);
3668 }
3669 return nullptr;
3670 }
3671
3672 case Intrinsic::nvvm_div_rm_f:
3673 case Intrinsic::nvvm_div_rn_f:
3674 case Intrinsic::nvvm_div_rp_f:
3675 case Intrinsic::nvvm_div_rz_f:
3676 case Intrinsic::nvvm_div_rm_d:
3677 case Intrinsic::nvvm_div_rn_d:
3678 case Intrinsic::nvvm_div_rp_d:
3679 case Intrinsic::nvvm_div_rz_d:
3680 case Intrinsic::nvvm_div_rm_ftz_f:
3681 case Intrinsic::nvvm_div_rn_ftz_f:
3682 case Intrinsic::nvvm_div_rp_ftz_f:
3683 case Intrinsic::nvvm_div_rz_ftz_f: {
3684 bool IsFTZ = nvvm::FDivShouldFTZ(IntrinsicID);
3685 APFloat A = IsFTZ ? FTZPreserveSign(V: Op1V) : Op1V;
3686 APFloat B = IsFTZ ? FTZPreserveSign(V: Op2V) : Op2V;
3687 APFloat::roundingMode RoundMode =
3688 nvvm::GetFDivRoundingMode(IntrinsicID);
3689
3690 APFloat Res = A;
3691 APFloat::opStatus Status = Res.divide(RHS: B, RM: RoundMode);
3692 if (!Res.isNaN() &&
3693 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
3694 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
3695 return ConstantFP::get(Ty, V: Res);
3696 }
3697 return nullptr;
3698 }
3699 }
3700
3701 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
3702 return nullptr;
3703
3704 switch (IntrinsicID) {
3705 default:
3706 break;
3707 case Intrinsic::pow:
3708 return ConstantFoldBinaryFP(NativeFP: pow, V: Op1V, W: Op2V, Ty);
3709 case Intrinsic::amdgcn_fmul_legacy:
3710 // The legacy behaviour is that multiplying +/- 0.0 by anything, even
3711 // NaN or infinity, gives +0.0.
3712 if (Op1V.isZero() || Op2V.isZero())
3713 return ConstantFP::getZero(Ty);
3714 return ConstantFP::get(Ty, V: Op1V * Op2V);
3715 }
3716
3717 } else if (auto *Op2C = dyn_cast<ConstantInt>(Val: Operands[1])) {
3718 switch (IntrinsicID) {
3719 case Intrinsic::ldexp: {
3720 // APFloat::scalbn takes the exponent as `int`. Clamp wider integer
3721 // exponents into [INT_MIN, INT_MAX] so values still saturate the
3722 // result to +/-inf or +/-0.
3723 APInt Exp = Op2C->getValue();
3724 Exp = Exp.getBitWidth() < 32 ? Exp.sext(width: 32) : Exp.truncSSat(width: 32);
3725 return ConstantFP::get(
3726 Context&: Ty->getContext(),
3727 V: scalbn(X: Op1V, Exp: Exp.getSExtValue(), RM: APFloat::rmNearestTiesToEven));
3728 }
3729 case Intrinsic::is_fpclass: {
3730 FPClassTest Mask = static_cast<FPClassTest>(Op2C->getZExtValue());
3731 bool Result =
3732 ((Mask & fcSNan) && Op1V.isNaN() && Op1V.isSignaling()) ||
3733 ((Mask & fcQNan) && Op1V.isNaN() && !Op1V.isSignaling()) ||
3734 ((Mask & fcNegInf) && Op1V.isNegInfinity()) ||
3735 ((Mask & fcNegNormal) && Op1V.isNormal() && Op1V.isNegative()) ||
3736 ((Mask & fcNegSubnormal) && Op1V.isDenormal() && Op1V.isNegative()) ||
3737 ((Mask & fcNegZero) && Op1V.isZero() && Op1V.isNegative()) ||
3738 ((Mask & fcPosZero) && Op1V.isZero() && !Op1V.isNegative()) ||
3739 ((Mask & fcPosSubnormal) && Op1V.isDenormal() && !Op1V.isNegative()) ||
3740 ((Mask & fcPosNormal) && Op1V.isNormal() && !Op1V.isNegative()) ||
3741 ((Mask & fcPosInf) && Op1V.isPosInfinity());
3742 return ConstantInt::get(Ty, V: Result);
3743 }
3744 case Intrinsic::powi: {
3745 int Exp = static_cast<int>(Op2C->getSExtValue());
3746 switch (Ty->getTypeID()) {
3747 case Type::HalfTyID:
3748 case Type::FloatTyID: {
3749 APFloat Res(static_cast<float>(std::pow(x: Op1V.convertToFloat(), y: Exp)));
3750 if (Ty->isHalfTy()) {
3751 bool Unused;
3752 Res.convert(ToSemantics: APFloat::IEEEhalf(), RM: APFloat::rmNearestTiesToEven,
3753 losesInfo: &Unused);
3754 }
3755 return ConstantFP::get(Ty, V: Res);
3756 }
3757 case Type::DoubleTyID:
3758 return ConstantFP::get(Ty, V: std::pow(x: Op1V.convertToDouble(), y: Exp));
3759 default:
3760 return nullptr;
3761 }
3762 }
3763 default:
3764 break;
3765 }
3766 }
3767 return nullptr;
3768 }
3769
3770 if (Operands[0]->getType()->isIntegerTy() &&
3771 Operands[1]->getType()->isIntegerTy()) {
3772 const APInt *C0, *C1;
3773 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
3774 !getConstIntOrUndef(Op: Operands[1], C&: C1))
3775 return nullptr;
3776
3777 switch (IntrinsicID) {
3778 default: break;
3779 case Intrinsic::smax:
3780 case Intrinsic::smin:
3781 case Intrinsic::umax:
3782 case Intrinsic::umin:
3783 if (!C0 || !C1)
3784 return MinMaxIntrinsic::getSaturationPoint(ID: IntrinsicID, Ty);
3785 return ConstantInt::get(
3786 Ty, V: ICmpInst::compare(LHS: *C0, RHS: *C1,
3787 Pred: MinMaxIntrinsic::getPredicate(ID: IntrinsicID))
3788 ? *C0
3789 : *C1);
3790
3791 case Intrinsic::scmp:
3792 case Intrinsic::ucmp:
3793 if (!C0 || !C1)
3794 return ConstantInt::get(Ty, V: 0);
3795
3796 int Res;
3797 if (IntrinsicID == Intrinsic::scmp)
3798 Res = C0->sgt(RHS: *C1) ? 1 : C0->slt(RHS: *C1) ? -1 : 0;
3799 else
3800 Res = C0->ugt(RHS: *C1) ? 1 : C0->ult(RHS: *C1) ? -1 : 0;
3801 return ConstantInt::get(Ty, V: Res, /*IsSigned=*/true);
3802
3803 case Intrinsic::usub_with_overflow:
3804 case Intrinsic::ssub_with_overflow:
3805 // X - undef -> { 0, false }
3806 // undef - X -> { 0, false }
3807 if (!C0 || !C1)
3808 return Constant::getNullValue(Ty);
3809 [[fallthrough]];
3810 case Intrinsic::uadd_with_overflow:
3811 case Intrinsic::sadd_with_overflow:
3812 // X + undef -> { -1, false }
3813 // undef + x -> { -1, false }
3814 if (!C0 || !C1) {
3815 return ConstantStruct::get(
3816 T: cast<StructType>(Val: Ty),
3817 V: {Constant::getAllOnesValue(Ty: Ty->getStructElementType(N: 0)),
3818 Constant::getNullValue(Ty: Ty->getStructElementType(N: 1))});
3819 }
3820 [[fallthrough]];
3821 case Intrinsic::smul_with_overflow:
3822 case Intrinsic::umul_with_overflow: {
3823 // undef * X -> { 0, false }
3824 // X * undef -> { 0, false }
3825 if (!C0 || !C1)
3826 return Constant::getNullValue(Ty);
3827
3828 APInt Res;
3829 bool Overflow;
3830 switch (IntrinsicID) {
3831 default: llvm_unreachable("Invalid case");
3832 case Intrinsic::sadd_with_overflow:
3833 Res = C0->sadd_ov(RHS: *C1, Overflow);
3834 break;
3835 case Intrinsic::uadd_with_overflow:
3836 Res = C0->uadd_ov(RHS: *C1, Overflow);
3837 break;
3838 case Intrinsic::ssub_with_overflow:
3839 Res = C0->ssub_ov(RHS: *C1, Overflow);
3840 break;
3841 case Intrinsic::usub_with_overflow:
3842 Res = C0->usub_ov(RHS: *C1, Overflow);
3843 break;
3844 case Intrinsic::smul_with_overflow:
3845 Res = C0->smul_ov(RHS: *C1, Overflow);
3846 break;
3847 case Intrinsic::umul_with_overflow:
3848 Res = C0->umul_ov(RHS: *C1, Overflow);
3849 break;
3850 }
3851 Constant *Ops[] = {
3852 ConstantInt::get(Context&: Ty->getContext(), V: Res),
3853 ConstantInt::get(Ty: Type::getInt1Ty(C&: Ty->getContext()), V: Overflow)
3854 };
3855 return ConstantStruct::get(T: cast<StructType>(Val: Ty), V: Ops);
3856 }
3857 case Intrinsic::uadd_sat:
3858 case Intrinsic::sadd_sat:
3859 if (!C0 || !C1)
3860 return Constant::getAllOnesValue(Ty);
3861 if (IntrinsicID == Intrinsic::uadd_sat)
3862 return ConstantInt::get(Ty, V: C0->uadd_sat(RHS: *C1));
3863 else
3864 return ConstantInt::get(Ty, V: C0->sadd_sat(RHS: *C1));
3865 case Intrinsic::usub_sat:
3866 case Intrinsic::ssub_sat:
3867 if (!C0 || !C1)
3868 return Constant::getNullValue(Ty);
3869 if (IntrinsicID == Intrinsic::usub_sat)
3870 return ConstantInt::get(Ty, V: C0->usub_sat(RHS: *C1));
3871 else
3872 return ConstantInt::get(Ty, V: C0->ssub_sat(RHS: *C1));
3873 case Intrinsic::cttz:
3874 case Intrinsic::ctlz:
3875 assert(C1 && "Must be constant int");
3876
3877 // cttz(0, 1) and ctlz(0, 1) are poison.
3878 if (C1->isOne() && (!C0 || C0->isZero()))
3879 return PoisonValue::get(T: Ty);
3880 if (!C0)
3881 return Constant::getNullValue(Ty);
3882 if (IntrinsicID == Intrinsic::cttz)
3883 return ConstantInt::get(Ty, V: C0->countr_zero());
3884 else
3885 return ConstantInt::get(Ty, V: C0->countl_zero());
3886
3887 case Intrinsic::abs:
3888 assert(C1 && "Must be constant int");
3889 assert((C1->isOne() || C1->isZero()) && "Must be 0 or 1");
3890
3891 // Undef or minimum val operand with poison min --> poison
3892 if (C1->isOne() && (!C0 || C0->isMinSignedValue()))
3893 return PoisonValue::get(T: Ty);
3894
3895 // Undef operand with no poison min --> 0 (sign bit must be clear)
3896 if (!C0)
3897 return Constant::getNullValue(Ty);
3898
3899 return ConstantInt::get(Ty, V: C0->abs());
3900 case Intrinsic::clmul:
3901 if (!C0 || !C1)
3902 return Constant::getNullValue(Ty);
3903 return ConstantInt::get(Ty, V: APIntOps::clmul(LHS: *C0, RHS: *C1));
3904 case Intrinsic::pdep:
3905 if (!C0 || !C1)
3906 return Constant::getNullValue(Ty);
3907 return ConstantInt::get(Ty, V: APIntOps::pdep(Val: *C0, Mask: *C1));
3908 case Intrinsic::pext:
3909 if (!C0 || !C1)
3910 return Constant::getNullValue(Ty);
3911 return ConstantInt::get(Ty, V: APIntOps::pext(Val: *C0, Mask: *C1));
3912 case Intrinsic::amdgcn_wave_reduce_umin:
3913 case Intrinsic::amdgcn_wave_reduce_umax:
3914 case Intrinsic::amdgcn_wave_reduce_max:
3915 case Intrinsic::amdgcn_wave_reduce_min:
3916 case Intrinsic::amdgcn_wave_reduce_and:
3917 case Intrinsic::amdgcn_wave_reduce_or:
3918 return Operands[0];
3919 }
3920
3921 return nullptr;
3922 }
3923
3924 // Support ConstantVector in case we have an Undef in the top.
3925 if ((isa<ConstantVector>(Val: Operands[0]) ||
3926 isa<ConstantDataVector>(Val: Operands[0])) &&
3927 // Check for default rounding mode.
3928 // FIXME: Support other rounding modes?
3929 isa<ConstantInt>(Val: Operands[1]) &&
3930 cast<ConstantInt>(Val: Operands[1])->getValue() == 4) {
3931 auto *Op = cast<Constant>(Val: Operands[0]);
3932 switch (IntrinsicID) {
3933 default: break;
3934 case Intrinsic::x86_avx512_vcvtss2si32:
3935 case Intrinsic::x86_avx512_vcvtss2si64:
3936 case Intrinsic::x86_avx512_vcvtsd2si32:
3937 case Intrinsic::x86_avx512_vcvtsd2si64:
3938 if (ConstantFP *FPOp =
3939 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3940 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3941 /*roundTowardZero=*/false, Ty,
3942 /*IsSigned*/true);
3943 break;
3944 case Intrinsic::x86_avx512_vcvtss2usi32:
3945 case Intrinsic::x86_avx512_vcvtss2usi64:
3946 case Intrinsic::x86_avx512_vcvtsd2usi32:
3947 case Intrinsic::x86_avx512_vcvtsd2usi64:
3948 if (ConstantFP *FPOp =
3949 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3950 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3951 /*roundTowardZero=*/false, Ty,
3952 /*IsSigned*/false);
3953 break;
3954 case Intrinsic::x86_avx512_cvttss2si:
3955 case Intrinsic::x86_avx512_cvttss2si64:
3956 case Intrinsic::x86_avx512_cvttsd2si:
3957 case Intrinsic::x86_avx512_cvttsd2si64:
3958 if (ConstantFP *FPOp =
3959 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3960 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3961 /*roundTowardZero=*/true, Ty,
3962 /*IsSigned*/true);
3963 break;
3964 case Intrinsic::x86_avx512_cvttss2usi:
3965 case Intrinsic::x86_avx512_cvttss2usi64:
3966 case Intrinsic::x86_avx512_cvttsd2usi:
3967 case Intrinsic::x86_avx512_cvttsd2usi64:
3968 if (ConstantFP *FPOp =
3969 dyn_cast_or_null<ConstantFP>(Val: Op->getAggregateElement(Elt: 0U)))
3970 return ConstantFoldSSEConvertToInt(Val: FPOp->getValueAPF(),
3971 /*roundTowardZero=*/true, Ty,
3972 /*IsSigned*/false);
3973 break;
3974 }
3975 }
3976
3977 if (IntrinsicID == Intrinsic::experimental_cttz_elts) {
3978 auto *FVTy = dyn_cast<FixedVectorType>(Val: Operands[0]->getType());
3979 bool ZeroIsPoison = cast<ConstantInt>(Val: Operands[1])->isOne();
3980 if (!FVTy)
3981 return nullptr;
3982 unsigned Width = Ty->getIntegerBitWidth();
3983 if (APInt::getMaxValue(numBits: Width).ult(RHS: FVTy->getNumElements()))
3984 return PoisonValue::get(T: Ty);
3985 for (unsigned I = 0; I < FVTy->getNumElements(); ++I) {
3986 Constant *Elt = Operands[0]->getAggregateElement(Elt: I);
3987 if (!Elt)
3988 return nullptr;
3989 if (isa<UndefValue>(Val: Elt) || Elt->isNullValue())
3990 continue;
3991 return ConstantInt::get(Ty, V: I);
3992 }
3993 if (ZeroIsPoison)
3994 return PoisonValue::get(T: Ty);
3995 return ConstantInt::get(Ty, V: FVTy->getNumElements());
3996 }
3997 return nullptr;
3998}
3999
4000static APFloat ConstantFoldAMDGCNCubeIntrinsic(Intrinsic::ID IntrinsicID,
4001 const APFloat &S0,
4002 const APFloat &S1,
4003 const APFloat &S2) {
4004 unsigned ID;
4005 const fltSemantics &Sem = S0.getSemantics();
4006 APFloat MA(Sem), SC(Sem), TC(Sem);
4007 if (abs(X: S2) >= abs(X: S0) && abs(X: S2) >= abs(X: S1)) {
4008 if (S2.isNegative() && S2.isNonZero() && !S2.isNaN()) {
4009 // S2 < 0
4010 ID = 5;
4011 SC = -S0;
4012 } else {
4013 ID = 4;
4014 SC = S0;
4015 }
4016 MA = S2;
4017 TC = -S1;
4018 } else if (abs(X: S1) >= abs(X: S0)) {
4019 if (S1.isNegative() && S1.isNonZero() && !S1.isNaN()) {
4020 // S1 < 0
4021 ID = 3;
4022 TC = -S2;
4023 } else {
4024 ID = 2;
4025 TC = S2;
4026 }
4027 MA = S1;
4028 SC = S0;
4029 } else {
4030 if (S0.isNegative() && S0.isNonZero() && !S0.isNaN()) {
4031 // S0 < 0
4032 ID = 1;
4033 SC = S2;
4034 } else {
4035 ID = 0;
4036 SC = -S2;
4037 }
4038 MA = S0;
4039 TC = -S1;
4040 }
4041 switch (IntrinsicID) {
4042 default:
4043 llvm_unreachable("unhandled amdgcn cube intrinsic");
4044 case Intrinsic::amdgcn_cubeid:
4045 return APFloat(Sem, ID);
4046 case Intrinsic::amdgcn_cubema:
4047 return MA + MA;
4048 case Intrinsic::amdgcn_cubesc:
4049 return SC;
4050 case Intrinsic::amdgcn_cubetc:
4051 return TC;
4052 }
4053}
4054
4055static Constant *ConstantFoldAMDGCNPermIntrinsic(ArrayRef<Constant *> Operands,
4056 Type *Ty) {
4057 const APInt *C0, *C1, *C2;
4058 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
4059 !getConstIntOrUndef(Op: Operands[1], C&: C1) ||
4060 !getConstIntOrUndef(Op: Operands[2], C&: C2))
4061 return nullptr;
4062
4063 if (!C2)
4064 return UndefValue::get(T: Ty);
4065
4066 APInt Val(32, 0);
4067 unsigned NumUndefBytes = 0;
4068 for (unsigned I = 0; I < 32; I += 8) {
4069 unsigned Sel = C2->extractBitsAsZExtValue(numBits: 8, bitPosition: I);
4070 unsigned B = 0;
4071
4072 if (Sel >= 13)
4073 B = 0xff;
4074 else if (Sel == 12)
4075 B = 0x00;
4076 else {
4077 const APInt *Src = ((Sel & 10) == 10 || (Sel & 12) == 4) ? C0 : C1;
4078 if (!Src)
4079 ++NumUndefBytes;
4080 else if (Sel < 8)
4081 B = Src->extractBitsAsZExtValue(numBits: 8, bitPosition: (Sel & 3) * 8);
4082 else
4083 B = Src->extractBitsAsZExtValue(numBits: 1, bitPosition: (Sel & 1) ? 31 : 15) * 0xff;
4084 }
4085
4086 Val.insertBits(SubBits: B, bitPosition: I, numBits: 8);
4087 }
4088
4089 if (NumUndefBytes == 4)
4090 return UndefValue::get(T: Ty);
4091
4092 return ConstantInt::get(Ty, V: Val);
4093}
4094
4095static Constant *ConstantFoldScalarCall3(StringRef Name,
4096 Intrinsic::ID IntrinsicID, Type *Ty,
4097 ArrayRef<Constant *> Operands,
4098 const TargetLibraryInfo *TLI = nullptr,
4099 const CallBase *Call = nullptr) {
4100 assert(Operands.size() == 3 && "Wrong number of operands.");
4101
4102 if (const auto *Op1 = dyn_cast<ConstantFP>(Val: Operands[0])) {
4103 if (const auto *Op2 = dyn_cast<ConstantFP>(Val: Operands[1])) {
4104 if (const auto *Op3 = dyn_cast<ConstantFP>(Val: Operands[2])) {
4105 const APFloat &C1 = Op1->getValueAPF();
4106 const APFloat &C2 = Op2->getValueAPF();
4107 const APFloat &C3 = Op3->getValueAPF();
4108
4109 if (const auto *ConstrIntr =
4110 dyn_cast_or_null<ConstrainedFPIntrinsic>(Val: Call)) {
4111 RoundingMode RM = getEvaluationRoundingMode(CI: ConstrIntr);
4112 APFloat Res = C1;
4113 APFloat::opStatus St;
4114 switch (IntrinsicID) {
4115 default:
4116 return nullptr;
4117 case Intrinsic::experimental_constrained_fma:
4118 case Intrinsic::experimental_constrained_fmuladd:
4119 St = Res.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM);
4120 break;
4121 }
4122 if (mayFoldConstrained(
4123 CI: const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), St))
4124 return ConstantFP::get(Ty, V: Res);
4125 return nullptr;
4126 }
4127
4128 switch (IntrinsicID) {
4129 default: break;
4130 case Intrinsic::amdgcn_fma_legacy: {
4131 // The legacy behaviour is that multiplying +/- 0.0 by anything, even
4132 // NaN or infinity, gives +0.0.
4133 if (C1.isZero() || C2.isZero()) {
4134 // It's tempting to just return C3 here, but that would give the
4135 // wrong result if C3 was -0.0.
4136 return ConstantFP::get(Ty, V: APFloat(0.0f) + C3);
4137 }
4138 [[fallthrough]];
4139 }
4140 case Intrinsic::fma:
4141 case Intrinsic::fmuladd: {
4142 APFloat V = C1;
4143 V.fusedMultiplyAdd(Multiplicand: C2, Addend: C3, RM: APFloat::rmNearestTiesToEven);
4144 return ConstantFP::get(Ty, V);
4145 }
4146
4147 case Intrinsic::nvvm_fma_rm_f:
4148 case Intrinsic::nvvm_fma_rn_f:
4149 case Intrinsic::nvvm_fma_rp_f:
4150 case Intrinsic::nvvm_fma_rz_f:
4151 case Intrinsic::nvvm_fma_rm_d:
4152 case Intrinsic::nvvm_fma_rn_d:
4153 case Intrinsic::nvvm_fma_rp_d:
4154 case Intrinsic::nvvm_fma_rz_d:
4155 case Intrinsic::nvvm_fma_rm_ftz_f:
4156 case Intrinsic::nvvm_fma_rn_ftz_f:
4157 case Intrinsic::nvvm_fma_rp_ftz_f:
4158 case Intrinsic::nvvm_fma_rz_ftz_f: {
4159 bool IsFTZ = nvvm::FMAShouldFTZ(IntrinsicID);
4160 APFloat A = IsFTZ ? FTZPreserveSign(V: C1) : C1;
4161 APFloat B = IsFTZ ? FTZPreserveSign(V: C2) : C2;
4162 APFloat C = IsFTZ ? FTZPreserveSign(V: C3) : C3;
4163
4164 APFloat::roundingMode RoundMode =
4165 nvvm::GetFMARoundingMode(IntrinsicID);
4166
4167 APFloat Res = A;
4168 APFloat::opStatus Status = Res.fusedMultiplyAdd(Multiplicand: B, Addend: C, RM: RoundMode);
4169
4170 if (!Res.isNaN() &&
4171 (Status == APFloat::opOK || Status == APFloat::opInexact)) {
4172 Res = IsFTZ ? FTZPreserveSign(V: Res) : Res;
4173 return ConstantFP::get(Ty, V: Res);
4174 }
4175 return nullptr;
4176 }
4177
4178 case Intrinsic::amdgcn_cubeid:
4179 case Intrinsic::amdgcn_cubema:
4180 case Intrinsic::amdgcn_cubesc:
4181 case Intrinsic::amdgcn_cubetc: {
4182 APFloat V = ConstantFoldAMDGCNCubeIntrinsic(IntrinsicID, S0: C1, S1: C2, S2: C3);
4183 return ConstantFP::get(Ty, V);
4184 }
4185 }
4186 }
4187 }
4188 }
4189
4190 if (IntrinsicID == Intrinsic::smul_fix ||
4191 IntrinsicID == Intrinsic::smul_fix_sat) {
4192 const APInt *C0, *C1;
4193 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
4194 !getConstIntOrUndef(Op: Operands[1], C&: C1))
4195 return nullptr;
4196
4197 // undef * C -> 0
4198 // C * undef -> 0
4199 if (!C0 || !C1)
4200 return Constant::getNullValue(Ty);
4201
4202 // This code performs rounding towards negative infinity in case the result
4203 // cannot be represented exactly for the given scale. Targets that do care
4204 // about rounding should use a target hook for specifying how rounding
4205 // should be done, and provide their own folding to be consistent with
4206 // rounding. This is the same approach as used by
4207 // DAGTypeLegalizer::ExpandIntRes_MULFIX.
4208 unsigned Scale = cast<ConstantInt>(Val: Operands[2])->getZExtValue();
4209 unsigned Width = C0->getBitWidth();
4210 assert(Scale < Width && "Illegal scale.");
4211 unsigned ExtendedWidth = Width * 2;
4212 APInt Product =
4213 (C0->sext(width: ExtendedWidth) * C1->sext(width: ExtendedWidth)).ashr(ShiftAmt: Scale);
4214 if (IntrinsicID == Intrinsic::smul_fix_sat) {
4215 APInt Max = APInt::getSignedMaxValue(numBits: Width).sext(width: ExtendedWidth);
4216 APInt Min = APInt::getSignedMinValue(numBits: Width).sext(width: ExtendedWidth);
4217 Product = APIntOps::smin(A: Product, B: Max);
4218 Product = APIntOps::smax(A: Product, B: Min);
4219 }
4220 return ConstantInt::get(Context&: Ty->getContext(), V: Product.sextOrTrunc(width: Width));
4221 }
4222
4223 if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {
4224 const APInt *C0, *C1, *C2;
4225 if (!getConstIntOrUndef(Op: Operands[0], C&: C0) ||
4226 !getConstIntOrUndef(Op: Operands[1], C&: C1) ||
4227 !getConstIntOrUndef(Op: Operands[2], C&: C2))
4228 return nullptr;
4229
4230 bool IsRight = IntrinsicID == Intrinsic::fshr;
4231 if (!C2)
4232 return Operands[IsRight ? 1 : 0];
4233 if (!C0 && !C1)
4234 return UndefValue::get(T: Ty);
4235
4236 // The shift amount is interpreted as modulo the bitwidth. If the shift
4237 // amount is effectively 0, avoid UB due to oversized inverse shift below.
4238 unsigned BitWidth = C2->getBitWidth();
4239 unsigned ShAmt = C2->urem(RHS: BitWidth);
4240 if (!ShAmt)
4241 return Operands[IsRight ? 1 : 0];
4242
4243 // (C0 << ShlAmt) | (C1 >> LshrAmt)
4244 unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt;
4245 unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt;
4246 if (!C0)
4247 return ConstantInt::get(Ty, V: C1->lshr(shiftAmt: LshrAmt));
4248 if (!C1)
4249 return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt));
4250 return ConstantInt::get(Ty, V: C0->shl(shiftAmt: ShlAmt) | C1->lshr(shiftAmt: LshrAmt));
4251 }
4252
4253 if (IntrinsicID == Intrinsic::amdgcn_perm)
4254 return ConstantFoldAMDGCNPermIntrinsic(Operands, Ty);
4255
4256 return nullptr;
4257}
4258
4259static Constant *ConstantFoldScalarCall(StringRef Name,
4260 Intrinsic::ID IntrinsicID, Type *Ty,
4261 ArrayRef<Constant *> Operands,
4262 const TargetLibraryInfo *TLI = nullptr,
4263 const CallBase *Call = nullptr) {
4264 if (IntrinsicID != Intrinsic::not_intrinsic &&
4265 any_of(Range&: Operands, P: IsaPred<PoisonValue>) &&
4266 intrinsicPropagatesPoison(IID: IntrinsicID))
4267 return PoisonValue::get(T: Ty);
4268
4269 if (Operands.size() == 1)
4270 return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call);
4271
4272 if (Operands.size() == 2) {
4273 if (Constant *FoldedLibCall =
4274 ConstantFoldLibCall2(Name, Ty, Operands, TLI)) {
4275 return FoldedLibCall;
4276 }
4277 return ConstantFoldIntrinsicCall2(IntrinsicID, Ty, Operands, Call);
4278 }
4279
4280 if (Operands.size() == 3)
4281 return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call);
4282
4283 return nullptr;
4284}
4285
4286static Constant *ConstantFoldFixedVectorCall(
4287 StringRef Name, Intrinsic::ID IntrinsicID, FixedVectorType *FVTy,
4288 ArrayRef<Constant *> Operands, const DataLayout &DL,
4289 const TargetLibraryInfo *TLI, const CallBase *Call) {
4290 SmallVector<Constant *, 4> Result(FVTy->getNumElements());
4291 SmallVector<Constant *, 4> Lane(Operands.size());
4292 Type *Ty = FVTy->getElementType();
4293
4294 switch (IntrinsicID) {
4295 case Intrinsic::masked_load: {
4296 auto *SrcPtr = Operands[0];
4297 auto *Mask = Operands[1];
4298 auto *Passthru = Operands[2];
4299
4300 Constant *VecData = ConstantFoldLoadFromConstPtr(C: SrcPtr, Ty: FVTy, DL);
4301
4302 SmallVector<Constant *, 32> NewElements;
4303 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
4304 auto *MaskElt = Mask->getAggregateElement(Elt: I);
4305 if (!MaskElt)
4306 break;
4307 auto *PassthruElt = Passthru->getAggregateElement(Elt: I);
4308 auto *VecElt = VecData ? VecData->getAggregateElement(Elt: I) : nullptr;
4309 if (isa<UndefValue>(Val: MaskElt)) {
4310 if (PassthruElt)
4311 NewElements.push_back(Elt: PassthruElt);
4312 else if (VecElt)
4313 NewElements.push_back(Elt: VecElt);
4314 else
4315 return nullptr;
4316 }
4317 if (MaskElt->isNullValue()) {
4318 if (!PassthruElt)
4319 return nullptr;
4320 NewElements.push_back(Elt: PassthruElt);
4321 } else if (MaskElt->isOneValue()) {
4322 if (!VecElt)
4323 return nullptr;
4324 NewElements.push_back(Elt: VecElt);
4325 } else {
4326 return nullptr;
4327 }
4328 }
4329 if (NewElements.size() != FVTy->getNumElements())
4330 return nullptr;
4331 return ConstantVector::get(V: NewElements);
4332 }
4333 case Intrinsic::arm_mve_vctp8:
4334 case Intrinsic::arm_mve_vctp16:
4335 case Intrinsic::arm_mve_vctp32:
4336 case Intrinsic::arm_mve_vctp64: {
4337 if (auto *Op = dyn_cast<ConstantInt>(Val: Operands[0])) {
4338 unsigned Lanes = FVTy->getNumElements();
4339 uint64_t Limit = Op->getZExtValue();
4340
4341 SmallVector<Constant *, 16> NCs;
4342 for (unsigned i = 0; i < Lanes; i++) {
4343 if (i < Limit)
4344 NCs.push_back(Elt: ConstantInt::getTrue(Ty));
4345 else
4346 NCs.push_back(Elt: ConstantInt::getFalse(Ty));
4347 }
4348 return ConstantVector::get(V: NCs);
4349 }
4350 return nullptr;
4351 }
4352 case Intrinsic::get_active_lane_mask: {
4353 auto *Op0 = dyn_cast<ConstantInt>(Val: Operands[0]);
4354 auto *Op1 = dyn_cast<ConstantInt>(Val: Operands[1]);
4355 if (Op0 && Op1) {
4356 unsigned Lanes = FVTy->getNumElements();
4357 uint64_t Base = Op0->getZExtValue();
4358 uint64_t Limit = Op1->getZExtValue();
4359
4360 SmallVector<Constant *, 16> NCs;
4361 for (unsigned i = 0; i < Lanes; i++) {
4362 if (Base + i < Limit)
4363 NCs.push_back(Elt: ConstantInt::getTrue(Ty));
4364 else
4365 NCs.push_back(Elt: ConstantInt::getFalse(Ty));
4366 }
4367 return ConstantVector::get(V: NCs);
4368 }
4369 return nullptr;
4370 }
4371 case Intrinsic::vector_extract: {
4372 auto *Idx = dyn_cast<ConstantInt>(Val: Operands[1]);
4373 Constant *Vec = Operands[0];
4374 if (!Idx || !isa<FixedVectorType>(Val: Vec->getType()))
4375 return nullptr;
4376
4377 unsigned NumElements = FVTy->getNumElements();
4378 unsigned VecNumElements =
4379 cast<FixedVectorType>(Val: Vec->getType())->getNumElements();
4380 unsigned StartingIndex = Idx->getZExtValue();
4381
4382 // Extracting entire vector is nop
4383 if (NumElements == VecNumElements && StartingIndex == 0)
4384 return Vec;
4385
4386 for (unsigned I = StartingIndex, E = StartingIndex + NumElements; I < E;
4387 ++I) {
4388 Constant *Elt = Vec->getAggregateElement(Elt: I);
4389 if (!Elt)
4390 return nullptr;
4391 Result[I - StartingIndex] = Elt;
4392 }
4393
4394 return ConstantVector::get(V: Result);
4395 }
4396 case Intrinsic::vector_insert: {
4397 Constant *Vec = Operands[0];
4398 Constant *SubVec = Operands[1];
4399 auto *Idx = dyn_cast<ConstantInt>(Val: Operands[2]);
4400 if (!Idx || !isa<FixedVectorType>(Val: Vec->getType()))
4401 return nullptr;
4402
4403 unsigned SubVecNumElements =
4404 cast<FixedVectorType>(Val: SubVec->getType())->getNumElements();
4405 unsigned VecNumElements =
4406 cast<FixedVectorType>(Val: Vec->getType())->getNumElements();
4407 unsigned IdxN = Idx->getZExtValue();
4408 // Replacing entire vector with a subvec is nop
4409 if (SubVecNumElements == VecNumElements && IdxN == 0)
4410 return SubVec;
4411
4412 for (unsigned I = 0; I < VecNumElements; ++I) {
4413 Constant *Elt;
4414 if (I < IdxN + SubVecNumElements)
4415 Elt = SubVec->getAggregateElement(Elt: I - IdxN);
4416 else
4417 Elt = Vec->getAggregateElement(Elt: I);
4418 if (!Elt)
4419 return nullptr;
4420 Result[I] = Elt;
4421 }
4422 return ConstantVector::get(V: Result);
4423 }
4424 case Intrinsic::vector_interleave2:
4425 case Intrinsic::vector_interleave3:
4426 case Intrinsic::vector_interleave4:
4427 case Intrinsic::vector_interleave5:
4428 case Intrinsic::vector_interleave6:
4429 case Intrinsic::vector_interleave7:
4430 case Intrinsic::vector_interleave8: {
4431 unsigned NumElements =
4432 cast<FixedVectorType>(Val: Operands[0]->getType())->getNumElements();
4433 unsigned NumOperands = Operands.size();
4434 for (unsigned I = 0; I < NumElements; ++I) {
4435 for (unsigned J = 0; J < NumOperands; ++J) {
4436 Constant *Elt = Operands[J]->getAggregateElement(Elt: I);
4437 if (!Elt)
4438 return nullptr;
4439 Result[NumOperands * I + J] = Elt;
4440 }
4441 }
4442 return ConstantVector::get(V: Result);
4443 }
4444 case Intrinsic::wasm_dot: {
4445 unsigned NumElements =
4446 cast<FixedVectorType>(Val: Operands[0]->getType())->getNumElements();
4447
4448 assert(NumElements == 8 && Result.size() == 4 &&
4449 "wasm dot takes i16x8 and produces i32x4");
4450 assert(Ty->isIntegerTy());
4451 int32_t MulVector[8];
4452
4453 for (unsigned I = 0; I < NumElements; ++I) {
4454 ConstantInt *Elt0 =
4455 cast<ConstantInt>(Val: Operands[0]->getAggregateElement(Elt: I));
4456 ConstantInt *Elt1 =
4457 cast<ConstantInt>(Val: Operands[1]->getAggregateElement(Elt: I));
4458
4459 MulVector[I] = Elt0->getSExtValue() * Elt1->getSExtValue();
4460 }
4461 for (unsigned I = 0; I < Result.size(); I++) {
4462 int64_t IAdd = (int64_t)MulVector[I * 2] + (int64_t)MulVector[I * 2 + 1];
4463 Result[I] = ConstantInt::getSigned(Ty, V: IAdd, /*ImplicitTrunc=*/true);
4464 }
4465
4466 return ConstantVector::get(V: Result);
4467 }
4468 default:
4469 break;
4470 }
4471
4472 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
4473 // Gather a column of constants.
4474 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
4475 // Some intrinsics use a scalar type for certain arguments.
4476 if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: J, /*TTI=*/nullptr)) {
4477 Lane[J] = Operands[J];
4478 continue;
4479 }
4480
4481 Constant *Agg = Operands[J]->getAggregateElement(Elt: I);
4482 if (!Agg)
4483 return nullptr;
4484
4485 Lane[J] = Agg;
4486 }
4487
4488 // Use the regular scalar folding to simplify this column.
4489 Constant *Folded =
4490 ConstantFoldScalarCall(Name, IntrinsicID, Ty, Operands: Lane, TLI, Call);
4491 if (!Folded)
4492 return nullptr;
4493 Result[I] = Folded;
4494 }
4495
4496 return ConstantVector::get(V: Result);
4497}
4498
4499static Constant *ConstantFoldScalableVectorCall(
4500 StringRef Name, Intrinsic::ID IntrinsicID, ScalableVectorType *SVTy,
4501 ArrayRef<Constant *> Operands, const DataLayout &DL,
4502 const TargetLibraryInfo *TLI, const CallBase *Call) {
4503 switch (IntrinsicID) {
4504 case Intrinsic::aarch64_sve_convert_from_svbool: {
4505 Constant *Src = Operands[0];
4506 if (!Src->isNullValue())
4507 break;
4508
4509 return ConstantInt::getFalse(Ty: SVTy);
4510 }
4511 case Intrinsic::get_active_lane_mask: {
4512 auto *Op0 = dyn_cast<ConstantInt>(Val: Operands[0]);
4513 auto *Op1 = dyn_cast<ConstantInt>(Val: Operands[1]);
4514 if (Op0 && Op1 && Op0->getValue().uge(RHS: Op1->getValue()))
4515 return ConstantVector::getNullValue(Ty: SVTy);
4516 break;
4517 }
4518 case Intrinsic::vector_interleave2:
4519 case Intrinsic::vector_interleave3:
4520 case Intrinsic::vector_interleave4:
4521 case Intrinsic::vector_interleave5:
4522 case Intrinsic::vector_interleave6:
4523 case Intrinsic::vector_interleave7:
4524 case Intrinsic::vector_interleave8: {
4525 Constant *SplatVal = Operands[0]->getSplatValue();
4526 if (!SplatVal)
4527 return nullptr;
4528
4529 if (!llvm::all_equal(Range&: Operands))
4530 return nullptr;
4531
4532 return ConstantVector::getSplat(EC: SVTy->getElementCount(), Elt: SplatVal);
4533 }
4534 default:
4535 break;
4536 }
4537
4538 // If trivially vectorizable, try folding it via the scalar call if all
4539 // operands are splats.
4540
4541 // TODO: ConstantFoldFixedVectorCall should probably check this too?
4542 if (!isTriviallyVectorizable(ID: IntrinsicID))
4543 return nullptr;
4544
4545 SmallVector<Constant *, 4> SplatOps;
4546 for (auto [I, Op] : enumerate(First&: Operands)) {
4547 if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinsicID, ScalarOpdIdx: I, /*TTI=*/nullptr)) {
4548 SplatOps.push_back(Elt: Op);
4549 continue;
4550 }
4551 Constant *Splat = Op->getSplatValue();
4552 if (!Splat)
4553 return nullptr;
4554 SplatOps.push_back(Elt: Splat);
4555 }
4556 Constant *Folded = ConstantFoldScalarCall(
4557 Name, IntrinsicID, Ty: SVTy->getElementType(), Operands: SplatOps, TLI, Call);
4558 if (!Folded)
4559 return nullptr;
4560 return ConstantVector::getSplat(EC: SVTy->getElementCount(), Elt: Folded);
4561}
4562
4563static std::pair<Constant *, Constant *>
4564ConstantFoldScalarFrexpCall(Constant *Op, Type *IntTy) {
4565 auto *ConstFP = dyn_cast<ConstantFP>(Val: Op);
4566 if (!ConstFP)
4567 return {};
4568
4569 const APFloat &U = ConstFP->getValueAPF();
4570 int FrexpExp;
4571 APFloat FrexpMant = frexp(X: U, Exp&: FrexpExp, RM: APFloat::rmNearestTiesToEven);
4572 Constant *Result0 = ConstantFP::get(Ty: ConstFP->getType(), V: FrexpMant);
4573
4574 // The exponent is an "unspecified value" for inf/nan. We use zero to avoid
4575 // using undef.
4576 Constant *Result1 = FrexpMant.isFinite()
4577 ? ConstantInt::getSigned(Ty: IntTy, V: FrexpExp)
4578 : ConstantInt::getNullValue(Ty: IntTy);
4579 return {Result0, Result1};
4580}
4581
4582/// Handle intrinsics that return tuples, which may be tuples of vectors.
4583static Constant *
4584ConstantFoldStructCall(StringRef Name, Intrinsic::ID IntrinsicID,
4585 StructType *StTy, ArrayRef<Constant *> Operands,
4586 const DataLayout &DL, const TargetLibraryInfo *TLI,
4587 const CallBase *Call) {
4588
4589 switch (IntrinsicID) {
4590 case Intrinsic::frexp: {
4591 Type *Ty0 = StTy->getContainedType(i: 0);
4592 Type *Ty1 = StTy->getContainedType(i: 1)->getScalarType();
4593
4594 if (auto *FVTy0 = dyn_cast<FixedVectorType>(Val: Ty0)) {
4595 SmallVector<Constant *, 4> Results0(FVTy0->getNumElements());
4596 SmallVector<Constant *, 4> Results1(FVTy0->getNumElements());
4597
4598 for (unsigned I = 0, E = FVTy0->getNumElements(); I != E; ++I) {
4599 Constant *Lane = Operands[0]->getAggregateElement(Elt: I);
4600 std::tie(args&: Results0[I], args&: Results1[I]) =
4601 ConstantFoldScalarFrexpCall(Op: Lane, IntTy: Ty1);
4602 if (!Results0[I])
4603 return nullptr;
4604 }
4605
4606 return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: Results0),
4607 Vs: ConstantVector::get(V: Results1));
4608 }
4609
4610 auto [Result0, Result1] = ConstantFoldScalarFrexpCall(Op: Operands[0], IntTy: Ty1);
4611 if (!Result0)
4612 return nullptr;
4613 return ConstantStruct::get(T: StTy, Vs: Result0, Vs: Result1);
4614 }
4615 case Intrinsic::sincos: {
4616 Type *Ty = StTy->getContainedType(i: 0);
4617 Type *TyScalar = Ty->getScalarType();
4618
4619 auto ConstantFoldScalarSincosCall =
4620 [&](Constant *Op) -> std::pair<Constant *, Constant *> {
4621 Constant *SinResult =
4622 ConstantFoldScalarCall(Name, IntrinsicID: Intrinsic::sin, Ty: TyScalar, Operands: Op, TLI, Call);
4623 Constant *CosResult =
4624 ConstantFoldScalarCall(Name, IntrinsicID: Intrinsic::cos, Ty: TyScalar, Operands: Op, TLI, Call);
4625 return std::make_pair(x&: SinResult, y&: CosResult);
4626 };
4627
4628 if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty)) {
4629 SmallVector<Constant *> SinResults(FVTy->getNumElements());
4630 SmallVector<Constant *> CosResults(FVTy->getNumElements());
4631
4632 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
4633 Constant *Lane = Operands[0]->getAggregateElement(Elt: I);
4634 std::tie(args&: SinResults[I], args&: CosResults[I]) =
4635 ConstantFoldScalarSincosCall(Lane);
4636 if (!SinResults[I] || !CosResults[I])
4637 return nullptr;
4638 }
4639
4640 return ConstantStruct::get(T: StTy, Vs: ConstantVector::get(V: SinResults),
4641 Vs: ConstantVector::get(V: CosResults));
4642 }
4643
4644 if (!Ty->isFloatingPointTy())
4645 return nullptr;
4646
4647 auto [SinResult, CosResult] = ConstantFoldScalarSincosCall(Operands[0]);
4648 if (!SinResult || !CosResult)
4649 return nullptr;
4650 return ConstantStruct::get(T: StTy, Vs: SinResult, Vs: CosResult);
4651 }
4652 case Intrinsic::vector_deinterleave2:
4653 case Intrinsic::vector_deinterleave3:
4654 case Intrinsic::vector_deinterleave4:
4655 case Intrinsic::vector_deinterleave5:
4656 case Intrinsic::vector_deinterleave6:
4657 case Intrinsic::vector_deinterleave7:
4658 case Intrinsic::vector_deinterleave8: {
4659 unsigned NumResults = StTy->getNumElements();
4660 auto *Vec = Operands[0];
4661 auto *VecTy = cast<VectorType>(Val: Vec->getType());
4662
4663 ElementCount ResultEC =
4664 VecTy->getElementCount().divideCoefficientBy(RHS: NumResults);
4665
4666 if (auto *EltC = Vec->getSplatValue()) {
4667 auto *ResultVec = ConstantVector::getSplat(EC: ResultEC, Elt: EltC);
4668 SmallVector<Constant *, 8> Results(NumResults, ResultVec);
4669 return ConstantStruct::get(T: StTy, V: Results);
4670 }
4671
4672 if (!ResultEC.isFixed())
4673 return nullptr;
4674
4675 unsigned NumElements = ResultEC.getFixedValue();
4676 SmallVector<Constant *, 8> Results(NumResults);
4677 SmallVector<Constant *> Elements(NumElements);
4678 for (unsigned I = 0; I != NumResults; ++I) {
4679 for (unsigned J = 0; J != NumElements; ++J) {
4680 Constant *Elt = Vec->getAggregateElement(Elt: J * NumResults + I);
4681 if (!Elt)
4682 return nullptr;
4683 Elements[J] = Elt;
4684 }
4685 Results[I] = ConstantVector::get(V: Elements);
4686 }
4687 return ConstantStruct::get(T: StTy, V: Results);
4688 }
4689 default:
4690 // TODO: Constant folding of vector intrinsics that fall through here does
4691 // not work (e.g. overflow intrinsics)
4692 return ConstantFoldScalarCall(Name, IntrinsicID, Ty: StTy, Operands, TLI, Call);
4693 }
4694
4695 return nullptr;
4696}
4697
4698} // end anonymous namespace
4699
4700Constant *llvm::ConstantFoldIntrinsic(Intrinsic::ID ID,
4701 ArrayRef<Constant *> Ops, Type *Ty) {
4702 return ConstantFoldScalarCall(Name: "", IntrinsicID: ID, Ty, Operands: Ops);
4703}
4704
4705Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F,
4706 ArrayRef<Constant *> Operands,
4707 const TargetLibraryInfo *TLI,
4708 bool AllowNonDeterministic) {
4709 if (Call->isNoBuiltin())
4710 return nullptr;
4711 if (!F->hasName())
4712 return nullptr;
4713
4714 // If this is not an intrinsic and not recognized as a library call, bail out.
4715 Intrinsic::ID IID = F->getIntrinsicID();
4716 if (IID == Intrinsic::not_intrinsic) {
4717 if (!TLI)
4718 return nullptr;
4719 LibFunc LibF;
4720 if (!TLI->getLibFunc(FDecl: *F, F&: LibF))
4721 return nullptr;
4722 }
4723
4724 // Conservatively assume that floating-point libcalls may be
4725 // non-deterministic.
4726 Type *Ty = F->getReturnType();
4727 if (!AllowNonDeterministic && Ty->isFPOrFPVectorTy())
4728 return nullptr;
4729
4730 StringRef Name = F->getName();
4731 if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Ty))
4732 return ConstantFoldFixedVectorCall(
4733 Name, IntrinsicID: IID, FVTy, Operands, DL: F->getDataLayout(), TLI, Call);
4734
4735 if (auto *SVTy = dyn_cast<ScalableVectorType>(Val: Ty))
4736 return ConstantFoldScalableVectorCall(
4737 Name, IntrinsicID: IID, SVTy, Operands, DL: F->getDataLayout(), TLI, Call);
4738
4739 if (auto *StTy = dyn_cast<StructType>(Val: Ty))
4740 return ConstantFoldStructCall(Name, IntrinsicID: IID, StTy, Operands,
4741 DL: F->getDataLayout(), TLI, Call);
4742
4743 // TODO: If this is a library function, we already discovered that above,
4744 // so we should pass the LibFunc, not the name (and it might be better
4745 // still to separate intrinsic handling from libcalls).
4746 return ConstantFoldScalarCall(Name, IntrinsicID: IID, Ty, Operands, TLI, Call);
4747}
4748
4749bool llvm::isMathLibCallNoop(const CallBase *Call,
4750 const TargetLibraryInfo *TLI) {
4751 // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
4752 // (and to some extent ConstantFoldScalarCall).
4753 if (Call->isNoBuiltin() || Call->isStrictFP())
4754 return false;
4755 Function *F = Call->getCalledFunction();
4756 if (!F)
4757 return false;
4758
4759 LibFunc Func;
4760 if (!TLI || !TLI->getLibFunc(FDecl: *F, F&: Func))
4761 return false;
4762
4763 if (Call->arg_size() == 1) {
4764 if (ConstantFP *OpC = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0))) {
4765 const APFloat &Op = OpC->getValueAPF();
4766 switch (Func) {
4767 case LibFunc_logl:
4768 case LibFunc_log:
4769 case LibFunc_logf:
4770 case LibFunc_log2l:
4771 case LibFunc_log2:
4772 case LibFunc_log2f:
4773 case LibFunc_log10l:
4774 case LibFunc_log10:
4775 case LibFunc_log10f:
4776 return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
4777
4778 case LibFunc_ilogb:
4779 return !Op.isNaN() && !Op.isZero() && !Op.isInfinity();
4780
4781 case LibFunc_expl:
4782 case LibFunc_exp:
4783 case LibFunc_expf:
4784 // FIXME: These boundaries are slightly conservative.
4785 if (OpC->getType()->isDoubleTy())
4786 return !(Op < APFloat(-745.0) || Op > APFloat(709.0));
4787 if (OpC->getType()->isFloatTy())
4788 return !(Op < APFloat(-103.0f) || Op > APFloat(88.0f));
4789 break;
4790
4791 case LibFunc_exp2l:
4792 case LibFunc_exp2:
4793 case LibFunc_exp2f:
4794 // FIXME: These boundaries are slightly conservative.
4795 if (OpC->getType()->isDoubleTy())
4796 return !(Op < APFloat(-1074.0) || Op > APFloat(1023.0));
4797 if (OpC->getType()->isFloatTy())
4798 return !(Op < APFloat(-149.0f) || Op > APFloat(127.0f));
4799 break;
4800
4801 case LibFunc_sinl:
4802 case LibFunc_sin:
4803 case LibFunc_sinf:
4804 case LibFunc_cosl:
4805 case LibFunc_cos:
4806 case LibFunc_cosf:
4807 return !Op.isInfinity();
4808
4809 case LibFunc_tanl:
4810 case LibFunc_tan:
4811 case LibFunc_tanf: {
4812 // FIXME: Stop using the host math library.
4813 // FIXME: The computation isn't done in the right precision.
4814 Type *Ty = OpC->getType();
4815 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy())
4816 return ConstantFoldFP(NativeFP: tan, V: OpC->getValueAPF(), Ty) != nullptr;
4817 break;
4818 }
4819
4820 case LibFunc_atan:
4821 case LibFunc_atanf:
4822 case LibFunc_atanl:
4823 // Per POSIX, this MAY fail if Op is denormal. We choose not failing.
4824 return true;
4825
4826 case LibFunc_asinl:
4827 case LibFunc_asin:
4828 case LibFunc_asinf:
4829 case LibFunc_acosl:
4830 case LibFunc_acos:
4831 case LibFunc_acosf:
4832 return !(Op < APFloat::getOne(Sem: Op.getSemantics(), Negative: true) ||
4833 Op > APFloat::getOne(Sem: Op.getSemantics()));
4834
4835 case LibFunc_sinh:
4836 case LibFunc_cosh:
4837 case LibFunc_sinhf:
4838 case LibFunc_coshf:
4839 case LibFunc_sinhl:
4840 case LibFunc_coshl:
4841 // FIXME: These boundaries are slightly conservative.
4842 if (OpC->getType()->isDoubleTy())
4843 return !(Op < APFloat(-710.0) || Op > APFloat(710.0));
4844 if (OpC->getType()->isFloatTy())
4845 return !(Op < APFloat(-89.0f) || Op > APFloat(89.0f));
4846 break;
4847
4848 case LibFunc_sqrtl:
4849 case LibFunc_sqrt:
4850 case LibFunc_sqrtf:
4851 return Op.isNaN() || Op.isZero() || !Op.isNegative();
4852
4853 // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
4854 // maybe others?
4855 default:
4856 break;
4857 }
4858 }
4859 }
4860
4861 if (Call->arg_size() == 2) {
4862 ConstantFP *Op0C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 0));
4863 ConstantFP *Op1C = dyn_cast<ConstantFP>(Val: Call->getArgOperand(i: 1));
4864 if (Op0C && Op1C) {
4865 const APFloat &Op0 = Op0C->getValueAPF();
4866 const APFloat &Op1 = Op1C->getValueAPF();
4867
4868 switch (Func) {
4869 case LibFunc_powl:
4870 case LibFunc_pow:
4871 case LibFunc_powf: {
4872 // FIXME: Stop using the host math library.
4873 // FIXME: The computation isn't done in the right precision.
4874 Type *Ty = Op0C->getType();
4875 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
4876 if (Ty == Op1C->getType())
4877 return ConstantFoldBinaryFP(NativeFP: pow, V: Op0, W: Op1, Ty) != nullptr;
4878 }
4879 break;
4880 }
4881
4882 case LibFunc_fmodl:
4883 case LibFunc_fmod:
4884 case LibFunc_fmodf:
4885 case LibFunc_remainderl:
4886 case LibFunc_remainder:
4887 case LibFunc_remainderf:
4888 return Op0.isNaN() || Op1.isNaN() ||
4889 (!Op0.isInfinity() && !Op1.isZero());
4890
4891 case LibFunc_atan2:
4892 case LibFunc_atan2f:
4893 case LibFunc_atan2l:
4894 // Although IEEE-754 says atan2(+/-0.0, +/-0.0) are well-defined, and
4895 // GLIBC and MSVC do not appear to raise an error on those, we
4896 // cannot rely on that behavior. POSIX and C11 say that a domain error
4897 // may occur, so allow for that possibility.
4898 return !Op0.isZero() || !Op1.isZero();
4899
4900 case LibFunc_nextafter:
4901 case LibFunc_nextafterf:
4902 case LibFunc_nextafterl:
4903 case LibFunc_nexttoward:
4904 case LibFunc_nexttowardf:
4905 case LibFunc_nexttowardl: {
4906 return ConstantFoldNextToward(Op0, Op1, RetTy: F->getReturnType()) != nullptr;
4907 }
4908 default:
4909 break;
4910 }
4911 }
4912 }
4913
4914 return false;
4915}
4916
4917Constant *llvm::getLosslessInvCast(Constant *C, Type *InvCastTo,
4918 unsigned CastOp, const DataLayout &DL,
4919 PreservedCastFlags *Flags) {
4920 switch (CastOp) {
4921 case Instruction::BitCast:
4922 // Bitcast is always lossless.
4923 return ConstantFoldCastOperand(Opcode: Instruction::BitCast, C, DestTy: InvCastTo, DL);
4924 case Instruction::Trunc: {
4925 auto *ZExtC = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C, DestTy: InvCastTo, DL);
4926 if (Flags) {
4927 // Truncation back on ZExt value is always NUW.
4928 Flags->NUW = true;
4929 // Test positivity of C.
4930 auto *SExtC =
4931 ConstantFoldCastOperand(Opcode: Instruction::SExt, C, DestTy: InvCastTo, DL);
4932 Flags->NSW = ZExtC == SExtC;
4933 }
4934 return ZExtC;
4935 }
4936 case Instruction::SExt:
4937 case Instruction::ZExt: {
4938 auto *InvC = ConstantExpr::getTrunc(C, Ty: InvCastTo);
4939 auto *CastInvC = ConstantFoldCastOperand(Opcode: CastOp, C: InvC, DestTy: C->getType(), DL);
4940 // Must satisfy CastOp(InvC) == C.
4941 if (!CastInvC || CastInvC != C)
4942 return nullptr;
4943 if (Flags && CastOp == Instruction::ZExt) {
4944 auto *SExtInvC =
4945 ConstantFoldCastOperand(Opcode: Instruction::SExt, C: InvC, DestTy: C->getType(), DL);
4946 // Test positivity of InvC.
4947 Flags->NNeg = CastInvC == SExtInvC;
4948 }
4949 return InvC;
4950 }
4951 case Instruction::FPExt: {
4952 Constant *InvC =
4953 ConstantFoldCastOperand(Opcode: Instruction::FPTrunc, C, DestTy: InvCastTo, DL);
4954 if (InvC) {
4955 Constant *CastInvC =
4956 ConstantFoldCastOperand(Opcode: CastOp, C: InvC, DestTy: C->getType(), DL);
4957 if (CastInvC == C)
4958 return InvC;
4959 }
4960 return nullptr;
4961 }
4962 default:
4963 return nullptr;
4964 }
4965}
4966
4967Constant *llvm::getLosslessUnsignedTrunc(Constant *C, Type *DestTy,
4968 const DataLayout &DL,
4969 PreservedCastFlags *Flags) {
4970 return getLosslessInvCast(C, InvCastTo: DestTy, CastOp: Instruction::ZExt, DL, Flags);
4971}
4972
4973Constant *llvm::getLosslessSignedTrunc(Constant *C, Type *DestTy,
4974 const DataLayout &DL,
4975 PreservedCastFlags *Flags) {
4976 return getLosslessInvCast(C, InvCastTo: DestTy, CastOp: Instruction::SExt, DL, Flags);
4977}
4978
4979void TargetFolder::anchor() {}
4980