Bug Summary

File:llvm/lib/Analysis/ConstantFolding.cpp
Warning:line 680, column 39
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ConstantFolding.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Analysis -I /build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/lib/Analysis -I include -I /build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/= -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2022-01-19-134126-35450-1 -x c++ /build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/lib/Analysis/ConstantFolding.cpp

/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/lib/Analysis/ConstantFolding.cpp

1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines routines for folding instructions into constants.
10//
11// Also, to supplement the basic IR ConstantExpr simplifications,
12// this file defines some additional folding routines that can make use of
13// DataLayout information. These functions cannot go in IR due to library
14// dependency issues.
15//
16//===----------------------------------------------------------------------===//
17
18#include "llvm/Analysis/ConstantFolding.h"
19#include "llvm/ADT/APFloat.h"
20#include "llvm/ADT/APInt.h"
21#include "llvm/ADT/APSInt.h"
22#include "llvm/ADT/ArrayRef.h"
23#include "llvm/ADT/DenseMap.h"
24#include "llvm/ADT/STLExtras.h"
25#include "llvm/ADT/SmallVector.h"
26#include "llvm/ADT/StringRef.h"
27#include "llvm/Analysis/TargetFolder.h"
28#include "llvm/Analysis/TargetLibraryInfo.h"
29#include "llvm/Analysis/ValueTracking.h"
30#include "llvm/Analysis/VectorUtils.h"
31#include "llvm/Config/config.h"
32#include "llvm/IR/Constant.h"
33#include "llvm/IR/Constants.h"
34#include "llvm/IR/DataLayout.h"
35#include "llvm/IR/DerivedTypes.h"
36#include "llvm/IR/Function.h"
37#include "llvm/IR/GlobalValue.h"
38#include "llvm/IR/GlobalVariable.h"
39#include "llvm/IR/InstrTypes.h"
40#include "llvm/IR/Instruction.h"
41#include "llvm/IR/Instructions.h"
42#include "llvm/IR/IntrinsicInst.h"
43#include "llvm/IR/Intrinsics.h"
44#include "llvm/IR/IntrinsicsAArch64.h"
45#include "llvm/IR/IntrinsicsAMDGPU.h"
46#include "llvm/IR/IntrinsicsARM.h"
47#include "llvm/IR/IntrinsicsWebAssembly.h"
48#include "llvm/IR/IntrinsicsX86.h"
49#include "llvm/IR/Operator.h"
50#include "llvm/IR/Type.h"
51#include "llvm/IR/Value.h"
52#include "llvm/Support/Casting.h"
53#include "llvm/Support/ErrorHandling.h"
54#include "llvm/Support/KnownBits.h"
55#include "llvm/Support/MathExtras.h"
56#include <cassert>
57#include <cerrno>
58#include <cfenv>
59#include <cmath>
60#include <cstddef>
61#include <cstdint>
62
63using namespace llvm;
64
65namespace {
66
67//===----------------------------------------------------------------------===//
68// Constant Folding internal helper functions
69//===----------------------------------------------------------------------===//
70
71static Constant *foldConstVectorToAPInt(APInt &Result, Type *DestTy,
72 Constant *C, Type *SrcEltTy,
73 unsigned NumSrcElts,
74 const DataLayout &DL) {
75 // Now that we know that the input value is a vector of integers, just shift
76 // and insert them into our result.
77 unsigned BitShift = DL.getTypeSizeInBits(SrcEltTy);
78 for (unsigned i = 0; i != NumSrcElts; ++i) {
79 Constant *Element;
80 if (DL.isLittleEndian())
81 Element = C->getAggregateElement(NumSrcElts - i - 1);
82 else
83 Element = C->getAggregateElement(i);
84
85 if (Element && isa<UndefValue>(Element)) {
86 Result <<= BitShift;
87 continue;
88 }
89
90 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
91 if (!ElementCI)
92 return ConstantExpr::getBitCast(C, DestTy);
93
94 Result <<= BitShift;
95 Result |= ElementCI->getValue().zextOrSelf(Result.getBitWidth());
96 }
97
98 return nullptr;
99}
100
101/// Constant fold bitcast, symbolically evaluating it with DataLayout.
102/// This always returns a non-null constant, but it may be a
103/// ConstantExpr if unfoldable.
104Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
105 assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) &&(static_cast <bool> (CastInst::castIsValid(Instruction::
BitCast, C, DestTy) && "Invalid constantexpr bitcast!"
) ? void (0) : __assert_fail ("CastInst::castIsValid(Instruction::BitCast, C, DestTy) && \"Invalid constantexpr bitcast!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 106, __extension__
__PRETTY_FUNCTION__))
106 "Invalid constantexpr bitcast!")(static_cast <bool> (CastInst::castIsValid(Instruction::
BitCast, C, DestTy) && "Invalid constantexpr bitcast!"
) ? void (0) : __assert_fail ("CastInst::castIsValid(Instruction::BitCast, C, DestTy) && \"Invalid constantexpr bitcast!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 106, __extension__
__PRETTY_FUNCTION__))
;
107
108 // Catch the obvious splat cases.
109 if (Constant *Res = ConstantFoldLoadFromUniformValue(C, DestTy))
110 return Res;
111
112 if (auto *VTy = dyn_cast<VectorType>(C->getType())) {
113 // Handle a vector->scalar integer/fp cast.
114 if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {
115 unsigned NumSrcElts = cast<FixedVectorType>(VTy)->getNumElements();
116 Type *SrcEltTy = VTy->getElementType();
117
118 // If the vector is a vector of floating point, convert it to vector of int
119 // to simplify things.
120 if (SrcEltTy->isFloatingPointTy()) {
121 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
122 auto *SrcIVTy = FixedVectorType::get(
123 IntegerType::get(C->getContext(), FPWidth), NumSrcElts);
124 // Ask IR to do the conversion now that #elts line up.
125 C = ConstantExpr::getBitCast(C, SrcIVTy);
126 }
127
128 APInt Result(DL.getTypeSizeInBits(DestTy), 0);
129 if (Constant *CE = foldConstVectorToAPInt(Result, DestTy, C,
130 SrcEltTy, NumSrcElts, DL))
131 return CE;
132
133 if (isa<IntegerType>(DestTy))
134 return ConstantInt::get(DestTy, Result);
135
136 APFloat FP(DestTy->getFltSemantics(), Result);
137 return ConstantFP::get(DestTy->getContext(), FP);
138 }
139 }
140
141 // The code below only handles casts to vectors currently.
142 auto *DestVTy = dyn_cast<VectorType>(DestTy);
143 if (!DestVTy)
144 return ConstantExpr::getBitCast(C, DestTy);
145
146 // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
147 // vector so the code below can handle it uniformly.
148 if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
149 Constant *Ops = C; // don't take the address of C!
150 return FoldBitCast(ConstantVector::get(Ops), DestTy, DL);
151 }
152
153 // If this is a bitcast from constant vector -> vector, fold it.
154 if (!isa<ConstantDataVector>(C) && !isa<ConstantVector>(C))
155 return ConstantExpr::getBitCast(C, DestTy);
156
157 // If the element types match, IR can fold it.
158 unsigned NumDstElt = cast<FixedVectorType>(DestVTy)->getNumElements();
159 unsigned NumSrcElt = cast<FixedVectorType>(C->getType())->getNumElements();
160 if (NumDstElt == NumSrcElt)
161 return ConstantExpr::getBitCast(C, DestTy);
162
163 Type *SrcEltTy = cast<VectorType>(C->getType())->getElementType();
164 Type *DstEltTy = DestVTy->getElementType();
165
166 // Otherwise, we're changing the number of elements in a vector, which
167 // requires endianness information to do the right thing. For example,
168 // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
169 // folds to (little endian):
170 // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
171 // and to (big endian):
172 // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
173
174 // First thing is first. We only want to think about integer here, so if
175 // we have something in FP form, recast it as integer.
176 if (DstEltTy->isFloatingPointTy()) {
177 // Fold to an vector of integers with same size as our FP type.
178 unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
179 auto *DestIVTy = FixedVectorType::get(
180 IntegerType::get(C->getContext(), FPWidth), NumDstElt);
181 // Recursively handle this integer conversion, if possible.
182 C = FoldBitCast(C, DestIVTy, DL);
183
184 // Finally, IR can handle this now that #elts line up.
185 return ConstantExpr::getBitCast(C, DestTy);
186 }
187
188 // Okay, we know the destination is integer, if the input is FP, convert
189 // it to integer first.
190 if (SrcEltTy->isFloatingPointTy()) {
191 unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
192 auto *SrcIVTy = FixedVectorType::get(
193 IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
194 // Ask IR to do the conversion now that #elts line up.
195 C = ConstantExpr::getBitCast(C, SrcIVTy);
196 // If IR wasn't able to fold it, bail out.
197 if (!isa<ConstantVector>(C) && // FIXME: Remove ConstantVector.
198 !isa<ConstantDataVector>(C))
199 return C;
200 }
201
202 // Now we know that the input and output vectors are both integer vectors
203 // of the same size, and that their #elements is not the same. Do the
204 // conversion here, which depends on whether the input or output has
205 // more elements.
206 bool isLittleEndian = DL.isLittleEndian();
207
208 SmallVector<Constant*, 32> Result;
209 if (NumDstElt < NumSrcElt) {
210 // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
211 Constant *Zero = Constant::getNullValue(DstEltTy);
212 unsigned Ratio = NumSrcElt/NumDstElt;
213 unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
214 unsigned SrcElt = 0;
215 for (unsigned i = 0; i != NumDstElt; ++i) {
216 // Build each element of the result.
217 Constant *Elt = Zero;
218 unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
219 for (unsigned j = 0; j != Ratio; ++j) {
220 Constant *Src = C->getAggregateElement(SrcElt++);
221 if (Src && isa<UndefValue>(Src))
222 Src = Constant::getNullValue(
223 cast<VectorType>(C->getType())->getElementType());
224 else
225 Src = dyn_cast_or_null<ConstantInt>(Src);
226 if (!Src) // Reject constantexpr elements.
227 return ConstantExpr::getBitCast(C, DestTy);
228
229 // Zero extend the element to the right size.
230 Src = ConstantExpr::getZExt(Src, Elt->getType());
231
232 // Shift it to the right place, depending on endianness.
233 Src = ConstantExpr::getShl(Src,
234 ConstantInt::get(Src->getType(), ShiftAmt));
235 ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
236
237 // Mix it in.
238 Elt = ConstantExpr::getOr(Elt, Src);
239 }
240 Result.push_back(Elt);
241 }
242 return ConstantVector::get(Result);
243 }
244
245 // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
246 unsigned Ratio = NumDstElt/NumSrcElt;
247 unsigned DstBitSize = DL.getTypeSizeInBits(DstEltTy);
248
249 // Loop over each source value, expanding into multiple results.
250 for (unsigned i = 0; i != NumSrcElt; ++i) {
251 auto *Element = C->getAggregateElement(i);
252
253 if (!Element) // Reject constantexpr elements.
254 return ConstantExpr::getBitCast(C, DestTy);
255
256 if (isa<UndefValue>(Element)) {
257 // Correctly Propagate undef values.
258 Result.append(Ratio, UndefValue::get(DstEltTy));
259 continue;
260 }
261
262 auto *Src = dyn_cast<ConstantInt>(Element);
263 if (!Src)
264 return ConstantExpr::getBitCast(C, DestTy);
265
266 unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
267 for (unsigned j = 0; j != Ratio; ++j) {
268 // Shift the piece of the value into the right place, depending on
269 // endianness.
270 Constant *Elt = ConstantExpr::getLShr(Src,
271 ConstantInt::get(Src->getType(), ShiftAmt));
272 ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
273
274 // Truncate the element to an integer with the same pointer size and
275 // convert the element back to a pointer using a inttoptr.
276 if (DstEltTy->isPointerTy()) {
277 IntegerType *DstIntTy = Type::getIntNTy(C->getContext(), DstBitSize);
278 Constant *CE = ConstantExpr::getTrunc(Elt, DstIntTy);
279 Result.push_back(ConstantExpr::getIntToPtr(CE, DstEltTy));
280 continue;
281 }
282
283 // Truncate and remember this piece.
284 Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
285 }
286 }
287
288 return ConstantVector::get(Result);
289}
290
291} // end anonymous namespace
292
293/// If this constant is a constant offset from a global, return the global and
294/// the constant. Because of constantexprs, this function is recursive.
295bool llvm::IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
296 APInt &Offset, const DataLayout &DL,
297 DSOLocalEquivalent **DSOEquiv) {
298 if (DSOEquiv)
299 *DSOEquiv = nullptr;
300
301 // Trivial case, constant is the global.
302 if ((GV = dyn_cast<GlobalValue>(C))) {
303 unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());
304 Offset = APInt(BitWidth, 0);
305 return true;
306 }
307
308 if (auto *FoundDSOEquiv = dyn_cast<DSOLocalEquivalent>(C)) {
309 if (DSOEquiv)
310 *DSOEquiv = FoundDSOEquiv;
311 GV = FoundDSOEquiv->getGlobalValue();
312 unsigned BitWidth = DL.getIndexTypeSizeInBits(GV->getType());
313 Offset = APInt(BitWidth, 0);
314 return true;
315 }
316
317 // Otherwise, if this isn't a constant expr, bail out.
318 auto *CE = dyn_cast<ConstantExpr>(C);
319 if (!CE) return false;
320
321 // Look through ptr->int and ptr->ptr casts.
322 if (CE->getOpcode() == Instruction::PtrToInt ||
323 CE->getOpcode() == Instruction::BitCast)
324 return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, DL,
325 DSOEquiv);
326
327 // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
328 auto *GEP = dyn_cast<GEPOperator>(CE);
329 if (!GEP)
330 return false;
331
332 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
333 APInt TmpOffset(BitWidth, 0);
334
335 // If the base isn't a global+constant, we aren't either.
336 if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, TmpOffset, DL,
337 DSOEquiv))
338 return false;
339
340 // Otherwise, add any offset that our operands provide.
341 if (!GEP->accumulateConstantOffset(DL, TmpOffset))
342 return false;
343
344 Offset = TmpOffset;
345 return true;
346}
347
348Constant *llvm::ConstantFoldLoadThroughBitcast(Constant *C, Type *DestTy,
349 const DataLayout &DL) {
350 do {
351 Type *SrcTy = C->getType();
352 if (SrcTy == DestTy)
353 return C;
354
355 TypeSize DestSize = DL.getTypeSizeInBits(DestTy);
356 TypeSize SrcSize = DL.getTypeSizeInBits(SrcTy);
357 if (!TypeSize::isKnownGE(SrcSize, DestSize))
358 return nullptr;
359
360 // Catch the obvious splat cases (since all-zeros can coerce non-integral
361 // pointers legally).
362 if (Constant *Res = ConstantFoldLoadFromUniformValue(C, DestTy))
363 return Res;
364
365 // If the type sizes are the same and a cast is legal, just directly
366 // cast the constant.
367 // But be careful not to coerce non-integral pointers illegally.
368 if (SrcSize == DestSize &&
369 DL.isNonIntegralPointerType(SrcTy->getScalarType()) ==
370 DL.isNonIntegralPointerType(DestTy->getScalarType())) {
371 Instruction::CastOps Cast = Instruction::BitCast;
372 // If we are going from a pointer to int or vice versa, we spell the cast
373 // differently.
374 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
375 Cast = Instruction::IntToPtr;
376 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
377 Cast = Instruction::PtrToInt;
378
379 if (CastInst::castIsValid(Cast, C, DestTy))
380 return ConstantExpr::getCast(Cast, C, DestTy);
381 }
382
383 // If this isn't an aggregate type, there is nothing we can do to drill down
384 // and find a bitcastable constant.
385 if (!SrcTy->isAggregateType() && !SrcTy->isVectorTy())
386 return nullptr;
387
388 // We're simulating a load through a pointer that was bitcast to point to
389 // a different type, so we can try to walk down through the initial
390 // elements of an aggregate to see if some part of the aggregate is
391 // castable to implement the "load" semantic model.
392 if (SrcTy->isStructTy()) {
393 // Struct types might have leading zero-length elements like [0 x i32],
394 // which are certainly not what we are looking for, so skip them.
395 unsigned Elem = 0;
396 Constant *ElemC;
397 do {
398 ElemC = C->getAggregateElement(Elem++);
399 } while (ElemC && DL.getTypeSizeInBits(ElemC->getType()).isZero());
400 C = ElemC;
401 } else {
402 // For non-byte-sized vector elements, the first element is not
403 // necessarily located at the vector base address.
404 if (auto *VT = dyn_cast<VectorType>(SrcTy))
405 if (!DL.typeSizeEqualsStoreSize(VT->getElementType()))
406 return nullptr;
407
408 C = C->getAggregateElement(0u);
409 }
410 } while (C);
411
412 return nullptr;
413}
414
415namespace {
416
417/// Recursive helper to read bits out of global. C is the constant being copied
418/// out of. ByteOffset is an offset into C. CurPtr is the pointer to copy
419/// results into and BytesLeft is the number of bytes left in
420/// the CurPtr buffer. DL is the DataLayout.
421bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset, unsigned char *CurPtr,
422 unsigned BytesLeft, const DataLayout &DL) {
423 assert(ByteOffset <= DL.getTypeAllocSize(C->getType()) &&(static_cast <bool> (ByteOffset <= DL.getTypeAllocSize
(C->getType()) && "Out of range access") ? void (0
) : __assert_fail ("ByteOffset <= DL.getTypeAllocSize(C->getType()) && \"Out of range access\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 424, __extension__
__PRETTY_FUNCTION__))
424 "Out of range access")(static_cast <bool> (ByteOffset <= DL.getTypeAllocSize
(C->getType()) && "Out of range access") ? void (0
) : __assert_fail ("ByteOffset <= DL.getTypeAllocSize(C->getType()) && \"Out of range access\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 424, __extension__
__PRETTY_FUNCTION__))
;
425
426 // If this element is zero or undefined, we can just return since *CurPtr is
427 // zero initialized.
428 if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
429 return true;
430
431 if (auto *CI = dyn_cast<ConstantInt>(C)) {
432 if (CI->getBitWidth() > 64 ||
433 (CI->getBitWidth() & 7) != 0)
434 return false;
435
436 uint64_t Val = CI->getZExtValue();
437 unsigned IntBytes = unsigned(CI->getBitWidth()/8);
438
439 for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
440 int n = ByteOffset;
441 if (!DL.isLittleEndian())
442 n = IntBytes - n - 1;
443 CurPtr[i] = (unsigned char)(Val >> (n * 8));
444 ++ByteOffset;
445 }
446 return true;
447 }
448
449 if (auto *CFP = dyn_cast<ConstantFP>(C)) {
450 if (CFP->getType()->isDoubleTy()) {
451 C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), DL);
452 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
453 }
454 if (CFP->getType()->isFloatTy()){
455 C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), DL);
456 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
457 }
458 if (CFP->getType()->isHalfTy()){
459 C = FoldBitCast(C, Type::getInt16Ty(C->getContext()), DL);
460 return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, DL);
461 }
462 return false;
463 }
464
465 if (auto *CS = dyn_cast<ConstantStruct>(C)) {
466 const StructLayout *SL = DL.getStructLayout(CS->getType());
467 unsigned Index = SL->getElementContainingOffset(ByteOffset);
468 uint64_t CurEltOffset = SL->getElementOffset(Index);
469 ByteOffset -= CurEltOffset;
470
471 while (true) {
472 // If the element access is to the element itself and not to tail padding,
473 // read the bytes from the element.
474 uint64_t EltSize = DL.getTypeAllocSize(CS->getOperand(Index)->getType());
475
476 if (ByteOffset < EltSize &&
477 !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
478 BytesLeft, DL))
479 return false;
480
481 ++Index;
482
483 // Check to see if we read from the last struct element, if so we're done.
484 if (Index == CS->getType()->getNumElements())
485 return true;
486
487 // If we read all of the bytes we needed from this element we're done.
488 uint64_t NextEltOffset = SL->getElementOffset(Index);
489
490 if (BytesLeft <= NextEltOffset - CurEltOffset - ByteOffset)
491 return true;
492
493 // Move to the next element of the struct.
494 CurPtr += NextEltOffset - CurEltOffset - ByteOffset;
495 BytesLeft -= NextEltOffset - CurEltOffset - ByteOffset;
496 ByteOffset = 0;
497 CurEltOffset = NextEltOffset;
498 }
499 // not reached.
500 }
501
502 if (isa<ConstantArray>(C) || isa<ConstantVector>(C) ||
503 isa<ConstantDataSequential>(C)) {
504 uint64_t NumElts;
505 Type *EltTy;
506 if (auto *AT = dyn_cast<ArrayType>(C->getType())) {
507 NumElts = AT->getNumElements();
508 EltTy = AT->getElementType();
509 } else {
510 NumElts = cast<FixedVectorType>(C->getType())->getNumElements();
511 EltTy = cast<FixedVectorType>(C->getType())->getElementType();
512 }
513 uint64_t EltSize = DL.getTypeAllocSize(EltTy);
514 uint64_t Index = ByteOffset / EltSize;
515 uint64_t Offset = ByteOffset - Index * EltSize;
516
517 for (; Index != NumElts; ++Index) {
518 if (!ReadDataFromGlobal(C->getAggregateElement(Index), Offset, CurPtr,
519 BytesLeft, DL))
520 return false;
521
522 uint64_t BytesWritten = EltSize - Offset;
523 assert(BytesWritten <= EltSize && "Not indexing into this element?")(static_cast <bool> (BytesWritten <= EltSize &&
"Not indexing into this element?") ? void (0) : __assert_fail
("BytesWritten <= EltSize && \"Not indexing into this element?\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 523, __extension__
__PRETTY_FUNCTION__))
;
524 if (BytesWritten >= BytesLeft)
525 return true;
526
527 Offset = 0;
528 BytesLeft -= BytesWritten;
529 CurPtr += BytesWritten;
530 }
531 return true;
532 }
533
534 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
535 if (CE->getOpcode() == Instruction::IntToPtr &&
536 CE->getOperand(0)->getType() == DL.getIntPtrType(CE->getType())) {
537 return ReadDataFromGlobal(CE->getOperand(0), ByteOffset, CurPtr,
538 BytesLeft, DL);
539 }
540 }
541
542 // Otherwise, unknown initializer type.
543 return false;
544}
545
546Constant *FoldReinterpretLoadFromConst(Constant *C, Type *LoadTy,
547 int64_t Offset, const DataLayout &DL) {
548 // Bail out early. Not expect to load from scalable global variable.
549 if (isa<ScalableVectorType>(LoadTy))
550 return nullptr;
551
552 auto *IntType = dyn_cast<IntegerType>(LoadTy);
553
554 // If this isn't an integer load we can't fold it directly.
555 if (!IntType) {
556 // If this is a float/double load, we can try folding it as an int32/64 load
557 // and then bitcast the result. This can be useful for union cases. Note
558 // that address spaces don't matter here since we're not going to result in
559 // an actual new load.
560 Type *MapTy;
561 if (LoadTy->isHalfTy())
562 MapTy = Type::getInt16Ty(C->getContext());
563 else if (LoadTy->isFloatTy())
564 MapTy = Type::getInt32Ty(C->getContext());
565 else if (LoadTy->isDoubleTy())
566 MapTy = Type::getInt64Ty(C->getContext());
567 else if (LoadTy->isVectorTy()) {
568 MapTy = PointerType::getIntNTy(
569 C->getContext(), DL.getTypeSizeInBits(LoadTy).getFixedSize());
570 } else
571 return nullptr;
572
573 if (Constant *Res = FoldReinterpretLoadFromConst(C, MapTy, Offset, DL)) {
574 if (Res->isNullValue() && !LoadTy->isX86_MMXTy() &&
575 !LoadTy->isX86_AMXTy())
576 // Materializing a zero can be done trivially without a bitcast
577 return Constant::getNullValue(LoadTy);
578 Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy;
579 Res = FoldBitCast(Res, CastTy, DL);
580 if (LoadTy->isPtrOrPtrVectorTy()) {
581 // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr
582 if (Res->isNullValue() && !LoadTy->isX86_MMXTy() &&
583 !LoadTy->isX86_AMXTy())
584 return Constant::getNullValue(LoadTy);
585 if (DL.isNonIntegralPointerType(LoadTy->getScalarType()))
586 // Be careful not to replace a load of an addrspace value with an inttoptr here
587 return nullptr;
588 Res = ConstantExpr::getCast(Instruction::IntToPtr, Res, LoadTy);
589 }
590 return Res;
591 }
592 return nullptr;
593 }
594
595 unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
596 if (BytesLoaded > 32 || BytesLoaded == 0)
597 return nullptr;
598
599 int64_t InitializerSize = DL.getTypeAllocSize(C->getType()).getFixedSize();
600
601 // If we're not accessing anything in this constant, the result is undefined.
602 if (Offset <= -1 * static_cast<int64_t>(BytesLoaded))
603 return UndefValue::get(IntType);
604
605 // If we're not accessing anything in this constant, the result is undefined.
606 if (Offset >= InitializerSize)
607 return UndefValue::get(IntType);
608
609 unsigned char RawBytes[32] = {0};
610 unsigned char *CurPtr = RawBytes;
611 unsigned BytesLeft = BytesLoaded;
612
613 // If we're loading off the beginning of the global, some bytes may be valid.
614 if (Offset < 0) {
615 CurPtr += -Offset;
616 BytesLeft += Offset;
617 Offset = 0;
618 }
619
620 if (!ReadDataFromGlobal(C, Offset, CurPtr, BytesLeft, DL))
621 return nullptr;
622
623 APInt ResultVal = APInt(IntType->getBitWidth(), 0);
624 if (DL.isLittleEndian()) {
625 ResultVal = RawBytes[BytesLoaded - 1];
626 for (unsigned i = 1; i != BytesLoaded; ++i) {
627 ResultVal <<= 8;
628 ResultVal |= RawBytes[BytesLoaded - 1 - i];
629 }
630 } else {
631 ResultVal = RawBytes[0];
632 for (unsigned i = 1; i != BytesLoaded; ++i) {
633 ResultVal <<= 8;
634 ResultVal |= RawBytes[i];
635 }
636 }
637
638 return ConstantInt::get(IntType->getContext(), ResultVal);
639}
640
641/// If this Offset points exactly to the start of an aggregate element, return
642/// that element, otherwise return nullptr.
643Constant *getConstantAtOffset(Constant *Base, APInt Offset,
644 const DataLayout &DL) {
645 if (Offset.isZero())
4
Calling 'APInt::isZero'
7
Returning from 'APInt::isZero'
8
Taking true branch
646 return Base;
9
Returning pointer (loaded from 'Base'), which participates in a condition later
647
648 if (!isa<ConstantAggregate>(Base) && !isa<ConstantDataSequential>(Base))
649 return nullptr;
650
651 Type *ElemTy = Base->getType();
652 SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset);
653 if (!Offset.isZero() || !Indices[0].isZero())
654 return nullptr;
655
656 Constant *C = Base;
657 for (const APInt &Index : drop_begin(Indices)) {
658 if (Index.isNegative() || Index.getActiveBits() >= 32)
659 return nullptr;
660
661 C = C->getAggregateElement(Index.getZExtValue());
662 if (!C)
663 return nullptr;
664 }
665
666 return C;
667}
668
669} // end anonymous namespace
670
671Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
672 const APInt &Offset,
673 const DataLayout &DL) {
674 if (Constant *AtOffset = getConstantAtOffset(C, Offset, DL))
3
Calling 'getConstantAtOffset'
10
Returning from 'getConstantAtOffset'
11
Assuming 'AtOffset' is null
12
Taking false branch
675 if (Constant *Result = ConstantFoldLoadThroughBitcast(AtOffset, Ty, DL))
676 return Result;
677
678 // Explicitly check for out-of-bounds access, so we return undef even if the
679 // constant is a uniform value.
680 TypeSize Size = DL.getTypeAllocSize(C->getType());
13
Called C++ object pointer is null
681 if (!Size.isScalable() && Offset.sge(Size.getFixedSize()))
682 return UndefValue::get(Ty);
683
684 // Try an offset-independent fold of a uniform value.
685 if (Constant *Result = ConstantFoldLoadFromUniformValue(C, Ty))
686 return Result;
687
688 // Try hard to fold loads from bitcasted strange and non-type-safe things.
689 if (Offset.getMinSignedBits() <= 64)
690 if (Constant *Result =
691 FoldReinterpretLoadFromConst(C, Ty, Offset.getSExtValue(), DL))
692 return Result;
693
694 return nullptr;
695}
696
697Constant *llvm::ConstantFoldLoadFromConst(Constant *C, Type *Ty,
698 const DataLayout &DL) {
699 return ConstantFoldLoadFromConst(C, Ty, APInt(64, 0), DL);
1
Passing value via 1st parameter 'C'
2
Calling 'ConstantFoldLoadFromConst'
700}
701
702Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
703 APInt Offset,
704 const DataLayout &DL) {
705 C = cast<Constant>(C->stripAndAccumulateConstantOffsets(
706 DL, Offset, /* AllowNonInbounds */ true));
707
708 if (auto *GV = dyn_cast<GlobalVariable>(C))
709 if (GV->isConstant() && GV->hasDefinitiveInitializer())
710 if (Constant *Result = ConstantFoldLoadFromConst(GV->getInitializer(), Ty,
711 Offset, DL))
712 return Result;
713
714 // If this load comes from anywhere in a uniform constant global, the value
715 // is always the same, regardless of the loaded offset.
716 if (auto *GV = dyn_cast<GlobalVariable>(getUnderlyingObject(C))) {
717 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
718 if (Constant *Res =
719 ConstantFoldLoadFromUniformValue(GV->getInitializer(), Ty))
720 return Res;
721 }
722 }
723
724 return nullptr;
725}
726
727Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty,
728 const DataLayout &DL) {
729 APInt Offset(DL.getIndexTypeSizeInBits(C->getType()), 0);
730 return ConstantFoldLoadFromConstPtr(C, Ty, Offset, DL);
731}
732
733Constant *llvm::ConstantFoldLoadFromUniformValue(Constant *C, Type *Ty) {
734 if (isa<PoisonValue>(C))
735 return PoisonValue::get(Ty);
736 if (isa<UndefValue>(C))
737 return UndefValue::get(Ty);
738 if (C->isNullValue() && !Ty->isX86_MMXTy() && !Ty->isX86_AMXTy())
739 return Constant::getNullValue(Ty);
740 if (C->isAllOnesValue() &&
741 (Ty->isIntOrIntVectorTy() || Ty->isFPOrFPVectorTy()))
742 return Constant::getAllOnesValue(Ty);
743 return nullptr;
744}
745
746namespace {
747
748/// One of Op0/Op1 is a constant expression.
749/// Attempt to symbolically evaluate the result of a binary operator merging
750/// these together. If target data info is available, it is provided as DL,
751/// otherwise DL is null.
752Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0, Constant *Op1,
753 const DataLayout &DL) {
754 // SROA
755
756 // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
757 // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
758 // bits.
759
760 if (Opc == Instruction::And) {
761 KnownBits Known0 = computeKnownBits(Op0, DL);
762 KnownBits Known1 = computeKnownBits(Op1, DL);
763 if ((Known1.One | Known0.Zero).isAllOnes()) {
764 // All the bits of Op0 that the 'and' could be masking are already zero.
765 return Op0;
766 }
767 if ((Known0.One | Known1.Zero).isAllOnes()) {
768 // All the bits of Op1 that the 'and' could be masking are already zero.
769 return Op1;
770 }
771
772 Known0 &= Known1;
773 if (Known0.isConstant())
774 return ConstantInt::get(Op0->getType(), Known0.getConstant());
775 }
776
777 // If the constant expr is something like &A[123] - &A[4].f, fold this into a
778 // constant. This happens frequently when iterating over a global array.
779 if (Opc == Instruction::Sub) {
780 GlobalValue *GV1, *GV2;
781 APInt Offs1, Offs2;
782
783 if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, DL))
784 if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, DL) && GV1 == GV2) {
785 unsigned OpSize = DL.getTypeSizeInBits(Op0->getType());
786
787 // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
788 // PtrToInt may change the bitwidth so we have convert to the right size
789 // first.
790 return ConstantInt::get(Op0->getType(), Offs1.zextOrTrunc(OpSize) -
791 Offs2.zextOrTrunc(OpSize));
792 }
793 }
794
795 return nullptr;
796}
797
798/// If array indices are not pointer-sized integers, explicitly cast them so
799/// that they aren't implicitly casted by the getelementptr.
800Constant *CastGEPIndices(Type *SrcElemTy, ArrayRef<Constant *> Ops,
801 Type *ResultTy, Optional<unsigned> InRangeIndex,
802 const DataLayout &DL, const TargetLibraryInfo *TLI) {
803 Type *IntIdxTy = DL.getIndexType(ResultTy);
804 Type *IntIdxScalarTy = IntIdxTy->getScalarType();
805
806 bool Any = false;
807 SmallVector<Constant*, 32> NewIdxs;
808 for (unsigned i = 1, e = Ops.size(); i != e; ++i) {
809 if ((i == 1 ||
810 !isa<StructType>(GetElementPtrInst::getIndexedType(
811 SrcElemTy, Ops.slice(1, i - 1)))) &&
812 Ops[i]->getType()->getScalarType() != IntIdxScalarTy) {
813 Any = true;
814 Type *NewType = Ops[i]->getType()->isVectorTy()
815 ? IntIdxTy
816 : IntIdxScalarTy;
817 NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
818 true,
819 NewType,
820 true),
821 Ops[i], NewType));
822 } else
823 NewIdxs.push_back(Ops[i]);
824 }
825
826 if (!Any)
827 return nullptr;
828
829 Constant *C = ConstantExpr::getGetElementPtr(
830 SrcElemTy, Ops[0], NewIdxs, /*InBounds=*/false, InRangeIndex);
831 return ConstantFoldConstant(C, DL, TLI);
832}
833
834/// Strip the pointer casts, but preserve the address space information.
835Constant *StripPtrCastKeepAS(Constant *Ptr) {
836 assert(Ptr->getType()->isPointerTy() && "Not a pointer type")(static_cast <bool> (Ptr->getType()->isPointerTy(
) && "Not a pointer type") ? void (0) : __assert_fail
("Ptr->getType()->isPointerTy() && \"Not a pointer type\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 836, __extension__
__PRETTY_FUNCTION__))
;
837 auto *OldPtrTy = cast<PointerType>(Ptr->getType());
838 Ptr = cast<Constant>(Ptr->stripPointerCasts());
839 auto *NewPtrTy = cast<PointerType>(Ptr->getType());
840
841 // Preserve the address space number of the pointer.
842 if (NewPtrTy->getAddressSpace() != OldPtrTy->getAddressSpace()) {
843 Ptr = ConstantExpr::getPointerCast(
844 Ptr, PointerType::getWithSamePointeeType(NewPtrTy,
845 OldPtrTy->getAddressSpace()));
846 }
847 return Ptr;
848}
849
850/// If we can symbolically evaluate the GEP constant expression, do so.
851Constant *SymbolicallyEvaluateGEP(const GEPOperator *GEP,
852 ArrayRef<Constant *> Ops,
853 const DataLayout &DL,
854 const TargetLibraryInfo *TLI) {
855 const GEPOperator *InnermostGEP = GEP;
856 bool InBounds = GEP->isInBounds();
857
858 Type *SrcElemTy = GEP->getSourceElementType();
859 Type *ResElemTy = GEP->getResultElementType();
860 Type *ResTy = GEP->getType();
861 if (!SrcElemTy->isSized() || isa<ScalableVectorType>(SrcElemTy))
862 return nullptr;
863
864 if (Constant *C = CastGEPIndices(SrcElemTy, Ops, ResTy,
865 GEP->getInRangeIndex(), DL, TLI))
866 return C;
867
868 Constant *Ptr = Ops[0];
869 if (!Ptr->getType()->isPointerTy())
870 return nullptr;
871
872 Type *IntIdxTy = DL.getIndexType(Ptr->getType());
873
874 // If this is "gep i8* Ptr, (sub 0, V)", fold this as:
875 // "inttoptr (sub (ptrtoint Ptr), V)"
876 if (Ops.size() == 2 && ResElemTy->isIntegerTy(8)) {
877 auto *CE = dyn_cast<ConstantExpr>(Ops[1]);
878 assert((!CE || CE->getType() == IntIdxTy) &&(static_cast <bool> ((!CE || CE->getType() == IntIdxTy
) && "CastGEPIndices didn't canonicalize index types!"
) ? void (0) : __assert_fail ("(!CE || CE->getType() == IntIdxTy) && \"CastGEPIndices didn't canonicalize index types!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 879, __extension__
__PRETTY_FUNCTION__))
879 "CastGEPIndices didn't canonicalize index types!")(static_cast <bool> ((!CE || CE->getType() == IntIdxTy
) && "CastGEPIndices didn't canonicalize index types!"
) ? void (0) : __assert_fail ("(!CE || CE->getType() == IntIdxTy) && \"CastGEPIndices didn't canonicalize index types!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 879, __extension__
__PRETTY_FUNCTION__))
;
880 if (CE && CE->getOpcode() == Instruction::Sub &&
881 CE->getOperand(0)->isNullValue()) {
882 Constant *Res = ConstantExpr::getPtrToInt(Ptr, CE->getType());
883 Res = ConstantExpr::getSub(Res, CE->getOperand(1));
884 Res = ConstantExpr::getIntToPtr(Res, ResTy);
885 return ConstantFoldConstant(Res, DL, TLI);
886 }
887 }
888
889 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
890 if (!isa<ConstantInt>(Ops[i]))
891 return nullptr;
892
893 unsigned BitWidth = DL.getTypeSizeInBits(IntIdxTy);
894 APInt Offset =
895 APInt(BitWidth,
896 DL.getIndexedOffsetInType(
897 SrcElemTy,
898 makeArrayRef((Value * const *)Ops.data() + 1, Ops.size() - 1)));
899 Ptr = StripPtrCastKeepAS(Ptr);
900
901 // If this is a GEP of a GEP, fold it all into a single GEP.
902 while (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
903 InnermostGEP = GEP;
904 InBounds &= GEP->isInBounds();
905
906 SmallVector<Value *, 4> NestedOps(llvm::drop_begin(GEP->operands()));
907
908 // Do not try the incorporate the sub-GEP if some index is not a number.
909 bool AllConstantInt = true;
910 for (Value *NestedOp : NestedOps)
911 if (!isa<ConstantInt>(NestedOp)) {
912 AllConstantInt = false;
913 break;
914 }
915 if (!AllConstantInt)
916 break;
917
918 Ptr = cast<Constant>(GEP->getOperand(0));
919 SrcElemTy = GEP->getSourceElementType();
920 Offset += APInt(BitWidth, DL.getIndexedOffsetInType(SrcElemTy, NestedOps));
921 Ptr = StripPtrCastKeepAS(Ptr);
922 }
923
924 // If the base value for this address is a literal integer value, fold the
925 // getelementptr to the resulting integer value casted to the pointer type.
926 APInt BasePtr(BitWidth, 0);
927 if (auto *CE = dyn_cast<ConstantExpr>(Ptr)) {
928 if (CE->getOpcode() == Instruction::IntToPtr) {
929 if (auto *Base = dyn_cast<ConstantInt>(CE->getOperand(0)))
930 BasePtr = Base->getValue().zextOrTrunc(BitWidth);
931 }
932 }
933
934 auto *PTy = cast<PointerType>(Ptr->getType());
935 if ((Ptr->isNullValue() || BasePtr != 0) &&
936 !DL.isNonIntegralPointerType(PTy)) {
937 Constant *C = ConstantInt::get(Ptr->getContext(), Offset + BasePtr);
938 return ConstantExpr::getIntToPtr(C, ResTy);
939 }
940
941 // Otherwise form a regular getelementptr. Recompute the indices so that
942 // we eliminate over-indexing of the notional static type array bounds.
943 // This makes it easy to determine if the getelementptr is "inbounds".
944 // Also, this helps GlobalOpt do SROA on GlobalVariables.
945
946 // For GEPs of GlobalValues, use the value type even for opaque pointers.
947 // Otherwise use an i8 GEP.
948 if (auto *GV = dyn_cast<GlobalValue>(Ptr))
949 SrcElemTy = GV->getValueType();
950 else if (!PTy->isOpaque())
951 SrcElemTy = PTy->getElementType();
952 else
953 SrcElemTy = Type::getInt8Ty(Ptr->getContext());
954
955 if (!SrcElemTy->isSized())
956 return nullptr;
957
958 Type *ElemTy = SrcElemTy;
959 SmallVector<APInt> Indices = DL.getGEPIndicesForOffset(ElemTy, Offset);
960 if (Offset != 0)
961 return nullptr;
962
963 // Try to add additional zero indices to reach the desired result element
964 // type.
965 // TODO: Should we avoid extra zero indices if ResElemTy can't be reached and
966 // we'll have to insert a bitcast anyway?
967 while (ElemTy != ResElemTy) {
968 Type *NextTy = GetElementPtrInst::getTypeAtIndex(ElemTy, (uint64_t)0);
969 if (!NextTy)
970 break;
971
972 Indices.push_back(APInt::getZero(isa<StructType>(ElemTy) ? 32 : BitWidth));
973 ElemTy = NextTy;
974 }
975
976 SmallVector<Constant *, 32> NewIdxs;
977 for (const APInt &Index : Indices)
978 NewIdxs.push_back(ConstantInt::get(
979 Type::getIntNTy(Ptr->getContext(), Index.getBitWidth()), Index));
980
981 // Preserve the inrange index from the innermost GEP if possible. We must
982 // have calculated the same indices up to and including the inrange index.
983 Optional<unsigned> InRangeIndex;
984 if (Optional<unsigned> LastIRIndex = InnermostGEP->getInRangeIndex())
985 if (SrcElemTy == InnermostGEP->getSourceElementType() &&
986 NewIdxs.size() > *LastIRIndex) {
987 InRangeIndex = LastIRIndex;
988 for (unsigned I = 0; I <= *LastIRIndex; ++I)
989 if (NewIdxs[I] != InnermostGEP->getOperand(I + 1))
990 return nullptr;
991 }
992
993 // Create a GEP.
994 Constant *C = ConstantExpr::getGetElementPtr(SrcElemTy, Ptr, NewIdxs,
995 InBounds, InRangeIndex);
996 assert((static_cast <bool> (cast<PointerType>(C->getType
())->isOpaqueOrPointeeTypeMatches(ElemTy) && "Computed GetElementPtr has unexpected type!"
) ? void (0) : __assert_fail ("cast<PointerType>(C->getType())->isOpaqueOrPointeeTypeMatches(ElemTy) && \"Computed GetElementPtr has unexpected type!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 998, __extension__
__PRETTY_FUNCTION__))
997 cast<PointerType>(C->getType())->isOpaqueOrPointeeTypeMatches(ElemTy) &&(static_cast <bool> (cast<PointerType>(C->getType
())->isOpaqueOrPointeeTypeMatches(ElemTy) && "Computed GetElementPtr has unexpected type!"
) ? void (0) : __assert_fail ("cast<PointerType>(C->getType())->isOpaqueOrPointeeTypeMatches(ElemTy) && \"Computed GetElementPtr has unexpected type!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 998, __extension__
__PRETTY_FUNCTION__))
998 "Computed GetElementPtr has unexpected type!")(static_cast <bool> (cast<PointerType>(C->getType
())->isOpaqueOrPointeeTypeMatches(ElemTy) && "Computed GetElementPtr has unexpected type!"
) ? void (0) : __assert_fail ("cast<PointerType>(C->getType())->isOpaqueOrPointeeTypeMatches(ElemTy) && \"Computed GetElementPtr has unexpected type!\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 998, __extension__
__PRETTY_FUNCTION__))
;
999
1000 // If we ended up indexing a member with a type that doesn't match
1001 // the type of what the original indices indexed, add a cast.
1002 if (C->getType() != ResTy)
1003 C = FoldBitCast(C, ResTy, DL);
1004
1005 return C;
1006}
1007
1008/// Attempt to constant fold an instruction with the
1009/// specified opcode and operands. If successful, the constant result is
1010/// returned, if not, null is returned. Note that this function can fail when
1011/// attempting to fold instructions like loads and stores, which have no
1012/// constant expression form.
1013Constant *ConstantFoldInstOperandsImpl(const Value *InstOrCE, unsigned Opcode,
1014 ArrayRef<Constant *> Ops,
1015 const DataLayout &DL,
1016 const TargetLibraryInfo *TLI) {
1017 Type *DestTy = InstOrCE->getType();
1018
1019 if (Instruction::isUnaryOp(Opcode))
1020 return ConstantFoldUnaryOpOperand(Opcode, Ops[0], DL);
1021
1022 if (Instruction::isBinaryOp(Opcode))
1023 return ConstantFoldBinaryOpOperands(Opcode, Ops[0], Ops[1], DL);
1024
1025 if (Instruction::isCast(Opcode))
1026 return ConstantFoldCastOperand(Opcode, Ops[0], DestTy, DL);
1027
1028 if (auto *GEP = dyn_cast<GEPOperator>(InstOrCE)) {
1029 if (Constant *C = SymbolicallyEvaluateGEP(GEP, Ops, DL, TLI))
1030 return C;
1031
1032 return ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), Ops[0],
1033 Ops.slice(1), GEP->isInBounds(),
1034 GEP->getInRangeIndex());
1035 }
1036
1037 if (auto *CE = dyn_cast<ConstantExpr>(InstOrCE))
1038 return CE->getWithOperands(Ops);
1039
1040 switch (Opcode) {
1041 default: return nullptr;
1042 case Instruction::ICmp:
1043 case Instruction::FCmp: llvm_unreachable("Invalid for compares")::llvm::llvm_unreachable_internal("Invalid for compares", "llvm/lib/Analysis/ConstantFolding.cpp"
, 1043)
;
1044 case Instruction::Freeze:
1045 return isGuaranteedNotToBeUndefOrPoison(Ops[0]) ? Ops[0] : nullptr;
1046 case Instruction::Call:
1047 if (auto *F = dyn_cast<Function>(Ops.back())) {
1048 const auto *Call = cast<CallBase>(InstOrCE);
1049 if (canConstantFoldCallTo(Call, F))
1050 return ConstantFoldCall(Call, F, Ops.slice(0, Ops.size() - 1), TLI);
1051 }
1052 return nullptr;
1053 case Instruction::Select:
1054 return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
1055 case Instruction::ExtractElement:
1056 return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
1057 case Instruction::ExtractValue:
1058 return ConstantExpr::getExtractValue(
1059 Ops[0], cast<ExtractValueInst>(InstOrCE)->getIndices());
1060 case Instruction::InsertElement:
1061 return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
1062 case Instruction::ShuffleVector:
1063 return ConstantExpr::getShuffleVector(
1064 Ops[0], Ops[1], cast<ShuffleVectorInst>(InstOrCE)->getShuffleMask());
1065 }
1066}
1067
1068} // end anonymous namespace
1069
1070//===----------------------------------------------------------------------===//
1071// Constant Folding public APIs
1072//===----------------------------------------------------------------------===//
1073
1074namespace {
1075
1076Constant *
1077ConstantFoldConstantImpl(const Constant *C, const DataLayout &DL,
1078 const TargetLibraryInfo *TLI,
1079 SmallDenseMap<Constant *, Constant *> &FoldedOps) {
1080 if (!isa<ConstantVector>(C) && !isa<ConstantExpr>(C))
1081 return const_cast<Constant *>(C);
1082
1083 SmallVector<Constant *, 8> Ops;
1084 for (const Use &OldU : C->operands()) {
1085 Constant *OldC = cast<Constant>(&OldU);
1086 Constant *NewC = OldC;
1087 // Recursively fold the ConstantExpr's operands. If we have already folded
1088 // a ConstantExpr, we don't have to process it again.
1089 if (isa<ConstantVector>(OldC) || isa<ConstantExpr>(OldC)) {
1090 auto It = FoldedOps.find(OldC);
1091 if (It == FoldedOps.end()) {
1092 NewC = ConstantFoldConstantImpl(OldC, DL, TLI, FoldedOps);
1093 FoldedOps.insert({OldC, NewC});
1094 } else {
1095 NewC = It->second;
1096 }
1097 }
1098 Ops.push_back(NewC);
1099 }
1100
1101 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1102 if (CE->isCompare())
1103 return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
1104 DL, TLI);
1105
1106 return ConstantFoldInstOperandsImpl(CE, CE->getOpcode(), Ops, DL, TLI);
1107 }
1108
1109 assert(isa<ConstantVector>(C))(static_cast <bool> (isa<ConstantVector>(C)) ? void
(0) : __assert_fail ("isa<ConstantVector>(C)", "llvm/lib/Analysis/ConstantFolding.cpp"
, 1109, __extension__ __PRETTY_FUNCTION__))
;
1110 return ConstantVector::get(Ops);
1111}
1112
1113} // end anonymous namespace
1114
1115Constant *llvm::ConstantFoldInstruction(Instruction *I, const DataLayout &DL,
1116 const TargetLibraryInfo *TLI) {
1117 // Handle PHI nodes quickly here...
1118 if (auto *PN = dyn_cast<PHINode>(I)) {
1119 Constant *CommonValue = nullptr;
1120
1121 SmallDenseMap<Constant *, Constant *> FoldedOps;
1122 for (Value *Incoming : PN->incoming_values()) {
1123 // If the incoming value is undef then skip it. Note that while we could
1124 // skip the value if it is equal to the phi node itself we choose not to
1125 // because that would break the rule that constant folding only applies if
1126 // all operands are constants.
1127 if (isa<UndefValue>(Incoming))
1128 continue;
1129 // If the incoming value is not a constant, then give up.
1130 auto *C = dyn_cast<Constant>(Incoming);
1131 if (!C)
1132 return nullptr;
1133 // Fold the PHI's operands.
1134 C = ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1135 // If the incoming value is a different constant to
1136 // the one we saw previously, then give up.
1137 if (CommonValue && C != CommonValue)
1138 return nullptr;
1139 CommonValue = C;
1140 }
1141
1142 // If we reach here, all incoming values are the same constant or undef.
1143 return CommonValue ? CommonValue : UndefValue::get(PN->getType());
1144 }
1145
1146 // Scan the operand list, checking to see if they are all constants, if so,
1147 // hand off to ConstantFoldInstOperandsImpl.
1148 if (!all_of(I->operands(), [](Use &U) { return isa<Constant>(U); }))
1149 return nullptr;
1150
1151 SmallDenseMap<Constant *, Constant *> FoldedOps;
1152 SmallVector<Constant *, 8> Ops;
1153 for (const Use &OpU : I->operands()) {
1154 auto *Op = cast<Constant>(&OpU);
1155 // Fold the Instruction's operands.
1156 Op = ConstantFoldConstantImpl(Op, DL, TLI, FoldedOps);
1157 Ops.push_back(Op);
1158 }
1159
1160 if (const auto *CI = dyn_cast<CmpInst>(I))
1161 return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
1162 DL, TLI);
1163
1164 if (const auto *LI = dyn_cast<LoadInst>(I)) {
1165 if (LI->isVolatile())
1166 return nullptr;
1167 return ConstantFoldLoadFromConstPtr(Ops[0], LI->getType(), DL);
1168 }
1169
1170 if (auto *IVI = dyn_cast<InsertValueInst>(I))
1171 return ConstantExpr::getInsertValue(Ops[0], Ops[1], IVI->getIndices());
1172
1173 if (auto *EVI = dyn_cast<ExtractValueInst>(I))
1174 return ConstantExpr::getExtractValue(Ops[0], EVI->getIndices());
1175
1176 return ConstantFoldInstOperands(I, Ops, DL, TLI);
1177}
1178
1179Constant *llvm::ConstantFoldConstant(const Constant *C, const DataLayout &DL,
1180 const TargetLibraryInfo *TLI) {
1181 SmallDenseMap<Constant *, Constant *> FoldedOps;
1182 return ConstantFoldConstantImpl(C, DL, TLI, FoldedOps);
1183}
1184
1185Constant *llvm::ConstantFoldInstOperands(Instruction *I,
1186 ArrayRef<Constant *> Ops,
1187 const DataLayout &DL,
1188 const TargetLibraryInfo *TLI) {
1189 return ConstantFoldInstOperandsImpl(I, I->getOpcode(), Ops, DL, TLI);
1190}
1191
1192Constant *llvm::ConstantFoldCompareInstOperands(unsigned IntPredicate,
1193 Constant *Ops0, Constant *Ops1,
1194 const DataLayout &DL,
1195 const TargetLibraryInfo *TLI) {
1196 CmpInst::Predicate Predicate = (CmpInst::Predicate)IntPredicate;
1197 // fold: icmp (inttoptr x), null -> icmp x, 0
1198 // fold: icmp null, (inttoptr x) -> icmp 0, x
1199 // fold: icmp (ptrtoint x), 0 -> icmp x, null
1200 // fold: icmp 0, (ptrtoint x) -> icmp null, x
1201 // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
1202 // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
1203 //
1204 // FIXME: The following comment is out of data and the DataLayout is here now.
1205 // ConstantExpr::getCompare cannot do this, because it doesn't have DL
1206 // around to know if bit truncation is happening.
1207 if (auto *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
1208 if (Ops1->isNullValue()) {
1209 if (CE0->getOpcode() == Instruction::IntToPtr) {
1210 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1211 // Convert the integer value to the right size to ensure we get the
1212 // proper extension or truncation.
1213 Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1214 IntPtrTy, false);
1215 Constant *Null = Constant::getNullValue(C->getType());
1216 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1217 }
1218
1219 // Only do this transformation if the int is intptrty in size, otherwise
1220 // there is a truncation or extension that we aren't modeling.
1221 if (CE0->getOpcode() == Instruction::PtrToInt) {
1222 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1223 if (CE0->getType() == IntPtrTy) {
1224 Constant *C = CE0->getOperand(0);
1225 Constant *Null = Constant::getNullValue(C->getType());
1226 return ConstantFoldCompareInstOperands(Predicate, C, Null, DL, TLI);
1227 }
1228 }
1229 }
1230
1231 if (auto *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
1232 if (CE0->getOpcode() == CE1->getOpcode()) {
1233 if (CE0->getOpcode() == Instruction::IntToPtr) {
1234 Type *IntPtrTy = DL.getIntPtrType(CE0->getType());
1235
1236 // Convert the integer value to the right size to ensure we get the
1237 // proper extension or truncation.
1238 Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
1239 IntPtrTy, false);
1240 Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
1241 IntPtrTy, false);
1242 return ConstantFoldCompareInstOperands(Predicate, C0, C1, DL, TLI);
1243 }
1244
1245 // Only do this transformation if the int is intptrty in size, otherwise
1246 // there is a truncation or extension that we aren't modeling.
1247 if (CE0->getOpcode() == Instruction::PtrToInt) {
1248 Type *IntPtrTy = DL.getIntPtrType(CE0->getOperand(0)->getType());
1249 if (CE0->getType() == IntPtrTy &&
1250 CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()) {
1251 return ConstantFoldCompareInstOperands(
1252 Predicate, CE0->getOperand(0), CE1->getOperand(0), DL, TLI);
1253 }
1254 }
1255 }
1256 }
1257
1258 // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
1259 // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
1260 if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
1261 CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
1262 Constant *LHS = ConstantFoldCompareInstOperands(
1263 Predicate, CE0->getOperand(0), Ops1, DL, TLI);
1264 Constant *RHS = ConstantFoldCompareInstOperands(
1265 Predicate, CE0->getOperand(1), Ops1, DL, TLI);
1266 unsigned OpC =
1267 Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1268 return ConstantFoldBinaryOpOperands(OpC, LHS, RHS, DL);
1269 }
1270
1271 // Convert pointer comparison (base+offset1) pred (base+offset2) into
1272 // offset1 pred offset2, for the case where the offset is inbounds. This
1273 // only works for equality and unsigned comparison, as inbounds permits
1274 // crossing the sign boundary. However, the offset comparison itself is
1275 // signed.
1276 if (Ops0->getType()->isPointerTy() && !ICmpInst::isSigned(Predicate)) {
1277 unsigned IndexWidth = DL.getIndexTypeSizeInBits(Ops0->getType());
1278 APInt Offset0(IndexWidth, 0);
1279 Value *Stripped0 =
1280 Ops0->stripAndAccumulateInBoundsConstantOffsets(DL, Offset0);
1281 APInt Offset1(IndexWidth, 0);
1282 Value *Stripped1 =
1283 Ops1->stripAndAccumulateInBoundsConstantOffsets(DL, Offset1);
1284 if (Stripped0 == Stripped1)
1285 return ConstantExpr::getCompare(
1286 ICmpInst::getSignedPredicate(Predicate),
1287 ConstantInt::get(CE0->getContext(), Offset0),
1288 ConstantInt::get(CE0->getContext(), Offset1));
1289 }
1290 } else if (isa<ConstantExpr>(Ops1)) {
1291 // If RHS is a constant expression, but the left side isn't, swap the
1292 // operands and try again.
1293 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1294 return ConstantFoldCompareInstOperands(Predicate, Ops1, Ops0, DL, TLI);
1295 }
1296
1297 return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
1298}
1299
1300Constant *llvm::ConstantFoldUnaryOpOperand(unsigned Opcode, Constant *Op,
1301 const DataLayout &DL) {
1302 assert(Instruction::isUnaryOp(Opcode))(static_cast <bool> (Instruction::isUnaryOp(Opcode)) ? void
(0) : __assert_fail ("Instruction::isUnaryOp(Opcode)", "llvm/lib/Analysis/ConstantFolding.cpp"
, 1302, __extension__ __PRETTY_FUNCTION__))
;
1303
1304 return ConstantExpr::get(Opcode, Op);
1305}
1306
1307Constant *llvm::ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS,
1308 Constant *RHS,
1309 const DataLayout &DL) {
1310 assert(Instruction::isBinaryOp(Opcode))(static_cast <bool> (Instruction::isBinaryOp(Opcode)) ?
void (0) : __assert_fail ("Instruction::isBinaryOp(Opcode)",
"llvm/lib/Analysis/ConstantFolding.cpp", 1310, __extension__
__PRETTY_FUNCTION__))
;
1311 if (isa<ConstantExpr>(LHS) || isa<ConstantExpr>(RHS))
1312 if (Constant *C = SymbolicallyEvaluateBinop(Opcode, LHS, RHS, DL))
1313 return C;
1314
1315 return ConstantExpr::get(Opcode, LHS, RHS);
1316}
1317
1318Constant *llvm::ConstantFoldCastOperand(unsigned Opcode, Constant *C,
1319 Type *DestTy, const DataLayout &DL) {
1320 assert(Instruction::isCast(Opcode))(static_cast <bool> (Instruction::isCast(Opcode)) ? void
(0) : __assert_fail ("Instruction::isCast(Opcode)", "llvm/lib/Analysis/ConstantFolding.cpp"
, 1320, __extension__ __PRETTY_FUNCTION__))
;
1321 switch (Opcode) {
1322 default:
1323 llvm_unreachable("Missing case")::llvm::llvm_unreachable_internal("Missing case", "llvm/lib/Analysis/ConstantFolding.cpp"
, 1323)
;
1324 case Instruction::PtrToInt:
1325 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1326 Constant *FoldedValue = nullptr;
1327 // If the input is a inttoptr, eliminate the pair. This requires knowing
1328 // the width of a pointer, so it can't be done in ConstantExpr::getCast.
1329 if (CE->getOpcode() == Instruction::IntToPtr) {
1330 // zext/trunc the inttoptr to pointer size.
1331 FoldedValue = ConstantExpr::getIntegerCast(
1332 CE->getOperand(0), DL.getIntPtrType(CE->getType()),
1333 /*IsSigned=*/false);
1334 } else if (auto *GEP = dyn_cast<GEPOperator>(CE)) {
1335 // If we have GEP, we can perform the following folds:
1336 // (ptrtoint (gep null, x)) -> x
1337 // (ptrtoint (gep (gep null, x), y) -> x + y, etc.
1338 unsigned BitWidth = DL.getIndexTypeSizeInBits(GEP->getType());
1339 APInt BaseOffset(BitWidth, 0);
1340 auto *Base = cast<Constant>(GEP->stripAndAccumulateConstantOffsets(
1341 DL, BaseOffset, /*AllowNonInbounds=*/true));
1342 if (Base->isNullValue()) {
1343 FoldedValue = ConstantInt::get(CE->getContext(), BaseOffset);
1344 }
1345 }
1346 if (FoldedValue) {
1347 // Do a zext or trunc to get to the ptrtoint dest size.
1348 return ConstantExpr::getIntegerCast(FoldedValue, DestTy,
1349 /*IsSigned=*/false);
1350 }
1351 }
1352 return ConstantExpr::getCast(Opcode, C, DestTy);
1353 case Instruction::IntToPtr:
1354 // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
1355 // the int size is >= the ptr size and the address spaces are the same.
1356 // This requires knowing the width of a pointer, so it can't be done in
1357 // ConstantExpr::getCast.
1358 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
1359 if (CE->getOpcode() == Instruction::PtrToInt) {
1360 Constant *SrcPtr = CE->getOperand(0);
1361 unsigned SrcPtrSize = DL.getPointerTypeSizeInBits(SrcPtr->getType());
1362 unsigned MidIntSize = CE->getType()->getScalarSizeInBits();
1363
1364 if (MidIntSize >= SrcPtrSize) {
1365 unsigned SrcAS = SrcPtr->getType()->getPointerAddressSpace();
1366 if (SrcAS == DestTy->getPointerAddressSpace())
1367 return FoldBitCast(CE->getOperand(0), DestTy, DL);
1368 }
1369 }
1370 }
1371
1372 return ConstantExpr::getCast(Opcode, C, DestTy);
1373 case Instruction::Trunc:
1374 case Instruction::ZExt:
1375 case Instruction::SExt:
1376 case Instruction::FPTrunc:
1377 case Instruction::FPExt:
1378 case Instruction::UIToFP:
1379 case Instruction::SIToFP:
1380 case Instruction::FPToUI:
1381 case Instruction::FPToSI:
1382 case Instruction::AddrSpaceCast:
1383 return ConstantExpr::getCast(Opcode, C, DestTy);
1384 case Instruction::BitCast:
1385 return FoldBitCast(C, DestTy, DL);
1386 }
1387}
1388
1389//===----------------------------------------------------------------------===//
1390// Constant Folding for Calls
1391//
1392
1393bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
1394 if (Call->isNoBuiltin())
1395 return false;
1396 switch (F->getIntrinsicID()) {
1397 // Operations that do not operate floating-point numbers and do not depend on
1398 // FP environment can be folded even in strictfp functions.
1399 case Intrinsic::bswap:
1400 case Intrinsic::ctpop:
1401 case Intrinsic::ctlz:
1402 case Intrinsic::cttz:
1403 case Intrinsic::fshl:
1404 case Intrinsic::fshr:
1405 case Intrinsic::launder_invariant_group:
1406 case Intrinsic::strip_invariant_group:
1407 case Intrinsic::masked_load:
1408 case Intrinsic::get_active_lane_mask:
1409 case Intrinsic::abs:
1410 case Intrinsic::smax:
1411 case Intrinsic::smin:
1412 case Intrinsic::umax:
1413 case Intrinsic::umin:
1414 case Intrinsic::sadd_with_overflow:
1415 case Intrinsic::uadd_with_overflow:
1416 case Intrinsic::ssub_with_overflow:
1417 case Intrinsic::usub_with_overflow:
1418 case Intrinsic::smul_with_overflow:
1419 case Intrinsic::umul_with_overflow:
1420 case Intrinsic::sadd_sat:
1421 case Intrinsic::uadd_sat:
1422 case Intrinsic::ssub_sat:
1423 case Intrinsic::usub_sat:
1424 case Intrinsic::smul_fix:
1425 case Intrinsic::smul_fix_sat:
1426 case Intrinsic::bitreverse:
1427 case Intrinsic::is_constant:
1428 case Intrinsic::vector_reduce_add:
1429 case Intrinsic::vector_reduce_mul:
1430 case Intrinsic::vector_reduce_and:
1431 case Intrinsic::vector_reduce_or:
1432 case Intrinsic::vector_reduce_xor:
1433 case Intrinsic::vector_reduce_smin:
1434 case Intrinsic::vector_reduce_smax:
1435 case Intrinsic::vector_reduce_umin:
1436 case Intrinsic::vector_reduce_umax:
1437 // Target intrinsics
1438 case Intrinsic::amdgcn_perm:
1439 case Intrinsic::arm_mve_vctp8:
1440 case Intrinsic::arm_mve_vctp16:
1441 case Intrinsic::arm_mve_vctp32:
1442 case Intrinsic::arm_mve_vctp64:
1443 case Intrinsic::aarch64_sve_convert_from_svbool:
1444 // WebAssembly float semantics are always known
1445 case Intrinsic::wasm_trunc_signed:
1446 case Intrinsic::wasm_trunc_unsigned:
1447 return true;
1448
1449 // Floating point operations cannot be folded in strictfp functions in
1450 // general case. They can be folded if FP environment is known to compiler.
1451 case Intrinsic::minnum:
1452 case Intrinsic::maxnum:
1453 case Intrinsic::minimum:
1454 case Intrinsic::maximum:
1455 case Intrinsic::log:
1456 case Intrinsic::log2:
1457 case Intrinsic::log10:
1458 case Intrinsic::exp:
1459 case Intrinsic::exp2:
1460 case Intrinsic::sqrt:
1461 case Intrinsic::sin:
1462 case Intrinsic::cos:
1463 case Intrinsic::pow:
1464 case Intrinsic::powi:
1465 case Intrinsic::fma:
1466 case Intrinsic::fmuladd:
1467 case Intrinsic::fptoui_sat:
1468 case Intrinsic::fptosi_sat:
1469 case Intrinsic::convert_from_fp16:
1470 case Intrinsic::convert_to_fp16:
1471 case Intrinsic::amdgcn_cos:
1472 case Intrinsic::amdgcn_cubeid:
1473 case Intrinsic::amdgcn_cubema:
1474 case Intrinsic::amdgcn_cubesc:
1475 case Intrinsic::amdgcn_cubetc:
1476 case Intrinsic::amdgcn_fmul_legacy:
1477 case Intrinsic::amdgcn_fma_legacy:
1478 case Intrinsic::amdgcn_fract:
1479 case Intrinsic::amdgcn_ldexp:
1480 case Intrinsic::amdgcn_sin:
1481 // The intrinsics below depend on rounding mode in MXCSR.
1482 case Intrinsic::x86_sse_cvtss2si:
1483 case Intrinsic::x86_sse_cvtss2si64:
1484 case Intrinsic::x86_sse_cvttss2si:
1485 case Intrinsic::x86_sse_cvttss2si64:
1486 case Intrinsic::x86_sse2_cvtsd2si:
1487 case Intrinsic::x86_sse2_cvtsd2si64:
1488 case Intrinsic::x86_sse2_cvttsd2si:
1489 case Intrinsic::x86_sse2_cvttsd2si64:
1490 case Intrinsic::x86_avx512_vcvtss2si32:
1491 case Intrinsic::x86_avx512_vcvtss2si64:
1492 case Intrinsic::x86_avx512_cvttss2si:
1493 case Intrinsic::x86_avx512_cvttss2si64:
1494 case Intrinsic::x86_avx512_vcvtsd2si32:
1495 case Intrinsic::x86_avx512_vcvtsd2si64:
1496 case Intrinsic::x86_avx512_cvttsd2si:
1497 case Intrinsic::x86_avx512_cvttsd2si64:
1498 case Intrinsic::x86_avx512_vcvtss2usi32:
1499 case Intrinsic::x86_avx512_vcvtss2usi64:
1500 case Intrinsic::x86_avx512_cvttss2usi:
1501 case Intrinsic::x86_avx512_cvttss2usi64:
1502 case Intrinsic::x86_avx512_vcvtsd2usi32:
1503 case Intrinsic::x86_avx512_vcvtsd2usi64:
1504 case Intrinsic::x86_avx512_cvttsd2usi:
1505 case Intrinsic::x86_avx512_cvttsd2usi64:
1506 return !Call->isStrictFP();
1507
1508 // Sign operations are actually bitwise operations, they do not raise
1509 // exceptions even for SNANs.
1510 case Intrinsic::fabs:
1511 case Intrinsic::copysign:
1512 // Non-constrained variants of rounding operations means default FP
1513 // environment, they can be folded in any case.
1514 case Intrinsic::ceil:
1515 case Intrinsic::floor:
1516 case Intrinsic::round:
1517 case Intrinsic::roundeven:
1518 case Intrinsic::trunc:
1519 case Intrinsic::nearbyint:
1520 case Intrinsic::rint:
1521 // Constrained intrinsics can be folded if FP environment is known
1522 // to compiler.
1523 case Intrinsic::experimental_constrained_fma:
1524 case Intrinsic::experimental_constrained_fmuladd:
1525 case Intrinsic::experimental_constrained_fadd:
1526 case Intrinsic::experimental_constrained_fsub:
1527 case Intrinsic::experimental_constrained_fmul:
1528 case Intrinsic::experimental_constrained_fdiv:
1529 case Intrinsic::experimental_constrained_frem:
1530 case Intrinsic::experimental_constrained_ceil:
1531 case Intrinsic::experimental_constrained_floor:
1532 case Intrinsic::experimental_constrained_round:
1533 case Intrinsic::experimental_constrained_roundeven:
1534 case Intrinsic::experimental_constrained_trunc:
1535 case Intrinsic::experimental_constrained_nearbyint:
1536 case Intrinsic::experimental_constrained_rint:
1537 return true;
1538 default:
1539 return false;
1540 case Intrinsic::not_intrinsic: break;
1541 }
1542
1543 if (!F->hasName() || Call->isStrictFP())
1544 return false;
1545
1546 // In these cases, the check of the length is required. We don't want to
1547 // return true for a name like "cos\0blah" which strcmp would return equal to
1548 // "cos", but has length 8.
1549 StringRef Name = F->getName();
1550 switch (Name[0]) {
1551 default:
1552 return false;
1553 case 'a':
1554 return Name == "acos" || Name == "acosf" ||
1555 Name == "asin" || Name == "asinf" ||
1556 Name == "atan" || Name == "atanf" ||
1557 Name == "atan2" || Name == "atan2f";
1558 case 'c':
1559 return Name == "ceil" || Name == "ceilf" ||
1560 Name == "cos" || Name == "cosf" ||
1561 Name == "cosh" || Name == "coshf";
1562 case 'e':
1563 return Name == "exp" || Name == "expf" ||
1564 Name == "exp2" || Name == "exp2f";
1565 case 'f':
1566 return Name == "fabs" || Name == "fabsf" ||
1567 Name == "floor" || Name == "floorf" ||
1568 Name == "fmod" || Name == "fmodf";
1569 case 'l':
1570 return Name == "log" || Name == "logf" ||
1571 Name == "log2" || Name == "log2f" ||
1572 Name == "log10" || Name == "log10f";
1573 case 'n':
1574 return Name == "nearbyint" || Name == "nearbyintf";
1575 case 'p':
1576 return Name == "pow" || Name == "powf";
1577 case 'r':
1578 return Name == "remainder" || Name == "remainderf" ||
1579 Name == "rint" || Name == "rintf" ||
1580 Name == "round" || Name == "roundf";
1581 case 's':
1582 return Name == "sin" || Name == "sinf" ||
1583 Name == "sinh" || Name == "sinhf" ||
1584 Name == "sqrt" || Name == "sqrtf";
1585 case 't':
1586 return Name == "tan" || Name == "tanf" ||
1587 Name == "tanh" || Name == "tanhf" ||
1588 Name == "trunc" || Name == "truncf";
1589 case '_':
1590 // Check for various function names that get used for the math functions
1591 // when the header files are preprocessed with the macro
1592 // __FINITE_MATH_ONLY__ enabled.
1593 // The '12' here is the length of the shortest name that can match.
1594 // We need to check the size before looking at Name[1] and Name[2]
1595 // so we may as well check a limit that will eliminate mismatches.
1596 if (Name.size() < 12 || Name[1] != '_')
1597 return false;
1598 switch (Name[2]) {
1599 default:
1600 return false;
1601 case 'a':
1602 return Name == "__acos_finite" || Name == "__acosf_finite" ||
1603 Name == "__asin_finite" || Name == "__asinf_finite" ||
1604 Name == "__atan2_finite" || Name == "__atan2f_finite";
1605 case 'c':
1606 return Name == "__cosh_finite" || Name == "__coshf_finite";
1607 case 'e':
1608 return Name == "__exp_finite" || Name == "__expf_finite" ||
1609 Name == "__exp2_finite" || Name == "__exp2f_finite";
1610 case 'l':
1611 return Name == "__log_finite" || Name == "__logf_finite" ||
1612 Name == "__log10_finite" || Name == "__log10f_finite";
1613 case 'p':
1614 return Name == "__pow_finite" || Name == "__powf_finite";
1615 case 's':
1616 return Name == "__sinh_finite" || Name == "__sinhf_finite";
1617 }
1618 }
1619}
1620
1621namespace {
1622
1623Constant *GetConstantFoldFPValue(double V, Type *Ty) {
1624 if (Ty->isHalfTy() || Ty->isFloatTy()) {
1625 APFloat APF(V);
1626 bool unused;
1627 APF.convert(Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &unused);
1628 return ConstantFP::get(Ty->getContext(), APF);
1629 }
1630 if (Ty->isDoubleTy())
1631 return ConstantFP::get(Ty->getContext(), APFloat(V));
1632 llvm_unreachable("Can only constant fold half/float/double")::llvm::llvm_unreachable_internal("Can only constant fold half/float/double"
, "llvm/lib/Analysis/ConstantFolding.cpp", 1632)
;
1633}
1634
1635/// Clear the floating-point exception state.
1636inline void llvm_fenv_clearexcept() {
1637#if defined(HAVE_FENV_H1) && HAVE_DECL_FE_ALL_EXCEPT1
1638 feclearexcept(FE_ALL_EXCEPT(0x20 | 0x04 | 0x10 | 0x08 | 0x01));
1639#endif
1640 errno(*__errno_location ()) = 0;
1641}
1642
1643/// Test if a floating-point exception was raised.
1644inline bool llvm_fenv_testexcept() {
1645 int errno_val = errno(*__errno_location ());
1646 if (errno_val == ERANGE34 || errno_val == EDOM33)
1647 return true;
1648#if defined(HAVE_FENV_H1) && HAVE_DECL_FE_ALL_EXCEPT1 && HAVE_DECL_FE_INEXACT1
1649 if (fetestexcept(FE_ALL_EXCEPT(0x20 | 0x04 | 0x10 | 0x08 | 0x01) & ~FE_INEXACT0x20))
1650 return true;
1651#endif
1652 return false;
1653}
1654
1655Constant *ConstantFoldFP(double (*NativeFP)(double), const APFloat &V,
1656 Type *Ty) {
1657 llvm_fenv_clearexcept();
1658 double Result = NativeFP(V.convertToDouble());
1659 if (llvm_fenv_testexcept()) {
1660 llvm_fenv_clearexcept();
1661 return nullptr;
1662 }
1663
1664 return GetConstantFoldFPValue(Result, Ty);
1665}
1666
1667Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
1668 const APFloat &V, const APFloat &W, Type *Ty) {
1669 llvm_fenv_clearexcept();
1670 double Result = NativeFP(V.convertToDouble(), W.convertToDouble());
1671 if (llvm_fenv_testexcept()) {
1672 llvm_fenv_clearexcept();
1673 return nullptr;
1674 }
1675
1676 return GetConstantFoldFPValue(Result, Ty);
1677}
1678
1679Constant *constantFoldVectorReduce(Intrinsic::ID IID, Constant *Op) {
1680 FixedVectorType *VT = dyn_cast<FixedVectorType>(Op->getType());
1681 if (!VT)
1682 return nullptr;
1683
1684 // This isn't strictly necessary, but handle the special/common case of zero:
1685 // all integer reductions of a zero input produce zero.
1686 if (isa<ConstantAggregateZero>(Op))
1687 return ConstantInt::get(VT->getElementType(), 0);
1688
1689 // This is the same as the underlying binops - poison propagates.
1690 if (isa<PoisonValue>(Op) || Op->containsPoisonElement())
1691 return PoisonValue::get(VT->getElementType());
1692
1693 // TODO: Handle undef.
1694 if (!isa<ConstantVector>(Op) && !isa<ConstantDataVector>(Op))
1695 return nullptr;
1696
1697 auto *EltC = dyn_cast<ConstantInt>(Op->getAggregateElement(0U));
1698 if (!EltC)
1699 return nullptr;
1700
1701 APInt Acc = EltC->getValue();
1702 for (unsigned I = 1, E = VT->getNumElements(); I != E; I++) {
1703 if (!(EltC = dyn_cast<ConstantInt>(Op->getAggregateElement(I))))
1704 return nullptr;
1705 const APInt &X = EltC->getValue();
1706 switch (IID) {
1707 case Intrinsic::vector_reduce_add:
1708 Acc = Acc + X;
1709 break;
1710 case Intrinsic::vector_reduce_mul:
1711 Acc = Acc * X;
1712 break;
1713 case Intrinsic::vector_reduce_and:
1714 Acc = Acc & X;
1715 break;
1716 case Intrinsic::vector_reduce_or:
1717 Acc = Acc | X;
1718 break;
1719 case Intrinsic::vector_reduce_xor:
1720 Acc = Acc ^ X;
1721 break;
1722 case Intrinsic::vector_reduce_smin:
1723 Acc = APIntOps::smin(Acc, X);
1724 break;
1725 case Intrinsic::vector_reduce_smax:
1726 Acc = APIntOps::smax(Acc, X);
1727 break;
1728 case Intrinsic::vector_reduce_umin:
1729 Acc = APIntOps::umin(Acc, X);
1730 break;
1731 case Intrinsic::vector_reduce_umax:
1732 Acc = APIntOps::umax(Acc, X);
1733 break;
1734 }
1735 }
1736
1737 return ConstantInt::get(Op->getContext(), Acc);
1738}
1739
1740/// Attempt to fold an SSE floating point to integer conversion of a constant
1741/// floating point. If roundTowardZero is false, the default IEEE rounding is
1742/// used (toward nearest, ties to even). This matches the behavior of the
1743/// non-truncating SSE instructions in the default rounding mode. The desired
1744/// integer type Ty is used to select how many bits are available for the
1745/// result. Returns null if the conversion cannot be performed, otherwise
1746/// returns the Constant value resulting from the conversion.
1747Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
1748 Type *Ty, bool IsSigned) {
1749 // All of these conversion intrinsics form an integer of at most 64bits.
1750 unsigned ResultWidth = Ty->getIntegerBitWidth();
1751 assert(ResultWidth <= 64 &&(static_cast <bool> (ResultWidth <= 64 && "Can only constant fold conversions to 64 and 32 bit ints"
) ? void (0) : __assert_fail ("ResultWidth <= 64 && \"Can only constant fold conversions to 64 and 32 bit ints\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 1752, __extension__
__PRETTY_FUNCTION__))
1752 "Can only constant fold conversions to 64 and 32 bit ints")(static_cast <bool> (ResultWidth <= 64 && "Can only constant fold conversions to 64 and 32 bit ints"
) ? void (0) : __assert_fail ("ResultWidth <= 64 && \"Can only constant fold conversions to 64 and 32 bit ints\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 1752, __extension__
__PRETTY_FUNCTION__))
;
1753
1754 uint64_t UIntVal;
1755 bool isExact = false;
1756 APFloat::roundingMode mode = roundTowardZero? APFloat::rmTowardZero
1757 : APFloat::rmNearestTiesToEven;
1758 APFloat::opStatus status =
1759 Val.convertToInteger(makeMutableArrayRef(UIntVal), ResultWidth,
1760 IsSigned, mode, &isExact);
1761 if (status != APFloat::opOK &&
1762 (!roundTowardZero || status != APFloat::opInexact))
1763 return nullptr;
1764 return ConstantInt::get(Ty, UIntVal, IsSigned);
1765}
1766
1767double getValueAsDouble(ConstantFP *Op) {
1768 Type *Ty = Op->getType();
1769
1770 if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())
1771 return Op->getValueAPF().convertToDouble();
1772
1773 bool unused;
1774 APFloat APF = Op->getValueAPF();
1775 APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &unused);
1776 return APF.convertToDouble();
1777}
1778
1779static bool getConstIntOrUndef(Value *Op, const APInt *&C) {
1780 if (auto *CI = dyn_cast<ConstantInt>(Op)) {
1781 C = &CI->getValue();
1782 return true;
1783 }
1784 if (isa<UndefValue>(Op)) {
1785 C = nullptr;
1786 return true;
1787 }
1788 return false;
1789}
1790
1791/// Checks if the given intrinsic call, which evaluates to constant, is allowed
1792/// to be folded.
1793///
1794/// \param CI Constrained intrinsic call.
1795/// \param St Exception flags raised during constant evaluation.
1796static bool mayFoldConstrained(ConstrainedFPIntrinsic *CI,
1797 APFloat::opStatus St) {
1798 Optional<RoundingMode> ORM = CI->getRoundingMode();
1799 Optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
1800
1801 // If the operation does not change exception status flags, it is safe
1802 // to fold.
1803 if (St == APFloat::opStatus::opOK)
1804 return true;
1805
1806 // If evaluation raised FP exception, the result can depend on rounding
1807 // mode. If the latter is unknown, folding is not possible.
1808 if (!ORM || *ORM == RoundingMode::Dynamic)
1809 return false;
1810
1811 // If FP exceptions are ignored, fold the call, even if such exception is
1812 // raised.
1813 if (!EB || *EB != fp::ExceptionBehavior::ebStrict)
1814 return true;
1815
1816 // Leave the calculation for runtime so that exception flags be correctly set
1817 // in hardware.
1818 return false;
1819}
1820
1821/// Returns the rounding mode that should be used for constant evaluation.
1822static RoundingMode
1823getEvaluationRoundingMode(const ConstrainedFPIntrinsic *CI) {
1824 Optional<RoundingMode> ORM = CI->getRoundingMode();
1825 if (!ORM || *ORM == RoundingMode::Dynamic)
1826 // Even if the rounding mode is unknown, try evaluating the operation.
1827 // If it does not raise inexact exception, rounding was not applied,
1828 // so the result is exact and does not depend on rounding mode. Whether
1829 // other FP exceptions are raised, it does not depend on rounding mode.
1830 return RoundingMode::NearestTiesToEven;
1831 return *ORM;
1832}
1833
1834static Constant *ConstantFoldScalarCall1(StringRef Name,
1835 Intrinsic::ID IntrinsicID,
1836 Type *Ty,
1837 ArrayRef<Constant *> Operands,
1838 const TargetLibraryInfo *TLI,
1839 const CallBase *Call) {
1840 assert(Operands.size() == 1 && "Wrong number of operands.")(static_cast <bool> (Operands.size() == 1 && "Wrong number of operands."
) ? void (0) : __assert_fail ("Operands.size() == 1 && \"Wrong number of operands.\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 1840, __extension__
__PRETTY_FUNCTION__))
;
1841
1842 if (IntrinsicID == Intrinsic::is_constant) {
1843 // We know we have a "Constant" argument. But we want to only
1844 // return true for manifest constants, not those that depend on
1845 // constants with unknowable values, e.g. GlobalValue or BlockAddress.
1846 if (Operands[0]->isManifestConstant())
1847 return ConstantInt::getTrue(Ty->getContext());
1848 return nullptr;
1849 }
1850 if (isa<UndefValue>(Operands[0])) {
1851 // cosine(arg) is between -1 and 1. cosine(invalid arg) is NaN.
1852 // ctpop() is between 0 and bitwidth, pick 0 for undef.
1853 // fptoui.sat and fptosi.sat can always fold to zero (for a zero input).
1854 if (IntrinsicID == Intrinsic::cos ||
1855 IntrinsicID == Intrinsic::ctpop ||
1856 IntrinsicID == Intrinsic::fptoui_sat ||
1857 IntrinsicID == Intrinsic::fptosi_sat)
1858 return Constant::getNullValue(Ty);
1859 if (IntrinsicID == Intrinsic::bswap ||
1860 IntrinsicID == Intrinsic::bitreverse ||
1861 IntrinsicID == Intrinsic::launder_invariant_group ||
1862 IntrinsicID == Intrinsic::strip_invariant_group)
1863 return Operands[0];
1864 }
1865
1866 if (isa<ConstantPointerNull>(Operands[0])) {
1867 // launder(null) == null == strip(null) iff in addrspace 0
1868 if (IntrinsicID == Intrinsic::launder_invariant_group ||
1869 IntrinsicID == Intrinsic::strip_invariant_group) {
1870 // If instruction is not yet put in a basic block (e.g. when cloning
1871 // a function during inlining), Call's caller may not be available.
1872 // So check Call's BB first before querying Call->getCaller.
1873 const Function *Caller =
1874 Call->getParent() ? Call->getCaller() : nullptr;
1875 if (Caller &&
1876 !NullPointerIsDefined(
1877 Caller, Operands[0]->getType()->getPointerAddressSpace())) {
1878 return Operands[0];
1879 }
1880 return nullptr;
1881 }
1882 }
1883
1884 if (auto *Op = dyn_cast<ConstantFP>(Operands[0])) {
1885 if (IntrinsicID == Intrinsic::convert_to_fp16) {
1886 APFloat Val(Op->getValueAPF());
1887
1888 bool lost = false;
1889 Val.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &lost);
1890
1891 return ConstantInt::get(Ty->getContext(), Val.bitcastToAPInt());
1892 }
1893
1894 APFloat U = Op->getValueAPF();
1895
1896 if (IntrinsicID == Intrinsic::wasm_trunc_signed ||
1897 IntrinsicID == Intrinsic::wasm_trunc_unsigned) {
1898 bool Signed = IntrinsicID == Intrinsic::wasm_trunc_signed;
1899
1900 if (U.isNaN())
1901 return nullptr;
1902
1903 unsigned Width = Ty->getIntegerBitWidth();
1904 APSInt Int(Width, !Signed);
1905 bool IsExact = false;
1906 APFloat::opStatus Status =
1907 U.convertToInteger(Int, APFloat::rmTowardZero, &IsExact);
1908
1909 if (Status == APFloat::opOK || Status == APFloat::opInexact)
1910 return ConstantInt::get(Ty, Int);
1911
1912 return nullptr;
1913 }
1914
1915 if (IntrinsicID == Intrinsic::fptoui_sat ||
1916 IntrinsicID == Intrinsic::fptosi_sat) {
1917 // convertToInteger() already has the desired saturation semantics.
1918 APSInt Int(Ty->getIntegerBitWidth(),
1919 IntrinsicID == Intrinsic::fptoui_sat);
1920 bool IsExact;
1921 U.convertToInteger(Int, APFloat::rmTowardZero, &IsExact);
1922 return ConstantInt::get(Ty, Int);
1923 }
1924
1925 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
1926 return nullptr;
1927
1928 // Use internal versions of these intrinsics.
1929
1930 if (IntrinsicID == Intrinsic::nearbyint || IntrinsicID == Intrinsic::rint) {
1931 U.roundToIntegral(APFloat::rmNearestTiesToEven);
1932 return ConstantFP::get(Ty->getContext(), U);
1933 }
1934
1935 if (IntrinsicID == Intrinsic::round) {
1936 U.roundToIntegral(APFloat::rmNearestTiesToAway);
1937 return ConstantFP::get(Ty->getContext(), U);
1938 }
1939
1940 if (IntrinsicID == Intrinsic::roundeven) {
1941 U.roundToIntegral(APFloat::rmNearestTiesToEven);
1942 return ConstantFP::get(Ty->getContext(), U);
1943 }
1944
1945 if (IntrinsicID == Intrinsic::ceil) {
1946 U.roundToIntegral(APFloat::rmTowardPositive);
1947 return ConstantFP::get(Ty->getContext(), U);
1948 }
1949
1950 if (IntrinsicID == Intrinsic::floor) {
1951 U.roundToIntegral(APFloat::rmTowardNegative);
1952 return ConstantFP::get(Ty->getContext(), U);
1953 }
1954
1955 if (IntrinsicID == Intrinsic::trunc) {
1956 U.roundToIntegral(APFloat::rmTowardZero);
1957 return ConstantFP::get(Ty->getContext(), U);
1958 }
1959
1960 if (IntrinsicID == Intrinsic::fabs) {
1961 U.clearSign();
1962 return ConstantFP::get(Ty->getContext(), U);
1963 }
1964
1965 if (IntrinsicID == Intrinsic::amdgcn_fract) {
1966 // The v_fract instruction behaves like the OpenCL spec, which defines
1967 // fract(x) as fmin(x - floor(x), 0x1.fffffep-1f): "The min() operator is
1968 // there to prevent fract(-small) from returning 1.0. It returns the
1969 // largest positive floating-point number less than 1.0."
1970 APFloat FloorU(U);
1971 FloorU.roundToIntegral(APFloat::rmTowardNegative);
1972 APFloat FractU(U - FloorU);
1973 APFloat AlmostOne(U.getSemantics(), 1);
1974 AlmostOne.next(/*nextDown*/ true);
1975 return ConstantFP::get(Ty->getContext(), minimum(FractU, AlmostOne));
1976 }
1977
1978 // Rounding operations (floor, trunc, ceil, round and nearbyint) do not
1979 // raise FP exceptions, unless the argument is signaling NaN.
1980
1981 Optional<APFloat::roundingMode> RM;
1982 switch (IntrinsicID) {
1983 default:
1984 break;
1985 case Intrinsic::experimental_constrained_nearbyint:
1986 case Intrinsic::experimental_constrained_rint: {
1987 auto CI = cast<ConstrainedFPIntrinsic>(Call);
1988 RM = CI->getRoundingMode();
1989 if (!RM || RM.getValue() == RoundingMode::Dynamic)
1990 return nullptr;
1991 break;
1992 }
1993 case Intrinsic::experimental_constrained_round:
1994 RM = APFloat::rmNearestTiesToAway;
1995 break;
1996 case Intrinsic::experimental_constrained_ceil:
1997 RM = APFloat::rmTowardPositive;
1998 break;
1999 case Intrinsic::experimental_constrained_floor:
2000 RM = APFloat::rmTowardNegative;
2001 break;
2002 case Intrinsic::experimental_constrained_trunc:
2003 RM = APFloat::rmTowardZero;
2004 break;
2005 }
2006 if (RM) {
2007 auto CI = cast<ConstrainedFPIntrinsic>(Call);
2008 if (U.isFinite()) {
2009 APFloat::opStatus St = U.roundToIntegral(*RM);
2010 if (IntrinsicID == Intrinsic::experimental_constrained_rint &&
2011 St == APFloat::opInexact) {
2012 Optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2013 if (EB && *EB == fp::ebStrict)
2014 return nullptr;
2015 }
2016 } else if (U.isSignaling()) {
2017 Optional<fp::ExceptionBehavior> EB = CI->getExceptionBehavior();
2018 if (EB && *EB != fp::ebIgnore)
2019 return nullptr;
2020 U = APFloat::getQNaN(U.getSemantics());
2021 }
2022 return ConstantFP::get(Ty->getContext(), U);
2023 }
2024
2025 /// We only fold functions with finite arguments. Folding NaN and inf is
2026 /// likely to be aborted with an exception anyway, and some host libms
2027 /// have known errors raising exceptions.
2028 if (!U.isFinite())
2029 return nullptr;
2030
2031 /// Currently APFloat versions of these functions do not exist, so we use
2032 /// the host native double versions. Float versions are not called
2033 /// directly but for all these it is true (float)(f((double)arg)) ==
2034 /// f(arg). Long double not supported yet.
2035 const APFloat &APF = Op->getValueAPF();
2036
2037 switch (IntrinsicID) {
2038 default: break;
2039 case Intrinsic::log:
2040 return ConstantFoldFP(log, APF, Ty);
2041 case Intrinsic::log2:
2042 // TODO: What about hosts that lack a C99 library?
2043 return ConstantFoldFP(Log2, APF, Ty);
2044 case Intrinsic::log10:
2045 // TODO: What about hosts that lack a C99 library?
2046 return ConstantFoldFP(log10, APF, Ty);
2047 case Intrinsic::exp:
2048 return ConstantFoldFP(exp, APF, Ty);
2049 case Intrinsic::exp2:
2050 // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
2051 return ConstantFoldBinaryFP(pow, APFloat(2.0), APF, Ty);
2052 case Intrinsic::sin:
2053 return ConstantFoldFP(sin, APF, Ty);
2054 case Intrinsic::cos:
2055 return ConstantFoldFP(cos, APF, Ty);
2056 case Intrinsic::sqrt:
2057 return ConstantFoldFP(sqrt, APF, Ty);
2058 case Intrinsic::amdgcn_cos:
2059 case Intrinsic::amdgcn_sin: {
2060 double V = getValueAsDouble(Op);
2061 if (V < -256.0 || V > 256.0)
2062 // The gfx8 and gfx9 architectures handle arguments outside the range
2063 // [-256, 256] differently. This should be a rare case so bail out
2064 // rather than trying to handle the difference.
2065 return nullptr;
2066 bool IsCos = IntrinsicID == Intrinsic::amdgcn_cos;
2067 double V4 = V * 4.0;
2068 if (V4 == floor(V4)) {
2069 // Force exact results for quarter-integer inputs.
2070 const double SinVals[4] = { 0.0, 1.0, 0.0, -1.0 };
2071 V = SinVals[((int)V4 + (IsCos ? 1 : 0)) & 3];
2072 } else {
2073 if (IsCos)
2074 V = cos(V * 2.0 * numbers::pi);
2075 else
2076 V = sin(V * 2.0 * numbers::pi);
2077 }
2078 return GetConstantFoldFPValue(V, Ty);
2079 }
2080 }
2081
2082 if (!TLI)
2083 return nullptr;
2084
2085 LibFunc Func = NotLibFunc;
2086 if (!TLI->getLibFunc(Name, Func))
2087 return nullptr;
2088
2089 switch (Func) {
2090 default:
2091 break;
2092 case LibFunc_acos:
2093 case LibFunc_acosf:
2094 case LibFunc_acos_finite:
2095 case LibFunc_acosf_finite:
2096 if (TLI->has(Func))
2097 return ConstantFoldFP(acos, APF, Ty);
2098 break;
2099 case LibFunc_asin:
2100 case LibFunc_asinf:
2101 case LibFunc_asin_finite:
2102 case LibFunc_asinf_finite:
2103 if (TLI->has(Func))
2104 return ConstantFoldFP(asin, APF, Ty);
2105 break;
2106 case LibFunc_atan:
2107 case LibFunc_atanf:
2108 if (TLI->has(Func))
2109 return ConstantFoldFP(atan, APF, Ty);
2110 break;
2111 case LibFunc_ceil:
2112 case LibFunc_ceilf:
2113 if (TLI->has(Func)) {
2114 U.roundToIntegral(APFloat::rmTowardPositive);
2115 return ConstantFP::get(Ty->getContext(), U);
2116 }
2117 break;
2118 case LibFunc_cos:
2119 case LibFunc_cosf:
2120 if (TLI->has(Func))
2121 return ConstantFoldFP(cos, APF, Ty);
2122 break;
2123 case LibFunc_cosh:
2124 case LibFunc_coshf:
2125 case LibFunc_cosh_finite:
2126 case LibFunc_coshf_finite:
2127 if (TLI->has(Func))
2128 return ConstantFoldFP(cosh, APF, Ty);
2129 break;
2130 case LibFunc_exp:
2131 case LibFunc_expf:
2132 case LibFunc_exp_finite:
2133 case LibFunc_expf_finite:
2134 if (TLI->has(Func))
2135 return ConstantFoldFP(exp, APF, Ty);
2136 break;
2137 case LibFunc_exp2:
2138 case LibFunc_exp2f:
2139 case LibFunc_exp2_finite:
2140 case LibFunc_exp2f_finite:
2141 if (TLI->has(Func))
2142 // Fold exp2(x) as pow(2, x), in case the host lacks a C99 library.
2143 return ConstantFoldBinaryFP(pow, APFloat(2.0), APF, Ty);
2144 break;
2145 case LibFunc_fabs:
2146 case LibFunc_fabsf:
2147 if (TLI->has(Func)) {
2148 U.clearSign();
2149 return ConstantFP::get(Ty->getContext(), U);
2150 }
2151 break;
2152 case LibFunc_floor:
2153 case LibFunc_floorf:
2154 if (TLI->has(Func)) {
2155 U.roundToIntegral(APFloat::rmTowardNegative);
2156 return ConstantFP::get(Ty->getContext(), U);
2157 }
2158 break;
2159 case LibFunc_log:
2160 case LibFunc_logf:
2161 case LibFunc_log_finite:
2162 case LibFunc_logf_finite:
2163 if (!APF.isNegative() && !APF.isZero() && TLI->has(Func))
2164 return ConstantFoldFP(log, APF, Ty);
2165 break;
2166 case LibFunc_log2:
2167 case LibFunc_log2f:
2168 case LibFunc_log2_finite:
2169 case LibFunc_log2f_finite:
2170 if (!APF.isNegative() && !APF.isZero() && TLI->has(Func))
2171 // TODO: What about hosts that lack a C99 library?
2172 return ConstantFoldFP(Log2, APF, Ty);
2173 break;
2174 case LibFunc_log10:
2175 case LibFunc_log10f:
2176 case LibFunc_log10_finite:
2177 case LibFunc_log10f_finite:
2178 if (!APF.isNegative() && !APF.isZero() && TLI->has(Func))
2179 // TODO: What about hosts that lack a C99 library?
2180 return ConstantFoldFP(log10, APF, Ty);
2181 break;
2182 case LibFunc_nearbyint:
2183 case LibFunc_nearbyintf:
2184 case LibFunc_rint:
2185 case LibFunc_rintf:
2186 if (TLI->has(Func)) {
2187 U.roundToIntegral(APFloat::rmNearestTiesToEven);
2188 return ConstantFP::get(Ty->getContext(), U);
2189 }
2190 break;
2191 case LibFunc_round:
2192 case LibFunc_roundf:
2193 if (TLI->has(Func)) {
2194 U.roundToIntegral(APFloat::rmNearestTiesToAway);
2195 return ConstantFP::get(Ty->getContext(), U);
2196 }
2197 break;
2198 case LibFunc_sin:
2199 case LibFunc_sinf:
2200 if (TLI->has(Func))
2201 return ConstantFoldFP(sin, APF, Ty);
2202 break;
2203 case LibFunc_sinh:
2204 case LibFunc_sinhf:
2205 case LibFunc_sinh_finite:
2206 case LibFunc_sinhf_finite:
2207 if (TLI->has(Func))
2208 return ConstantFoldFP(sinh, APF, Ty);
2209 break;
2210 case LibFunc_sqrt:
2211 case LibFunc_sqrtf:
2212 if (!APF.isNegative() && TLI->has(Func))
2213 return ConstantFoldFP(sqrt, APF, Ty);
2214 break;
2215 case LibFunc_tan:
2216 case LibFunc_tanf:
2217 if (TLI->has(Func))
2218 return ConstantFoldFP(tan, APF, Ty);
2219 break;
2220 case LibFunc_tanh:
2221 case LibFunc_tanhf:
2222 if (TLI->has(Func))
2223 return ConstantFoldFP(tanh, APF, Ty);
2224 break;
2225 case LibFunc_trunc:
2226 case LibFunc_truncf:
2227 if (TLI->has(Func)) {
2228 U.roundToIntegral(APFloat::rmTowardZero);
2229 return ConstantFP::get(Ty->getContext(), U);
2230 }
2231 break;
2232 }
2233 return nullptr;
2234 }
2235
2236 if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
2237 switch (IntrinsicID) {
2238 case Intrinsic::bswap:
2239 return ConstantInt::get(Ty->getContext(), Op->getValue().byteSwap());
2240 case Intrinsic::ctpop:
2241 return ConstantInt::get(Ty, Op->getValue().countPopulation());
2242 case Intrinsic::bitreverse:
2243 return ConstantInt::get(Ty->getContext(), Op->getValue().reverseBits());
2244 case Intrinsic::convert_from_fp16: {
2245 APFloat Val(APFloat::IEEEhalf(), Op->getValue());
2246
2247 bool lost = false;
2248 APFloat::opStatus status = Val.convert(
2249 Ty->getFltSemantics(), APFloat::rmNearestTiesToEven, &lost);
2250
2251 // Conversion is always precise.
2252 (void)status;
2253 assert(status == APFloat::opOK && !lost &&(static_cast <bool> (status == APFloat::opOK &&
!lost && "Precision lost during fp16 constfolding") ?
void (0) : __assert_fail ("status == APFloat::opOK && !lost && \"Precision lost during fp16 constfolding\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2254, __extension__
__PRETTY_FUNCTION__))
2254 "Precision lost during fp16 constfolding")(static_cast <bool> (status == APFloat::opOK &&
!lost && "Precision lost during fp16 constfolding") ?
void (0) : __assert_fail ("status == APFloat::opOK && !lost && \"Precision lost during fp16 constfolding\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2254, __extension__
__PRETTY_FUNCTION__))
;
2255
2256 return ConstantFP::get(Ty->getContext(), Val);
2257 }
2258 default:
2259 return nullptr;
2260 }
2261 }
2262
2263 switch (IntrinsicID) {
2264 default: break;
2265 case Intrinsic::vector_reduce_add:
2266 case Intrinsic::vector_reduce_mul:
2267 case Intrinsic::vector_reduce_and:
2268 case Intrinsic::vector_reduce_or:
2269 case Intrinsic::vector_reduce_xor:
2270 case Intrinsic::vector_reduce_smin:
2271 case Intrinsic::vector_reduce_smax:
2272 case Intrinsic::vector_reduce_umin:
2273 case Intrinsic::vector_reduce_umax:
2274 if (Constant *C = constantFoldVectorReduce(IntrinsicID, Operands[0]))
2275 return C;
2276 break;
2277 }
2278
2279 // Support ConstantVector in case we have an Undef in the top.
2280 if (isa<ConstantVector>(Operands[0]) ||
2281 isa<ConstantDataVector>(Operands[0])) {
2282 auto *Op = cast<Constant>(Operands[0]);
2283 switch (IntrinsicID) {
2284 default: break;
2285 case Intrinsic::x86_sse_cvtss2si:
2286 case Intrinsic::x86_sse_cvtss2si64:
2287 case Intrinsic::x86_sse2_cvtsd2si:
2288 case Intrinsic::x86_sse2_cvtsd2si64:
2289 if (ConstantFP *FPOp =
2290 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2291 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2292 /*roundTowardZero=*/false, Ty,
2293 /*IsSigned*/true);
2294 break;
2295 case Intrinsic::x86_sse_cvttss2si:
2296 case Intrinsic::x86_sse_cvttss2si64:
2297 case Intrinsic::x86_sse2_cvttsd2si:
2298 case Intrinsic::x86_sse2_cvttsd2si64:
2299 if (ConstantFP *FPOp =
2300 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2301 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2302 /*roundTowardZero=*/true, Ty,
2303 /*IsSigned*/true);
2304 break;
2305 }
2306 }
2307
2308 return nullptr;
2309}
2310
2311static Constant *ConstantFoldScalarCall2(StringRef Name,
2312 Intrinsic::ID IntrinsicID,
2313 Type *Ty,
2314 ArrayRef<Constant *> Operands,
2315 const TargetLibraryInfo *TLI,
2316 const CallBase *Call) {
2317 assert(Operands.size() == 2 && "Wrong number of operands.")(static_cast <bool> (Operands.size() == 2 && "Wrong number of operands."
) ? void (0) : __assert_fail ("Operands.size() == 2 && \"Wrong number of operands.\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2317, __extension__
__PRETTY_FUNCTION__))
;
2318
2319 if (Ty->isFloatingPointTy()) {
2320 // TODO: We should have undef handling for all of the FP intrinsics that
2321 // are attempted to be folded in this function.
2322 bool IsOp0Undef = isa<UndefValue>(Operands[0]);
2323 bool IsOp1Undef = isa<UndefValue>(Operands[1]);
2324 switch (IntrinsicID) {
2325 case Intrinsic::maxnum:
2326 case Intrinsic::minnum:
2327 case Intrinsic::maximum:
2328 case Intrinsic::minimum:
2329 // If one argument is undef, return the other argument.
2330 if (IsOp0Undef)
2331 return Operands[1];
2332 if (IsOp1Undef)
2333 return Operands[0];
2334 break;
2335 }
2336 }
2337
2338 if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
2339 if (!Ty->isFloatingPointTy())
2340 return nullptr;
2341 const APFloat &Op1V = Op1->getValueAPF();
2342
2343 if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
2344 if (Op2->getType() != Op1->getType())
2345 return nullptr;
2346 const APFloat &Op2V = Op2->getValueAPF();
2347
2348 if (const auto *ConstrIntr = dyn_cast<ConstrainedFPIntrinsic>(Call)) {
2349 RoundingMode RM = getEvaluationRoundingMode(ConstrIntr);
2350 APFloat Res = Op1V;
2351 APFloat::opStatus St;
2352 switch (IntrinsicID) {
2353 default:
2354 return nullptr;
2355 case Intrinsic::experimental_constrained_fadd:
2356 St = Res.add(Op2V, RM);
2357 break;
2358 case Intrinsic::experimental_constrained_fsub:
2359 St = Res.subtract(Op2V, RM);
2360 break;
2361 case Intrinsic::experimental_constrained_fmul:
2362 St = Res.multiply(Op2V, RM);
2363 break;
2364 case Intrinsic::experimental_constrained_fdiv:
2365 St = Res.divide(Op2V, RM);
2366 break;
2367 case Intrinsic::experimental_constrained_frem:
2368 St = Res.mod(Op2V);
2369 break;
2370 }
2371 if (mayFoldConstrained(const_cast<ConstrainedFPIntrinsic *>(ConstrIntr),
2372 St))
2373 return ConstantFP::get(Ty->getContext(), Res);
2374 return nullptr;
2375 }
2376
2377 switch (IntrinsicID) {
2378 default:
2379 break;
2380 case Intrinsic::copysign:
2381 return ConstantFP::get(Ty->getContext(), APFloat::copySign(Op1V, Op2V));
2382 case Intrinsic::minnum:
2383 return ConstantFP::get(Ty->getContext(), minnum(Op1V, Op2V));
2384 case Intrinsic::maxnum:
2385 return ConstantFP::get(Ty->getContext(), maxnum(Op1V, Op2V));
2386 case Intrinsic::minimum:
2387 return ConstantFP::get(Ty->getContext(), minimum(Op1V, Op2V));
2388 case Intrinsic::maximum:
2389 return ConstantFP::get(Ty->getContext(), maximum(Op1V, Op2V));
2390 }
2391
2392 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
2393 return nullptr;
2394
2395 switch (IntrinsicID) {
2396 default:
2397 break;
2398 case Intrinsic::pow:
2399 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
2400 case Intrinsic::amdgcn_fmul_legacy:
2401 // The legacy behaviour is that multiplying +/- 0.0 by anything, even
2402 // NaN or infinity, gives +0.0.
2403 if (Op1V.isZero() || Op2V.isZero())
2404 return ConstantFP::getNullValue(Ty);
2405 return ConstantFP::get(Ty->getContext(), Op1V * Op2V);
2406 }
2407
2408 if (!TLI)
2409 return nullptr;
2410
2411 LibFunc Func = NotLibFunc;
2412 if (!TLI->getLibFunc(Name, Func))
2413 return nullptr;
2414
2415 switch (Func) {
2416 default:
2417 break;
2418 case LibFunc_pow:
2419 case LibFunc_powf:
2420 case LibFunc_pow_finite:
2421 case LibFunc_powf_finite:
2422 if (TLI->has(Func))
2423 return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
2424 break;
2425 case LibFunc_fmod:
2426 case LibFunc_fmodf:
2427 if (TLI->has(Func)) {
2428 APFloat V = Op1->getValueAPF();
2429 if (APFloat::opStatus::opOK == V.mod(Op2->getValueAPF()))
2430 return ConstantFP::get(Ty->getContext(), V);
2431 }
2432 break;
2433 case LibFunc_remainder:
2434 case LibFunc_remainderf:
2435 if (TLI->has(Func)) {
2436 APFloat V = Op1->getValueAPF();
2437 if (APFloat::opStatus::opOK == V.remainder(Op2->getValueAPF()))
2438 return ConstantFP::get(Ty->getContext(), V);
2439 }
2440 break;
2441 case LibFunc_atan2:
2442 case LibFunc_atan2f:
2443 case LibFunc_atan2_finite:
2444 case LibFunc_atan2f_finite:
2445 if (TLI->has(Func))
2446 return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
2447 break;
2448 }
2449 } else if (auto *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
2450 if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
2451 return nullptr;
2452 if (IntrinsicID == Intrinsic::powi && Ty->isHalfTy())
2453 return ConstantFP::get(
2454 Ty->getContext(),
2455 APFloat((float)std::pow((float)Op1V.convertToDouble(),
2456 (int)Op2C->getZExtValue())));
2457 if (IntrinsicID == Intrinsic::powi && Ty->isFloatTy())
2458 return ConstantFP::get(
2459 Ty->getContext(),
2460 APFloat((float)std::pow((float)Op1V.convertToDouble(),
2461 (int)Op2C->getZExtValue())));
2462 if (IntrinsicID == Intrinsic::powi && Ty->isDoubleTy())
2463 return ConstantFP::get(
2464 Ty->getContext(),
2465 APFloat((double)std::pow(Op1V.convertToDouble(),
2466 (int)Op2C->getZExtValue())));
2467
2468 if (IntrinsicID == Intrinsic::amdgcn_ldexp) {
2469 // FIXME: Should flush denorms depending on FP mode, but that's ignored
2470 // everywhere else.
2471
2472 // scalbn is equivalent to ldexp with float radix 2
2473 APFloat Result = scalbn(Op1->getValueAPF(), Op2C->getSExtValue(),
2474 APFloat::rmNearestTiesToEven);
2475 return ConstantFP::get(Ty->getContext(), Result);
2476 }
2477 }
2478 return nullptr;
2479 }
2480
2481 if (Operands[0]->getType()->isIntegerTy() &&
2482 Operands[1]->getType()->isIntegerTy()) {
2483 const APInt *C0, *C1;
2484 if (!getConstIntOrUndef(Operands[0], C0) ||
2485 !getConstIntOrUndef(Operands[1], C1))
2486 return nullptr;
2487
2488 switch (IntrinsicID) {
2489 default: break;
2490 case Intrinsic::smax:
2491 case Intrinsic::smin:
2492 case Intrinsic::umax:
2493 case Intrinsic::umin:
2494 if (!C0 && !C1)
2495 return UndefValue::get(Ty);
2496 if (!C0 || !C1)
2497 return MinMaxIntrinsic::getSaturationPoint(IntrinsicID, Ty);
2498 return ConstantInt::get(
2499 Ty, ICmpInst::compare(*C0, *C1,
2500 MinMaxIntrinsic::getPredicate(IntrinsicID))
2501 ? *C0
2502 : *C1);
2503
2504 case Intrinsic::usub_with_overflow:
2505 case Intrinsic::ssub_with_overflow:
2506 // X - undef -> { 0, false }
2507 // undef - X -> { 0, false }
2508 if (!C0 || !C1)
2509 return Constant::getNullValue(Ty);
2510 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2511 case Intrinsic::uadd_with_overflow:
2512 case Intrinsic::sadd_with_overflow:
2513 // X + undef -> { -1, false }
2514 // undef + x -> { -1, false }
2515 if (!C0 || !C1) {
2516 return ConstantStruct::get(
2517 cast<StructType>(Ty),
2518 {Constant::getAllOnesValue(Ty->getStructElementType(0)),
2519 Constant::getNullValue(Ty->getStructElementType(1))});
2520 }
2521 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2522 case Intrinsic::smul_with_overflow:
2523 case Intrinsic::umul_with_overflow: {
2524 // undef * X -> { 0, false }
2525 // X * undef -> { 0, false }
2526 if (!C0 || !C1)
2527 return Constant::getNullValue(Ty);
2528
2529 APInt Res;
2530 bool Overflow;
2531 switch (IntrinsicID) {
2532 default: llvm_unreachable("Invalid case")::llvm::llvm_unreachable_internal("Invalid case", "llvm/lib/Analysis/ConstantFolding.cpp"
, 2532)
;
2533 case Intrinsic::sadd_with_overflow:
2534 Res = C0->sadd_ov(*C1, Overflow);
2535 break;
2536 case Intrinsic::uadd_with_overflow:
2537 Res = C0->uadd_ov(*C1, Overflow);
2538 break;
2539 case Intrinsic::ssub_with_overflow:
2540 Res = C0->ssub_ov(*C1, Overflow);
2541 break;
2542 case Intrinsic::usub_with_overflow:
2543 Res = C0->usub_ov(*C1, Overflow);
2544 break;
2545 case Intrinsic::smul_with_overflow:
2546 Res = C0->smul_ov(*C1, Overflow);
2547 break;
2548 case Intrinsic::umul_with_overflow:
2549 Res = C0->umul_ov(*C1, Overflow);
2550 break;
2551 }
2552 Constant *Ops[] = {
2553 ConstantInt::get(Ty->getContext(), Res),
2554 ConstantInt::get(Type::getInt1Ty(Ty->getContext()), Overflow)
2555 };
2556 return ConstantStruct::get(cast<StructType>(Ty), Ops);
2557 }
2558 case Intrinsic::uadd_sat:
2559 case Intrinsic::sadd_sat:
2560 if (!C0 && !C1)
2561 return UndefValue::get(Ty);
2562 if (!C0 || !C1)
2563 return Constant::getAllOnesValue(Ty);
2564 if (IntrinsicID == Intrinsic::uadd_sat)
2565 return ConstantInt::get(Ty, C0->uadd_sat(*C1));
2566 else
2567 return ConstantInt::get(Ty, C0->sadd_sat(*C1));
2568 case Intrinsic::usub_sat:
2569 case Intrinsic::ssub_sat:
2570 if (!C0 && !C1)
2571 return UndefValue::get(Ty);
2572 if (!C0 || !C1)
2573 return Constant::getNullValue(Ty);
2574 if (IntrinsicID == Intrinsic::usub_sat)
2575 return ConstantInt::get(Ty, C0->usub_sat(*C1));
2576 else
2577 return ConstantInt::get(Ty, C0->ssub_sat(*C1));
2578 case Intrinsic::cttz:
2579 case Intrinsic::ctlz:
2580 assert(C1 && "Must be constant int")(static_cast <bool> (C1 && "Must be constant int"
) ? void (0) : __assert_fail ("C1 && \"Must be constant int\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2580, __extension__
__PRETTY_FUNCTION__))
;
2581
2582 // cttz(0, 1) and ctlz(0, 1) are undef.
2583 if (C1->isOne() && (!C0 || C0->isZero()))
2584 return UndefValue::get(Ty);
2585 if (!C0)
2586 return Constant::getNullValue(Ty);
2587 if (IntrinsicID == Intrinsic::cttz)
2588 return ConstantInt::get(Ty, C0->countTrailingZeros());
2589 else
2590 return ConstantInt::get(Ty, C0->countLeadingZeros());
2591
2592 case Intrinsic::abs:
2593 assert(C1 && "Must be constant int")(static_cast <bool> (C1 && "Must be constant int"
) ? void (0) : __assert_fail ("C1 && \"Must be constant int\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2593, __extension__
__PRETTY_FUNCTION__))
;
2594 assert((C1->isOne() || C1->isZero()) && "Must be 0 or 1")(static_cast <bool> ((C1->isOne() || C1->isZero()
) && "Must be 0 or 1") ? void (0) : __assert_fail ("(C1->isOne() || C1->isZero()) && \"Must be 0 or 1\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2594, __extension__
__PRETTY_FUNCTION__))
;
2595
2596 // Undef or minimum val operand with poison min --> undef
2597 if (C1->isOne() && (!C0 || C0->isMinSignedValue()))
2598 return UndefValue::get(Ty);
2599
2600 // Undef operand with no poison min --> 0 (sign bit must be clear)
2601 if (!C0)
2602 return Constant::getNullValue(Ty);
2603
2604 return ConstantInt::get(Ty, C0->abs());
2605 }
2606
2607 return nullptr;
2608 }
2609
2610 // Support ConstantVector in case we have an Undef in the top.
2611 if ((isa<ConstantVector>(Operands[0]) ||
2612 isa<ConstantDataVector>(Operands[0])) &&
2613 // Check for default rounding mode.
2614 // FIXME: Support other rounding modes?
2615 isa<ConstantInt>(Operands[1]) &&
2616 cast<ConstantInt>(Operands[1])->getValue() == 4) {
2617 auto *Op = cast<Constant>(Operands[0]);
2618 switch (IntrinsicID) {
2619 default: break;
2620 case Intrinsic::x86_avx512_vcvtss2si32:
2621 case Intrinsic::x86_avx512_vcvtss2si64:
2622 case Intrinsic::x86_avx512_vcvtsd2si32:
2623 case Intrinsic::x86_avx512_vcvtsd2si64:
2624 if (ConstantFP *FPOp =
2625 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2626 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2627 /*roundTowardZero=*/false, Ty,
2628 /*IsSigned*/true);
2629 break;
2630 case Intrinsic::x86_avx512_vcvtss2usi32:
2631 case Intrinsic::x86_avx512_vcvtss2usi64:
2632 case Intrinsic::x86_avx512_vcvtsd2usi32:
2633 case Intrinsic::x86_avx512_vcvtsd2usi64:
2634 if (ConstantFP *FPOp =
2635 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2636 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2637 /*roundTowardZero=*/false, Ty,
2638 /*IsSigned*/false);
2639 break;
2640 case Intrinsic::x86_avx512_cvttss2si:
2641 case Intrinsic::x86_avx512_cvttss2si64:
2642 case Intrinsic::x86_avx512_cvttsd2si:
2643 case Intrinsic::x86_avx512_cvttsd2si64:
2644 if (ConstantFP *FPOp =
2645 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2646 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2647 /*roundTowardZero=*/true, Ty,
2648 /*IsSigned*/true);
2649 break;
2650 case Intrinsic::x86_avx512_cvttss2usi:
2651 case Intrinsic::x86_avx512_cvttss2usi64:
2652 case Intrinsic::x86_avx512_cvttsd2usi:
2653 case Intrinsic::x86_avx512_cvttsd2usi64:
2654 if (ConstantFP *FPOp =
2655 dyn_cast_or_null<ConstantFP>(Op->getAggregateElement(0U)))
2656 return ConstantFoldSSEConvertToInt(FPOp->getValueAPF(),
2657 /*roundTowardZero=*/true, Ty,
2658 /*IsSigned*/false);
2659 break;
2660 }
2661 }
2662 return nullptr;
2663}
2664
2665static APFloat ConstantFoldAMDGCNCubeIntrinsic(Intrinsic::ID IntrinsicID,
2666 const APFloat &S0,
2667 const APFloat &S1,
2668 const APFloat &S2) {
2669 unsigned ID;
2670 const fltSemantics &Sem = S0.getSemantics();
2671 APFloat MA(Sem), SC(Sem), TC(Sem);
2672 if (abs(S2) >= abs(S0) && abs(S2) >= abs(S1)) {
2673 if (S2.isNegative() && S2.isNonZero() && !S2.isNaN()) {
2674 // S2 < 0
2675 ID = 5;
2676 SC = -S0;
2677 } else {
2678 ID = 4;
2679 SC = S0;
2680 }
2681 MA = S2;
2682 TC = -S1;
2683 } else if (abs(S1) >= abs(S0)) {
2684 if (S1.isNegative() && S1.isNonZero() && !S1.isNaN()) {
2685 // S1 < 0
2686 ID = 3;
2687 TC = -S2;
2688 } else {
2689 ID = 2;
2690 TC = S2;
2691 }
2692 MA = S1;
2693 SC = S0;
2694 } else {
2695 if (S0.isNegative() && S0.isNonZero() && !S0.isNaN()) {
2696 // S0 < 0
2697 ID = 1;
2698 SC = S2;
2699 } else {
2700 ID = 0;
2701 SC = -S2;
2702 }
2703 MA = S0;
2704 TC = -S1;
2705 }
2706 switch (IntrinsicID) {
2707 default:
2708 llvm_unreachable("unhandled amdgcn cube intrinsic")::llvm::llvm_unreachable_internal("unhandled amdgcn cube intrinsic"
, "llvm/lib/Analysis/ConstantFolding.cpp", 2708)
;
2709 case Intrinsic::amdgcn_cubeid:
2710 return APFloat(Sem, ID);
2711 case Intrinsic::amdgcn_cubema:
2712 return MA + MA;
2713 case Intrinsic::amdgcn_cubesc:
2714 return SC;
2715 case Intrinsic::amdgcn_cubetc:
2716 return TC;
2717 }
2718}
2719
2720static Constant *ConstantFoldAMDGCNPermIntrinsic(ArrayRef<Constant *> Operands,
2721 Type *Ty) {
2722 const APInt *C0, *C1, *C2;
2723 if (!getConstIntOrUndef(Operands[0], C0) ||
2724 !getConstIntOrUndef(Operands[1], C1) ||
2725 !getConstIntOrUndef(Operands[2], C2))
2726 return nullptr;
2727
2728 if (!C2)
2729 return UndefValue::get(Ty);
2730
2731 APInt Val(32, 0);
2732 unsigned NumUndefBytes = 0;
2733 for (unsigned I = 0; I < 32; I += 8) {
2734 unsigned Sel = C2->extractBitsAsZExtValue(8, I);
2735 unsigned B = 0;
2736
2737 if (Sel >= 13)
2738 B = 0xff;
2739 else if (Sel == 12)
2740 B = 0x00;
2741 else {
2742 const APInt *Src = ((Sel & 10) == 10 || (Sel & 12) == 4) ? C0 : C1;
2743 if (!Src)
2744 ++NumUndefBytes;
2745 else if (Sel < 8)
2746 B = Src->extractBitsAsZExtValue(8, (Sel & 3) * 8);
2747 else
2748 B = Src->extractBitsAsZExtValue(1, (Sel & 1) ? 31 : 15) * 0xff;
2749 }
2750
2751 Val.insertBits(B, I, 8);
2752 }
2753
2754 if (NumUndefBytes == 4)
2755 return UndefValue::get(Ty);
2756
2757 return ConstantInt::get(Ty, Val);
2758}
2759
2760static Constant *ConstantFoldScalarCall3(StringRef Name,
2761 Intrinsic::ID IntrinsicID,
2762 Type *Ty,
2763 ArrayRef<Constant *> Operands,
2764 const TargetLibraryInfo *TLI,
2765 const CallBase *Call) {
2766 assert(Operands.size() == 3 && "Wrong number of operands.")(static_cast <bool> (Operands.size() == 3 && "Wrong number of operands."
) ? void (0) : __assert_fail ("Operands.size() == 3 && \"Wrong number of operands.\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2766, __extension__
__PRETTY_FUNCTION__))
;
2767
2768 if (const auto *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
2769 if (const auto *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
2770 if (const auto *Op3 = dyn_cast<ConstantFP>(Operands[2])) {
2771 const APFloat &C1 = Op1->getValueAPF();
2772 const APFloat &C2 = Op2->getValueAPF();
2773 const APFloat &C3 = Op3->getValueAPF();
2774
2775 if (const auto *ConstrIntr = dyn_cast<ConstrainedFPIntrinsic>(Call)) {
2776 RoundingMode RM = getEvaluationRoundingMode(ConstrIntr);
2777 APFloat Res = C1;
2778 APFloat::opStatus St;
2779 switch (IntrinsicID) {
2780 default:
2781 return nullptr;
2782 case Intrinsic::experimental_constrained_fma:
2783 case Intrinsic::experimental_constrained_fmuladd:
2784 St = Res.fusedMultiplyAdd(C2, C3, RM);
2785 break;
2786 }
2787 if (mayFoldConstrained(
2788 const_cast<ConstrainedFPIntrinsic *>(ConstrIntr), St))
2789 return ConstantFP::get(Ty->getContext(), Res);
2790 return nullptr;
2791 }
2792
2793 switch (IntrinsicID) {
2794 default: break;
2795 case Intrinsic::amdgcn_fma_legacy: {
2796 // The legacy behaviour is that multiplying +/- 0.0 by anything, even
2797 // NaN or infinity, gives +0.0.
2798 if (C1.isZero() || C2.isZero()) {
2799 // It's tempting to just return C3 here, but that would give the
2800 // wrong result if C3 was -0.0.
2801 return ConstantFP::get(Ty->getContext(), APFloat(0.0f) + C3);
2802 }
2803 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2804 }
2805 case Intrinsic::fma:
2806 case Intrinsic::fmuladd: {
2807 APFloat V = C1;
2808 V.fusedMultiplyAdd(C2, C3, APFloat::rmNearestTiesToEven);
2809 return ConstantFP::get(Ty->getContext(), V);
2810 }
2811 case Intrinsic::amdgcn_cubeid:
2812 case Intrinsic::amdgcn_cubema:
2813 case Intrinsic::amdgcn_cubesc:
2814 case Intrinsic::amdgcn_cubetc: {
2815 APFloat V = ConstantFoldAMDGCNCubeIntrinsic(IntrinsicID, C1, C2, C3);
2816 return ConstantFP::get(Ty->getContext(), V);
2817 }
2818 }
2819 }
2820 }
2821 }
2822
2823 if (IntrinsicID == Intrinsic::smul_fix ||
2824 IntrinsicID == Intrinsic::smul_fix_sat) {
2825 // poison * C -> poison
2826 // C * poison -> poison
2827 if (isa<PoisonValue>(Operands[0]) || isa<PoisonValue>(Operands[1]))
2828 return PoisonValue::get(Ty);
2829
2830 const APInt *C0, *C1;
2831 if (!getConstIntOrUndef(Operands[0], C0) ||
2832 !getConstIntOrUndef(Operands[1], C1))
2833 return nullptr;
2834
2835 // undef * C -> 0
2836 // C * undef -> 0
2837 if (!C0 || !C1)
2838 return Constant::getNullValue(Ty);
2839
2840 // This code performs rounding towards negative infinity in case the result
2841 // cannot be represented exactly for the given scale. Targets that do care
2842 // about rounding should use a target hook for specifying how rounding
2843 // should be done, and provide their own folding to be consistent with
2844 // rounding. This is the same approach as used by
2845 // DAGTypeLegalizer::ExpandIntRes_MULFIX.
2846 unsigned Scale = cast<ConstantInt>(Operands[2])->getZExtValue();
2847 unsigned Width = C0->getBitWidth();
2848 assert(Scale < Width && "Illegal scale.")(static_cast <bool> (Scale < Width && "Illegal scale."
) ? void (0) : __assert_fail ("Scale < Width && \"Illegal scale.\""
, "llvm/lib/Analysis/ConstantFolding.cpp", 2848, __extension__
__PRETTY_FUNCTION__))
;
2849 unsigned ExtendedWidth = Width * 2;
2850 APInt Product = (C0->sextOrSelf(ExtendedWidth) *
2851 C1->sextOrSelf(ExtendedWidth)).ashr(Scale);
2852 if (IntrinsicID == Intrinsic::smul_fix_sat) {
2853 APInt Max = APInt::getSignedMaxValue(Width).sextOrSelf(ExtendedWidth);
2854 APInt Min = APInt::getSignedMinValue(Width).sextOrSelf(ExtendedWidth);
2855 Product = APIntOps::smin(Product, Max);
2856 Product = APIntOps::smax(Product, Min);
2857 }
2858 return ConstantInt::get(Ty->getContext(), Product.sextOrTrunc(Width));
2859 }
2860
2861 if (IntrinsicID == Intrinsic::fshl || IntrinsicID == Intrinsic::fshr) {
2862 const APInt *C0, *C1, *C2;
2863 if (!getConstIntOrUndef(Operands[0], C0) ||
2864 !getConstIntOrUndef(Operands[1], C1) ||
2865 !getConstIntOrUndef(Operands[2], C2))
2866 return nullptr;
2867
2868 bool IsRight = IntrinsicID == Intrinsic::fshr;
2869 if (!C2)
2870 return Operands[IsRight ? 1 : 0];
2871 if (!C0 && !C1)
2872 return UndefValue::get(Ty);
2873
2874 // The shift amount is interpreted as modulo the bitwidth. If the shift
2875 // amount is effectively 0, avoid UB due to oversized inverse shift below.
2876 unsigned BitWidth = C2->getBitWidth();
2877 unsigned ShAmt = C2->urem(BitWidth);
2878 if (!ShAmt)
2879 return Operands[IsRight ? 1 : 0];
2880
2881 // (C0 << ShlAmt) | (C1 >> LshrAmt)
2882 unsigned LshrAmt = IsRight ? ShAmt : BitWidth - ShAmt;
2883 unsigned ShlAmt = !IsRight ? ShAmt : BitWidth - ShAmt;
2884 if (!C0)
2885 return ConstantInt::get(Ty, C1->lshr(LshrAmt));
2886 if (!C1)
2887 return ConstantInt::get(Ty, C0->shl(ShlAmt));
2888 return ConstantInt::get(Ty, C0->shl(ShlAmt) | C1->lshr(LshrAmt));
2889 }
2890
2891 if (IntrinsicID == Intrinsic::amdgcn_perm)
2892 return ConstantFoldAMDGCNPermIntrinsic(Operands, Ty);
2893
2894 return nullptr;
2895}
2896
2897static Constant *ConstantFoldScalarCall(StringRef Name,
2898 Intrinsic::ID IntrinsicID,
2899 Type *Ty,
2900 ArrayRef<Constant *> Operands,
2901 const TargetLibraryInfo *TLI,
2902 const CallBase *Call) {
2903 if (Operands.size() == 1)
2904 return ConstantFoldScalarCall1(Name, IntrinsicID, Ty, Operands, TLI, Call);
2905
2906 if (Operands.size() == 2)
2907 return ConstantFoldScalarCall2(Name, IntrinsicID, Ty, Operands, TLI, Call);
2908
2909 if (Operands.size() == 3)
2910 return ConstantFoldScalarCall3(Name, IntrinsicID, Ty, Operands, TLI, Call);
2911
2912 return nullptr;
2913}
2914
2915static Constant *ConstantFoldFixedVectorCall(
2916 StringRef Name, Intrinsic::ID IntrinsicID, FixedVectorType *FVTy,
2917 ArrayRef<Constant *> Operands, const DataLayout &DL,
2918 const TargetLibraryInfo *TLI, const CallBase *Call) {
2919 SmallVector<Constant *, 4> Result(FVTy->getNumElements());
2920 SmallVector<Constant *, 4> Lane(Operands.size());
2921 Type *Ty = FVTy->getElementType();
2922
2923 switch (IntrinsicID) {
2924 case Intrinsic::masked_load: {
2925 auto *SrcPtr = Operands[0];
2926 auto *Mask = Operands[2];
2927 auto *Passthru = Operands[3];
2928
2929 Constant *VecData = ConstantFoldLoadFromConstPtr(SrcPtr, FVTy, DL);
2930
2931 SmallVector<Constant *, 32> NewElements;
2932 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
2933 auto *MaskElt = Mask->getAggregateElement(I);
2934 if (!MaskElt)
2935 break;
2936 auto *PassthruElt = Passthru->getAggregateElement(I);
2937 auto *VecElt = VecData ? VecData->getAggregateElement(I) : nullptr;
2938 if (isa<UndefValue>(MaskElt)) {
2939 if (PassthruElt)
2940 NewElements.push_back(PassthruElt);
2941 else if (VecElt)
2942 NewElements.push_back(VecElt);
2943 else
2944 return nullptr;
2945 }
2946 if (MaskElt->isNullValue()) {
2947 if (!PassthruElt)
2948 return nullptr;
2949 NewElements.push_back(PassthruElt);
2950 } else if (MaskElt->isOneValue()) {
2951 if (!VecElt)
2952 return nullptr;
2953 NewElements.push_back(VecElt);
2954 } else {
2955 return nullptr;
2956 }
2957 }
2958 if (NewElements.size() != FVTy->getNumElements())
2959 return nullptr;
2960 return ConstantVector::get(NewElements);
2961 }
2962 case Intrinsic::arm_mve_vctp8:
2963 case Intrinsic::arm_mve_vctp16:
2964 case Intrinsic::arm_mve_vctp32:
2965 case Intrinsic::arm_mve_vctp64: {
2966 if (auto *Op = dyn_cast<ConstantInt>(Operands[0])) {
2967 unsigned Lanes = FVTy->getNumElements();
2968 uint64_t Limit = Op->getZExtValue();
2969
2970 SmallVector<Constant *, 16> NCs;
2971 for (unsigned i = 0; i < Lanes; i++) {
2972 if (i < Limit)
2973 NCs.push_back(ConstantInt::getTrue(Ty));
2974 else
2975 NCs.push_back(ConstantInt::getFalse(Ty));
2976 }
2977 return ConstantVector::get(NCs);
2978 }
2979 break;
2980 }
2981 case Intrinsic::get_active_lane_mask: {
2982 auto *Op0 = dyn_cast<ConstantInt>(Operands[0]);
2983 auto *Op1 = dyn_cast<ConstantInt>(Operands[1]);
2984 if (Op0 && Op1) {
2985 unsigned Lanes = FVTy->getNumElements();
2986 uint64_t Base = Op0->getZExtValue();
2987 uint64_t Limit = Op1->getZExtValue();
2988
2989 SmallVector<Constant *, 16> NCs;
2990 for (unsigned i = 0; i < Lanes; i++) {
2991 if (Base + i < Limit)
2992 NCs.push_back(ConstantInt::getTrue(Ty));
2993 else
2994 NCs.push_back(ConstantInt::getFalse(Ty));
2995 }
2996 return ConstantVector::get(NCs);
2997 }
2998 break;
2999 }
3000 default:
3001 break;
3002 }
3003
3004 for (unsigned I = 0, E = FVTy->getNumElements(); I != E; ++I) {
3005 // Gather a column of constants.
3006 for (unsigned J = 0, JE = Operands.size(); J != JE; ++J) {
3007 // Some intrinsics use a scalar type for certain arguments.
3008 if (hasVectorInstrinsicScalarOpd(IntrinsicID, J)) {
3009 Lane[J] = Operands[J];
3010 continue;
3011 }
3012
3013 Constant *Agg = Operands[J]->getAggregateElement(I);
3014 if (!Agg)
3015 return nullptr;
3016
3017 Lane[J] = Agg;
3018 }
3019
3020 // Use the regular scalar folding to simplify this column.
3021 Constant *Folded =
3022 ConstantFoldScalarCall(Name, IntrinsicID, Ty, Lane, TLI, Call);
3023 if (!Folded)
3024 return nullptr;
3025 Result[I] = Folded;
3026 }
3027
3028 return ConstantVector::get(Result);
3029}
3030
3031static Constant *ConstantFoldScalableVectorCall(
3032 StringRef Name, Intrinsic::ID IntrinsicID, ScalableVectorType *SVTy,
3033 ArrayRef<Constant *> Operands, const DataLayout &DL,
3034 const TargetLibraryInfo *TLI, const CallBase *Call) {
3035 switch (IntrinsicID) {
3036 case Intrinsic::aarch64_sve_convert_from_svbool: {
3037 auto *Src = dyn_cast<Constant>(Operands[0]);
3038 if (!Src || !Src->isNullValue())
3039 break;
3040
3041 return ConstantInt::getFalse(SVTy);
3042 }
3043 default:
3044 break;
3045 }
3046 return nullptr;
3047}
3048
3049} // end anonymous namespace
3050
3051Constant *llvm::ConstantFoldCall(const CallBase *Call, Function *F,
3052 ArrayRef<Constant *> Operands,
3053 const TargetLibraryInfo *TLI) {
3054 if (Call->isNoBuiltin())
3055 return nullptr;
3056 if (!F->hasName())
3057 return nullptr;
3058
3059 // If this is not an intrinsic and not recognized as a library call, bail out.
3060 if (F->getIntrinsicID() == Intrinsic::not_intrinsic) {
3061 if (!TLI)
3062 return nullptr;
3063 LibFunc LibF;
3064 if (!TLI->getLibFunc(*F, LibF))
3065 return nullptr;
3066 }
3067
3068 StringRef Name = F->getName();
3069 Type *Ty = F->getReturnType();
3070 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty))
3071 return ConstantFoldFixedVectorCall(
3072 Name, F->getIntrinsicID(), FVTy, Operands,
3073 F->getParent()->getDataLayout(), TLI, Call);
3074
3075 if (auto *SVTy = dyn_cast<ScalableVectorType>(Ty))
3076 return ConstantFoldScalableVectorCall(
3077 Name, F->getIntrinsicID(), SVTy, Operands,
3078 F->getParent()->getDataLayout(), TLI, Call);
3079
3080 // TODO: If this is a library function, we already discovered that above,
3081 // so we should pass the LibFunc, not the name (and it might be better
3082 // still to separate intrinsic handling from libcalls).
3083 return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI,
3084 Call);
3085}
3086
3087bool llvm::isMathLibCallNoop(const CallBase *Call,
3088 const TargetLibraryInfo *TLI) {
3089 // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
3090 // (and to some extent ConstantFoldScalarCall).
3091 if (Call->isNoBuiltin() || Call->isStrictFP())
3092 return false;
3093 Function *F = Call->getCalledFunction();
3094 if (!F)
3095 return false;
3096
3097 LibFunc Func;
3098 if (!TLI || !TLI->getLibFunc(*F, Func))
3099 return false;
3100
3101 if (Call->arg_size() == 1) {
3102 if (ConstantFP *OpC = dyn_cast<ConstantFP>(Call->getArgOperand(0))) {
3103 const APFloat &Op = OpC->getValueAPF();
3104 switch (Func) {
3105 case LibFunc_logl:
3106 case LibFunc_log:
3107 case LibFunc_logf:
3108 case LibFunc_log2l:
3109 case LibFunc_log2:
3110 case LibFunc_log2f:
3111 case LibFunc_log10l:
3112 case LibFunc_log10:
3113 case LibFunc_log10f:
3114 return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
3115
3116 case LibFunc_expl:
3117 case LibFunc_exp:
3118 case LibFunc_expf:
3119 // FIXME: These boundaries are slightly conservative.
3120 if (OpC->getType()->isDoubleTy())
3121 return !(Op < APFloat(-745.0) || Op > APFloat(709.0));
3122 if (OpC->getType()->isFloatTy())
3123 return !(Op < APFloat(-103.0f) || Op > APFloat(88.0f));
3124 break;
3125
3126 case LibFunc_exp2l:
3127 case LibFunc_exp2:
3128 case LibFunc_exp2f:
3129 // FIXME: These boundaries are slightly conservative.
3130 if (OpC->getType()->isDoubleTy())
3131 return !(Op < APFloat(-1074.0) || Op > APFloat(1023.0));
3132 if (OpC->getType()->isFloatTy())
3133 return !(Op < APFloat(-149.0f) || Op > APFloat(127.0f));
3134 break;
3135
3136 case LibFunc_sinl:
3137 case LibFunc_sin:
3138 case LibFunc_sinf:
3139 case LibFunc_cosl:
3140 case LibFunc_cos:
3141 case LibFunc_cosf:
3142 return !Op.isInfinity();
3143
3144 case LibFunc_tanl:
3145 case LibFunc_tan:
3146 case LibFunc_tanf: {
3147 // FIXME: Stop using the host math library.
3148 // FIXME: The computation isn't done in the right precision.
3149 Type *Ty = OpC->getType();
3150 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy())
3151 return ConstantFoldFP(tan, OpC->getValueAPF(), Ty) != nullptr;
3152 break;
3153 }
3154
3155 case LibFunc_asinl:
3156 case LibFunc_asin:
3157 case LibFunc_asinf:
3158 case LibFunc_acosl:
3159 case LibFunc_acos:
3160 case LibFunc_acosf:
3161 return !(Op < APFloat(Op.getSemantics(), "-1") ||
3162 Op > APFloat(Op.getSemantics(), "1"));
3163
3164 case LibFunc_sinh:
3165 case LibFunc_cosh:
3166 case LibFunc_sinhf:
3167 case LibFunc_coshf:
3168 case LibFunc_sinhl:
3169 case LibFunc_coshl:
3170 // FIXME: These boundaries are slightly conservative.
3171 if (OpC->getType()->isDoubleTy())
3172 return !(Op < APFloat(-710.0) || Op > APFloat(710.0));
3173 if (OpC->getType()->isFloatTy())
3174 return !(Op < APFloat(-89.0f) || Op > APFloat(89.0f));
3175 break;
3176
3177 case LibFunc_sqrtl:
3178 case LibFunc_sqrt:
3179 case LibFunc_sqrtf:
3180 return Op.isNaN() || Op.isZero() || !Op.isNegative();
3181
3182 // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
3183 // maybe others?
3184 default:
3185 break;
3186 }
3187 }
3188 }
3189
3190 if (Call->arg_size() == 2) {
3191 ConstantFP *Op0C = dyn_cast<ConstantFP>(Call->getArgOperand(0));
3192 ConstantFP *Op1C = dyn_cast<ConstantFP>(Call->getArgOperand(1));
3193 if (Op0C && Op1C) {
3194 const APFloat &Op0 = Op0C->getValueAPF();
3195 const APFloat &Op1 = Op1C->getValueAPF();
3196
3197 switch (Func) {
3198 case LibFunc_powl:
3199 case LibFunc_pow:
3200 case LibFunc_powf: {
3201 // FIXME: Stop using the host math library.
3202 // FIXME: The computation isn't done in the right precision.
3203 Type *Ty = Op0C->getType();
3204 if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
3205 if (Ty == Op1C->getType())
3206 return ConstantFoldBinaryFP(pow, Op0, Op1, Ty) != nullptr;
3207 }
3208 break;
3209 }
3210
3211 case LibFunc_fmodl:
3212 case LibFunc_fmod:
3213 case LibFunc_fmodf:
3214 case LibFunc_remainderl:
3215 case LibFunc_remainder:
3216 case LibFunc_remainderf:
3217 return Op0.isNaN() || Op1.isNaN() ||
3218 (!Op0.isInfinity() && !Op1.isZero());
3219
3220 default:
3221 break;
3222 }
3223 }
3224 }
3225
3226 return false;
3227}
3228
3229void TargetFolder::anchor() {}

/build/llvm-toolchain-snapshot-14~++20220119111520+da61cb019eb2/llvm/include/llvm/ADT/APInt.h

1//===-- llvm/ADT/APInt.h - For Arbitrary Precision Integer -----*- C++ -*--===//
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/// \file
10/// This file implements a class to represent arbitrary precision
11/// integral constant values and operations on them.
12///
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_ADT_APINT_H
16#define LLVM_ADT_APINT_H
17
18#include "llvm/Support/Compiler.h"
19#include "llvm/Support/MathExtras.h"
20#include <cassert>
21#include <climits>
22#include <cstring>
23#include <utility>
24
25namespace llvm {
26class FoldingSetNodeID;
27class StringRef;
28class hash_code;
29class raw_ostream;
30
31template <typename T> class SmallVectorImpl;
32template <typename T> class ArrayRef;
33template <typename T> class Optional;
34template <typename T, typename Enable> struct DenseMapInfo;
35
36class APInt;
37
38inline APInt operator-(APInt);
39
40//===----------------------------------------------------------------------===//
41// APInt Class
42//===----------------------------------------------------------------------===//
43
44/// Class for arbitrary precision integers.
45///
46/// APInt is a functional replacement for common case unsigned integer type like
47/// "unsigned", "unsigned long" or "uint64_t", but also allows non-byte-width
48/// integer sizes and large integer value types such as 3-bits, 15-bits, or more
49/// than 64-bits of precision. APInt provides a variety of arithmetic operators
50/// and methods to manipulate integer values of any bit-width. It supports both
51/// the typical integer arithmetic and comparison operations as well as bitwise
52/// manipulation.
53///
54/// The class has several invariants worth noting:
55/// * All bit, byte, and word positions are zero-based.
56/// * Once the bit width is set, it doesn't change except by the Truncate,
57/// SignExtend, or ZeroExtend operations.
58/// * All binary operators must be on APInt instances of the same bit width.
59/// Attempting to use these operators on instances with different bit
60/// widths will yield an assertion.
61/// * The value is stored canonically as an unsigned value. For operations
62/// where it makes a difference, there are both signed and unsigned variants
63/// of the operation. For example, sdiv and udiv. However, because the bit
64/// widths must be the same, operations such as Mul and Add produce the same
65/// results regardless of whether the values are interpreted as signed or
66/// not.
67/// * In general, the class tries to follow the style of computation that LLVM
68/// uses in its IR. This simplifies its use for LLVM.
69/// * APInt supports zero-bit-width values, but operations that require bits
70/// are not defined on it (e.g. you cannot ask for the sign of a zero-bit
71/// integer). This means that operations like zero extension and logical
72/// shifts are defined, but sign extension and ashr is not. Zero bit values
73/// compare and hash equal to themselves, and countLeadingZeros returns 0.
74///
75class LLVM_NODISCARD[[clang::warn_unused_result]] APInt {
76public:
77 typedef uint64_t WordType;
78
79 /// This enum is used to hold the constants we needed for APInt.
80 enum : unsigned {
81 /// Byte size of a word.
82 APINT_WORD_SIZE = sizeof(WordType),
83 /// Bits in a word.
84 APINT_BITS_PER_WORD = APINT_WORD_SIZE * CHAR_BIT8
85 };
86
87 enum class Rounding {
88 DOWN,
89 TOWARD_ZERO,
90 UP,
91 };
92
93 static constexpr WordType WORDTYPE_MAX = ~WordType(0);
94
95 /// \name Constructors
96 /// @{
97
98 /// Create a new APInt of numBits width, initialized as val.
99 ///
100 /// If isSigned is true then val is treated as if it were a signed value
101 /// (i.e. as an int64_t) and the appropriate sign extension to the bit width
102 /// will be done. Otherwise, no sign extension occurs (high order bits beyond
103 /// the range of val are zero filled).
104 ///
105 /// \param numBits the bit width of the constructed APInt
106 /// \param val the initial value of the APInt
107 /// \param isSigned how to treat signedness of val
108 APInt(unsigned numBits, uint64_t val, bool isSigned = false)
109 : BitWidth(numBits) {
110 if (isSingleWord()) {
111 U.VAL = val;
112 clearUnusedBits();
113 } else {
114 initSlowCase(val, isSigned);
115 }
116 }
117
118 /// Construct an APInt of numBits width, initialized as bigVal[].
119 ///
120 /// Note that bigVal.size() can be smaller or larger than the corresponding
121 /// bit width but any extraneous bits will be dropped.
122 ///
123 /// \param numBits the bit width of the constructed APInt
124 /// \param bigVal a sequence of words to form the initial value of the APInt
125 APInt(unsigned numBits, ArrayRef<uint64_t> bigVal);
126
127 /// Equivalent to APInt(numBits, ArrayRef<uint64_t>(bigVal, numWords)), but
128 /// deprecated because this constructor is prone to ambiguity with the
129 /// APInt(unsigned, uint64_t, bool) constructor.
130 ///
131 /// If this overload is ever deleted, care should be taken to prevent calls
132 /// from being incorrectly captured by the APInt(unsigned, uint64_t, bool)
133 /// constructor.
134 APInt(unsigned numBits, unsigned numWords, const uint64_t bigVal[]);
135
136 /// Construct an APInt from a string representation.
137 ///
138 /// This constructor interprets the string \p str in the given radix. The
139 /// interpretation stops when the first character that is not suitable for the
140 /// radix is encountered, or the end of the string. Acceptable radix values
141 /// are 2, 8, 10, 16, and 36. It is an error for the value implied by the
142 /// string to require more bits than numBits.
143 ///
144 /// \param numBits the bit width of the constructed APInt
145 /// \param str the string to be interpreted
146 /// \param radix the radix to use for the conversion
147 APInt(unsigned numBits, StringRef str, uint8_t radix);
148
149 /// Default constructor that creates an APInt with a 1-bit zero value.
150 explicit APInt() : BitWidth(1) { U.VAL = 0; }
151
152 /// Copy Constructor.
153 APInt(const APInt &that) : BitWidth(that.BitWidth) {
154 if (isSingleWord())
155 U.VAL = that.U.VAL;
156 else
157 initSlowCase(that);
158 }
159
160 /// Move Constructor.
161 APInt(APInt &&that) : BitWidth(that.BitWidth) {
162 memcpy(&U, &that.U, sizeof(U));
163 that.BitWidth = 0;
164 }
165
166 /// Destructor.
167 ~APInt() {
168 if (needsCleanup())
169 delete[] U.pVal;
170 }
171
172 /// @}
173 /// \name Value Generators
174 /// @{
175
176 /// Get the '0' value for the specified bit-width.
177 static APInt getZero(unsigned numBits) { return APInt(numBits, 0); }
178
179 /// NOTE: This is soft-deprecated. Please use `getZero()` instead.
180 static APInt getNullValue(unsigned numBits) { return getZero(numBits); }
181
182 /// Return an APInt zero bits wide.
183 static APInt getZeroWidth() { return getZero(0); }
184
185 /// Gets maximum unsigned value of APInt for specific bit width.
186 static APInt getMaxValue(unsigned numBits) { return getAllOnes(numBits); }
187
188 /// Gets maximum signed value of APInt for a specific bit width.
189 static APInt getSignedMaxValue(unsigned numBits) {
190 APInt API = getAllOnes(numBits);
191 API.clearBit(numBits - 1);
192 return API;
193 }
194
195 /// Gets minimum unsigned value of APInt for a specific bit width.
196 static APInt getMinValue(unsigned numBits) { return APInt(numBits, 0); }
197
198 /// Gets minimum signed value of APInt for a specific bit width.
199 static APInt getSignedMinValue(unsigned numBits) {
200 APInt API(numBits, 0);
201 API.setBit(numBits - 1);
202 return API;
203 }
204
205 /// Get the SignMask for a specific bit width.
206 ///
207 /// This is just a wrapper function of getSignedMinValue(), and it helps code
208 /// readability when we want to get a SignMask.
209 static APInt getSignMask(unsigned BitWidth) {
210 return getSignedMinValue(BitWidth);
211 }
212
213 /// Return an APInt of a specified width with all bits set.
214 static APInt getAllOnes(unsigned numBits) {
215 return APInt(numBits, WORDTYPE_MAX, true);
216 }
217
218 /// NOTE: This is soft-deprecated. Please use `getAllOnes()` instead.
219 static APInt getAllOnesValue(unsigned numBits) { return getAllOnes(numBits); }
220
221 /// Return an APInt with exactly one bit set in the result.
222 static APInt getOneBitSet(unsigned numBits, unsigned BitNo) {
223 APInt Res(numBits, 0);
224 Res.setBit(BitNo);
225 return Res;
226 }
227
228 /// Get a value with a block of bits set.
229 ///
230 /// Constructs an APInt value that has a contiguous range of bits set. The
231 /// bits from loBit (inclusive) to hiBit (exclusive) will be set. All other
232 /// bits will be zero. For example, with parameters(32, 0, 16) you would get
233 /// 0x0000FFFF. Please call getBitsSetWithWrap if \p loBit may be greater than
234 /// \p hiBit.
235 ///
236 /// \param numBits the intended bit width of the result
237 /// \param loBit the index of the lowest bit set.
238 /// \param hiBit the index of the highest bit set.
239 ///
240 /// \returns An APInt value with the requested bits set.
241 static APInt getBitsSet(unsigned numBits, unsigned loBit, unsigned hiBit) {
242 APInt Res(numBits, 0);
243 Res.setBits(loBit, hiBit);
244 return Res;
245 }
246
247 /// Wrap version of getBitsSet.
248 /// If \p hiBit is bigger than \p loBit, this is same with getBitsSet.
249 /// If \p hiBit is not bigger than \p loBit, the set bits "wrap". For example,
250 /// with parameters (32, 28, 4), you would get 0xF000000F.
251 /// If \p hiBit is equal to \p loBit, you would get a result with all bits
252 /// set.
253 static APInt getBitsSetWithWrap(unsigned numBits, unsigned loBit,
254 unsigned hiBit) {
255 APInt Res(numBits, 0);
256 Res.setBitsWithWrap(loBit, hiBit);
257 return Res;
258 }
259
260 /// Constructs an APInt value that has a contiguous range of bits set. The
261 /// bits from loBit (inclusive) to numBits (exclusive) will be set. All other
262 /// bits will be zero. For example, with parameters(32, 12) you would get
263 /// 0xFFFFF000.
264 ///
265 /// \param numBits the intended bit width of the result
266 /// \param loBit the index of the lowest bit to set.
267 ///
268 /// \returns An APInt value with the requested bits set.
269 static APInt getBitsSetFrom(unsigned numBits, unsigned loBit) {
270 APInt Res(numBits, 0);
271 Res.setBitsFrom(loBit);
272 return Res;
273 }
274
275 /// Constructs an APInt value that has the top hiBitsSet bits set.
276 ///
277 /// \param numBits the bitwidth of the result
278 /// \param hiBitsSet the number of high-order bits set in the result.
279 static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet) {
280 APInt Res(numBits, 0);
281 Res.setHighBits(hiBitsSet);
282 return Res;
283 }
284
285 /// Constructs an APInt value that has the bottom loBitsSet bits set.
286 ///
287 /// \param numBits the bitwidth of the result
288 /// \param loBitsSet the number of low-order bits set in the result.
289 static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet) {
290 APInt Res(numBits, 0);
291 Res.setLowBits(loBitsSet);
292 return Res;
293 }
294
295 /// Return a value containing V broadcasted over NewLen bits.
296 static APInt getSplat(unsigned NewLen, const APInt &V);
297
298 /// @}
299 /// \name Value Tests
300 /// @{
301
302 /// Determine if this APInt just has one word to store value.
303 ///
304 /// \returns true if the number of bits <= 64, false otherwise.
305 bool isSingleWord() const { return BitWidth <= APINT_BITS_PER_WORD; }
306
307 /// Determine sign of this APInt.
308 ///
309 /// This tests the high bit of this APInt to determine if it is set.
310 ///
311 /// \returns true if this APInt is negative, false otherwise
312 bool isNegative() const { return (*this)[BitWidth - 1]; }
313
314 /// Determine if this APInt Value is non-negative (>= 0)
315 ///
316 /// This tests the high bit of the APInt to determine if it is unset.
317 bool isNonNegative() const { return !isNegative(); }
318
319 /// Determine if sign bit of this APInt is set.
320 ///
321 /// This tests the high bit of this APInt to determine if it is set.
322 ///
323 /// \returns true if this APInt has its sign bit set, false otherwise.
324 bool isSignBitSet() const { return (*this)[BitWidth - 1]; }
325
326 /// Determine if sign bit of this APInt is clear.
327 ///
328 /// This tests the high bit of this APInt to determine if it is clear.
329 ///
330 /// \returns true if this APInt has its sign bit clear, false otherwise.
331 bool isSignBitClear() const { return !isSignBitSet(); }
332
333 /// Determine if this APInt Value is positive.
334 ///
335 /// This tests if the value of this APInt is positive (> 0). Note
336 /// that 0 is not a positive value.
337 ///
338 /// \returns true if this APInt is positive.
339 bool isStrictlyPositive() const { return isNonNegative() && !isZero(); }
340
341 /// Determine if this APInt Value is non-positive (<= 0).
342 ///
343 /// \returns true if this APInt is non-positive.
344 bool isNonPositive() const { return !isStrictlyPositive(); }
345
346 /// Determine if all bits are set. This is true for zero-width values.
347 bool isAllOnes() const {
348 if (BitWidth == 0)
349 return true;
350 if (isSingleWord())
351 return U.VAL == WORDTYPE_MAX >> (APINT_BITS_PER_WORD - BitWidth);
352 return countTrailingOnesSlowCase() == BitWidth;
353 }
354
355 /// NOTE: This is soft-deprecated. Please use `isAllOnes()` instead.
356 bool isAllOnesValue() const { return isAllOnes(); }
357
358 /// Determine if this value is zero, i.e. all bits are clear.
359 bool isZero() const {
360 if (isSingleWord())
5
Taking true branch
361 return U.VAL == 0;
6
Returning the value 1, which participates in a condition later
362 return countLeadingZerosSlowCase() == BitWidth;
363 }
364
365 /// NOTE: This is soft-deprecated. Please use `isZero()` instead.
366 bool isNullValue() const { return isZero(); }
367
368 /// Determine if this is a value of 1.
369 ///
370 /// This checks to see if the value of this APInt is one.
371 bool isOne() const {
372 if (isSingleWord())
373 return U.VAL == 1;
374 return countLeadingZerosSlowCase() == BitWidth - 1;
375 }
376
377 /// NOTE: This is soft-deprecated. Please use `isOne()` instead.
378 bool isOneValue() const { return isOne(); }
379
380 /// Determine if this is the largest unsigned value.
381 ///
382 /// This checks to see if the value of this APInt is the maximum unsigned
383 /// value for the APInt's bit width.
384 bool isMaxValue() const { return isAllOnes(); }
385
386 /// Determine if this is the largest signed value.
387 ///
388 /// This checks to see if the value of this APInt is the maximum signed
389 /// value for the APInt's bit width.
390 bool isMaxSignedValue() const {
391 if (isSingleWord()) {
392 assert(BitWidth && "zero width values not allowed")(static_cast <bool> (BitWidth && "zero width values not allowed"
) ? void (0) : __assert_fail ("BitWidth && \"zero width values not allowed\""
, "llvm/include/llvm/ADT/APInt.h", 392, __extension__ __PRETTY_FUNCTION__
))
;
393 return U.VAL == ((WordType(1) << (BitWidth - 1)) - 1);
394 }
395 return !isNegative() && countTrailingOnesSlowCase() == BitWidth - 1;
396 }
397
398 /// Determine if this is the smallest unsigned value.
399 ///
400 /// This checks to see if the value of this APInt is the minimum unsigned
401 /// value for the APInt's bit width.
402 bool isMinValue() const { return isZero(); }
403
404 /// Determine if this is the smallest signed value.
405 ///
406 /// This checks to see if the value of this APInt is the minimum signed
407 /// value for the APInt's bit width.
408 bool isMinSignedValue() const {
409 if (isSingleWord()) {
410 assert(BitWidth && "zero width values not allowed")(static_cast <bool> (BitWidth && "zero width values not allowed"
) ? void (0) : __assert_fail ("BitWidth && \"zero width values not allowed\""
, "llvm/include/llvm/ADT/APInt.h", 410, __extension__ __PRETTY_FUNCTION__
))
;
411 return U.VAL == (WordType(1) << (BitWidth - 1));
412 }
413 return isNegative() && countTrailingZerosSlowCase() == BitWidth - 1;
414 }
415
416 /// Check if this APInt has an N-bits unsigned integer value.
417 bool isIntN(unsigned N) const { return getActiveBits() <= N; }
418
419 /// Check if this APInt has an N-bits signed integer value.
420 bool isSignedIntN(unsigned N) const { return getSignificantBits() <= N; }
421
422 /// Check if this APInt's value is a power of two greater than zero.
423 ///
424 /// \returns true if the argument APInt value is a power of two > 0.
425 bool isPowerOf2() const {
426 if (isSingleWord()) {
427 assert(BitWidth && "zero width values not allowed")(static_cast <bool> (BitWidth && "zero width values not allowed"
) ? void (0) : __assert_fail ("BitWidth && \"zero width values not allowed\""
, "llvm/include/llvm/ADT/APInt.h", 427, __extension__ __PRETTY_FUNCTION__
))
;
428 return isPowerOf2_64(U.VAL);
429 }
430 return countPopulationSlowCase() == 1;
431 }
432
433 /// Check if this APInt's negated value is a power of two greater than zero.
434 bool isNegatedPowerOf2() const {
435 assert(BitWidth && "zero width values not allowed")(static_cast <bool> (BitWidth && "zero width values not allowed"
) ? void (0) : __assert_fail ("BitWidth && \"zero width values not allowed\""
, "llvm/include/llvm/ADT/APInt.h", 435, __extension__ __PRETTY_FUNCTION__
))
;
436 if (isNonNegative())
437 return false;
438 // NegatedPowerOf2 - shifted mask in the top bits.
439 unsigned LO = countLeadingOnes();
440 unsigned TZ = countTrailingZeros();
441 return (LO + TZ) == BitWidth;
442 }
443
444 /// Check if the APInt's value is returned by getSignMask.
445 ///
446 /// \returns true if this is the value returned by getSignMask.
447 bool isSignMask() const { return isMinSignedValue(); }
448
449 /// Convert APInt to a boolean value.
450 ///
451 /// This converts the APInt to a boolean value as a test against zero.
452 bool getBoolValue() const { return !isZero(); }
453
454 /// If this value is smaller than the specified limit, return it, otherwise
455 /// return the limit value. This causes the value to saturate to the limit.
456 uint64_t getLimitedValue(uint64_t Limit = UINT64_MAX(18446744073709551615UL)) const {
457 return ugt(Limit) ? Limit : getZExtValue();
458 }
459
460 /// Check if the APInt consists of a repeated bit pattern.
461 ///
462 /// e.g. 0x01010101 satisfies isSplat(8).
463 /// \param SplatSizeInBits The size of the pattern in bits. Must divide bit
464 /// width without remainder.
465 bool isSplat(unsigned SplatSizeInBits) const;
466
467 /// \returns true if this APInt value is a sequence of \param numBits ones
468 /// starting at the least significant bit with the remainder zero.
469 bool isMask(unsigned numBits) const {
470 assert(numBits != 0 && "numBits must be non-zero")(static_cast <bool> (numBits != 0 && "numBits must be non-zero"
) ? void (0) : __assert_fail ("numBits != 0 && \"numBits must be non-zero\""
, "llvm/include/llvm/ADT/APInt.h", 470, __extension__ __PRETTY_FUNCTION__
))
;
471 assert(numBits <= BitWidth && "numBits out of range")(static_cast <bool> (numBits <= BitWidth && "numBits out of range"
) ? void (0) : __assert_fail ("numBits <= BitWidth && \"numBits out of range\""
, "llvm/include/llvm/ADT/APInt.h", 471, __extension__ __PRETTY_FUNCTION__
))
;
472 if (isSingleWord())
473 return U.VAL == (WORDTYPE_MAX >> (APINT_BITS_PER_WORD - numBits));
474 unsigned Ones = countTrailingOnesSlowCase();
475 return (numBits == Ones) &&
476 ((Ones + countLeadingZerosSlowCase()) == BitWidth);
477 }
478
479 /// \returns true if this APInt is a non-empty sequence of ones starting at
480 /// the least significant bit with the remainder zero.
481 /// Ex. isMask(0x0000FFFFU) == true.
482 bool isMask() const {
483 if (isSingleWord())
484 return isMask_64(U.VAL);
485 unsigned Ones = countTrailingOnesSlowCase();
486 return (Ones > 0) && ((Ones + countLeadingZerosSlowCase()) == BitWidth);
487 }
488
489 /// Return true if this APInt value contains a sequence of ones with
490 /// the remainder zero.
491 bool isShiftedMask() const {
492 if (isSingleWord())
493 return isShiftedMask_64(U.VAL);
494 unsigned Ones = countPopulationSlowCase();
495 unsigned LeadZ = countLeadingZerosSlowCase();
496 return (Ones + LeadZ + countTrailingZeros()) == BitWidth;
497 }
498
499 /// Compute an APInt containing numBits highbits from this APInt.
500 ///
501 /// Get an APInt with the same BitWidth as this APInt, just zero mask the low
502 /// bits and right shift to the least significant bit.
503 ///
504 /// \returns the high "numBits" bits of this APInt.
505 APInt getHiBits(unsigned numBits) const;
506
507 /// Compute an APInt containing numBits lowbits from this APInt.
508 ///
509 /// Get an APInt with the same BitWidth as this APInt, just zero mask the high
510 /// bits.
511 ///
512 /// \returns the low "numBits" bits of this APInt.
513 APInt getLoBits(unsigned numBits) const;
514
515 /// Determine if two APInts have the same value, after zero-extending
516 /// one of them (if needed!) to ensure that the bit-widths match.
517 static bool isSameValue(const APInt &I1, const APInt &I2) {
518 if (I1.getBitWidth() == I2.getBitWidth())
519 return I1 == I2;
520
521 if (I1.getBitWidth() > I2.getBitWidth())
522 return I1 == I2.zext(I1.getBitWidth());
523
524 return I1.zext(I2.getBitWidth()) == I2;
525 }
526
527 /// Overload to compute a hash_code for an APInt value.
528 friend hash_code hash_value(const APInt &Arg);
529
530 /// This function returns a pointer to the internal storage of the APInt.
531 /// This is useful for writing out the APInt in binary form without any
532 /// conversions.
533 const uint64_t *getRawData() const {
534 if (isSingleWord())
535 return &U.VAL;
536 return &U.pVal[0];
537 }
538
539 /// @}
540 /// \name Unary Operators
541 /// @{
542
543 /// Postfix increment operator. Increment *this by 1.
544 ///
545 /// \returns a new APInt value representing the original value of *this.
546 APInt operator++(int) {
547 APInt API(*this);
548 ++(*this);
549 return API;
550 }
551
552 /// Prefix increment operator.
553 ///
554 /// \returns *this incremented by one
555 APInt &operator++();
556
557 /// Postfix decrement operator. Decrement *this by 1.
558 ///
559 /// \returns a new APInt value representing the original value of *this.
560 APInt operator--(int) {
561 APInt API(*this);
562 --(*this);
563 return API;
564 }
565
566 /// Prefix decrement operator.
567 ///
568 /// \returns *this decremented by one.
569 APInt &operator--();
570
571 /// Logical negation operation on this APInt returns true if zero, like normal
572 /// integers.
573 bool operator!() const { return isZero(); }
574
575 /// @}
576 /// \name Assignment Operators
577 /// @{
578
579 /// Copy assignment operator.
580 ///
581 /// \returns *this after assignment of RHS.
582 APInt &operator=(const APInt &RHS) {
583 // The common case (both source or dest being inline) doesn't require
584 // allocation or deallocation.
585 if (isSingleWord() && RHS.isSingleWord()) {
586 U.VAL = RHS.U.VAL;
587 BitWidth = RHS.BitWidth;
588 return *this;
589 }
590
591 assignSlowCase(RHS);
592 return *this;
593 }
594
595 /// Move assignment operator.
596 APInt &operator=(APInt &&that) {
597#ifdef EXPENSIVE_CHECKS
598 // Some std::shuffle implementations still do self-assignment.
599 if (this == &that)
600 return *this;
601#endif
602 assert(this != &that && "Self-move not supported")(static_cast <bool> (this != &that && "Self-move not supported"
) ? void (0) : __assert_fail ("this != &that && \"Self-move not supported\""
, "llvm/include/llvm/ADT/APInt.h", 602, __extension__ __PRETTY_FUNCTION__
))
;
603 if (!isSingleWord())
604 delete[] U.pVal;
605
606 // Use memcpy so that type based alias analysis sees both VAL and pVal
607 // as modified.
608 memcpy(&U, &that.U, sizeof(U));
609
610 BitWidth = that.BitWidth;
611 that.BitWidth = 0;
612 return *this;
613 }
614
615 /// Assignment operator.
616 ///
617 /// The RHS value is assigned to *this. If the significant bits in RHS exceed
618 /// the bit width, the excess bits are truncated. If the bit width is larger
619 /// than 64, the value is zero filled in the unspecified high order bits.
620 ///
621 /// \returns *this after assignment of RHS value.
622 APInt &operator=(uint64_t RHS) {
623 if (isSingleWord()) {
624 U.VAL = RHS;
625 return clearUnusedBits();
626 }
627 U.pVal[0] = RHS;
628 memset(U.pVal + 1, 0, (getNumWords() - 1) * APINT_WORD_SIZE);
629 return *this;
630 }
631
632 /// Bitwise AND assignment operator.
633 ///
634 /// Performs a bitwise AND operation on this APInt and RHS. The result is
635 /// assigned to *this.
636 ///
637 /// \returns *this after ANDing with RHS.
638 APInt &operator&=(const APInt &RHS) {
639 assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth &&
"Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\""
, "llvm/include/llvm/ADT/APInt.h", 639, __extension__ __PRETTY_FUNCTION__
))
;
640 if (isSingleWord())
641 U.VAL &= RHS.U.VAL;
642 else
643 andAssignSlowCase(RHS);
644 return *this;
645 }
646
647 /// Bitwise AND assignment operator.
648 ///
649 /// Performs a bitwise AND operation on this APInt and RHS. RHS is
650 /// logically zero-extended or truncated to match the bit-width of
651 /// the LHS.
652 APInt &operator&=(uint64_t RHS) {
653 if (isSingleWord()) {
654 U.VAL &= RHS;
655 return *this;
656 }
657 U.pVal[0] &= RHS;
658 memset(U.pVal + 1, 0, (getNumWords() - 1) * APINT_WORD_SIZE);
659 return *this;
660 }
661
662 /// Bitwise OR assignment operator.
663 ///
664 /// Performs a bitwise OR operation on this APInt and RHS. The result is
665 /// assigned *this;
666 ///
667 /// \returns *this after ORing with RHS.
668 APInt &operator|=(const APInt &RHS) {
669 assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth &&
"Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\""
, "llvm/include/llvm/ADT/APInt.h", 669, __extension__ __PRETTY_FUNCTION__
))
;
670 if (isSingleWord())
671 U.VAL |= RHS.U.VAL;
672 else
673 orAssignSlowCase(RHS);
674 return *this;
675 }
676
677 /// Bitwise OR assignment operator.
678 ///
679 /// Performs a bitwise OR operation on this APInt and RHS. RHS is
680 /// logically zero-extended or truncated to match the bit-width of
681 /// the LHS.
682 APInt &operator|=(uint64_t RHS) {
683 if (isSingleWord()) {
684 U.VAL |= RHS;
685 return clearUnusedBits();
686 }
687 U.pVal[0] |= RHS;
688 return *this;
689 }
690
691 /// Bitwise XOR assignment operator.
692 ///
693 /// Performs a bitwise XOR operation on this APInt and RHS. The result is
694 /// assigned to *this.
695 ///
696 /// \returns *this after XORing with RHS.
697 APInt &operator^=(const APInt &RHS) {
698 assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth &&
"Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\""
, "llvm/include/llvm/ADT/APInt.h", 698, __extension__ __PRETTY_FUNCTION__
))
;
699 if (isSingleWord())
700 U.VAL ^= RHS.U.VAL;
701 else
702 xorAssignSlowCase(RHS);
703 return *this;
704 }
705
706 /// Bitwise XOR assignment operator.
707 ///
708 /// Performs a bitwise XOR operation on this APInt and RHS. RHS is
709 /// logically zero-extended or truncated to match the bit-width of
710 /// the LHS.
711 APInt &operator^=(uint64_t RHS) {
712 if (isSingleWord()) {
713 U.VAL ^= RHS;
714 return clearUnusedBits();
715 }
716 U.pVal[0] ^= RHS;
717 return *this;
718 }
719
720 /// Multiplication assignment operator.
721 ///
722 /// Multiplies this APInt by RHS and assigns the result to *this.
723 ///
724 /// \returns *this
725 APInt &operator*=(const APInt &RHS);
726 APInt &operator*=(uint64_t RHS);
727
728 /// Addition assignment operator.
729 ///
730 /// Adds RHS to *this and assigns the result to *this.
731 ///
732 /// \returns *this
733 APInt &operator+=(const APInt &RHS);
734 APInt &operator+=(uint64_t RHS);
735
736 /// Subtraction assignment operator.
737 ///
738 /// Subtracts RHS from *this and assigns the result to *this.
739 ///
740 /// \returns *this
741 APInt &operator-=(const APInt &RHS);
742 APInt &operator-=(uint64_t RHS);
743
744 /// Left-shift assignment function.
745 ///
746 /// Shifts *this left by shiftAmt and assigns the result to *this.
747 ///
748 /// \returns *this after shifting left by ShiftAmt
749 APInt &operator<<=(unsigned ShiftAmt) {
750 assert(ShiftAmt <= BitWidth && "Invalid shift amount")(static_cast <bool> (ShiftAmt <= BitWidth &&
"Invalid shift amount") ? void (0) : __assert_fail ("ShiftAmt <= BitWidth && \"Invalid shift amount\""
, "llvm/include/llvm/ADT/APInt.h", 750, __extension__ __PRETTY_FUNCTION__
))
;
751 if (isSingleWord()) {
752 if (ShiftAmt == BitWidth)
753 U.VAL = 0;
754 else
755 U.VAL <<= ShiftAmt;
756 return clearUnusedBits();
757 }
758 shlSlowCase(ShiftAmt);
759 return *this;
760 }
761
762 /// Left-shift assignment function.
763 ///
764 /// Shifts *this left by shiftAmt and assigns the result to *this.
765 ///
766 /// \returns *this after shifting left by ShiftAmt
767 APInt &operator<<=(const APInt &ShiftAmt);
768
769 /// @}
770 /// \name Binary Operators
771 /// @{
772
773 /// Multiplication operator.
774 ///
775 /// Multiplies this APInt by RHS and returns the result.
776 APInt operator*(const APInt &RHS) const;
777
778 /// Left logical shift operator.
779 ///
780 /// Shifts this APInt left by \p Bits and returns the result.
781 APInt operator<<(unsigned Bits) const { return shl(Bits); }
782
783 /// Left logical shift operator.
784 ///
785 /// Shifts this APInt left by \p Bits and returns the result.
786 APInt operator<<(const APInt &Bits) const { return shl(Bits); }
787
788 /// Arithmetic right-shift function.
789 ///
790 /// Arithmetic right-shift this APInt by shiftAmt.
791 APInt ashr(unsigned ShiftAmt) const {
792 APInt R(*this);
793 R.ashrInPlace(ShiftAmt);
794 return R;
795 }
796
797 /// Arithmetic right-shift this APInt by ShiftAmt in place.
798 void ashrInPlace(unsigned ShiftAmt) {
799 assert(ShiftAmt <= BitWidth && "Invalid shift amount")(static_cast <bool> (ShiftAmt <= BitWidth &&
"Invalid shift amount") ? void (0) : __assert_fail ("ShiftAmt <= BitWidth && \"Invalid shift amount\""
, "llvm/include/llvm/ADT/APInt.h", 799, __extension__ __PRETTY_FUNCTION__
))
;
800 if (isSingleWord()) {
801 int64_t SExtVAL = SignExtend64(U.VAL, BitWidth);
802 if (ShiftAmt == BitWidth)
803 U.VAL = SExtVAL >> (APINT_BITS_PER_WORD - 1); // Fill with sign bit.
804 else
805 U.VAL = SExtVAL >> ShiftAmt;
806 clearUnusedBits();
807 return;
808 }
809 ashrSlowCase(ShiftAmt);
810 }
811
812 /// Logical right-shift function.
813 ///
814 /// Logical right-shift this APInt by shiftAmt.
815 APInt lshr(unsigned shiftAmt) const {
816 APInt R(*this);
817 R.lshrInPlace(shiftAmt);
818 return R;
819 }
820
821 /// Logical right-shift this APInt by ShiftAmt in place.
822 void lshrInPlace(unsigned ShiftAmt) {
823 assert(ShiftAmt <= BitWidth && "Invalid shift amount")(static_cast <bool> (ShiftAmt <= BitWidth &&
"Invalid shift amount") ? void (0) : __assert_fail ("ShiftAmt <= BitWidth && \"Invalid shift amount\""
, "llvm/include/llvm/ADT/APInt.h", 823, __extension__ __PRETTY_FUNCTION__
))
;
824 if (isSingleWord()) {
825 if (ShiftAmt == BitWidth)
826 U.VAL = 0;
827 else
828 U.VAL >>= ShiftAmt;
829 return;
830 }
831 lshrSlowCase(ShiftAmt);
832 }
833
834 /// Left-shift function.
835 ///
836 /// Left-shift this APInt by shiftAmt.
837 APInt shl(unsigned shiftAmt) const {
838 APInt R(*this);
839 R <<= shiftAmt;
840 return R;
841 }
842
843 /// Rotate left by rotateAmt.
844 APInt rotl(unsigned rotateAmt) const;
845
846 /// Rotate right by rotateAmt.
847 APInt rotr(unsigned rotateAmt) const;
848
849 /// Arithmetic right-shift function.
850 ///
851 /// Arithmetic right-shift this APInt by shiftAmt.
852 APInt ashr(const APInt &ShiftAmt) const {
853 APInt R(*this);
854 R.ashrInPlace(ShiftAmt);
855 return R;
856 }
857
858 /// Arithmetic right-shift this APInt by shiftAmt in place.
859 void ashrInPlace(const APInt &shiftAmt);
860
861 /// Logical right-shift function.
862 ///
863 /// Logical right-shift this APInt by shiftAmt.
864 APInt lshr(const APInt &ShiftAmt) const {
865 APInt R(*this);
866 R.lshrInPlace(ShiftAmt);
867 return R;
868 }
869
870 /// Logical right-shift this APInt by ShiftAmt in place.
871 void lshrInPlace(const APInt &ShiftAmt);
872
873 /// Left-shift function.
874 ///
875 /// Left-shift this APInt by shiftAmt.
876 APInt shl(const APInt &ShiftAmt) const {
877 APInt R(*this);
878 R <<= ShiftAmt;
879 return R;
880 }
881
882 /// Rotate left by rotateAmt.
883 APInt rotl(const APInt &rotateAmt) const;
884
885 /// Rotate right by rotateAmt.
886 APInt rotr(const APInt &rotateAmt) const;
887
888 /// Concatenate the bits from "NewLSB" onto the bottom of *this. This is
889 /// equivalent to:
890 /// (this->zext(NewWidth) << NewLSB.getBitWidth()) | NewLSB.zext(NewWidth)
891 APInt concat(const APInt &NewLSB) const {
892 /// If the result will be small, then both the merged values are small.
893 unsigned NewWidth = getBitWidth() + NewLSB.getBitWidth();
894 if (NewWidth <= APINT_BITS_PER_WORD)
895 return APInt(NewWidth, (U.VAL << NewLSB.getBitWidth()) | NewLSB.U.VAL);
896 return concatSlowCase(NewLSB);
897 }
898
899 /// Unsigned division operation.
900 ///
901 /// Perform an unsigned divide operation on this APInt by RHS. Both this and
902 /// RHS are treated as unsigned quantities for purposes of this division.
903 ///
904 /// \returns a new APInt value containing the division result, rounded towards
905 /// zero.
906 APInt udiv(const APInt &RHS) const;
907 APInt udiv(uint64_t RHS) const;
908
909 /// Signed division function for APInt.
910 ///
911 /// Signed divide this APInt by APInt RHS.
912 ///
913 /// The result is rounded towards zero.
914 APInt sdiv(const APInt &RHS) const;
915 APInt sdiv(int64_t RHS) const;
916
917 /// Unsigned remainder operation.
918 ///
919 /// Perform an unsigned remainder operation on this APInt with RHS being the
920 /// divisor. Both this and RHS are treated as unsigned quantities for purposes
921 /// of this operation. Note that this is a true remainder operation and not a
922 /// modulo operation because the sign follows the sign of the dividend which
923 /// is *this.
924 ///
925 /// \returns a new APInt value containing the remainder result
926 APInt urem(const APInt &RHS) const;
927 uint64_t urem(uint64_t RHS) const;
928
929 /// Function for signed remainder operation.
930 ///
931 /// Signed remainder operation on APInt.
932 APInt srem(const APInt &RHS) const;
933 int64_t srem(int64_t RHS) const;
934
935 /// Dual division/remainder interface.
936 ///
937 /// Sometimes it is convenient to divide two APInt values and obtain both the
938 /// quotient and remainder. This function does both operations in the same
939 /// computation making it a little more efficient. The pair of input arguments
940 /// may overlap with the pair of output arguments. It is safe to call
941 /// udivrem(X, Y, X, Y), for example.
942 static void udivrem(const APInt &LHS, const APInt &RHS, APInt &Quotient,
943 APInt &Remainder);
944 static void udivrem(const APInt &LHS, uint64_t RHS, APInt &Quotient,
945 uint64_t &Remainder);
946
947 static void sdivrem(const APInt &LHS, const APInt &RHS, APInt &Quotient,
948 APInt &Remainder);
949 static void sdivrem(const APInt &LHS, int64_t RHS, APInt &Quotient,
950 int64_t &Remainder);
951
952 // Operations that return overflow indicators.
953 APInt sadd_ov(const APInt &RHS, bool &Overflow) const;
954 APInt uadd_ov(const APInt &RHS, bool &Overflow) const;
955 APInt ssub_ov(const APInt &RHS, bool &Overflow) const;
956 APInt usub_ov(const APInt &RHS, bool &Overflow) const;
957 APInt sdiv_ov(const APInt &RHS, bool &Overflow) const;
958 APInt smul_ov(const APInt &RHS, bool &Overflow) const;
959 APInt umul_ov(const APInt &RHS, bool &Overflow) const;
960 APInt sshl_ov(const APInt &Amt, bool &Overflow) const;
961 APInt ushl_ov(const APInt &Amt, bool &Overflow) const;
962
963 // Operations that saturate
964 APInt sadd_sat(const APInt &RHS) const;
965 APInt uadd_sat(const APInt &RHS) const;
966 APInt ssub_sat(const APInt &RHS) const;
967 APInt usub_sat(const APInt &RHS) const;
968 APInt smul_sat(const APInt &RHS) const;
969 APInt umul_sat(const APInt &RHS) const;
970 APInt sshl_sat(const APInt &RHS) const;
971 APInt ushl_sat(const APInt &RHS) const;
972
973 /// Array-indexing support.
974 ///
975 /// \returns the bit value at bitPosition
976 bool operator[](unsigned bitPosition) const {
977 assert(bitPosition < getBitWidth() && "Bit position out of bounds!")(static_cast <bool> (bitPosition < getBitWidth() &&
"Bit position out of bounds!") ? void (0) : __assert_fail ("bitPosition < getBitWidth() && \"Bit position out of bounds!\""
, "llvm/include/llvm/ADT/APInt.h", 977, __extension__ __PRETTY_FUNCTION__
))
;
978 return (maskBit(bitPosition) & getWord(bitPosition)) != 0;
979 }
980
981 /// @}
982 /// \name Comparison Operators
983 /// @{
984
985 /// Equality operator.
986 ///
987 /// Compares this APInt with RHS for the validity of the equality
988 /// relationship.
989 bool operator==(const APInt &RHS) const {
990 assert(BitWidth == RHS.BitWidth && "Comparison requires equal bit widths")(static_cast <bool> (BitWidth == RHS.BitWidth &&
"Comparison requires equal bit widths") ? void (0) : __assert_fail
("BitWidth == RHS.BitWidth && \"Comparison requires equal bit widths\""
, "llvm/include/llvm/ADT/APInt.h", 990, __extension__ __PRETTY_FUNCTION__
))
;
991 if (isSingleWord())
992 return U.VAL == RHS.U.VAL;
993 return equalSlowCase(RHS);
994 }
995
996 /// Equality operator.
997 ///
998 /// Compares this APInt with a uint64_t for the validity of the equality
999 /// relationship.
1000 ///
1001 /// \returns true if *this == Val
1002 bool operator==(uint64_t Val) const {
1003 return (isSingleWord() || getActiveBits() <= 64) && getZExtValue() == Val;
1004 }
1005
1006 /// Equality comparison.
1007 ///
1008 /// Compares this APInt with RHS for the validity of the equality
1009 /// relationship.
1010 ///
1011 /// \returns true if *this == Val
1012 bool eq(const APInt &RHS) const { return (*this) == RHS; }
1013
1014 /// Inequality operator.
1015 ///
1016 /// Compares this APInt with RHS for the validity of the inequality
1017 /// relationship.
1018 ///
1019 /// \returns true if *this != Val
1020 bool operator!=(const APInt &RHS) const { return !((*this) == RHS); }
1021
1022 /// Inequality operator.
1023 ///
1024 /// Compares this APInt with a uint64_t for the validity of the inequality
1025 /// relationship.
1026 ///
1027 /// \returns true if *this != Val
1028 bool operator!=(uint64_t Val) const { return !((*this) == Val); }
1029
1030 /// Inequality comparison
1031 ///
1032 /// Compares this APInt with RHS for the validity of the inequality
1033 /// relationship.
1034 ///
1035 /// \returns true if *this != Val
1036 bool ne(const APInt &RHS) const { return !((*this) == RHS); }
1037
1038 /// Unsigned less than comparison
1039 ///
1040 /// Regards both *this and RHS as unsigned quantities and compares them for
1041 /// the validity of the less-than relationship.
1042 ///
1043 /// \returns true if *this < RHS when both are considered unsigned.
1044 bool ult(const APInt &RHS) const { return compare(RHS) < 0; }
1045
1046 /// Unsigned less than comparison
1047 ///
1048 /// Regards both *this as an unsigned quantity and compares it with RHS for
1049 /// the validity of the less-than relationship.
1050 ///
1051 /// \returns true if *this < RHS when considered unsigned.
1052 bool ult(uint64_t RHS) const {
1053 // Only need to check active bits if not a single word.
1054 return (isSingleWord() || getActiveBits() <= 64) && getZExtValue() < RHS;
1055 }
1056
1057 /// Signed less than comparison
1058 ///
1059 /// Regards both *this and RHS as signed quantities and compares them for
1060 /// validity of the less-than relationship.
1061 ///
1062 /// \returns true if *this < RHS when both are considered signed.
1063 bool slt(const APInt &RHS) const { return compareSigned(RHS) < 0; }
1064
1065 /// Signed less than comparison
1066 ///
1067 /// Regards both *this as a signed quantity and compares it with RHS for
1068 /// the validity of the less-than relationship.
1069 ///
1070 /// \returns true if *this < RHS when considered signed.
1071 bool slt(int64_t RHS) const {
1072 return (!isSingleWord() && getSignificantBits() > 64)
1073 ? isNegative()
1074 : getSExtValue() < RHS;
1075 }
1076
1077 /// Unsigned less or equal comparison
1078 ///
1079 /// Regards both *this and RHS as unsigned quantities and compares them for
1080 /// validity of the less-or-equal relationship.
1081 ///
1082 /// \returns true if *this <= RHS when both are considered unsigned.
1083 bool ule(const APInt &RHS) const { return compare(RHS) <= 0; }
1084
1085 /// Unsigned less or equal comparison
1086 ///
1087 /// Regards both *this as an unsigned quantity and compares it with RHS for
1088 /// the validity of the less-or-equal relationship.
1089 ///
1090 /// \returns true if *this <= RHS when considered unsigned.
1091 bool ule(uint64_t RHS) const { return !ugt(RHS); }
1092
1093 /// Signed less or equal comparison
1094 ///
1095 /// Regards both *this and RHS as signed quantities and compares them for
1096 /// validity of the less-or-equal relationship.
1097 ///
1098 /// \returns true if *this <= RHS when both are considered signed.
1099 bool sle(const APInt &RHS) const { return compareSigned(RHS) <= 0; }
1100
1101 /// Signed less or equal comparison
1102 ///
1103 /// Regards both *this as a signed quantity and compares it with RHS for the
1104 /// validity of the less-or-equal relationship.
1105 ///
1106 /// \returns true if *this <= RHS when considered signed.
1107 bool sle(uint64_t RHS) const { return !sgt(RHS); }
1108
1109 /// Unsigned greater than comparison
1110 ///
1111 /// Regards both *this and RHS as unsigned quantities and compares them for
1112 /// the validity of the greater-than relationship.
1113 ///
1114 /// \returns true if *this > RHS when both are considered unsigned.
1115 bool ugt(const APInt &RHS) const { return !ule(RHS); }
1116
1117 /// Unsigned greater than comparison
1118 ///
1119 /// Regards both *this as an unsigned quantity and compares it with RHS for
1120 /// the validity of the greater-than relationship.
1121 ///
1122 /// \returns true if *this > RHS when considered unsigned.
1123 bool ugt(uint64_t RHS) const {
1124 // Only need to check active bits if not a single word.
1125 return (!isSingleWord() && getActiveBits() > 64) || getZExtValue() > RHS;
1126 }
1127
1128 /// Signed greater than comparison
1129 ///
1130 /// Regards both *this and RHS as signed quantities and compares them for the
1131 /// validity of the greater-than relationship.
1132 ///
1133 /// \returns true if *this > RHS when both are considered signed.
1134 bool sgt(const APInt &RHS) const { return !sle(RHS); }
1135
1136 /// Signed greater than comparison
1137 ///
1138 /// Regards both *this as a signed quantity and compares it with RHS for
1139 /// the validity of the greater-than relationship.
1140 ///
1141 /// \returns true if *this > RHS when considered signed.
1142 bool sgt(int64_t RHS) const {
1143 return (!isSingleWord() && getSignificantBits() > 64)
1144 ? !isNegative()
1145 : getSExtValue() > RHS;
1146 }
1147
1148 /// Unsigned greater or equal comparison
1149 ///
1150 /// Regards both *this and RHS as unsigned quantities and compares them for
1151 /// validity of the greater-or-equal relationship.
1152 ///
1153 /// \returns true if *this >= RHS when both are considered unsigned.
1154 bool uge(const APInt &RHS) const { return !ult(RHS); }
1155
1156 /// Unsigned greater or equal comparison
1157 ///
1158 /// Regards both *this as an unsigned quantity and compares it with RHS for
1159 /// the validity of the greater-or-equal relationship.
1160 ///
1161 /// \returns true if *this >= RHS when considered unsigned.
1162 bool uge(uint64_t RHS) const { return !ult(RHS); }
1163
1164 /// Signed greater or equal comparison
1165 ///
1166 /// Regards both *this and RHS as signed quantities and compares them for
1167 /// validity of the greater-or-equal relationship.
1168 ///
1169 /// \returns true if *this >= RHS when both are considered signed.
1170 bool sge(const APInt &RHS) const { return !slt(RHS); }
1171
1172 /// Signed greater or equal comparison
1173 ///
1174 /// Regards both *this as a signed quantity and compares it with RHS for
1175 /// the validity of the greater-or-equal relationship.
1176 ///
1177 /// \returns true if *this >= RHS when considered signed.
1178 bool sge(int64_t RHS) const { return !slt(RHS); }
1179
1180 /// This operation tests if there are any pairs of corresponding bits
1181 /// between this APInt and RHS that are both set.
1182 bool intersects(const APInt &RHS) const {
1183 assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth &&
"Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\""
, "llvm/include/llvm/ADT/APInt.h", 1183, __extension__ __PRETTY_FUNCTION__
))
;
1184 if (isSingleWord())
1185 return (U.VAL & RHS.U.VAL) != 0;
1186 return intersectsSlowCase(RHS);
1187 }
1188
1189 /// This operation checks that all bits set in this APInt are also set in RHS.
1190 bool isSubsetOf(const APInt &RHS) const {
1191 assert(BitWidth == RHS.BitWidth && "Bit widths must be the same")(static_cast <bool> (BitWidth == RHS.BitWidth &&
"Bit widths must be the same") ? void (0) : __assert_fail ("BitWidth == RHS.BitWidth && \"Bit widths must be the same\""
, "llvm/include/llvm/ADT/APInt.h", 1191, __extension__ __PRETTY_FUNCTION__
))
;
1192 if (isSingleWord())
1193 return (U.VAL & ~RHS.U.VAL) == 0;
1194 return isSubsetOfSlowCase(RHS);
1195 }
1196
1197 /// @}
1198 /// \name Resizing Operators
1199 /// @{
1200
1201 /// Truncate to new width.
1202 ///
1203 /// Truncate the APInt to a specified width. It is an error to specify a width
1204 /// that is greater than or equal to the current width.
1205 APInt trunc(unsigned width) const;
1206
1207 /// Truncate to new width with unsigned saturation.
1208 ///
1209 /// If the APInt, treated as unsigned integer, can be losslessly truncated to
1210 /// the new bitwidth, then return truncated APInt. Else, return max value.
1211 APInt truncUSat(unsigned width) const;
1212
1213 /// Truncate to new width with signed saturation.
1214 ///
1215 /// If this APInt, treated as signed integer, can be losslessly truncated to
1216 /// the new bitwidth, then return truncated APInt. Else, return either
1217 /// signed min value if the APInt was negative, or signed max value.
1218 APInt truncSSat(unsigned width) const;
1219
1220 /// Sign extend to a new width.
1221 ///
1222 /// This operation sign extends the APInt to a new width. If the high order
1223 /// bit is set, the fill on the left will be done with 1 bits, otherwise zero.
1224 /// It is an error to specify a width that is less than or equal to the
1225 /// current width.
1226 APInt sext(unsigned width) const;
1227
1228 /// Zero extend to a new width.
1229 ///
1230 /// This operation zero extends the APInt to a new width. The high order bits
1231 /// are filled with 0 bits. It is an error to specify a width that is less
1232 /// than or equal to the current width.
1233 APInt zext(unsigned width) const;
1234
1235 /// Sign extend or truncate to width
1236 ///
1237 /// Make this APInt have the bit width given by \p width. The value is sign
1238 /// extended, truncated, or left alone to make it that width.
1239 APInt sextOrTrunc(unsigned width) const;
1240
1241 /// Zero extend or truncate to width
1242 ///
1243 /// Make this APInt have the bit width given by \p width. The value is zero
1244 /// extended, truncated, or left alone to make it that width.
1245 APInt zextOrTrunc(unsigned width) const;
1246
1247 /// Truncate to width
1248 ///
1249 /// Make this APInt have the bit width given by \p width. The value is
1250 /// truncated or left alone to make it that width.
1251 APInt truncOrSelf(unsigned width) const;
1252
1253 /// Sign extend or truncate to width
1254 ///
1255 /// Make this APInt have the bit width given by \p width. The value is sign
1256 /// extended, or left alone to make it that width.
1257 APInt sextOrSelf(unsigned width) const;
1258
1259 /// Zero extend or truncate to width
1260 ///
1261 /// Make this APInt have the bit width given by \p width. The value is zero
1262 /// extended, or left alone to make it that width.
1263 APInt zextOrSelf(unsigned width) const;
1264
1265 /// @}
1266 /// \name Bit Manipulation Operators
1267 /// @{
1268
1269 /// Set every bit to 1.
1270 void setAllBits() {
1271 if (isSingleWord())
1272 U.VAL = WORDTYPE_MAX;
1273 else
1274 // Set all the bits in all the words.
1275 memset(U.pVal, -1, getNumWords() * APINT_WORD_SIZE);
1276 // Clear the unused ones
1277 clearUnusedBits();
1278 }
1279
1280 /// Set the given bit to 1 whose position is given as "bitPosition".
1281 void setBit(unsigned BitPosition) {
1282 assert(BitPosition < BitWidth && "BitPosition out of range")(static_cast <bool> (BitPosition < BitWidth &&
"BitPosition out of range") ? void (0) : __assert_fail ("BitPosition < BitWidth && \"BitPosition out of range\""
, "llvm/include/llvm/ADT/APInt.h", 1282, __extension__ __PRETTY_FUNCTION__
))
;
1283 WordType Mask = maskBit(BitPosition);
1284 if (isSingleWord())
1285 U.VAL |= Mask;
1286 else
1287 U.pVal[whichWord(BitPosition)] |= Mask;
1288 }
1289
1290 /// Set the sign bit to 1.
1291 void setSignBit() { setBit(BitWidth - 1); }
1292
1293 /// Set a given bit to a given value.
1294 void setBitVal(unsigned BitPosition, bool BitValue) {
1295 if (BitValue)
1296 setBit(BitPosition);
1297 else
1298 clearBit(BitPosition);
1299 }
1300
1301 /// Set the bits from loBit (inclusive) to hiBit (exclusive) to 1.
1302 /// This function handles "wrap" case when \p loBit >= \p hiBit, and calls
1303 /// setBits when \p loBit < \p hiBit.
1304 /// For \p loBit == \p hiBit wrap case, set every bit to 1.
1305 void setBitsWithWrap(unsigned loBit, unsigned hiBit) {
1306 assert(hiBit <= BitWidth && "hiBit out of range")(static_cast <bool> (hiBit <= BitWidth && "hiBit out of range"
) ? void (0) : __assert_fail ("hiBit <= BitWidth && \"hiBit out of range\""
, "llvm/include/llvm/ADT/APInt.h", 1306, __extension__ __PRETTY_FUNCTION__
))
;
1307 assert(loBit <= BitWidth && "loBit out of range")(static_cast <bool> (loBit <= BitWidth && "loBit out of range"
) ? void (0) : __assert_fail ("loBit <= BitWidth && \"loBit out of range\""
, "llvm/include/llvm/ADT/APInt.h", 1307, __extension__ __PRETTY_FUNCTION__
))
;
1308 if (loBit < hiBit) {
1309 setBits(loBit, hiBit);
1310 return;
1311 }
1312 setLowBits(hiBit);
1313 setHighBits(BitWidth - loBit);
1314 }
1315
1316 /// Set the bits from loBit (inclusive) to hiBit (exclusive) to 1.
1317 /// This function handles case when \p loBit <= \p hiBit.
1318 void setBits(unsigned loBit, unsigned hiBit) {
1319 assert(hiBit <= BitWidth && "hiBit out of range")(static_cast <bool> (hiBit <= BitWidth && "hiBit out of range"
) ? void (0) : __assert_fail ("hiBit <= BitWidth && \"hiBit out of range\""
, "llvm/include/llvm/ADT/APInt.h", 1319, __extension__ __PRETTY_FUNCTION__
))
;
1320 assert(loBit <= BitWidth && "loBit out of range")(static_cast <bool> (loBit <= BitWidth && "loBit out of range"
) ? void (0) : __assert_fail ("loBit <= BitWidth && \"loBit out of range\""
, "llvm/include/llvm/ADT/APInt.h", 1320, __extension__ __PRETTY_FUNCTION__
))
;
1321 assert(loBit <= hiBit && "loBit greater than hiBit")(static_cast <bool> (loBit <= hiBit && "loBit greater than hiBit"
) ? void (0) : __assert_fail ("loBit <= hiBit && \"loBit greater than hiBit\""
, "llvm/include/llvm/ADT/APInt.h", 1321, __extension__ __PRETTY_FUNCTION__
))
;
1322 if (loBit == hiBit)
1323 return;
1324 if (loBit < APINT_BITS_PER_WORD && hiBit <= APINT_BITS_PER_WORD) {
1325 uint64_t mask = WORDTYPE_MAX >> (APINT_BITS_PER_WORD - (hiBit - loBit));
1326 mask <<= loBit;
1327 if (isSingleWord())
1328 U.VAL |= mask;
1329 else
1330 U.pVal[0] |= mask;
1331 } else {
1332 setBitsSlowCase(loBit, hiBit);
1333 }
1334 }
1335
1336 /// Set the top bits starting from loBit.
1337 void setBitsFrom(unsigned loBit) { return setBits(loBit, BitWidth); }
1338
1339 /// Set the bottom loBits bits.
1340 void setLowBits(unsigned loBits) { return setBits(0, loBits); }
1341
1342 /// Set the top hiBits bits.
1343 void setHighBits(unsigned hiBits) {
1344 return setBits(BitWidth - hiBits, BitWidth);
1345 }
1346
1347 /// Set every bit to 0.
1348 void clearAllBits() {
1349 if (isSingleWord())
1350 U.VAL = 0;
1351 else
1352 memset(U.pVal, 0, getNumWords() * APINT_WORD_SIZE);
1353 }
1354
1355 /// Set a given bit to 0.
1356 ///
1357 /// Set the given bit to 0 whose position is given as "bitPosition".
1358 void clearBit(unsigned BitPosition) {
1359 assert(BitPosition < BitWidth && "BitPosition out of range")(static_cast <bool> (BitPosition < BitWidth &&
"BitPosition out of range") ? void (0) : __assert_fail ("BitPosition < BitWidth && \"BitPosition out of range\""
, "llvm/include/llvm/ADT/APInt.h", 1359, __extension__ __PRETTY_FUNCTION__
))
;
1360 WordType Mask = ~maskBit(BitPosition);
1361 if (isSingleWord())
1362 U.VAL &= Mask;
1363 else
1364 U.pVal[whichWord(BitPosition)] &= Mask;
1365 }
1366
1367 /// Set bottom loBits bits to 0.
1368 void clearLowBits(unsigned loBits) {
1369 assert(loBits <= BitWidth && "More bits than bitwidth")(static_cast <bool> (loBits <= BitWidth && "More bits than bitwidth"
) ? void (0) : __assert_fail ("loBits <= BitWidth && \"More bits than bitwidth\""
, "llvm/include/llvm/ADT/APInt.h", 1369, __extension__ __PRETTY_FUNCTION__
))
;
1370 APInt Keep = getHighBitsSet(BitWidth, BitWidth - loBits);
1371 *this &= Keep;
1372 }
1373
1374 /// Set the sign bit to 0.
1375 void clearSignBit() { clearBit(BitWidth - 1); }
1376
1377 /// Toggle every bit to its opposite value.
1378 void flipAllBits() {
1379 if (isSingleWord()) {
1380 U.VAL ^= WORDTYPE_MAX;
1381 clearUnusedBits();
1382 } else {
1383 flipAllBitsSlowCase();
1384 }
1385 }
1386
1387 /// Toggles a given bit to its opposite value.
1388 ///
1389 /// Toggle a given bit to its opposite value whose position is given
1390 /// as "bitPosition".
1391 void flipBit(unsigned bitPosition);
1392
1393 /// Negate this APInt in place.
1394 void negate() {
1395 flipAllBits();
1396 ++(*this);
1397 }
1398
1399 /// Insert the bits from a smaller APInt starting at bitPosition.
1400 void insertBits(const APInt &SubBits, unsigned bitPosition);
1401 void insertBits(uint64_t SubBits, unsigned bitPosition, unsigned numBits);
1402
1403 /// Return an APInt with the extracted bits [bitPosition,bitPosition+numBits).
1404 APInt extractBits(unsigned numBits, unsigned bitPosition) const;
1405 uint64_t extractBitsAsZExtValue(unsigned numBits, unsigned bitPosition) const;
1406
1407 /// @}
1408 /// \name Value Characterization Functions
1409 /// @{
1410
1411 /// Return the number of bits in the APInt.
1412 unsigned getBitWidth() const { return BitWidth; }
1413
1414 /// Get the number of words.
1415 ///
1416 /// Here one word's bitwidth equals to that of uint64_t.
1417 ///
1418 /// \returns the number of words to hold the integer value of this APInt.
1419 unsigned getNumWords() const { return getNumWords(BitWidth); }
1420
1421 /// Get the number of words.
1422 ///
1423 /// *NOTE* Here one word's bitwidth equals to that of uint64_t.
1424 ///
1425 /// \returns the number of words to hold the integer value with a given bit
1426 /// width.
1427 static unsigned getNumWords(unsigned BitWidth) {
1428 return ((uint64_t)BitWidth + APINT_BITS_PER_WORD - 1) / APINT_BITS_PER_WORD;
1429 }
1430
1431 /// Compute the number of active bits in the value
1432 ///
1433 /// This function returns the number of active bits which is defined as the
1434 /// bit width minus the number of leading zeros. This is used in several
1435 /// computations to see how "wide" the value is.
1436 unsigned getActiveBits() const { return BitWidth - countLeadingZeros(); }
1437
1438 /// Compute the number of active words in the value of this APInt.
1439 ///
1440 /// This is used in conjunction with getActiveData to extract the raw value of
1441 /// the APInt.
1442 unsigned getActiveWords() const {
1443 unsigned numActiveBits = getActiveBits();
1444 return numActiveBits ? whichWord(numActiveBits - 1) + 1 : 1;
1445 }
1446
1447 /// Get the minimum bit size for this signed APInt
1448 ///
1449 /// Computes the minimum bit width for this APInt while considering it to be a
1450 /// signed (and probably negative) value. If the value is not negative, this
1451 /// function returns the same value as getActiveBits()+1. Otherwise, it
1452 /// returns the smallest bit width that will retain the negative value. For
1453 /// example, -1 can be written as 0b1 or 0xFFFFFFFFFF. 0b1 is shorter and so
1454 /// for -1, this function will always return 1.
1455 unsigned getSignificantBits() const {
1456 return BitWidth - getNumSignBits() + 1;
1457 }
1458
1459 /// NOTE: This is soft-deprecated. Please use `getSignificantBits()` instead.
1460 unsigned getMinSignedBits() const { return getSignificantBits(); }
1461
1462 /// Get zero extended value
1463 ///
1464 /// This method attempts to return the value of this APInt as a zero extended
1465 /// uint64_t. The bitwidth must be <= 64 or the value must fit within a
1466 /// uint64_t. Otherwise an assertion will result.
1467 uint64_t getZExtValue() const {
1468 if (isSingleWord())
1469 return U.VAL;
1470 assert(getActiveBits() <= 64 && "Too many bits for uint64_t")(static_cast <bool> (getActiveBits() <= 64 &&
"Too many bits for uint64_t") ? void (0) : __assert_fail ("getActiveBits() <= 64 && \"Too many bits for uint64_t\""
, "llvm/include/llvm/ADT/APInt.h", 1470, __extension__ __PRETTY_FUNCTION__
))
;
1471 return U.pVal[0];
1472 }
1473
1474 /// Get sign extended value
1475 ///
1476 /// This method attempts to return the value of this APInt as a sign extended
1477 /// int64_t. The bit width must be <= 64 or the value must fit within an
1478 /// int64_t. Otherwise an assertion will result.
1479 int64_t getSExtValue() const {
1480 if (isSingleWord())
1481 return SignExtend64(U.VAL, BitWidth);
1482 assert(getSignificantBits() <= 64 && "Too many bits for int64_t")(static_cast <bool> (getSignificantBits() <= 64 &&
"Too many bits for int64_t") ? void (0) : __assert_fail ("getSignificantBits() <= 64 && \"Too many bits for int64_t\""
, "llvm/include/llvm/ADT/APInt.h", 1482, __extension__ __PRETTY_FUNCTION__
))
;
1483 return int64_t(U.pVal[0]);
1484 }
1485
1486 /// Get bits required for string value.
1487 ///
1488 /// This method determines how many bits are required to hold the APInt
1489 /// equivalent of the string given by \p str.
1490 static unsigned getBitsNeeded(StringRef str, uint8_t radix);
1491
1492 /// The APInt version of the countLeadingZeros functions in
1493 /// MathExtras.h.
1494 ///
1495 /// It counts the number of zeros from the most significant bit to the first
1496 /// one bit.
1497 ///
1498 /// \returns BitWidth if the value is zero, otherwise returns the number of
1499 /// zeros from the most significant bit to the first one bits.
1500 unsigned countLeadingZeros() const {
1501 if (isSingleWord()) {
1502 unsigned unusedBits = APINT_BITS_PER_WORD - BitWidth;
1503 return llvm::countLeadingZeros(U.VAL) - unusedBits;
1504 }
1505 return countLeadingZerosSlowCase();
1506 }
1507
1508 /// Count the number of leading one bits.
1509 ///
1510 /// This function is an APInt version of the countLeadingOnes
1511 /// functions in MathExtras.h. It counts the number of ones from the most
1512 /// significant bit to the first zero bit.
1513 ///
1514 /// \returns 0 if the high order bit is not set, otherwise returns the number
1515 /// of 1 bits from the most significant to the least
1516 unsigned countLeadingOnes() const {
1517 if (isSingleWord()) {
1518 if (LLVM_UNLIKELY(BitWidth == 0)__builtin_expect((bool)(BitWidth == 0), false))
1519 return 0;
1520 return llvm::countLeadingOnes(U.VAL << (APINT_BITS_PER_WORD - BitWidth));
1521 }
1522 return countLeadingOnesSlowCase();
1523 }
1524
1525 /// Computes the number of leading bits of this APInt that are equal to its
1526 /// sign bit.
1527 unsigned getNumSignBits() const {
1528 return isNegative() ? countLeadingOnes() : countLeadingZeros();
1529 }
1530
1531 /// Count the number of trailing zero bits.
1532 ///
1533 /// This function is an APInt version of the countTrailingZeros
1534 /// functions in MathExtras.h. It counts the number of zeros from the least
1535 /// significant bit to the first set bit.
1536 ///
1537 /// \returns BitWidth if the value is zero, otherwise returns the number of
1538 /// zeros from the least significant bit to the first one bit.
1539 unsigned countTrailingZeros() const {
1540 if (isSingleWord()) {
1541 unsigned TrailingZeros = llvm::countTrailingZeros(U.VAL);
1542 return (TrailingZeros > BitWidth ? BitWidth : TrailingZeros);
1543 }
1544 return countTrailingZerosSlowCase();
1545 }
1546
1547 /// Count the number of trailing one bits.
1548 ///
1549 /// This function is an APInt version of the countTrailingOnes
1550 /// functions in MathExtras.h. It counts the number of ones from the least
1551 /// significant bit to the first zero bit.
1552 ///
1553 /// \returns BitWidth if the value is all ones, otherwise returns the number
1554 /// of ones from the least significant bit to the first zero bit.
1555 unsigned countTrailingOnes() const {
1556 if (isSingleWord())
1557 return llvm::countTrailingOnes(U.VAL);
1558 return countTrailingOnesSlowCase();
1559 }
1560
1561 /// Count the number of bits set.
1562 ///
1563 /// This function is an APInt version of the countPopulation functions
1564 /// in MathExtras.h. It counts the number of 1 bits in the APInt value.
1565 ///
1566 /// \returns 0 if the value is zero, otherwise returns the number of set bits.
1567 unsigned countPopulation() const {
1568 if (isSingleWord())
1569 return llvm::countPopulation(U.VAL);
1570 return countPopulationSlowCase();
1571 }
1572
1573 /// @}
1574 /// \name Conversion Functions
1575 /// @{
1576 void print(raw_ostream &OS, bool isSigned) const;
1577
1578 /// Converts an APInt to a string and append it to Str. Str is commonly a
1579 /// SmallString.
1580 void toString(SmallVectorImpl<char> &Str, unsigned Radix, bool Signed,
1581 bool formatAsCLiteral = false) const;
1582
1583 /// Considers the APInt to be unsigned and converts it into a string in the
1584 /// radix given. The radix can be 2, 8, 10 16, or 36.
1585 void toStringUnsigned(SmallVectorImpl<char> &Str, unsigned Radix = 10) const {
1586 toString(Str, Radix, false, false);
1587 }
1588
1589 /// Considers the APInt to be signed and converts it into a string in the
1590 /// radix given. The radix can be 2, 8, 10, 16, or 36.
1591 void toStringSigned(SmallVectorImpl<char> &Str, unsigned Radix = 10) const {
1592 toString(Str, Radix, true, false);
1593 }
1594
1595 /// \returns a byte-swapped representation of this APInt Value.
1596 APInt byteSwap() const;
1597
1598 /// \returns the value with the bit representation reversed of this APInt
1599 /// Value.
1600 APInt reverseBits() const;
1601
1602 /// Converts this APInt to a double value.
1603 double roundToDouble(bool isSigned) const;
1604
1605 /// Converts this unsigned APInt to a double value.
1606 double roundToDouble() const { return roundToDouble(false); }
1607
1608 /// Converts this signed APInt to a double value.
1609 double signedRoundToDouble() const { return roundToDouble(true); }
1610
1611 /// Converts APInt bits to a double
1612 ///
1613 /// The conversion does not do a translation from integer to double, it just
1614 /// re-interprets the bits as a double. Note that it is valid to do this on
1615 /// any bit width. Exactly 64 bits will be translated.
1616 double bitsToDouble() const { return BitsToDouble(getWord(0)); }
1617
1618 /// Converts APInt bits to a float
1619 ///
1620 /// The conversion does not do a translation from integer to float, it just
1621 /// re-interprets the bits as a float. Note that it is valid to do this on
1622 /// any bit width. Exactly 32 bits will be translated.
1623 float bitsToFloat() const {
1624 return BitsToFloat(static_cast<uint32_t>(getWord(0)));
1625 }
1626
1627 /// Converts a double to APInt bits.
1628 ///
1629 /// The conversion does not do a translation from double to integer, it just
1630 /// re-interprets the bits of the double.
1631 static APInt doubleToBits(double V) {
1632 return APInt(sizeof(double) * CHAR_BIT8, DoubleToBits(V));
1633 }
1634
1635 /// Converts a float to APInt bits.
1636 ///
1637 /// The conversion does not do a translation from float to integer, it just
1638 /// re-interprets the bits of the float.
1639 static APInt floatToBits(float V) {
1640 return APInt(sizeof(float) * CHAR_BIT8, FloatToBits(V));
1641 }
1642
1643 /// @}
1644 /// \name Mathematics Operations
1645 /// @{
1646
1647 /// \returns the floor log base 2 of this APInt.
1648 unsigned logBase2() const { return getActiveBits() - 1; }
1649
1650 /// \returns the ceil log base 2 of this APInt.
1651 unsigned ceilLogBase2() const {
1652 APInt temp(*this);
1653 --temp;
1654 return temp.getActiveBits();
1655 }
1656
1657 /// \returns the nearest log base 2 of this APInt. Ties round up.
1658 ///
1659 /// NOTE: When we have a BitWidth of 1, we define:
1660 ///
1661 /// log2(0) = UINT32_MAX
1662 /// log2(1) = 0
1663 ///
1664 /// to get around any mathematical concerns resulting from
1665 /// referencing 2 in a space where 2 does no exist.
1666 unsigned nearestLogBase2() const;
1667
1668 /// \returns the log base 2 of this APInt if its an exact power of two, -1
1669 /// otherwise
1670 int32_t exactLogBase2() const {
1671 if (!isPowerOf2())
1672 return -1;
1673 return logBase2();
1674 }
1675
1676 /// Compute the square root.
1677 APInt sqrt() const;
1678
1679 /// Get the absolute value. If *this is < 0 then return -(*this), otherwise
1680 /// *this. Note that the "most negative" signed number (e.g. -128 for 8 bit
1681 /// wide APInt) is unchanged due to how negation works.
1682 APInt abs() const {
1683 if (isNegative())
1684 return -(*this);
1685 return *this;
1686 }
1687
1688 /// \returns the multiplicative inverse for a given modulo.
1689 APInt multiplicativeInverse(const APInt &modulo) const;
1690
1691 /// @}
1692 /// \name Building-block Operations for APInt and APFloat
1693 /// @{
1694
1695 // These building block operations operate on a representation of arbitrary
1696 // precision, two's-complement, bignum integer values. They should be
1697 // sufficient to implement APInt and APFloat bignum requirements. Inputs are
1698 // generally a pointer to the base of an array of integer parts, representing
1699 // an unsigned bignum, and a count of how many parts there are.
1700
1701 /// Sets the least significant part of a bignum to the input value, and zeroes
1702 /// out higher parts.
1703 static void tcSet(WordType *, WordType, unsigned);
1704
1705 /// Assign one bignum to another.
1706 static void tcAssign(WordType *, const WordType *, unsigned);
1707
1708 /// Returns true if a bignum is zero, false otherwise.
1709 static bool tcIsZero(const WordType *, unsigned);
1710
1711 /// Extract the given bit of a bignum; returns 0 or 1. Zero-based.
1712 static int tcExtractBit(const WordType *, unsigned bit);
1713
1714 /// Copy the bit vector of width srcBITS from SRC, starting at bit srcLSB, to
1715 /// DST, of dstCOUNT parts, such that the bit srcLSB becomes the least
1716 /// significant bit of DST. All high bits above srcBITS in DST are
1717 /// zero-filled.
1718 static void tcExtract(WordType *, unsigned dstCount, const WordType *,
1719 unsigned srcBits, unsigned srcLSB);
1720
1721 /// Set the given bit of a bignum. Zero-based.
1722 static void tcSetBit(WordType *, unsigned bit);
1723
1724 /// Clear the given bit of a bignum. Zero-based.
1725 static void tcClearBit(WordType *, unsigned bit);
1726
1727 /// Returns the bit number of the least or most significant set bit of a
1728 /// number. If the input number has no bits set -1U is returned.
1729 static unsigned tcLSB(const WordType *, unsigned n);
1730 static unsigned tcMSB(const WordType *parts, unsigned n);
1731
1732 /// Negate a bignum in-place.
1733 static void tcNegate(WordType *, unsigned);
1734
1735 /// DST += RHS + CARRY where CARRY is zero or one. Returns the carry flag.
1736 static WordType tcAdd(WordType *, const WordType *, WordType carry, unsigned);
1737 /// DST += RHS. Returns the carry flag.
1738 static WordType tcAddPart(WordType *, WordType, unsigned);
1739
1740 /// DST -= RHS + CARRY where CARRY is zero or one. Returns the carry flag.
1741 static WordType tcSubtract(WordType *, const WordType *, WordType carry,
1742 unsigned);
1743 /// DST -= RHS. Returns the carry flag.
1744 static WordType tcSubtractPart(WordType *, WordType, unsigned);
1745
1746 /// DST += SRC * MULTIPLIER + PART if add is true
1747 /// DST = SRC * MULTIPLIER + PART if add is false
1748 ///
1749 /// Requires 0 <= DSTPARTS <= SRCPARTS + 1. If DST overlaps SRC they must
1750 /// start at the same point, i.e. DST == SRC.
1751 ///
1752 /// If DSTPARTS == SRC_PARTS + 1 no overflow occurs and zero is returned.
1753 /// Otherwise DST is filled with the least significant DSTPARTS parts of the
1754 /// result, and if all of the omitted higher parts were zero return zero,
1755 /// otherwise overflow occurred and return one.
1756 static int tcMultiplyPart(WordType *dst, const WordType *src,
1757 WordType multiplier, WordType carry,
1758 unsigned srcParts, unsigned dstParts, bool add);
1759
1760 /// DST = LHS * RHS, where DST has the same width as the operands and is
1761 /// filled with the least significant parts of the result. Returns one if
1762 /// overflow occurred, otherwise zero. DST must be disjoint from both
1763 /// operands.
1764 static int tcMultiply(WordType *, const WordType *, const WordType *,
1765 unsigned);
1766
1767 /// DST = LHS * RHS, where DST has width the sum of the widths of the
1768 /// operands. No overflow occurs. DST must be disjoint from both operands.
1769 static void tcFullMultiply(WordType *, const WordType *, const WordType *,
1770 unsigned, unsigned);
1771
1772 /// If RHS is zero LHS and REMAINDER are left unchanged, return one.
1773 /// Otherwise set LHS to LHS / RHS with the fractional part discarded, set
1774 /// REMAINDER to the remainder, return zero. i.e.
1775 ///
1776 /// OLD_LHS = RHS * LHS + REMAINDER
1777 ///
1778 /// SCRATCH is a bignum of the same size as the operands and result for use by
1779 /// the routine; its contents need not be initialized and are destroyed. LHS,
1780 /// REMAINDER and SCRATCH must be distinct.
1781 static int tcDivide(WordType *lhs, const WordType *rhs, WordType *remainder,
1782 WordType *scratch, unsigned parts);
1783
1784 /// Shift a bignum left Count bits. Shifted in bits are zero. There are no
1785 /// restrictions on Count.
1786 static void tcShiftLeft(WordType *, unsigned Words, unsigned Count);
1787
1788 /// Shift a bignum right Count bits. Shifted in bits are zero. There are no
1789 /// restrictions on Count.
1790 static void tcShiftRight(WordType *, unsigned Words, unsigned Count);
1791
1792 /// Comparison (unsigned) of two bignums.
1793 static int tcCompare(const WordType *, const WordType *, unsigned);
1794
1795 /// Increment a bignum in-place. Return the carry flag.
1796 static WordType tcIncrement(WordType *dst, unsigned parts) {
1797 return tcAddPart(dst, 1, parts);
1798 }
1799
1800 /// Decrement a bignum in-place. Return the borrow flag.
1801 static WordType tcDecrement(WordType *dst, unsigned parts) {
1802 return tcSubtractPart(dst, 1, parts);
1803 }
1804
1805 /// Used to insert APInt objects, or objects that contain APInt objects, into
1806 /// FoldingSets.
1807 void Profile(FoldingSetNodeID &id) const;
1808
1809 /// debug method
1810 void dump() const;
1811
1812 /// Returns whether this instance allocated memory.
1813 bool needsCleanup() const { return !isSingleWord(); }
1814
1815private:
1816 /// This union is used to store the integer value. When the
1817 /// integer bit-width <= 64, it uses VAL, otherwise it uses pVal.
1818 union {
1819 uint64_t VAL; ///< Used to store the <= 64 bits integer value.
1820 uint64_t *pVal; ///< Used to store the >64 bits integer value.
1821 } U;
1822
1823 unsigned BitWidth; ///< The number of bits in this APInt.
1824
1825 friend struct DenseMapInfo<APInt, void>;
1826 friend class APSInt;
1827
1828 /// This constructor is used only internally for speed of construction of
1829 /// temporaries. It is unsafe since it takes ownership of the pointer, so it
1830 /// is not public.
1831 APInt(uint64_t *val, unsigned bits) : BitWidth(bits) { U.pVal = val; }
1832
1833 /// Determine which word a bit is in.
1834 ///
1835 /// \returns the word position for the specified bit position.
1836 static unsigned whichWord(unsigned bitPosition) {
1837 return bitPosition / APINT_BITS_PER_WORD;
1838 }
1839
1840 /// Determine which bit in a word the specified bit position is in.
1841 static unsigned whichBit(unsigned bitPosition) {
1842 return bitPosition % APINT_BITS_PER_WORD;
1843 }
1844
1845 /// Get a single bit mask.
1846 ///
1847 /// \returns a uint64_t with only bit at "whichBit(bitPosition)" set
1848 /// This method generates and returns a uint64_t (word) mask for a single
1849 /// bit at a specific bit position. This is used to mask the bit in the
1850 /// corresponding word.
1851 static uint64_t maskBit(unsigned bitPosition) {
1852 return 1ULL << whichBit(bitPosition);
1853 }
1854
1855 /// Clear unused high order bits
1856 ///
1857 /// This method is used internally to clear the top "N" bits in the high order
1858 /// word that are not used by the APInt. This is needed after the most
1859 /// significant word is assigned a value to ensure that those bits are
1860 /// zero'd out.
1861 APInt &clearUnusedBits() {
1862 // Compute how many bits are used in the final word.
1863 unsigned WordBits = ((BitWidth - 1) % APINT_BITS_PER_WORD) + 1;
1864
1865 // Mask out the high bits.
1866 uint64_t mask = WORDTYPE_MAX >> (APINT_BITS_PER_WORD - WordBits);
1867 if (LLVM_UNLIKELY(BitWidth == 0)__builtin_expect((bool)(BitWidth == 0), false))
1868 mask = 0;
1869
1870 if (isSingleWord())
1871 U.VAL &= mask;
1872 else
1873 U.pVal[getNumWords() - 1] &= mask;
1874 return *this;
1875 }
1876
1877 /// Get the word corresponding to a bit position
1878 /// \returns the corresponding word for the specified bit position.
1879 uint64_t getWord(unsigned bitPosition) const {
1880 return isSingleWord() ? U.VAL : U.pVal[whichWord(bitPosition)];
1881 }
1882
1883 /// Utility method to change the bit width of this APInt to new bit width,
1884 /// allocating and/or deallocating as necessary. There is no guarantee on the
1885 /// value of any bits upon return. Caller should populate the bits after.
1886 void reallocate(unsigned NewBitWidth);
1887
1888 /// Convert a char array into an APInt
1889 ///
1890 /// \param radix 2, 8, 10, 16, or 36
1891 /// Converts a string into a number. The string must be non-empty
1892 /// and well-formed as a number of the given base. The bit-width
1893 /// must be sufficient to hold the result.
1894 ///
1895 /// This is used by the constructors that take string arguments.
1896 ///
1897 /// StringRef::getAsInteger is superficially similar but (1) does
1898 /// not assume that the string is well-formed and (2) grows the
1899 /// result to hold the input.
1900 void fromString(unsigned numBits, StringRef str, uint8_t radix);
1901
1902 /// An internal division function for dividing APInts.
1903 ///
1904 /// This is used by the toString method to divide by the radix. It simply
1905 /// provides a more convenient form of divide for internal use since KnuthDiv
1906 /// has specific constraints on its inputs. If those constraints are not met
1907 /// then it provides a simpler form of divide.
1908 static void divide(const WordType *LHS, unsigned lhsWords,
1909 const WordType *RHS, unsigned rhsWords, WordType *Quotient,
1910 WordType *Remainder);
1911
1912 /// out-of-line slow case for inline constructor
1913 void initSlowCase(uint64_t val, bool isSigned);
1914
1915 /// shared code between two array constructors
1916 void initFromArray(ArrayRef<uint64_t> array);
1917
1918 /// out-of-line slow case for inline copy constructor
1919 void initSlowCase(const APInt &that);
1920
1921 /// out-of-line slow case for shl
1922 void shlSlowCase(unsigned ShiftAmt);
1923
1924 /// out-of-line slow case for lshr.
1925 void lshrSlowCase(unsigned ShiftAmt);
1926
1927 /// out-of-line slow case for ashr.
1928 void ashrSlowCase(unsigned ShiftAmt);
1929
1930 /// out-of-line slow case for operator=
1931 void assignSlowCase(const APInt &RHS);
1932
1933 /// out-of-line slow case for operator==
1934 bool equalSlowCase(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__));
1935
1936 /// out-of-line slow case for countLeadingZeros
1937 unsigned countLeadingZerosSlowCase() const LLVM_READONLY__attribute__((__pure__));
1938
1939 /// out-of-line slow case for countLeadingOnes.
1940 unsigned countLeadingOnesSlowCase() const LLVM_READONLY__attribute__((__pure__));
1941
1942 /// out-of-line slow case for countTrailingZeros.
1943 unsigned countTrailingZerosSlowCase() const LLVM_READONLY__attribute__((__pure__));
1944
1945 /// out-of-line slow case for countTrailingOnes
1946 unsigned countTrailingOnesSlowCase() const LLVM_READONLY__attribute__((__pure__));
1947
1948 /// out-of-line slow case for countPopulation
1949 unsigned countPopulationSlowCase() const LLVM_READONLY__attribute__((__pure__));
1950
1951 /// out-of-line slow case for intersects.
1952 bool intersectsSlowCase(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__));
1953
1954 /// out-of-line slow case for isSubsetOf.
1955 bool isSubsetOfSlowCase(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__));
1956
1957 /// out-of-line slow case for setBits.
1958 void setBitsSlowCase(unsigned loBit, unsigned hiBit);
1959
1960 /// out-of-line slow case for flipAllBits.
1961 void flipAllBitsSlowCase();
1962
1963 /// out-of-line slow case for concat.
1964 APInt concatSlowCase(const APInt &NewLSB) const;
1965
1966 /// out-of-line slow case for operator&=.
1967 void andAssignSlowCase(const APInt &RHS);
1968
1969 /// out-of-line slow case for operator|=.
1970 void orAssignSlowCase(const APInt &RHS);
1971
1972 /// out-of-line slow case for operator^=.
1973 void xorAssignSlowCase(const APInt &RHS);
1974
1975 /// Unsigned comparison. Returns -1, 0, or 1 if this APInt is less than, equal
1976 /// to, or greater than RHS.
1977 int compare(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__));
1978
1979 /// Signed comparison. Returns -1, 0, or 1 if this APInt is less than, equal
1980 /// to, or greater than RHS.
1981 int compareSigned(const APInt &RHS) const LLVM_READONLY__attribute__((__pure__));
1982
1983 /// @}
1984};
1985
1986inline bool operator==(uint64_t V1, const APInt &V2) { return V2 == V1; }
1987
1988inline bool operator!=(uint64_t V1, const APInt &V2) { return V2 != V1; }
1989
1990/// Unary bitwise complement operator.
1991///
1992/// \returns an APInt that is the bitwise complement of \p v.
1993inline APInt operator~(APInt v) {
1994 v.flipAllBits();
1995 return v;
1996}
1997
1998inline APInt operator&(APInt a, const APInt &b) {
1999 a &= b;
2000 return a;
2001}
2002
2003inline APInt operator&(const APInt &a, APInt &&b) {
2004 b &= a;
2005 return std::move(b);
2006}
2007
2008inline APInt operator&(APInt a, uint64_t RHS) {
2009 a &= RHS;
2010 return a;
2011}
2012
2013inline APInt operator&(uint64_t LHS, APInt b) {
2014 b &= LHS;
2015 return b;
2016}
2017
2018inline APInt operator|(APInt a, const APInt &b) {
2019 a |= b;
2020 return a;
2021}
2022
2023inline APInt operator|(const APInt &a, APInt &&b) {
2024 b |= a;
2025 return std::move(b);
2026}
2027
2028inline APInt operator|(APInt a, uint64_t RHS) {
2029 a |= RHS;
2030 return a;
2031}
2032
2033inline APInt operator|(uint64_t LHS, APInt b) {
2034 b |= LHS;
2035 return b;
2036}
2037
2038inline APInt operator^(APInt a, const APInt &b) {
2039 a ^= b;
2040 return a;
2041}
2042
2043inline APInt operator^(const APInt &a, APInt &&b) {
2044 b ^= a;
2045 return std::move(b);
2046}
2047
2048inline APInt operator^(APInt a, uint64_t RHS) {
2049 a ^= RHS;
2050 return a;
2051}
2052
2053inline APInt operator^(uint64_t LHS, APInt b) {
2054 b ^= LHS;
2055 return b;
2056}
2057
2058inline raw_ostream &operator<<(raw_ostream &OS, const APInt &I) {
2059 I.print(OS, true);
2060 return OS;
2061}
2062
2063inline APInt operator-(APInt v) {
2064 v.negate();
2065 return v;
2066}
2067
2068inline APInt operator+(APInt a, const APInt &b) {
2069 a += b;
2070 return a;
2071}
2072
2073inline APInt operator+(const APInt &a, APInt &&b) {
2074 b += a;
2075 return std::move(b);
2076}
2077
2078inline APInt operator+(APInt a, uint64_t RHS) {
2079 a += RHS;
2080 return a;
2081}
2082
2083inline APInt operator+(uint64_t LHS, APInt b) {
2084 b += LHS;
2085 return b;
2086}
2087
2088inline APInt operator-(APInt a, const APInt &b) {
2089 a -= b;
2090 return a;
2091}
2092
2093inline APInt operator-(const APInt &a, APInt &&b) {
2094 b.negate();
2095 b += a;
2096 return std::move(b);
2097}
2098
2099inline APInt operator-(APInt a, uint64_t RHS) {
2100 a -= RHS;
2101 return a;
2102}
2103
2104inline APInt operator-(uint64_t LHS, APInt b) {
2105 b.negate();
2106 b += LHS;
2107 return b;
2108}
2109
2110inline APInt operator*(APInt a, uint64_t RHS) {
2111 a *= RHS;
2112 return a;
2113}
2114
2115inline APInt operator*(uint64_t LHS, APInt b) {
2116 b *= LHS;
2117 return b;
2118}
2119
2120namespace APIntOps {
2121
2122/// Determine the smaller of two APInts considered to be signed.
2123inline const APInt &smin(const APInt &A, const APInt &B) {
2124 return A.slt(B) ? A : B;
2125}
2126
2127/// Determine the larger of two APInts considered to be signed.
2128inline const APInt &smax(const APInt &A, const APInt &B) {
2129 return A.sgt(B) ? A : B;
2130}
2131
2132/// Determine the smaller of two APInts considered to be unsigned.
2133inline const APInt &umin(const APInt &A, const APInt &B) {
2134 return A.ult(B) ? A : B;
2135}
2136
2137/// Determine the larger of two APInts considered to be unsigned.
2138inline const APInt &umax(const APInt &A, const APInt &B) {
2139 return A.ugt(B) ? A : B;
2140}
2141
2142/// Compute GCD of two unsigned APInt values.
2143///
2144/// This function returns the greatest common divisor of the two APInt values
2145/// using Stein's algorithm.
2146///
2147/// \returns the greatest common divisor of A and B.
2148APInt GreatestCommonDivisor(APInt A, APInt B);
2149
2150/// Converts the given APInt to a double value.
2151///
2152/// Treats the APInt as an unsigned value for conversion purposes.
2153inline double RoundAPIntToDouble(const APInt &APIVal) {
2154 return APIVal.roundToDouble();
2155}
2156
2157/// Converts the given APInt to a double value.
2158///
2159/// Treats the APInt as a signed value for conversion purposes.
2160inline double RoundSignedAPIntToDouble(const APInt &APIVal) {
2161 return APIVal.signedRoundToDouble();
2162}
2163
2164/// Converts the given APInt to a float value.
2165inline float RoundAPIntToFloat(const APInt &APIVal) {
2166 return float(RoundAPIntToDouble(APIVal));
2167}
2168
2169/// Converts the given APInt to a float value.
2170///
2171/// Treats the APInt as a signed value for conversion purposes.
2172inline float RoundSignedAPIntToFloat(const APInt &APIVal) {
2173 return float(APIVal.signedRoundToDouble());
2174}
2175
2176/// Converts the given double value into a APInt.
2177///
2178/// This function convert a double value to an APInt value.
2179APInt RoundDoubleToAPInt(double Double, unsigned width);
2180
2181/// Converts a float value into a APInt.
2182///
2183/// Converts a float value into an APInt value.
2184inline APInt RoundFloatToAPInt(float Float, unsigned width) {
2185 return RoundDoubleToAPInt(double(Float), width);
2186}
2187
2188/// Return A unsign-divided by B, rounded by the given rounding mode.
2189APInt RoundingUDiv(const APInt &A, const APInt &B, APInt::Rounding RM);
2190
2191/// Return A sign-divided by B, rounded by the given rounding mode.
2192APInt RoundingSDiv(const APInt &A, const APInt &B, APInt::Rounding RM);
2193
2194/// Let q(n) = An^2 + Bn + C, and BW = bit width of the value range
2195/// (e.g. 32 for i32).
2196/// This function finds the smallest number n, such that
2197/// (a) n >= 0 and q(n) = 0, or
2198/// (b) n >= 1 and q(n-1) and q(n), when evaluated in the set of all
2199/// integers, belong to two different intervals [Rk, Rk+R),
2200/// where R = 2^BW, and k is an integer.
2201/// The idea here is to find when q(n) "overflows" 2^BW, while at the
2202/// same time "allowing" subtraction. In unsigned modulo arithmetic a
2203/// subtraction (treated as addition of negated numbers) would always
2204/// count as an overflow, but here we want to allow values to decrease
2205/// and increase as long as they are within the same interval.
2206/// Specifically, adding of two negative numbers should not cause an
2207/// overflow (as long as the magnitude does not exceed the bit width).
2208/// On the other hand, given a positive number, adding a negative
2209/// number to it can give a negative result, which would cause the
2210/// value to go from [-2^BW, 0) to [0, 2^BW). In that sense, zero is
2211/// treated as a special case of an overflow.
2212///
2213/// This function returns None if after finding k that minimizes the
2214/// positive solution to q(n) = kR, both solutions are contained between
2215/// two consecutive integers.
2216///
2217/// There are cases where q(n) > T, and q(n+1) < T (assuming evaluation
2218/// in arithmetic modulo 2^BW, and treating the values as signed) by the
2219/// virtue of *signed* overflow. This function will *not* find such an n,
2220/// however it may find a value of n satisfying the inequalities due to
2221/// an *unsigned* overflow (if the values are treated as unsigned).
2222/// To find a solution for a signed overflow, treat it as a problem of
2223/// finding an unsigned overflow with a range with of BW-1.
2224///
2225/// The returned value may have a different bit width from the input
2226/// coefficients.
2227Optional<APInt> SolveQuadraticEquationWrap(APInt A, APInt B, APInt C,
2228 unsigned RangeWidth);
2229
2230/// Compare two values, and if they are different, return the position of the
2231/// most significant bit that is different in the values.
2232Optional<unsigned> GetMostSignificantDifferentBit(const APInt &A,
2233 const APInt &B);
2234
2235/// Splat/Merge neighboring bits to widen/narrow the bitmask represented
2236/// by \param A to \param NewBitWidth bits.
2237///
2238/// e.g. ScaleBitMask(0b0101, 8) -> 0b00110011
2239/// e.g. ScaleBitMask(0b00011011, 4) -> 0b0111
2240/// A.getBitwidth() or NewBitWidth must be a whole multiples of the other.
2241///
2242/// TODO: Do we need a mode where all bits must be set when merging down?
2243APInt ScaleBitMask(const APInt &A, unsigned NewBitWidth);
2244} // namespace APIntOps
2245
2246// See friend declaration above. This additional declaration is required in
2247// order to compile LLVM with IBM xlC compiler.
2248hash_code hash_value(const APInt &Arg);
2249
2250/// StoreIntToMemory - Fills the StoreBytes bytes of memory starting from Dst
2251/// with the integer held in IntVal.
2252void StoreIntToMemory(const APInt &IntVal, uint8_t *Dst, unsigned StoreBytes);
2253
2254/// LoadIntFromMemory - Loads the integer stored in the LoadBytes bytes starting
2255/// from Src into IntVal, which is assumed to be wide enough and to hold zero.
2256void LoadIntFromMemory(APInt &IntVal, const uint8_t *Src, unsigned LoadBytes);
2257
2258/// Provide DenseMapInfo for APInt.
2259template <> struct DenseMapInfo<APInt, void> {
2260 static inline APInt getEmptyKey() {
2261 APInt V(nullptr, 0);
2262 V.U.VAL = 0;
2263 return V;
2264 }
2265
2266 static inline APInt getTombstoneKey() {
2267 APInt V(nullptr, 0);
2268 V.U.VAL = 1;
2269 return V;
2270 }
2271
2272 static unsigned getHashValue(const APInt &Key);
2273
2274 static bool isEqual(const APInt &LHS, const APInt &RHS) {
2275 return LHS.getBitWidth() == RHS.getBitWidth() && LHS == RHS;
2276 }
2277};
2278
2279} // namespace llvm
2280
2281#endif