Bug Summary

File:llvm/include/llvm/IR/Instructions.h
Warning:line 2525, column 17
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name ValueTracking.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 -mthread-model posix -mframe-pointer=none -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -target-cpu x86-64 -dwarf-column-info -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/lib/Analysis -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/include -I /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/build-llvm/lib/Analysis -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-01-13-084841-49055-1 -x c++ /build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp

/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp

1//===- ValueTracking.cpp - Walk computations to compute properties --------===//
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 contains routines that help analyze properties that chains of
10// computations have.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Analysis/ValueTracking.h"
15#include "llvm/ADT/APFloat.h"
16#include "llvm/ADT/APInt.h"
17#include "llvm/ADT/ArrayRef.h"
18#include "llvm/ADT/None.h"
19#include "llvm/ADT/Optional.h"
20#include "llvm/ADT/STLExtras.h"
21#include "llvm/ADT/SmallPtrSet.h"
22#include "llvm/ADT/SmallSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/StringRef.h"
25#include "llvm/ADT/iterator_range.h"
26#include "llvm/Analysis/AliasAnalysis.h"
27#include "llvm/Analysis/AssumptionCache.h"
28#include "llvm/Analysis/GuardUtils.h"
29#include "llvm/Analysis/InstructionSimplify.h"
30#include "llvm/Analysis/Loads.h"
31#include "llvm/Analysis/LoopInfo.h"
32#include "llvm/Analysis/OptimizationRemarkEmitter.h"
33#include "llvm/Analysis/TargetLibraryInfo.h"
34#include "llvm/IR/Argument.h"
35#include "llvm/IR/Attributes.h"
36#include "llvm/IR/BasicBlock.h"
37#include "llvm/IR/CallSite.h"
38#include "llvm/IR/Constant.h"
39#include "llvm/IR/ConstantRange.h"
40#include "llvm/IR/Constants.h"
41#include "llvm/IR/DerivedTypes.h"
42#include "llvm/IR/DiagnosticInfo.h"
43#include "llvm/IR/Dominators.h"
44#include "llvm/IR/Function.h"
45#include "llvm/IR/GetElementPtrTypeIterator.h"
46#include "llvm/IR/GlobalAlias.h"
47#include "llvm/IR/GlobalValue.h"
48#include "llvm/IR/GlobalVariable.h"
49#include "llvm/IR/InstrTypes.h"
50#include "llvm/IR/Instruction.h"
51#include "llvm/IR/Instructions.h"
52#include "llvm/IR/IntrinsicInst.h"
53#include "llvm/IR/Intrinsics.h"
54#include "llvm/IR/IntrinsicsAArch64.h"
55#include "llvm/IR/IntrinsicsX86.h"
56#include "llvm/IR/LLVMContext.h"
57#include "llvm/IR/Metadata.h"
58#include "llvm/IR/Module.h"
59#include "llvm/IR/Operator.h"
60#include "llvm/IR/PatternMatch.h"
61#include "llvm/IR/Type.h"
62#include "llvm/IR/User.h"
63#include "llvm/IR/Value.h"
64#include "llvm/Support/Casting.h"
65#include "llvm/Support/CommandLine.h"
66#include "llvm/Support/Compiler.h"
67#include "llvm/Support/ErrorHandling.h"
68#include "llvm/Support/KnownBits.h"
69#include "llvm/Support/MathExtras.h"
70#include <algorithm>
71#include <array>
72#include <cassert>
73#include <cstdint>
74#include <iterator>
75#include <utility>
76
77using namespace llvm;
78using namespace llvm::PatternMatch;
79
80const unsigned MaxDepth = 6;
81
82// Controls the number of uses of the value searched for possible
83// dominating comparisons.
84static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
85 cl::Hidden, cl::init(20));
86
87/// Returns the bitwidth of the given scalar or pointer type. For vector types,
88/// returns the element type's bitwidth.
89static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
90 if (unsigned BitWidth = Ty->getScalarSizeInBits())
91 return BitWidth;
92
93 return DL.getPointerTypeSizeInBits(Ty);
94}
95
96namespace {
97
98// Simplifying using an assume can only be done in a particular control-flow
99// context (the context instruction provides that context). If an assume and
100// the context instruction are not in the same block then the DT helps in
101// figuring out if we can use it.
102struct Query {
103 const DataLayout &DL;
104 AssumptionCache *AC;
105 const Instruction *CxtI;
106 const DominatorTree *DT;
107
108 // Unlike the other analyses, this may be a nullptr because not all clients
109 // provide it currently.
110 OptimizationRemarkEmitter *ORE;
111
112 /// Set of assumptions that should be excluded from further queries.
113 /// This is because of the potential for mutual recursion to cause
114 /// computeKnownBits to repeatedly visit the same assume intrinsic. The
115 /// classic case of this is assume(x = y), which will attempt to determine
116 /// bits in x from bits in y, which will attempt to determine bits in y from
117 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
118 /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo
119 /// (all of which can call computeKnownBits), and so on.
120 std::array<const Value *, MaxDepth> Excluded;
121
122 /// If true, it is safe to use metadata during simplification.
123 InstrInfoQuery IIQ;
124
125 unsigned NumExcluded = 0;
126
127 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
128 const DominatorTree *DT, bool UseInstrInfo,
129 OptimizationRemarkEmitter *ORE = nullptr)
130 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {}
131
132 Query(const Query &Q, const Value *NewExcl)
133 : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE), IIQ(Q.IIQ),
134 NumExcluded(Q.NumExcluded) {
135 Excluded = Q.Excluded;
136 Excluded[NumExcluded++] = NewExcl;
137 assert(NumExcluded <= Excluded.size())((NumExcluded <= Excluded.size()) ? static_cast<void>
(0) : __assert_fail ("NumExcluded <= Excluded.size()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 137, __PRETTY_FUNCTION__))
;
138 }
139
140 bool isExcluded(const Value *Value) const {
141 if (NumExcluded == 0)
142 return false;
143 auto End = Excluded.begin() + NumExcluded;
144 return std::find(Excluded.begin(), End, Value) != End;
145 }
146};
147
148} // end anonymous namespace
149
150// Given the provided Value and, potentially, a context instruction, return
151// the preferred context instruction (if any).
152static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
153 // If we've been provided with a context instruction, then use that (provided
154 // it has been inserted).
155 if (CxtI && CxtI->getParent())
156 return CxtI;
157
158 // If the value is really an already-inserted instruction, then use that.
159 CxtI = dyn_cast<Instruction>(V);
160 if (CxtI && CxtI->getParent())
161 return CxtI;
162
163 return nullptr;
164}
165
166static void computeKnownBits(const Value *V, KnownBits &Known,
167 unsigned Depth, const Query &Q);
168
169void llvm::computeKnownBits(const Value *V, KnownBits &Known,
170 const DataLayout &DL, unsigned Depth,
171 AssumptionCache *AC, const Instruction *CxtI,
172 const DominatorTree *DT,
173 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) {
174 ::computeKnownBits(V, Known, Depth,
175 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
176}
177
178static KnownBits computeKnownBits(const Value *V, unsigned Depth,
179 const Query &Q);
180
181KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL,
182 unsigned Depth, AssumptionCache *AC,
183 const Instruction *CxtI,
184 const DominatorTree *DT,
185 OptimizationRemarkEmitter *ORE,
186 bool UseInstrInfo) {
187 return ::computeKnownBits(
188 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE));
189}
190
191bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
192 const DataLayout &DL, AssumptionCache *AC,
193 const Instruction *CxtI, const DominatorTree *DT,
194 bool UseInstrInfo) {
195 assert(LHS->getType() == RHS->getType() &&((LHS->getType() == RHS->getType() && "LHS and RHS should have the same type"
) ? static_cast<void> (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"LHS and RHS should have the same type\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 196, __PRETTY_FUNCTION__))
196 "LHS and RHS should have the same type")((LHS->getType() == RHS->getType() && "LHS and RHS should have the same type"
) ? static_cast<void> (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"LHS and RHS should have the same type\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 196, __PRETTY_FUNCTION__))
;
197 assert(LHS->getType()->isIntOrIntVectorTy() &&((LHS->getType()->isIntOrIntVectorTy() && "LHS and RHS should be integers"
) ? static_cast<void> (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy() && \"LHS and RHS should be integers\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 198, __PRETTY_FUNCTION__))
198 "LHS and RHS should be integers")((LHS->getType()->isIntOrIntVectorTy() && "LHS and RHS should be integers"
) ? static_cast<void> (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy() && \"LHS and RHS should be integers\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 198, __PRETTY_FUNCTION__))
;
199 // Look for an inverted mask: (X & ~M) op (Y & M).
200 Value *M;
201 if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
202 match(RHS, m_c_And(m_Specific(M), m_Value())))
203 return true;
204 if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
205 match(LHS, m_c_And(m_Specific(M), m_Value())))
206 return true;
207 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
208 KnownBits LHSKnown(IT->getBitWidth());
209 KnownBits RHSKnown(IT->getBitWidth());
210 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
211 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo);
212 return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
213}
214
215bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) {
216 for (const User *U : CxtI->users()) {
217 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
218 if (IC->isEquality())
219 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
220 if (C->isNullValue())
221 continue;
222 return false;
223 }
224 return true;
225}
226
227static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
228 const Query &Q);
229
230bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
231 bool OrZero, unsigned Depth,
232 AssumptionCache *AC, const Instruction *CxtI,
233 const DominatorTree *DT, bool UseInstrInfo) {
234 return ::isKnownToBeAPowerOfTwo(
235 V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
236}
237
238static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
239
240bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
241 AssumptionCache *AC, const Instruction *CxtI,
242 const DominatorTree *DT, bool UseInstrInfo) {
243 return ::isKnownNonZero(V, Depth,
244 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
245}
246
247bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
248 unsigned Depth, AssumptionCache *AC,
249 const Instruction *CxtI, const DominatorTree *DT,
250 bool UseInstrInfo) {
251 KnownBits Known =
252 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
253 return Known.isNonNegative();
254}
255
256bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
257 AssumptionCache *AC, const Instruction *CxtI,
258 const DominatorTree *DT, bool UseInstrInfo) {
259 if (auto *CI = dyn_cast<ConstantInt>(V))
260 return CI->getValue().isStrictlyPositive();
261
262 // TODO: We'd doing two recursive queries here. We should factor this such
263 // that only a single query is needed.
264 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) &&
265 isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo);
266}
267
268bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
269 AssumptionCache *AC, const Instruction *CxtI,
270 const DominatorTree *DT, bool UseInstrInfo) {
271 KnownBits Known =
272 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo);
273 return Known.isNegative();
274}
275
276static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
277
278bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
279 const DataLayout &DL, AssumptionCache *AC,
280 const Instruction *CxtI, const DominatorTree *DT,
281 bool UseInstrInfo) {
282 return ::isKnownNonEqual(V1, V2,
283 Query(DL, AC, safeCxtI(V1, safeCxtI(V2, CxtI)), DT,
284 UseInstrInfo, /*ORE=*/nullptr));
285}
286
287static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
288 const Query &Q);
289
290bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
291 const DataLayout &DL, unsigned Depth,
292 AssumptionCache *AC, const Instruction *CxtI,
293 const DominatorTree *DT, bool UseInstrInfo) {
294 return ::MaskedValueIsZero(
295 V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
296}
297
298static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
299 const Query &Q);
300
301unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
302 unsigned Depth, AssumptionCache *AC,
303 const Instruction *CxtI,
304 const DominatorTree *DT, bool UseInstrInfo) {
305 return ::ComputeNumSignBits(
306 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo));
307}
308
309static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
310 bool NSW,
311 KnownBits &KnownOut, KnownBits &Known2,
312 unsigned Depth, const Query &Q) {
313 unsigned BitWidth = KnownOut.getBitWidth();
314
315 // If an initial sequence of bits in the result is not needed, the
316 // corresponding bits in the operands are not needed.
317 KnownBits LHSKnown(BitWidth);
318 computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
319 computeKnownBits(Op1, Known2, Depth + 1, Q);
320
321 KnownOut = KnownBits::computeForAddSub(Add, NSW, LHSKnown, Known2);
322}
323
324static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
325 KnownBits &Known, KnownBits &Known2,
326 unsigned Depth, const Query &Q) {
327 unsigned BitWidth = Known.getBitWidth();
328 computeKnownBits(Op1, Known, Depth + 1, Q);
329 computeKnownBits(Op0, Known2, Depth + 1, Q);
330
331 bool isKnownNegative = false;
332 bool isKnownNonNegative = false;
333 // If the multiplication is known not to overflow, compute the sign bit.
334 if (NSW) {
335 if (Op0 == Op1) {
336 // The product of a number with itself is non-negative.
337 isKnownNonNegative = true;
338 } else {
339 bool isKnownNonNegativeOp1 = Known.isNonNegative();
340 bool isKnownNonNegativeOp0 = Known2.isNonNegative();
341 bool isKnownNegativeOp1 = Known.isNegative();
342 bool isKnownNegativeOp0 = Known2.isNegative();
343 // The product of two numbers with the same sign is non-negative.
344 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
345 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
346 // The product of a negative number and a non-negative number is either
347 // negative or zero.
348 if (!isKnownNonNegative)
349 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
350 isKnownNonZero(Op0, Depth, Q)) ||
351 (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
352 isKnownNonZero(Op1, Depth, Q));
353 }
354 }
355
356 assert(!Known.hasConflict() && !Known2.hasConflict())((!Known.hasConflict() && !Known2.hasConflict()) ? static_cast
<void> (0) : __assert_fail ("!Known.hasConflict() && !Known2.hasConflict()"
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 356, __PRETTY_FUNCTION__))
;
357 // Compute a conservative estimate for high known-0 bits.
358 unsigned LeadZ = std::max(Known.countMinLeadingZeros() +
359 Known2.countMinLeadingZeros(),
360 BitWidth) - BitWidth;
361 LeadZ = std::min(LeadZ, BitWidth);
362
363 // The result of the bottom bits of an integer multiply can be
364 // inferred by looking at the bottom bits of both operands and
365 // multiplying them together.
366 // We can infer at least the minimum number of known trailing bits
367 // of both operands. Depending on number of trailing zeros, we can
368 // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming
369 // a and b are divisible by m and n respectively.
370 // We then calculate how many of those bits are inferrable and set
371 // the output. For example, the i8 mul:
372 // a = XXXX1100 (12)
373 // b = XXXX1110 (14)
374 // We know the bottom 3 bits are zero since the first can be divided by
375 // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4).
376 // Applying the multiplication to the trimmed arguments gets:
377 // XX11 (3)
378 // X111 (7)
379 // -------
380 // XX11
381 // XX11
382 // XX11
383 // XX11
384 // -------
385 // XXXXX01
386 // Which allows us to infer the 2 LSBs. Since we're multiplying the result
387 // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits.
388 // The proof for this can be described as:
389 // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) &&
390 // (C7 == (1 << (umin(countTrailingZeros(C1), C5) +
391 // umin(countTrailingZeros(C2), C6) +
392 // umin(C5 - umin(countTrailingZeros(C1), C5),
393 // C6 - umin(countTrailingZeros(C2), C6)))) - 1)
394 // %aa = shl i8 %a, C5
395 // %bb = shl i8 %b, C6
396 // %aaa = or i8 %aa, C1
397 // %bbb = or i8 %bb, C2
398 // %mul = mul i8 %aaa, %bbb
399 // %mask = and i8 %mul, C7
400 // =>
401 // %mask = i8 ((C1*C2)&C7)
402 // Where C5, C6 describe the known bits of %a, %b
403 // C1, C2 describe the known bottom bits of %a, %b.
404 // C7 describes the mask of the known bits of the result.
405 APInt Bottom0 = Known.One;
406 APInt Bottom1 = Known2.One;
407
408 // How many times we'd be able to divide each argument by 2 (shr by 1).
409 // This gives us the number of trailing zeros on the multiplication result.
410 unsigned TrailBitsKnown0 = (Known.Zero | Known.One).countTrailingOnes();
411 unsigned TrailBitsKnown1 = (Known2.Zero | Known2.One).countTrailingOnes();
412 unsigned TrailZero0 = Known.countMinTrailingZeros();
413 unsigned TrailZero1 = Known2.countMinTrailingZeros();
414 unsigned TrailZ = TrailZero0 + TrailZero1;
415
416 // Figure out the fewest known-bits operand.
417 unsigned SmallestOperand = std::min(TrailBitsKnown0 - TrailZero0,
418 TrailBitsKnown1 - TrailZero1);
419 unsigned ResultBitsKnown = std::min(SmallestOperand + TrailZ, BitWidth);
420
421 APInt BottomKnown = Bottom0.getLoBits(TrailBitsKnown0) *
422 Bottom1.getLoBits(TrailBitsKnown1);
423
424 Known.resetAll();
425 Known.Zero.setHighBits(LeadZ);
426 Known.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
427 Known.One |= BottomKnown.getLoBits(ResultBitsKnown);
428
429 // Only make use of no-wrap flags if we failed to compute the sign bit
430 // directly. This matters if the multiplication always overflows, in
431 // which case we prefer to follow the result of the direct computation,
432 // though as the program is invoking undefined behaviour we can choose
433 // whatever we like here.
434 if (isKnownNonNegative && !Known.isNegative())
435 Known.makeNonNegative();
436 else if (isKnownNegative && !Known.isNonNegative())
437 Known.makeNegative();
438}
439
440void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
441 KnownBits &Known) {
442 unsigned BitWidth = Known.getBitWidth();
443 unsigned NumRanges = Ranges.getNumOperands() / 2;
444 assert(NumRanges >= 1)((NumRanges >= 1) ? static_cast<void> (0) : __assert_fail
("NumRanges >= 1", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 444, __PRETTY_FUNCTION__))
;
445
446 Known.Zero.setAllBits();
447 Known.One.setAllBits();
448
449 for (unsigned i = 0; i < NumRanges; ++i) {
450 ConstantInt *Lower =
451 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
452 ConstantInt *Upper =
453 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
454 ConstantRange Range(Lower->getValue(), Upper->getValue());
455
456 // The first CommonPrefixBits of all values in Range are equal.
457 unsigned CommonPrefixBits =
458 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
459
460 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
461 Known.One &= Range.getUnsignedMax() & Mask;
462 Known.Zero &= ~Range.getUnsignedMax() & Mask;
463 }
464}
465
466static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
467 SmallVector<const Value *, 16> WorkSet(1, I);
468 SmallPtrSet<const Value *, 32> Visited;
469 SmallPtrSet<const Value *, 16> EphValues;
470
471 // The instruction defining an assumption's condition itself is always
472 // considered ephemeral to that assumption (even if it has other
473 // non-ephemeral users). See r246696's test case for an example.
474 if (is_contained(I->operands(), E))
475 return true;
476
477 while (!WorkSet.empty()) {
478 const Value *V = WorkSet.pop_back_val();
479 if (!Visited.insert(V).second)
480 continue;
481
482 // If all uses of this value are ephemeral, then so is this value.
483 if (llvm::all_of(V->users(), [&](const User *U) {
484 return EphValues.count(U);
485 })) {
486 if (V == E)
487 return true;
488
489 if (V == I || isSafeToSpeculativelyExecute(V)) {
490 EphValues.insert(V);
491 if (const User *U = dyn_cast<User>(V))
492 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
493 J != JE; ++J)
494 WorkSet.push_back(*J);
495 }
496 }
497 }
498
499 return false;
500}
501
502// Is this an intrinsic that cannot be speculated but also cannot trap?
503bool llvm::isAssumeLikeIntrinsic(const Instruction *I) {
504 if (const CallInst *CI = dyn_cast<CallInst>(I))
505 if (Function *F = CI->getCalledFunction())
506 switch (F->getIntrinsicID()) {
507 default: break;
508 // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
509 case Intrinsic::assume:
510 case Intrinsic::sideeffect:
511 case Intrinsic::dbg_declare:
512 case Intrinsic::dbg_value:
513 case Intrinsic::dbg_label:
514 case Intrinsic::invariant_start:
515 case Intrinsic::invariant_end:
516 case Intrinsic::lifetime_start:
517 case Intrinsic::lifetime_end:
518 case Intrinsic::objectsize:
519 case Intrinsic::ptr_annotation:
520 case Intrinsic::var_annotation:
521 return true;
522 }
523
524 return false;
525}
526
527bool llvm::isValidAssumeForContext(const Instruction *Inv,
528 const Instruction *CxtI,
529 const DominatorTree *DT) {
530 // There are two restrictions on the use of an assume:
531 // 1. The assume must dominate the context (or the control flow must
532 // reach the assume whenever it reaches the context).
533 // 2. The context must not be in the assume's set of ephemeral values
534 // (otherwise we will use the assume to prove that the condition
535 // feeding the assume is trivially true, thus causing the removal of
536 // the assume).
537
538 if (DT) {
539 if (DT->dominates(Inv, CxtI))
540 return true;
541 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
542 // We don't have a DT, but this trivially dominates.
543 return true;
544 }
545
546 // With or without a DT, the only remaining case we will check is if the
547 // instructions are in the same BB. Give up if that is not the case.
548 if (Inv->getParent() != CxtI->getParent())
549 return false;
550
551 // If we have a dom tree, then we now know that the assume doesn't dominate
552 // the other instruction. If we don't have a dom tree then we can check if
553 // the assume is first in the BB.
554 if (!DT) {
555 // Search forward from the assume until we reach the context (or the end
556 // of the block); the common case is that the assume will come first.
557 for (auto I = std::next(BasicBlock::const_iterator(Inv)),
558 IE = Inv->getParent()->end(); I != IE; ++I)
559 if (&*I == CxtI)
560 return true;
561 }
562
563 // Don't let an assume affect itself - this would cause the problems
564 // `isEphemeralValueOf` is trying to prevent, and it would also make
565 // the loop below go out of bounds.
566 if (Inv == CxtI)
567 return false;
568
569 // The context comes first, but they're both in the same block.
570 // Make sure there is nothing in between that might interrupt
571 // the control flow, not even CxtI itself.
572 for (BasicBlock::const_iterator I(CxtI), IE(Inv); I != IE; ++I)
573 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
574 return false;
575
576 return !isEphemeralValueOf(Inv, CxtI);
577}
578
579static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) {
580 // Use of assumptions is context-sensitive. If we don't have a context, we
581 // cannot use them!
582 if (!Q.AC || !Q.CxtI)
583 return false;
584
585 // Note that the patterns below need to be kept in sync with the code
586 // in AssumptionCache::updateAffectedValues.
587
588 auto CmpExcludesZero = [V](ICmpInst *Cmp) {
589 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V)));
590
591 Value *RHS;
592 CmpInst::Predicate Pred;
593 if (!match(Cmp, m_c_ICmp(Pred, m_V, m_Value(RHS))))
594 return false;
595 // Canonicalize 'v' to be on the LHS of the comparison.
596 if (Cmp->getOperand(1) != RHS)
597 Pred = CmpInst::getSwappedPredicate(Pred);
598
599 // assume(v u> y) -> assume(v != 0)
600 if (Pred == ICmpInst::ICMP_UGT)
601 return true;
602
603 // assume(v != 0)
604 // We special-case this one to ensure that we handle `assume(v != null)`.
605 if (Pred == ICmpInst::ICMP_NE)
606 return match(RHS, m_Zero());
607
608 // All other predicates - rely on generic ConstantRange handling.
609 ConstantInt *CI;
610 if (!match(RHS, m_ConstantInt(CI)))
611 return false;
612 ConstantRange RHSRange(CI->getValue());
613 ConstantRange TrueValues =
614 ConstantRange::makeAllowedICmpRegion(Pred, RHSRange);
615 return !TrueValues.contains(APInt::getNullValue(CI->getBitWidth()));
616 };
617
618 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
619 if (!AssumeVH)
620 continue;
621 CallInst *I = cast<CallInst>(AssumeVH);
622 assert(I->getFunction() == Q.CxtI->getFunction() &&((I->getFunction() == Q.CxtI->getFunction() && "Got assumption for the wrong function!"
) ? static_cast<void> (0) : __assert_fail ("I->getFunction() == Q.CxtI->getFunction() && \"Got assumption for the wrong function!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 623, __PRETTY_FUNCTION__))
623 "Got assumption for the wrong function!")((I->getFunction() == Q.CxtI->getFunction() && "Got assumption for the wrong function!"
) ? static_cast<void> (0) : __assert_fail ("I->getFunction() == Q.CxtI->getFunction() && \"Got assumption for the wrong function!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 623, __PRETTY_FUNCTION__))
;
624 if (Q.isExcluded(I))
625 continue;
626
627 // Warning: This loop can end up being somewhat performance sensitive.
628 // We're running this loop for once for each value queried resulting in a
629 // runtime of ~O(#assumes * #values).
630
631 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&((I->getCalledFunction()->getIntrinsicID() == Intrinsic
::assume && "must be an assume intrinsic") ? static_cast
<void> (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 632, __PRETTY_FUNCTION__))
632 "must be an assume intrinsic")((I->getCalledFunction()->getIntrinsicID() == Intrinsic
::assume && "must be an assume intrinsic") ? static_cast
<void> (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 632, __PRETTY_FUNCTION__))
;
633
634 Value *Arg = I->getArgOperand(0);
635 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg);
636 if (!Cmp)
637 continue;
638
639 if (CmpExcludesZero(Cmp) && isValidAssumeForContext(I, Q.CxtI, Q.DT))
640 return true;
641 }
642
643 return false;
644}
645
646static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
647 unsigned Depth, const Query &Q) {
648 // Use of assumptions is context-sensitive. If we don't have a context, we
649 // cannot use them!
650 if (!Q.AC || !Q.CxtI)
651 return;
652
653 unsigned BitWidth = Known.getBitWidth();
654
655 // Note that the patterns below need to be kept in sync with the code
656 // in AssumptionCache::updateAffectedValues.
657
658 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
659 if (!AssumeVH)
660 continue;
661 CallInst *I = cast<CallInst>(AssumeVH);
662 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&((I->getParent()->getParent() == Q.CxtI->getParent()
->getParent() && "Got assumption for the wrong function!"
) ? static_cast<void> (0) : __assert_fail ("I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 663, __PRETTY_FUNCTION__))
663 "Got assumption for the wrong function!")((I->getParent()->getParent() == Q.CxtI->getParent()
->getParent() && "Got assumption for the wrong function!"
) ? static_cast<void> (0) : __assert_fail ("I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 663, __PRETTY_FUNCTION__))
;
664 if (Q.isExcluded(I))
665 continue;
666
667 // Warning: This loop can end up being somewhat performance sensitive.
668 // We're running this loop for once for each value queried resulting in a
669 // runtime of ~O(#assumes * #values).
670
671 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&((I->getCalledFunction()->getIntrinsicID() == Intrinsic
::assume && "must be an assume intrinsic") ? static_cast
<void> (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 672, __PRETTY_FUNCTION__))
672 "must be an assume intrinsic")((I->getCalledFunction()->getIntrinsicID() == Intrinsic
::assume && "must be an assume intrinsic") ? static_cast
<void> (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 672, __PRETTY_FUNCTION__))
;
673
674 Value *Arg = I->getArgOperand(0);
675
676 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
677 assert(BitWidth == 1 && "assume operand is not i1?")((BitWidth == 1 && "assume operand is not i1?") ? static_cast
<void> (0) : __assert_fail ("BitWidth == 1 && \"assume operand is not i1?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 677, __PRETTY_FUNCTION__))
;
678 Known.setAllOnes();
679 return;
680 }
681 if (match(Arg, m_Not(m_Specific(V))) &&
682 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
683 assert(BitWidth == 1 && "assume operand is not i1?")((BitWidth == 1 && "assume operand is not i1?") ? static_cast
<void> (0) : __assert_fail ("BitWidth == 1 && \"assume operand is not i1?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 683, __PRETTY_FUNCTION__))
;
684 Known.setAllZero();
685 return;
686 }
687
688 // The remaining tests are all recursive, so bail out if we hit the limit.
689 if (Depth == MaxDepth)
690 continue;
691
692 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg);
693 if (!Cmp)
694 continue;
695
696 Value *A, *B;
697 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V)));
698
699 CmpInst::Predicate Pred;
700 uint64_t C;
701 switch (Cmp->getPredicate()) {
702 default:
703 break;
704 case ICmpInst::ICMP_EQ:
705 // assume(v = a)
706 if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) &&
707 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
708 KnownBits RHSKnown(BitWidth);
709 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
710 Known.Zero |= RHSKnown.Zero;
711 Known.One |= RHSKnown.One;
712 // assume(v & b = a)
713 } else if (match(Cmp,
714 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
715 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
716 KnownBits RHSKnown(BitWidth);
717 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
718 KnownBits MaskKnown(BitWidth);
719 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
720
721 // For those bits in the mask that are known to be one, we can propagate
722 // known bits from the RHS to V.
723 Known.Zero |= RHSKnown.Zero & MaskKnown.One;
724 Known.One |= RHSKnown.One & MaskKnown.One;
725 // assume(~(v & b) = a)
726 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
727 m_Value(A))) &&
728 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
729 KnownBits RHSKnown(BitWidth);
730 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
731 KnownBits MaskKnown(BitWidth);
732 computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
733
734 // For those bits in the mask that are known to be one, we can propagate
735 // inverted known bits from the RHS to V.
736 Known.Zero |= RHSKnown.One & MaskKnown.One;
737 Known.One |= RHSKnown.Zero & MaskKnown.One;
738 // assume(v | b = a)
739 } else if (match(Cmp,
740 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
741 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
742 KnownBits RHSKnown(BitWidth);
743 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
744 KnownBits BKnown(BitWidth);
745 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
746
747 // For those bits in B that are known to be zero, we can propagate known
748 // bits from the RHS to V.
749 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
750 Known.One |= RHSKnown.One & BKnown.Zero;
751 // assume(~(v | b) = a)
752 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
753 m_Value(A))) &&
754 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
755 KnownBits RHSKnown(BitWidth);
756 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
757 KnownBits BKnown(BitWidth);
758 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
759
760 // For those bits in B that are known to be zero, we can propagate
761 // inverted known bits from the RHS to V.
762 Known.Zero |= RHSKnown.One & BKnown.Zero;
763 Known.One |= RHSKnown.Zero & BKnown.Zero;
764 // assume(v ^ b = a)
765 } else if (match(Cmp,
766 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
767 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
768 KnownBits RHSKnown(BitWidth);
769 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
770 KnownBits BKnown(BitWidth);
771 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
772
773 // For those bits in B that are known to be zero, we can propagate known
774 // bits from the RHS to V. For those bits in B that are known to be one,
775 // we can propagate inverted known bits from the RHS to V.
776 Known.Zero |= RHSKnown.Zero & BKnown.Zero;
777 Known.One |= RHSKnown.One & BKnown.Zero;
778 Known.Zero |= RHSKnown.One & BKnown.One;
779 Known.One |= RHSKnown.Zero & BKnown.One;
780 // assume(~(v ^ b) = a)
781 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
782 m_Value(A))) &&
783 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
784 KnownBits RHSKnown(BitWidth);
785 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
786 KnownBits BKnown(BitWidth);
787 computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
788
789 // For those bits in B that are known to be zero, we can propagate
790 // inverted known bits from the RHS to V. For those bits in B that are
791 // known to be one, we can propagate known bits from the RHS to V.
792 Known.Zero |= RHSKnown.One & BKnown.Zero;
793 Known.One |= RHSKnown.Zero & BKnown.Zero;
794 Known.Zero |= RHSKnown.Zero & BKnown.One;
795 Known.One |= RHSKnown.One & BKnown.One;
796 // assume(v << c = a)
797 } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
798 m_Value(A))) &&
799 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
800 KnownBits RHSKnown(BitWidth);
801 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
802 // For those bits in RHS that are known, we can propagate them to known
803 // bits in V shifted to the right by C.
804 RHSKnown.Zero.lshrInPlace(C);
805 Known.Zero |= RHSKnown.Zero;
806 RHSKnown.One.lshrInPlace(C);
807 Known.One |= RHSKnown.One;
808 // assume(~(v << c) = a)
809 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
810 m_Value(A))) &&
811 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
812 KnownBits RHSKnown(BitWidth);
813 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
814 // For those bits in RHS that are known, we can propagate them inverted
815 // to known bits in V shifted to the right by C.
816 RHSKnown.One.lshrInPlace(C);
817 Known.Zero |= RHSKnown.One;
818 RHSKnown.Zero.lshrInPlace(C);
819 Known.One |= RHSKnown.Zero;
820 // assume(v >> c = a)
821 } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
822 m_Value(A))) &&
823 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
824 KnownBits RHSKnown(BitWidth);
825 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
826 // For those bits in RHS that are known, we can propagate them to known
827 // bits in V shifted to the right by C.
828 Known.Zero |= RHSKnown.Zero << C;
829 Known.One |= RHSKnown.One << C;
830 // assume(~(v >> c) = a)
831 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
832 m_Value(A))) &&
833 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) {
834 KnownBits RHSKnown(BitWidth);
835 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
836 // For those bits in RHS that are known, we can propagate them inverted
837 // to known bits in V shifted to the right by C.
838 Known.Zero |= RHSKnown.One << C;
839 Known.One |= RHSKnown.Zero << C;
840 }
841 break;
842 case ICmpInst::ICMP_SGE:
843 // assume(v >=_s c) where c is non-negative
844 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
845 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
846 KnownBits RHSKnown(BitWidth);
847 computeKnownBits(A, RHSKnown, Depth + 1, Query(Q, I));
848
849 if (RHSKnown.isNonNegative()) {
850 // We know that the sign bit is zero.
851 Known.makeNonNegative();
852 }
853 }
854 break;
855 case ICmpInst::ICMP_SGT:
856 // assume(v >_s c) where c is at least -1.
857 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
858 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
859 KnownBits RHSKnown(BitWidth);
860 computeKnownBits(A, RHSKnown, Depth + 1, Query(Q, I));
861
862 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
863 // We know that the sign bit is zero.
864 Known.makeNonNegative();
865 }
866 }
867 break;
868 case ICmpInst::ICMP_SLE:
869 // assume(v <=_s c) where c is negative
870 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
871 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
872 KnownBits RHSKnown(BitWidth);
873 computeKnownBits(A, RHSKnown, Depth + 1, Query(Q, I));
874
875 if (RHSKnown.isNegative()) {
876 // We know that the sign bit is one.
877 Known.makeNegative();
878 }
879 }
880 break;
881 case ICmpInst::ICMP_SLT:
882 // assume(v <_s c) where c is non-positive
883 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
884 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
885 KnownBits RHSKnown(BitWidth);
886 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
887
888 if (RHSKnown.isZero() || RHSKnown.isNegative()) {
889 // We know that the sign bit is one.
890 Known.makeNegative();
891 }
892 }
893 break;
894 case ICmpInst::ICMP_ULE:
895 // assume(v <=_u c)
896 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
897 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
898 KnownBits RHSKnown(BitWidth);
899 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
900
901 // Whatever high bits in c are zero are known to be zero.
902 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
903 }
904 break;
905 case ICmpInst::ICMP_ULT:
906 // assume(v <_u c)
907 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) &&
908 isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
909 KnownBits RHSKnown(BitWidth);
910 computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
911
912 // If the RHS is known zero, then this assumption must be wrong (nothing
913 // is unsigned less than zero). Signal a conflict and get out of here.
914 if (RHSKnown.isZero()) {
915 Known.Zero.setAllBits();
916 Known.One.setAllBits();
917 break;
918 }
919
920 // Whatever high bits in c are zero are known to be zero (if c is a power
921 // of 2, then one more).
922 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
923 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
924 else
925 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
926 }
927 break;
928 }
929 }
930
931 // If assumptions conflict with each other or previous known bits, then we
932 // have a logical fallacy. It's possible that the assumption is not reachable,
933 // so this isn't a real bug. On the other hand, the program may have undefined
934 // behavior, or we might have a bug in the compiler. We can't assert/crash, so
935 // clear out the known bits, try to warn the user, and hope for the best.
936 if (Known.Zero.intersects(Known.One)) {
937 Known.resetAll();
938
939 if (Q.ORE)
940 Q.ORE->emit([&]() {
941 auto *CxtI = const_cast<Instruction *>(Q.CxtI);
942 return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
943 CxtI)
944 << "Detected conflicting code assumptions. Program may "
945 "have undefined behavior, or compiler may have "
946 "internal error.";
947 });
948 }
949}
950
951/// Compute known bits from a shift operator, including those with a
952/// non-constant shift amount. Known is the output of this function. Known2 is a
953/// pre-allocated temporary with the same bit width as Known. KZF and KOF are
954/// operator-specific functions that, given the known-zero or known-one bits
955/// respectively, and a shift amount, compute the implied known-zero or
956/// known-one bits of the shift operator's result respectively for that shift
957/// amount. The results from calling KZF and KOF are conservatively combined for
958/// all permitted shift amounts.
959static void computeKnownBitsFromShiftOperator(
960 const Operator *I, KnownBits &Known, KnownBits &Known2,
961 unsigned Depth, const Query &Q,
962 function_ref<APInt(const APInt &, unsigned)> KZF,
963 function_ref<APInt(const APInt &, unsigned)> KOF) {
964 unsigned BitWidth = Known.getBitWidth();
965
966 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
967 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
968
969 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
970 Known.Zero = KZF(Known.Zero, ShiftAmt);
971 Known.One = KOF(Known.One, ShiftAmt);
972 // If the known bits conflict, this must be an overflowing left shift, so
973 // the shift result is poison. We can return anything we want. Choose 0 for
974 // the best folding opportunity.
975 if (Known.hasConflict())
976 Known.setAllZero();
977
978 return;
979 }
980
981 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
982
983 // If the shift amount could be greater than or equal to the bit-width of the
984 // LHS, the value could be poison, but bail out because the check below is
985 // expensive. TODO: Should we just carry on?
986 if (Known.getMaxValue().uge(BitWidth)) {
987 Known.resetAll();
988 return;
989 }
990
991 // Note: We cannot use Known.Zero.getLimitedValue() here, because if
992 // BitWidth > 64 and any upper bits are known, we'll end up returning the
993 // limit value (which implies all bits are known).
994 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
995 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
996
997 // It would be more-clearly correct to use the two temporaries for this
998 // calculation. Reusing the APInts here to prevent unnecessary allocations.
999 Known.resetAll();
1000
1001 // If we know the shifter operand is nonzero, we can sometimes infer more
1002 // known bits. However this is expensive to compute, so be lazy about it and
1003 // only compute it when absolutely necessary.
1004 Optional<bool> ShifterOperandIsNonZero;
1005
1006 // Early exit if we can't constrain any well-defined shift amount.
1007 if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
1008 !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
1009 ShifterOperandIsNonZero = isKnownNonZero(I->getOperand(1), Depth + 1, Q);
1010 if (!*ShifterOperandIsNonZero)
1011 return;
1012 }
1013
1014 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1015
1016 Known.Zero.setAllBits();
1017 Known.One.setAllBits();
1018 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
1019 // Combine the shifted known input bits only for those shift amounts
1020 // compatible with its known constraints.
1021 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
1022 continue;
1023 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
1024 continue;
1025 // If we know the shifter is nonzero, we may be able to infer more known
1026 // bits. This check is sunk down as far as possible to avoid the expensive
1027 // call to isKnownNonZero if the cheaper checks above fail.
1028 if (ShiftAmt == 0) {
1029 if (!ShifterOperandIsNonZero.hasValue())
1030 ShifterOperandIsNonZero =
1031 isKnownNonZero(I->getOperand(1), Depth + 1, Q);
1032 if (*ShifterOperandIsNonZero)
1033 continue;
1034 }
1035
1036 Known.Zero &= KZF(Known2.Zero, ShiftAmt);
1037 Known.One &= KOF(Known2.One, ShiftAmt);
1038 }
1039
1040 // If the known bits conflict, the result is poison. Return a 0 and hope the
1041 // caller can further optimize that.
1042 if (Known.hasConflict())
1043 Known.setAllZero();
1044}
1045
1046static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
1047 unsigned Depth, const Query &Q) {
1048 unsigned BitWidth = Known.getBitWidth();
1049
1050 KnownBits Known2(Known);
1051 switch (I->getOpcode()) {
1052 default: break;
1053 case Instruction::Load:
1054 if (MDNode *MD =
1055 Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range))
1056 computeKnownBitsFromRangeMetadata(*MD, Known);
1057 break;
1058 case Instruction::And: {
1059 // If either the LHS or the RHS are Zero, the result is zero.
1060 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1061 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1062
1063 // Output known-1 bits are only known if set in both the LHS & RHS.
1064 Known.One &= Known2.One;
1065 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1066 Known.Zero |= Known2.Zero;
1067
1068 // and(x, add (x, -1)) is a common idiom that always clears the low bit;
1069 // here we handle the more general case of adding any odd number by
1070 // matching the form add(x, add(x, y)) where y is odd.
1071 // TODO: This could be generalized to clearing any bit set in y where the
1072 // following bit is known to be unset in y.
1073 Value *X = nullptr, *Y = nullptr;
1074 if (!Known.Zero[0] && !Known.One[0] &&
1075 match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) {
1076 Known2.resetAll();
1077 computeKnownBits(Y, Known2, Depth + 1, Q);
1078 if (Known2.countMinTrailingOnes() > 0)
1079 Known.Zero.setBit(0);
1080 }
1081 break;
1082 }
1083 case Instruction::Or:
1084 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1085 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1086
1087 // Output known-0 bits are only known if clear in both the LHS & RHS.
1088 Known.Zero &= Known2.Zero;
1089 // Output known-1 are known to be set if set in either the LHS | RHS.
1090 Known.One |= Known2.One;
1091 break;
1092 case Instruction::Xor: {
1093 computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1094 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1095
1096 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1097 APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
1098 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1099 Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
1100 Known.Zero = std::move(KnownZeroOut);
1101 break;
1102 }
1103 case Instruction::Mul: {
1104 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1105 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
1106 Known2, Depth, Q);
1107 break;
1108 }
1109 case Instruction::UDiv: {
1110 // For the purposes of computing leading zeros we can conservatively
1111 // treat a udiv as a logical right shift by the power of 2 known to
1112 // be less than the denominator.
1113 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1114 unsigned LeadZ = Known2.countMinLeadingZeros();
1115
1116 Known2.resetAll();
1117 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1118 unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
1119 if (RHSMaxLeadingZeros != BitWidth)
1120 LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
1121
1122 Known.Zero.setHighBits(LeadZ);
1123 break;
1124 }
1125 case Instruction::Select: {
1126 const Value *LHS = nullptr, *RHS = nullptr;
1127 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
1128 if (SelectPatternResult::isMinOrMax(SPF)) {
1129 computeKnownBits(RHS, Known, Depth + 1, Q);
1130 computeKnownBits(LHS, Known2, Depth + 1, Q);
1131 } else {
1132 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1133 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1134 }
1135
1136 unsigned MaxHighOnes = 0;
1137 unsigned MaxHighZeros = 0;
1138 if (SPF == SPF_SMAX) {
1139 // If both sides are negative, the result is negative.
1140 if (Known.isNegative() && Known2.isNegative())
1141 // We can derive a lower bound on the result by taking the max of the
1142 // leading one bits.
1143 MaxHighOnes =
1144 std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1145 // If either side is non-negative, the result is non-negative.
1146 else if (Known.isNonNegative() || Known2.isNonNegative())
1147 MaxHighZeros = 1;
1148 } else if (SPF == SPF_SMIN) {
1149 // If both sides are non-negative, the result is non-negative.
1150 if (Known.isNonNegative() && Known2.isNonNegative())
1151 // We can derive an upper bound on the result by taking the max of the
1152 // leading zero bits.
1153 MaxHighZeros = std::max(Known.countMinLeadingZeros(),
1154 Known2.countMinLeadingZeros());
1155 // If either side is negative, the result is negative.
1156 else if (Known.isNegative() || Known2.isNegative())
1157 MaxHighOnes = 1;
1158 } else if (SPF == SPF_UMAX) {
1159 // We can derive a lower bound on the result by taking the max of the
1160 // leading one bits.
1161 MaxHighOnes =
1162 std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1163 } else if (SPF == SPF_UMIN) {
1164 // We can derive an upper bound on the result by taking the max of the
1165 // leading zero bits.
1166 MaxHighZeros =
1167 std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1168 } else if (SPF == SPF_ABS) {
1169 // RHS from matchSelectPattern returns the negation part of abs pattern.
1170 // If the negate has an NSW flag we can assume the sign bit of the result
1171 // will be 0 because that makes abs(INT_MIN) undefined.
1172 if (match(RHS, m_Neg(m_Specific(LHS))) &&
1173 Q.IIQ.hasNoSignedWrap(cast<Instruction>(RHS)))
1174 MaxHighZeros = 1;
1175 }
1176
1177 // Only known if known in both the LHS and RHS.
1178 Known.One &= Known2.One;
1179 Known.Zero &= Known2.Zero;
1180 if (MaxHighOnes > 0)
1181 Known.One.setHighBits(MaxHighOnes);
1182 if (MaxHighZeros > 0)
1183 Known.Zero.setHighBits(MaxHighZeros);
1184 break;
1185 }
1186 case Instruction::FPTrunc:
1187 case Instruction::FPExt:
1188 case Instruction::FPToUI:
1189 case Instruction::FPToSI:
1190 case Instruction::SIToFP:
1191 case Instruction::UIToFP:
1192 break; // Can't work with floating point.
1193 case Instruction::PtrToInt:
1194 case Instruction::IntToPtr:
1195 // Fall through and handle them the same as zext/trunc.
1196 LLVM_FALLTHROUGH[[gnu::fallthrough]];
1197 case Instruction::ZExt:
1198 case Instruction::Trunc: {
1199 Type *SrcTy = I->getOperand(0)->getType();
1200
1201 unsigned SrcBitWidth;
1202 // Note that we handle pointer operands here because of inttoptr/ptrtoint
1203 // which fall through here.
1204 Type *ScalarTy = SrcTy->getScalarType();
1205 SrcBitWidth = ScalarTy->isPointerTy() ?
1206 Q.DL.getPointerTypeSizeInBits(ScalarTy) :
1207 Q.DL.getTypeSizeInBits(ScalarTy);
1208
1209 assert(SrcBitWidth && "SrcBitWidth can't be zero")((SrcBitWidth && "SrcBitWidth can't be zero") ? static_cast
<void> (0) : __assert_fail ("SrcBitWidth && \"SrcBitWidth can't be zero\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1209, __PRETTY_FUNCTION__))
;
1210 Known = Known.zextOrTrunc(SrcBitWidth, false);
1211 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1212 Known = Known.zextOrTrunc(BitWidth, true /* ExtendedBitsAreKnownZero */);
1213 break;
1214 }
1215 case Instruction::BitCast: {
1216 Type *SrcTy = I->getOperand(0)->getType();
1217 if (SrcTy->isIntOrPtrTy() &&
1218 // TODO: For now, not handling conversions like:
1219 // (bitcast i64 %x to <2 x i32>)
1220 !I->getType()->isVectorTy()) {
1221 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1222 break;
1223 }
1224 break;
1225 }
1226 case Instruction::SExt: {
1227 // Compute the bits in the result that are not present in the input.
1228 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1229
1230 Known = Known.trunc(SrcBitWidth);
1231 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1232 // If the sign bit of the input is known set or clear, then we know the
1233 // top bits of the result.
1234 Known = Known.sext(BitWidth);
1235 break;
1236 }
1237 case Instruction::Shl: {
1238 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
1239 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1240 auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1241 APInt KZResult = KnownZero << ShiftAmt;
1242 KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1243 // If this shift has "nsw" keyword, then the result is either a poison
1244 // value or has the same sign bit as the first operand.
1245 if (NSW && KnownZero.isSignBitSet())
1246 KZResult.setSignBit();
1247 return KZResult;
1248 };
1249
1250 auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1251 APInt KOResult = KnownOne << ShiftAmt;
1252 if (NSW && KnownOne.isSignBitSet())
1253 KOResult.setSignBit();
1254 return KOResult;
1255 };
1256
1257 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1258 break;
1259 }
1260 case Instruction::LShr: {
1261 // (lshr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1262 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1263 APInt KZResult = KnownZero.lshr(ShiftAmt);
1264 // High bits known zero.
1265 KZResult.setHighBits(ShiftAmt);
1266 return KZResult;
1267 };
1268
1269 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1270 return KnownOne.lshr(ShiftAmt);
1271 };
1272
1273 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1274 break;
1275 }
1276 case Instruction::AShr: {
1277 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
1278 auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1279 return KnownZero.ashr(ShiftAmt);
1280 };
1281
1282 auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1283 return KnownOne.ashr(ShiftAmt);
1284 };
1285
1286 computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1287 break;
1288 }
1289 case Instruction::Sub: {
1290 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1291 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1292 Known, Known2, Depth, Q);
1293 break;
1294 }
1295 case Instruction::Add: {
1296 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I));
1297 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1298 Known, Known2, Depth, Q);
1299 break;
1300 }
1301 case Instruction::SRem:
1302 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1303 APInt RA = Rem->getValue().abs();
1304 if (RA.isPowerOf2()) {
1305 APInt LowBits = RA - 1;
1306 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1307
1308 // The low bits of the first operand are unchanged by the srem.
1309 Known.Zero = Known2.Zero & LowBits;
1310 Known.One = Known2.One & LowBits;
1311
1312 // If the first operand is non-negative or has all low bits zero, then
1313 // the upper bits are all zero.
1314 if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1315 Known.Zero |= ~LowBits;
1316
1317 // If the first operand is negative and not all low bits are zero, then
1318 // the upper bits are all one.
1319 if (Known2.isNegative() && LowBits.intersects(Known2.One))
1320 Known.One |= ~LowBits;
1321
1322 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?")(((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?"
) ? static_cast<void> (0) : __assert_fail ("(Known.Zero & Known.One) == 0 && \"Bits known to be one AND zero?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1322, __PRETTY_FUNCTION__))
;
1323 break;
1324 }
1325 }
1326
1327 // The sign bit is the LHS's sign bit, except when the result of the
1328 // remainder is zero.
1329 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1330 // If it's known zero, our sign bit is also zero.
1331 if (Known2.isNonNegative())
1332 Known.makeNonNegative();
1333
1334 break;
1335 case Instruction::URem: {
1336 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1337 const APInt &RA = Rem->getValue();
1338 if (RA.isPowerOf2()) {
1339 APInt LowBits = (RA - 1);
1340 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1341 Known.Zero |= ~LowBits;
1342 Known.One &= LowBits;
1343 break;
1344 }
1345 }
1346
1347 // Since the result is less than or equal to either operand, any leading
1348 // zero bits in either operand must also exist in the result.
1349 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1350 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1351
1352 unsigned Leaders =
1353 std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1354 Known.resetAll();
1355 Known.Zero.setHighBits(Leaders);
1356 break;
1357 }
1358
1359 case Instruction::Alloca: {
1360 const AllocaInst *AI = cast<AllocaInst>(I);
1361 unsigned Align = AI->getAlignment();
1362 if (Align == 0)
1363 Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1364
1365 if (Align > 0)
1366 Known.Zero.setLowBits(countTrailingZeros(Align));
1367 break;
1368 }
1369 case Instruction::GetElementPtr: {
1370 // Analyze all of the subscripts of this getelementptr instruction
1371 // to determine if we can prove known low zero bits.
1372 KnownBits LocalKnown(BitWidth);
1373 computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1374 unsigned TrailZ = LocalKnown.countMinTrailingZeros();
1375
1376 gep_type_iterator GTI = gep_type_begin(I);
1377 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1378 Value *Index = I->getOperand(i);
1379 if (StructType *STy = GTI.getStructTypeOrNull()) {
1380 // Handle struct member offset arithmetic.
1381
1382 // Handle case when index is vector zeroinitializer
1383 Constant *CIndex = cast<Constant>(Index);
1384 if (CIndex->isZeroValue())
1385 continue;
1386
1387 if (CIndex->getType()->isVectorTy())
1388 Index = CIndex->getSplatValue();
1389
1390 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1391 const StructLayout *SL = Q.DL.getStructLayout(STy);
1392 uint64_t Offset = SL->getElementOffset(Idx);
1393 TrailZ = std::min<unsigned>(TrailZ,
1394 countTrailingZeros(Offset));
1395 } else {
1396 // Handle array index arithmetic.
1397 Type *IndexedTy = GTI.getIndexedType();
1398 if (!IndexedTy->isSized()) {
1399 TrailZ = 0;
1400 break;
1401 }
1402 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1403 uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1404 LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1405 computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1406 TrailZ = std::min(TrailZ,
1407 unsigned(countTrailingZeros(TypeSize) +
1408 LocalKnown.countMinTrailingZeros()));
1409 }
1410 }
1411
1412 Known.Zero.setLowBits(TrailZ);
1413 break;
1414 }
1415 case Instruction::PHI: {
1416 const PHINode *P = cast<PHINode>(I);
1417 // Handle the case of a simple two-predecessor recurrence PHI.
1418 // There's a lot more that could theoretically be done here, but
1419 // this is sufficient to catch some interesting cases.
1420 if (P->getNumIncomingValues() == 2) {
1421 for (unsigned i = 0; i != 2; ++i) {
1422 Value *L = P->getIncomingValue(i);
1423 Value *R = P->getIncomingValue(!i);
1424 Instruction *RInst = P->getIncomingBlock(!i)->getTerminator();
1425 Instruction *LInst = P->getIncomingBlock(i)->getTerminator();
1426 Operator *LU = dyn_cast<Operator>(L);
1427 if (!LU)
1428 continue;
1429 unsigned Opcode = LU->getOpcode();
1430 // Check for operations that have the property that if
1431 // both their operands have low zero bits, the result
1432 // will have low zero bits.
1433 if (Opcode == Instruction::Add ||
1434 Opcode == Instruction::Sub ||
1435 Opcode == Instruction::And ||
1436 Opcode == Instruction::Or ||
1437 Opcode == Instruction::Mul) {
1438 Value *LL = LU->getOperand(0);
1439 Value *LR = LU->getOperand(1);
1440 // Find a recurrence.
1441 if (LL == I)
1442 L = LR;
1443 else if (LR == I)
1444 L = LL;
1445 else
1446 continue; // Check for recurrence with L and R flipped.
1447
1448 // Change the context instruction to the "edge" that flows into the
1449 // phi. This is important because that is where the value is actually
1450 // "evaluated" even though it is used later somewhere else. (see also
1451 // D69571).
1452 Query RecQ = Q;
1453
1454 // Ok, we have a PHI of the form L op= R. Check for low
1455 // zero bits.
1456 RecQ.CxtI = RInst;
1457 computeKnownBits(R, Known2, Depth + 1, RecQ);
1458
1459 // We need to take the minimum number of known bits
1460 KnownBits Known3(Known);
1461 RecQ.CxtI = LInst;
1462 computeKnownBits(L, Known3, Depth + 1, RecQ);
1463
1464 Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1465 Known3.countMinTrailingZeros()));
1466
1467 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1468 if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) {
1469 // If initial value of recurrence is nonnegative, and we are adding
1470 // a nonnegative number with nsw, the result can only be nonnegative
1471 // or poison value regardless of the number of times we execute the
1472 // add in phi recurrence. If initial value is negative and we are
1473 // adding a negative number with nsw, the result can only be
1474 // negative or poison value. Similar arguments apply to sub and mul.
1475 //
1476 // (add non-negative, non-negative) --> non-negative
1477 // (add negative, negative) --> negative
1478 if (Opcode == Instruction::Add) {
1479 if (Known2.isNonNegative() && Known3.isNonNegative())
1480 Known.makeNonNegative();
1481 else if (Known2.isNegative() && Known3.isNegative())
1482 Known.makeNegative();
1483 }
1484
1485 // (sub nsw non-negative, negative) --> non-negative
1486 // (sub nsw negative, non-negative) --> negative
1487 else if (Opcode == Instruction::Sub && LL == I) {
1488 if (Known2.isNonNegative() && Known3.isNegative())
1489 Known.makeNonNegative();
1490 else if (Known2.isNegative() && Known3.isNonNegative())
1491 Known.makeNegative();
1492 }
1493
1494 // (mul nsw non-negative, non-negative) --> non-negative
1495 else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1496 Known3.isNonNegative())
1497 Known.makeNonNegative();
1498 }
1499
1500 break;
1501 }
1502 }
1503 }
1504
1505 // Unreachable blocks may have zero-operand PHI nodes.
1506 if (P->getNumIncomingValues() == 0)
1507 break;
1508
1509 // Otherwise take the unions of the known bit sets of the operands,
1510 // taking conservative care to avoid excessive recursion.
1511 if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1512 // Skip if every incoming value references to ourself.
1513 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1514 break;
1515
1516 Known.Zero.setAllBits();
1517 Known.One.setAllBits();
1518 for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) {
1519 Value *IncValue = P->getIncomingValue(u);
1520 // Skip direct self references.
1521 if (IncValue == P) continue;
1522
1523 // Change the context instruction to the "edge" that flows into the
1524 // phi. This is important because that is where the value is actually
1525 // "evaluated" even though it is used later somewhere else. (see also
1526 // D69571).
1527 Query RecQ = Q;
1528 RecQ.CxtI = P->getIncomingBlock(u)->getTerminator();
1529
1530 Known2 = KnownBits(BitWidth);
1531 // Recurse, but cap the recursion to one level, because we don't
1532 // want to waste time spinning around in loops.
1533 computeKnownBits(IncValue, Known2, MaxDepth - 1, RecQ);
1534 Known.Zero &= Known2.Zero;
1535 Known.One &= Known2.One;
1536 // If all bits have been ruled out, there's no need to check
1537 // more operands.
1538 if (!Known.Zero && !Known.One)
1539 break;
1540 }
1541 }
1542 break;
1543 }
1544 case Instruction::Call:
1545 case Instruction::Invoke:
1546 // If range metadata is attached to this call, set known bits from that,
1547 // and then intersect with known bits based on other properties of the
1548 // function.
1549 if (MDNode *MD =
1550 Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range))
1551 computeKnownBitsFromRangeMetadata(*MD, Known);
1552 if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1553 computeKnownBits(RV, Known2, Depth + 1, Q);
1554 Known.Zero |= Known2.Zero;
1555 Known.One |= Known2.One;
1556 }
1557 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1558 switch (II->getIntrinsicID()) {
1559 default: break;
1560 case Intrinsic::bitreverse:
1561 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1562 Known.Zero |= Known2.Zero.reverseBits();
1563 Known.One |= Known2.One.reverseBits();
1564 break;
1565 case Intrinsic::bswap:
1566 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1567 Known.Zero |= Known2.Zero.byteSwap();
1568 Known.One |= Known2.One.byteSwap();
1569 break;
1570 case Intrinsic::ctlz: {
1571 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1572 // If we have a known 1, its position is our upper bound.
1573 unsigned PossibleLZ = Known2.One.countLeadingZeros();
1574 // If this call is undefined for 0, the result will be less than 2^n.
1575 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1576 PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1577 unsigned LowBits = Log2_32(PossibleLZ)+1;
1578 Known.Zero.setBitsFrom(LowBits);
1579 break;
1580 }
1581 case Intrinsic::cttz: {
1582 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1583 // If we have a known 1, its position is our upper bound.
1584 unsigned PossibleTZ = Known2.One.countTrailingZeros();
1585 // If this call is undefined for 0, the result will be less than 2^n.
1586 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1587 PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1588 unsigned LowBits = Log2_32(PossibleTZ)+1;
1589 Known.Zero.setBitsFrom(LowBits);
1590 break;
1591 }
1592 case Intrinsic::ctpop: {
1593 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1594 // We can bound the space the count needs. Also, bits known to be zero
1595 // can't contribute to the population.
1596 unsigned BitsPossiblySet = Known2.countMaxPopulation();
1597 unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1598 Known.Zero.setBitsFrom(LowBits);
1599 // TODO: we could bound KnownOne using the lower bound on the number
1600 // of bits which might be set provided by popcnt KnownOne2.
1601 break;
1602 }
1603 case Intrinsic::fshr:
1604 case Intrinsic::fshl: {
1605 const APInt *SA;
1606 if (!match(I->getOperand(2), m_APInt(SA)))
1607 break;
1608
1609 // Normalize to funnel shift left.
1610 uint64_t ShiftAmt = SA->urem(BitWidth);
1611 if (II->getIntrinsicID() == Intrinsic::fshr)
1612 ShiftAmt = BitWidth - ShiftAmt;
1613
1614 KnownBits Known3(Known);
1615 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1616 computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q);
1617
1618 Known.Zero =
1619 Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt);
1620 Known.One =
1621 Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt);
1622 break;
1623 }
1624 case Intrinsic::uadd_sat:
1625 case Intrinsic::usub_sat: {
1626 bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat;
1627 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1628 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1629
1630 // Add: Leading ones of either operand are preserved.
1631 // Sub: Leading zeros of LHS and leading ones of RHS are preserved
1632 // as leading zeros in the result.
1633 unsigned LeadingKnown;
1634 if (IsAdd)
1635 LeadingKnown = std::max(Known.countMinLeadingOnes(),
1636 Known2.countMinLeadingOnes());
1637 else
1638 LeadingKnown = std::max(Known.countMinLeadingZeros(),
1639 Known2.countMinLeadingOnes());
1640
1641 Known = KnownBits::computeForAddSub(
1642 IsAdd, /* NSW */ false, Known, Known2);
1643
1644 // We select between the operation result and all-ones/zero
1645 // respectively, so we can preserve known ones/zeros.
1646 if (IsAdd) {
1647 Known.One.setHighBits(LeadingKnown);
1648 Known.Zero.clearAllBits();
1649 } else {
1650 Known.Zero.setHighBits(LeadingKnown);
1651 Known.One.clearAllBits();
1652 }
1653 break;
1654 }
1655 case Intrinsic::x86_sse42_crc32_64_64:
1656 Known.Zero.setBitsFrom(32);
1657 break;
1658 }
1659 }
1660 break;
1661 case Instruction::ExtractElement:
1662 // Look through extract element. At the moment we keep this simple and skip
1663 // tracking the specific element. But at least we might find information
1664 // valid for all elements of the vector (for example if vector is sign
1665 // extended, shifted, etc).
1666 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1667 break;
1668 case Instruction::ExtractValue:
1669 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1670 const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1671 if (EVI->getNumIndices() != 1) break;
1672 if (EVI->getIndices()[0] == 0) {
1673 switch (II->getIntrinsicID()) {
1674 default: break;
1675 case Intrinsic::uadd_with_overflow:
1676 case Intrinsic::sadd_with_overflow:
1677 computeKnownBitsAddSub(true, II->getArgOperand(0),
1678 II->getArgOperand(1), false, Known, Known2,
1679 Depth, Q);
1680 break;
1681 case Intrinsic::usub_with_overflow:
1682 case Intrinsic::ssub_with_overflow:
1683 computeKnownBitsAddSub(false, II->getArgOperand(0),
1684 II->getArgOperand(1), false, Known, Known2,
1685 Depth, Q);
1686 break;
1687 case Intrinsic::umul_with_overflow:
1688 case Intrinsic::smul_with_overflow:
1689 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1690 Known, Known2, Depth, Q);
1691 break;
1692 }
1693 }
1694 }
1695 }
1696}
1697
1698/// Determine which bits of V are known to be either zero or one and return
1699/// them.
1700KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1701 KnownBits Known(getBitWidth(V->getType(), Q.DL));
1702 computeKnownBits(V, Known, Depth, Q);
1703 return Known;
1704}
1705
1706/// Determine which bits of V are known to be either zero or one and return
1707/// them in the Known bit set.
1708///
1709/// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
1710/// we cannot optimize based on the assumption that it is zero without changing
1711/// it to be an explicit zero. If we don't change it to zero, other code could
1712/// optimized based on the contradictory assumption that it is non-zero.
1713/// Because instcombine aggressively folds operations with undef args anyway,
1714/// this won't lose us code quality.
1715///
1716/// This function is defined on values with integer type, values with pointer
1717/// type, and vectors of integers. In the case
1718/// where V is a vector, known zero, and known one values are the
1719/// same width as the vector element, and the bit is set only if it is true
1720/// for all of the elements in the vector.
1721void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1722 const Query &Q) {
1723 assert(V && "No Value?")((V && "No Value?") ? static_cast<void> (0) : __assert_fail
("V && \"No Value?\"", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1723, __PRETTY_FUNCTION__))
;
1724 assert(Depth <= MaxDepth && "Limit Search Depth")((Depth <= MaxDepth && "Limit Search Depth") ? static_cast
<void> (0) : __assert_fail ("Depth <= MaxDepth && \"Limit Search Depth\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1724, __PRETTY_FUNCTION__))
;
1725 unsigned BitWidth = Known.getBitWidth();
1726
1727 assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||(((V->getType()->isIntOrIntVectorTy(BitWidth) || V->
getType()->isPtrOrPtrVectorTy()) && "Not integer or pointer type!"
) ? static_cast<void> (0) : __assert_fail ("(V->getType()->isIntOrIntVectorTy(BitWidth) || V->getType()->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1729, __PRETTY_FUNCTION__))
1728 V->getType()->isPtrOrPtrVectorTy()) &&(((V->getType()->isIntOrIntVectorTy(BitWidth) || V->
getType()->isPtrOrPtrVectorTy()) && "Not integer or pointer type!"
) ? static_cast<void> (0) : __assert_fail ("(V->getType()->isIntOrIntVectorTy(BitWidth) || V->getType()->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1729, __PRETTY_FUNCTION__))
1729 "Not integer or pointer type!")(((V->getType()->isIntOrIntVectorTy(BitWidth) || V->
getType()->isPtrOrPtrVectorTy()) && "Not integer or pointer type!"
) ? static_cast<void> (0) : __assert_fail ("(V->getType()->isIntOrIntVectorTy(BitWidth) || V->getType()->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1729, __PRETTY_FUNCTION__))
;
1730
1731 Type *ScalarTy = V->getType()->getScalarType();
1732 unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
1733 Q.DL.getPointerTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
1734 assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth")((ExpectedWidth == BitWidth && "V and Known should have same BitWidth"
) ? static_cast<void> (0) : __assert_fail ("ExpectedWidth == BitWidth && \"V and Known should have same BitWidth\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1734, __PRETTY_FUNCTION__))
;
1735 (void)BitWidth;
1736 (void)ExpectedWidth;
1737
1738 const APInt *C;
1739 if (match(V, m_APInt(C))) {
1740 // We know all of the bits for a scalar constant or a splat vector constant!
1741 Known.One = *C;
1742 Known.Zero = ~Known.One;
1743 return;
1744 }
1745 // Null and aggregate-zero are all-zeros.
1746 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1747 Known.setAllZero();
1748 return;
1749 }
1750 // Handle a constant vector by taking the intersection of the known bits of
1751 // each element.
1752 if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1753 // We know that CDS must be a vector of integers. Take the intersection of
1754 // each element.
1755 Known.Zero.setAllBits(); Known.One.setAllBits();
1756 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1757 APInt Elt = CDS->getElementAsAPInt(i);
1758 Known.Zero &= ~Elt;
1759 Known.One &= Elt;
1760 }
1761 return;
1762 }
1763
1764 if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1765 // We know that CV must be a vector of integers. Take the intersection of
1766 // each element.
1767 Known.Zero.setAllBits(); Known.One.setAllBits();
1768 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1769 Constant *Element = CV->getAggregateElement(i);
1770 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1771 if (!ElementCI) {
1772 Known.resetAll();
1773 return;
1774 }
1775 const APInt &Elt = ElementCI->getValue();
1776 Known.Zero &= ~Elt;
1777 Known.One &= Elt;
1778 }
1779 return;
1780 }
1781
1782 // Start out not knowing anything.
1783 Known.resetAll();
1784
1785 // We can't imply anything about undefs.
1786 if (isa<UndefValue>(V))
1787 return;
1788
1789 // There's no point in looking through other users of ConstantData for
1790 // assumptions. Confirm that we've handled them all.
1791 assert(!isa<ConstantData>(V) && "Unhandled constant data!")((!isa<ConstantData>(V) && "Unhandled constant data!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantData>(V) && \"Unhandled constant data!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1791, __PRETTY_FUNCTION__))
;
1792
1793 // Limit search depth.
1794 // All recursive calls that increase depth must come after this.
1795 if (Depth == MaxDepth)
1796 return;
1797
1798 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1799 // the bits of its aliasee.
1800 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1801 if (!GA->isInterposable())
1802 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1803 return;
1804 }
1805
1806 if (const Operator *I = dyn_cast<Operator>(V))
1807 computeKnownBitsFromOperator(I, Known, Depth, Q);
1808
1809 // Aligned pointers have trailing zeros - refine Known.Zero set
1810 if (V->getType()->isPointerTy()) {
1811 const MaybeAlign Align = V->getPointerAlignment(Q.DL);
1812 if (Align)
1813 Known.Zero.setLowBits(countTrailingZeros(Align->value()));
1814 }
1815
1816 // computeKnownBitsFromAssume strictly refines Known.
1817 // Therefore, we run them after computeKnownBitsFromOperator.
1818
1819 // Check whether a nearby assume intrinsic can determine some known bits.
1820 computeKnownBitsFromAssume(V, Known, Depth, Q);
1821
1822 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?")(((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?"
) ? static_cast<void> (0) : __assert_fail ("(Known.Zero & Known.One) == 0 && \"Bits known to be one AND zero?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1822, __PRETTY_FUNCTION__))
;
1823}
1824
1825/// Return true if the given value is known to have exactly one
1826/// bit set when defined. For vectors return true if every element is known to
1827/// be a power of two when defined. Supports values with integer or pointer
1828/// types and vectors of integers.
1829bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1830 const Query &Q) {
1831 assert(Depth <= MaxDepth && "Limit Search Depth")((Depth <= MaxDepth && "Limit Search Depth") ? static_cast
<void> (0) : __assert_fail ("Depth <= MaxDepth && \"Limit Search Depth\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1831, __PRETTY_FUNCTION__))
;
1832
1833 // Attempt to match against constants.
1834 if (OrZero && match(V, m_Power2OrZero()))
1835 return true;
1836 if (match(V, m_Power2()))
1837 return true;
1838
1839 // 1 << X is clearly a power of two if the one is not shifted off the end. If
1840 // it is shifted off the end then the result is undefined.
1841 if (match(V, m_Shl(m_One(), m_Value())))
1842 return true;
1843
1844 // (signmask) >>l X is clearly a power of two if the one is not shifted off
1845 // the bottom. If it is shifted off the bottom then the result is undefined.
1846 if (match(V, m_LShr(m_SignMask(), m_Value())))
1847 return true;
1848
1849 // The remaining tests are all recursive, so bail out if we hit the limit.
1850 if (Depth++ == MaxDepth)
1851 return false;
1852
1853 Value *X = nullptr, *Y = nullptr;
1854 // A shift left or a logical shift right of a power of two is a power of two
1855 // or zero.
1856 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1857 match(V, m_LShr(m_Value(X), m_Value()))))
1858 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1859
1860 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1861 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1862
1863 if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1864 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1865 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1866
1867 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1868 // A power of two and'd with anything is a power of two or zero.
1869 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1870 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1871 return true;
1872 // X & (-X) is always a power of two or zero.
1873 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1874 return true;
1875 return false;
1876 }
1877
1878 // Adding a power-of-two or zero to the same power-of-two or zero yields
1879 // either the original power-of-two, a larger power-of-two or zero.
1880 if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1881 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1882 if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) ||
1883 Q.IIQ.hasNoSignedWrap(VOBO)) {
1884 if (match(X, m_And(m_Specific(Y), m_Value())) ||
1885 match(X, m_And(m_Value(), m_Specific(Y))))
1886 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1887 return true;
1888 if (match(Y, m_And(m_Specific(X), m_Value())) ||
1889 match(Y, m_And(m_Value(), m_Specific(X))))
1890 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1891 return true;
1892
1893 unsigned BitWidth = V->getType()->getScalarSizeInBits();
1894 KnownBits LHSBits(BitWidth);
1895 computeKnownBits(X, LHSBits, Depth, Q);
1896
1897 KnownBits RHSBits(BitWidth);
1898 computeKnownBits(Y, RHSBits, Depth, Q);
1899 // If i8 V is a power of two or zero:
1900 // ZeroBits: 1 1 1 0 1 1 1 1
1901 // ~ZeroBits: 0 0 0 1 0 0 0 0
1902 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1903 // If OrZero isn't set, we cannot give back a zero result.
1904 // Make sure either the LHS or RHS has a bit set.
1905 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1906 return true;
1907 }
1908 }
1909
1910 // An exact divide or right shift can only shift off zero bits, so the result
1911 // is a power of two only if the first operand is a power of two and not
1912 // copying a sign bit (sdiv int_min, 2).
1913 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1914 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1915 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1916 Depth, Q);
1917 }
1918
1919 return false;
1920}
1921
1922/// Test whether a GEP's result is known to be non-null.
1923///
1924/// Uses properties inherent in a GEP to try to determine whether it is known
1925/// to be non-null.
1926///
1927/// Currently this routine does not support vector GEPs.
1928static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1929 const Query &Q) {
1930 const Function *F = nullptr;
1931 if (const Instruction *I = dyn_cast<Instruction>(GEP))
1932 F = I->getFunction();
1933
1934 if (!GEP->isInBounds() ||
1935 NullPointerIsDefined(F, GEP->getPointerAddressSpace()))
1936 return false;
1937
1938 // FIXME: Support vector-GEPs.
1939 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP")((GEP->getType()->isPointerTy() && "We only support plain pointer GEP"
) ? static_cast<void> (0) : __assert_fail ("GEP->getType()->isPointerTy() && \"We only support plain pointer GEP\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 1939, __PRETTY_FUNCTION__))
;
1940
1941 // If the base pointer is non-null, we cannot walk to a null address with an
1942 // inbounds GEP in address space zero.
1943 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1944 return true;
1945
1946 // Walk the GEP operands and see if any operand introduces a non-zero offset.
1947 // If so, then the GEP cannot produce a null pointer, as doing so would
1948 // inherently violate the inbounds contract within address space zero.
1949 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1950 GTI != GTE; ++GTI) {
1951 // Struct types are easy -- they must always be indexed by a constant.
1952 if (StructType *STy = GTI.getStructTypeOrNull()) {
1953 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1954 unsigned ElementIdx = OpC->getZExtValue();
1955 const StructLayout *SL = Q.DL.getStructLayout(STy);
1956 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1957 if (ElementOffset > 0)
1958 return true;
1959 continue;
1960 }
1961
1962 // If we have a zero-sized type, the index doesn't matter. Keep looping.
1963 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1964 continue;
1965
1966 // Fast path the constant operand case both for efficiency and so we don't
1967 // increment Depth when just zipping down an all-constant GEP.
1968 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1969 if (!OpC->isZero())
1970 return true;
1971 continue;
1972 }
1973
1974 // We post-increment Depth here because while isKnownNonZero increments it
1975 // as well, when we pop back up that increment won't persist. We don't want
1976 // to recurse 10k times just because we have 10k GEP operands. We don't
1977 // bail completely out because we want to handle constant GEPs regardless
1978 // of depth.
1979 if (Depth++ >= MaxDepth)
1980 continue;
1981
1982 if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1983 return true;
1984 }
1985
1986 return false;
1987}
1988
1989static bool isKnownNonNullFromDominatingCondition(const Value *V,
1990 const Instruction *CtxI,
1991 const DominatorTree *DT) {
1992 if (isa<Constant>(V))
1993 return false;
1994
1995 if (!CtxI || !DT)
1996 return false;
1997
1998 unsigned NumUsesExplored = 0;
1999 for (auto *U : V->users()) {
2000 // Avoid massive lists
2001 if (NumUsesExplored >= DomConditionsMaxUses)
2002 break;
2003 NumUsesExplored++;
2004
2005 // If the value is used as an argument to a call or invoke, then argument
2006 // attributes may provide an answer about null-ness.
2007 if (auto CS = ImmutableCallSite(U))
2008 if (auto *CalledFunc = CS.getCalledFunction())
2009 for (const Argument &Arg : CalledFunc->args())
2010 if (CS.getArgOperand(Arg.getArgNo()) == V &&
2011 Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
2012 return true;
2013
2014 // If the value is used as a load/store, then the pointer must be non null.
2015 if (V == getLoadStorePointerOperand(U)) {
2016 const Instruction *I = cast<Instruction>(U);
2017 if (!NullPointerIsDefined(I->getFunction(),
2018 V->getType()->getPointerAddressSpace()) &&
2019 DT->dominates(I, CtxI))
2020 return true;
2021 }
2022
2023 // Consider only compare instructions uniquely controlling a branch
2024 CmpInst::Predicate Pred;
2025 if (!match(const_cast<User *>(U),
2026 m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
2027 (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
2028 continue;
2029
2030 SmallVector<const User *, 4> WorkList;
2031 SmallPtrSet<const User *, 4> Visited;
2032 for (auto *CmpU : U->users()) {
2033 assert(WorkList.empty() && "Should be!")((WorkList.empty() && "Should be!") ? static_cast<
void> (0) : __assert_fail ("WorkList.empty() && \"Should be!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2033, __PRETTY_FUNCTION__))
;
2034 if (Visited.insert(CmpU).second)
2035 WorkList.push_back(CmpU);
2036
2037 while (!WorkList.empty()) {
2038 auto *Curr = WorkList.pop_back_val();
2039
2040 // If a user is an AND, add all its users to the work list. We only
2041 // propagate "pred != null" condition through AND because it is only
2042 // correct to assume that all conditions of AND are met in true branch.
2043 // TODO: Support similar logic of OR and EQ predicate?
2044 if (Pred == ICmpInst::ICMP_NE)
2045 if (auto *BO = dyn_cast<BinaryOperator>(Curr))
2046 if (BO->getOpcode() == Instruction::And) {
2047 for (auto *BOU : BO->users())
2048 if (Visited.insert(BOU).second)
2049 WorkList.push_back(BOU);
2050 continue;
2051 }
2052
2053 if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) {
2054 assert(BI->isConditional() && "uses a comparison!")((BI->isConditional() && "uses a comparison!") ? static_cast
<void> (0) : __assert_fail ("BI->isConditional() && \"uses a comparison!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2054, __PRETTY_FUNCTION__))
;
2055
2056 BasicBlock *NonNullSuccessor =
2057 BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
2058 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
2059 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
2060 return true;
2061 } else if (Pred == ICmpInst::ICMP_NE && isGuard(Curr) &&
2062 DT->dominates(cast<Instruction>(Curr), CtxI)) {
2063 return true;
2064 }
2065 }
2066 }
2067 }
2068
2069 return false;
2070}
2071
2072/// Does the 'Range' metadata (which must be a valid MD_range operand list)
2073/// ensure that the value it's attached to is never Value? 'RangeType' is
2074/// is the type of the value described by the range.
2075static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
2076 const unsigned NumRanges = Ranges->getNumOperands() / 2;
2077 assert(NumRanges >= 1)((NumRanges >= 1) ? static_cast<void> (0) : __assert_fail
("NumRanges >= 1", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2077, __PRETTY_FUNCTION__))
;
2078 for (unsigned i = 0; i < NumRanges; ++i) {
2079 ConstantInt *Lower =
2080 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
2081 ConstantInt *Upper =
2082 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
2083 ConstantRange Range(Lower->getValue(), Upper->getValue());
2084 if (Range.contains(Value))
2085 return false;
2086 }
2087 return true;
2088}
2089
2090/// Return true if the given value is known to be non-zero when defined. For
2091/// vectors, return true if every element is known to be non-zero when
2092/// defined. For pointers, if the context instruction and dominator tree are
2093/// specified, perform context-sensitive analysis and return true if the
2094/// pointer couldn't possibly be null at the specified instruction.
2095/// Supports values with integer or pointer type and vectors of integers.
2096bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
2097 if (auto *C = dyn_cast<Constant>(V)) {
2098 if (C->isNullValue())
2099 return false;
2100 if (isa<ConstantInt>(C))
2101 // Must be non-zero due to null test above.
2102 return true;
2103
2104 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
2105 // See the comment for IntToPtr/PtrToInt instructions below.
2106 if (CE->getOpcode() == Instruction::IntToPtr ||
2107 CE->getOpcode() == Instruction::PtrToInt)
2108 if (Q.DL.getTypeSizeInBits(CE->getOperand(0)->getType()) <=
2109 Q.DL.getTypeSizeInBits(CE->getType()))
2110 return isKnownNonZero(CE->getOperand(0), Depth, Q);
2111 }
2112
2113 // For constant vectors, check that all elements are undefined or known
2114 // non-zero to determine that the whole vector is known non-zero.
2115 if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
2116 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
2117 Constant *Elt = C->getAggregateElement(i);
2118 if (!Elt || Elt->isNullValue())
2119 return false;
2120 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
2121 return false;
2122 }
2123 return true;
2124 }
2125
2126 // A global variable in address space 0 is non null unless extern weak
2127 // or an absolute symbol reference. Other address spaces may have null as a
2128 // valid address for a global, so we can't assume anything.
2129 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
2130 if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
2131 GV->getType()->getAddressSpace() == 0)
2132 return true;
2133 } else
2134 return false;
2135 }
2136
2137 if (auto *I = dyn_cast<Instruction>(V)) {
2138 if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) {
2139 // If the possible ranges don't contain zero, then the value is
2140 // definitely non-zero.
2141 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
2142 const APInt ZeroValue(Ty->getBitWidth(), 0);
2143 if (rangeMetadataExcludesValue(Ranges, ZeroValue))
2144 return true;
2145 }
2146 }
2147 }
2148
2149 if (isKnownNonZeroFromAssume(V, Q))
2150 return true;
2151
2152 // Some of the tests below are recursive, so bail out if we hit the limit.
2153 if (Depth++ >= MaxDepth)
2154 return false;
2155
2156 // Check for pointer simplifications.
2157 if (V->getType()->isPointerTy()) {
2158 // Alloca never returns null, malloc might.
2159 if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
2160 return true;
2161
2162 // A byval, inalloca, or nonnull argument is never null.
2163 if (const Argument *A = dyn_cast<Argument>(V))
2164 if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
2165 return true;
2166
2167 // A Load tagged with nonnull metadata is never null.
2168 if (const LoadInst *LI = dyn_cast<LoadInst>(V))
2169 if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull))
2170 return true;
2171
2172 if (const auto *Call = dyn_cast<CallBase>(V)) {
2173 if (Call->isReturnNonNull())
2174 return true;
2175 if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true))
2176 return isKnownNonZero(RP, Depth, Q);
2177 }
2178 }
2179
2180 if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
2181 return true;
2182
2183 // Check for recursive pointer simplifications.
2184 if (V->getType()->isPointerTy()) {
2185 // Look through bitcast operations, GEPs, and int2ptr instructions as they
2186 // do not alter the value, or at least not the nullness property of the
2187 // value, e.g., int2ptr is allowed to zero/sign extend the value.
2188 //
2189 // Note that we have to take special care to avoid looking through
2190 // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well
2191 // as casts that can alter the value, e.g., AddrSpaceCasts.
2192 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
2193 if (isGEPKnownNonNull(GEP, Depth, Q))
2194 return true;
2195
2196 if (auto *BCO = dyn_cast<BitCastOperator>(V))
2197 return isKnownNonZero(BCO->getOperand(0), Depth, Q);
2198
2199 if (auto *I2P = dyn_cast<IntToPtrInst>(V))
2200 if (Q.DL.getTypeSizeInBits(I2P->getSrcTy()) <=
2201 Q.DL.getTypeSizeInBits(I2P->getDestTy()))
2202 return isKnownNonZero(I2P->getOperand(0), Depth, Q);
2203 }
2204
2205 // Similar to int2ptr above, we can look through ptr2int here if the cast
2206 // is a no-op or an extend and not a truncate.
2207 if (auto *P2I = dyn_cast<PtrToIntInst>(V))
2208 if (Q.DL.getTypeSizeInBits(P2I->getSrcTy()) <=
2209 Q.DL.getTypeSizeInBits(P2I->getDestTy()))
2210 return isKnownNonZero(P2I->getOperand(0), Depth, Q);
2211
2212 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
2213
2214 // X | Y != 0 if X != 0 or Y != 0.
2215 Value *X = nullptr, *Y = nullptr;
2216 if (match(V, m_Or(m_Value(X), m_Value(Y))))
2217 return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
2218
2219 // ext X != 0 if X != 0.
2220 if (isa<SExtInst>(V) || isa<ZExtInst>(V))
2221 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2222
2223 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined
2224 // if the lowest bit is shifted off the end.
2225 if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2226 // shl nuw can't remove any non-zero bits.
2227 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2228 if (Q.IIQ.hasNoUnsignedWrap(BO))
2229 return isKnownNonZero(X, Depth, Q);
2230
2231 KnownBits Known(BitWidth);
2232 computeKnownBits(X, Known, Depth, Q);
2233 if (Known.One[0])
2234 return true;
2235 }
2236 // shr X, Y != 0 if X is negative. Note that the value of the shift is not
2237 // defined if the sign bit is shifted off the end.
2238 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2239 // shr exact can only shift out zero bits.
2240 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2241 if (BO->isExact())
2242 return isKnownNonZero(X, Depth, Q);
2243
2244 KnownBits Known = computeKnownBits(X, Depth, Q);
2245 if (Known.isNegative())
2246 return true;
2247
2248 // If the shifter operand is a constant, and all of the bits shifted
2249 // out are known to be zero, and X is known non-zero then at least one
2250 // non-zero bit must remain.
2251 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2252 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2253 // Is there a known one in the portion not shifted out?
2254 if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2255 return true;
2256 // Are all the bits to be shifted out known zero?
2257 if (Known.countMinTrailingZeros() >= ShiftVal)
2258 return isKnownNonZero(X, Depth, Q);
2259 }
2260 }
2261 // div exact can only produce a zero if the dividend is zero.
2262 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2263 return isKnownNonZero(X, Depth, Q);
2264 }
2265 // X + Y.
2266 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2267 KnownBits XKnown = computeKnownBits(X, Depth, Q);
2268 KnownBits YKnown = computeKnownBits(Y, Depth, Q);
2269
2270 // If X and Y are both non-negative (as signed values) then their sum is not
2271 // zero unless both X and Y are zero.
2272 if (XKnown.isNonNegative() && YKnown.isNonNegative())
2273 if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
2274 return true;
2275
2276 // If X and Y are both negative (as signed values) then their sum is not
2277 // zero unless both X and Y equal INT_MIN.
2278 if (XKnown.isNegative() && YKnown.isNegative()) {
2279 APInt Mask = APInt::getSignedMaxValue(BitWidth);
2280 // The sign bit of X is set. If some other bit is set then X is not equal
2281 // to INT_MIN.
2282 if (XKnown.One.intersects(Mask))
2283 return true;
2284 // The sign bit of Y is set. If some other bit is set then Y is not equal
2285 // to INT_MIN.
2286 if (YKnown.One.intersects(Mask))
2287 return true;
2288 }
2289
2290 // The sum of a non-negative number and a power of two is not zero.
2291 if (XKnown.isNonNegative() &&
2292 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2293 return true;
2294 if (YKnown.isNonNegative() &&
2295 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2296 return true;
2297 }
2298 // X * Y.
2299 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2300 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2301 // If X and Y are non-zero then so is X * Y as long as the multiplication
2302 // does not overflow.
2303 if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) &&
2304 isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
2305 return true;
2306 }
2307 // (C ? X : Y) != 0 if X != 0 and Y != 0.
2308 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2309 if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
2310 isKnownNonZero(SI->getFalseValue(), Depth, Q))
2311 return true;
2312 }
2313 // PHI
2314 else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2315 // Try and detect a recurrence that monotonically increases from a
2316 // starting value, as these are common as induction variables.
2317 if (PN->getNumIncomingValues() == 2) {
2318 Value *Start = PN->getIncomingValue(0);
2319 Value *Induction = PN->getIncomingValue(1);
2320 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2321 std::swap(Start, Induction);
2322 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2323 if (!C->isZero() && !C->isNegative()) {
2324 ConstantInt *X;
2325 if (Q.IIQ.UseInstrInfo &&
2326 (match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2327 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2328 !X->isNegative())
2329 return true;
2330 }
2331 }
2332 }
2333 // Check if all incoming values are non-zero constant.
2334 bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
2335 return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
2336 });
2337 if (AllNonZeroConstants)
2338 return true;
2339 }
2340
2341 KnownBits Known(BitWidth);
2342 computeKnownBits(V, Known, Depth, Q);
2343 return Known.One != 0;
2344}
2345
2346/// Return true if V2 == V1 + X, where X is known non-zero.
2347static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
2348 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2349 if (!BO || BO->getOpcode() != Instruction::Add)
2350 return false;
2351 Value *Op = nullptr;
2352 if (V2 == BO->getOperand(0))
2353 Op = BO->getOperand(1);
2354 else if (V2 == BO->getOperand(1))
2355 Op = BO->getOperand(0);
2356 else
2357 return false;
2358 return isKnownNonZero(Op, 0, Q);
2359}
2360
2361/// Return true if it is known that V1 != V2.
2362static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
2363 if (V1 == V2)
2364 return false;
2365 if (V1->getType() != V2->getType())
2366 // We can't look through casts yet.
2367 return false;
2368 if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2369 return true;
2370
2371 if (V1->getType()->isIntOrIntVectorTy()) {
2372 // Are any known bits in V1 contradictory to known bits in V2? If V1
2373 // has a known zero where V2 has a known one, they must not be equal.
2374 KnownBits Known1 = computeKnownBits(V1, 0, Q);
2375 KnownBits Known2 = computeKnownBits(V2, 0, Q);
2376
2377 if (Known1.Zero.intersects(Known2.One) ||
2378 Known2.Zero.intersects(Known1.One))
2379 return true;
2380 }
2381 return false;
2382}
2383
2384/// Return true if 'V & Mask' is known to be zero. We use this predicate to
2385/// simplify operations downstream. Mask is known to be zero for bits that V
2386/// cannot have.
2387///
2388/// This function is defined on values with integer type, values with pointer
2389/// type, and vectors of integers. In the case
2390/// where V is a vector, the mask, known zero, and known one values are the
2391/// same width as the vector element, and the bit is set only if it is true
2392/// for all of the elements in the vector.
2393bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2394 const Query &Q) {
2395 KnownBits Known(Mask.getBitWidth());
2396 computeKnownBits(V, Known, Depth, Q);
2397 return Mask.isSubsetOf(Known.Zero);
2398}
2399
2400// Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow).
2401// Returns the input and lower/upper bounds.
2402static bool isSignedMinMaxClamp(const Value *Select, const Value *&In,
2403 const APInt *&CLow, const APInt *&CHigh) {
2404 assert(isa<Operator>(Select) &&((isa<Operator>(Select) && cast<Operator>
(Select)->getOpcode() == Instruction::Select && "Input should be a Select!"
) ? static_cast<void> (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2406, __PRETTY_FUNCTION__))
2405 cast<Operator>(Select)->getOpcode() == Instruction::Select &&((isa<Operator>(Select) && cast<Operator>
(Select)->getOpcode() == Instruction::Select && "Input should be a Select!"
) ? static_cast<void> (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2406, __PRETTY_FUNCTION__))
2406 "Input should be a Select!")((isa<Operator>(Select) && cast<Operator>
(Select)->getOpcode() == Instruction::Select && "Input should be a Select!"
) ? static_cast<void> (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2406, __PRETTY_FUNCTION__))
;
2407
2408 const Value *LHS = nullptr, *RHS = nullptr;
2409 SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor;
2410 if (SPF != SPF_SMAX && SPF != SPF_SMIN)
2411 return false;
2412
2413 if (!match(RHS, m_APInt(CLow)))
2414 return false;
2415
2416 const Value *LHS2 = nullptr, *RHS2 = nullptr;
2417 SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor;
2418 if (getInverseMinMaxFlavor(SPF) != SPF2)
2419 return false;
2420
2421 if (!match(RHS2, m_APInt(CHigh)))
2422 return false;
2423
2424 if (SPF == SPF_SMIN)
2425 std::swap(CLow, CHigh);
2426
2427 In = LHS2;
2428 return CLow->sle(*CHigh);
2429}
2430
2431/// For vector constants, loop over the elements and find the constant with the
2432/// minimum number of sign bits. Return 0 if the value is not a vector constant
2433/// or if any element was not analyzed; otherwise, return the count for the
2434/// element with the minimum number of sign bits.
2435static unsigned computeNumSignBitsVectorConstant(const Value *V,
2436 unsigned TyBits) {
2437 const auto *CV = dyn_cast<Constant>(V);
2438 if (!CV || !CV->getType()->isVectorTy())
2439 return 0;
2440
2441 unsigned MinSignBits = TyBits;
2442 unsigned NumElts = CV->getType()->getVectorNumElements();
2443 for (unsigned i = 0; i != NumElts; ++i) {
2444 // If we find a non-ConstantInt, bail out.
2445 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2446 if (!Elt)
2447 return 0;
2448
2449 MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2450 }
2451
2452 return MinSignBits;
2453}
2454
2455static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2456 const Query &Q);
2457
2458static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2459 const Query &Q) {
2460 unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2461 assert(Result > 0 && "At least one sign bit needs to be present!")((Result > 0 && "At least one sign bit needs to be present!"
) ? static_cast<void> (0) : __assert_fail ("Result > 0 && \"At least one sign bit needs to be present!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2461, __PRETTY_FUNCTION__))
;
2462 return Result;
2463}
2464
2465/// Return the number of times the sign bit of the register is replicated into
2466/// the other bits. We know that at least 1 bit is always equal to the sign bit
2467/// (itself), but other cases can give us information. For example, immediately
2468/// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2469/// other, so we return 3. For vectors, return the number of sign bits for the
2470/// vector element with the minimum number of known sign bits.
2471static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2472 const Query &Q) {
2473 assert(Depth <= MaxDepth && "Limit Search Depth")((Depth <= MaxDepth && "Limit Search Depth") ? static_cast
<void> (0) : __assert_fail ("Depth <= MaxDepth && \"Limit Search Depth\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2473, __PRETTY_FUNCTION__))
;
2474
2475 // We return the minimum number of sign bits that are guaranteed to be present
2476 // in V, so for undef we have to conservatively return 1. We don't have the
2477 // same behavior for poison though -- that's a FIXME today.
2478
2479 Type *ScalarTy = V->getType()->getScalarType();
2480 unsigned TyBits = ScalarTy->isPointerTy() ?
2481 Q.DL.getPointerTypeSizeInBits(ScalarTy) :
2482 Q.DL.getTypeSizeInBits(ScalarTy);
2483
2484 unsigned Tmp, Tmp2;
2485 unsigned FirstAnswer = 1;
2486
2487 // Note that ConstantInt is handled by the general computeKnownBits case
2488 // below.
2489
2490 if (Depth == MaxDepth)
2491 return 1; // Limit search depth.
2492
2493 if (auto *U = dyn_cast<Operator>(V)) {
2494 switch (Operator::getOpcode(V)) {
2495 default: break;
2496 case Instruction::SExt:
2497 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2498 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2499
2500 case Instruction::SDiv: {
2501 const APInt *Denominator;
2502 // sdiv X, C -> adds log(C) sign bits.
2503 if (match(U->getOperand(1), m_APInt(Denominator))) {
2504
2505 // Ignore non-positive denominator.
2506 if (!Denominator->isStrictlyPositive())
2507 break;
2508
2509 // Calculate the incoming numerator bits.
2510 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2511
2512 // Add floor(log(C)) bits to the numerator bits.
2513 return std::min(TyBits, NumBits + Denominator->logBase2());
2514 }
2515 break;
2516 }
2517
2518 case Instruction::SRem: {
2519 const APInt *Denominator;
2520 // srem X, C -> we know that the result is within [-C+1,C) when C is a
2521 // positive constant. This let us put a lower bound on the number of sign
2522 // bits.
2523 if (match(U->getOperand(1), m_APInt(Denominator))) {
2524
2525 // Ignore non-positive denominator.
2526 if (!Denominator->isStrictlyPositive())
2527 break;
2528
2529 // Calculate the incoming numerator bits. SRem by a positive constant
2530 // can't lower the number of sign bits.
2531 unsigned NumrBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2532
2533 // Calculate the leading sign bit constraints by examining the
2534 // denominator. Given that the denominator is positive, there are two
2535 // cases:
2536 //
2537 // 1. the numerator is positive. The result range is [0,C) and [0,C) u<
2538 // (1 << ceilLogBase2(C)).
2539 //
2540 // 2. the numerator is negative. Then the result range is (-C,0] and
2541 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2542 //
2543 // Thus a lower bound on the number of sign bits is `TyBits -
2544 // ceilLogBase2(C)`.
2545
2546 unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2547 return std::max(NumrBits, ResBits);
2548 }
2549 break;
2550 }
2551
2552 case Instruction::AShr: {
2553 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2554 // ashr X, C -> adds C sign bits. Vectors too.
2555 const APInt *ShAmt;
2556 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2557 if (ShAmt->uge(TyBits))
2558 break; // Bad shift.
2559 unsigned ShAmtLimited = ShAmt->getZExtValue();
2560 Tmp += ShAmtLimited;
2561 if (Tmp > TyBits) Tmp = TyBits;
2562 }
2563 return Tmp;
2564 }
2565 case Instruction::Shl: {
2566 const APInt *ShAmt;
2567 if (match(U->getOperand(1), m_APInt(ShAmt))) {
2568 // shl destroys sign bits.
2569 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2570 if (ShAmt->uge(TyBits) || // Bad shift.
2571 ShAmt->uge(Tmp)) break; // Shifted all sign bits out.
2572 Tmp2 = ShAmt->getZExtValue();
2573 return Tmp - Tmp2;
2574 }
2575 break;
2576 }
2577 case Instruction::And:
2578 case Instruction::Or:
2579 case Instruction::Xor: // NOT is handled here.
2580 // Logical binary ops preserve the number of sign bits at the worst.
2581 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2582 if (Tmp != 1) {
2583 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2584 FirstAnswer = std::min(Tmp, Tmp2);
2585 // We computed what we know about the sign bits as our first
2586 // answer. Now proceed to the generic code that uses
2587 // computeKnownBits, and pick whichever answer is better.
2588 }
2589 break;
2590
2591 case Instruction::Select: {
2592 // If we have a clamp pattern, we know that the number of sign bits will
2593 // be the minimum of the clamp min/max range.
2594 const Value *X;
2595 const APInt *CLow, *CHigh;
2596 if (isSignedMinMaxClamp(U, X, CLow, CHigh))
2597 return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits());
2598
2599 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2600 if (Tmp == 1) break;
2601 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2602 return std::min(Tmp, Tmp2);
2603 }
2604
2605 case Instruction::Add:
2606 // Add can have at most one carry bit. Thus we know that the output
2607 // is, at worst, one more bit than the inputs.
2608 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2609 if (Tmp == 1) break;
2610
2611 // Special case decrementing a value (ADD X, -1):
2612 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2613 if (CRHS->isAllOnesValue()) {
2614 KnownBits Known(TyBits);
2615 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2616
2617 // If the input is known to be 0 or 1, the output is 0/-1, which is
2618 // all sign bits set.
2619 if ((Known.Zero | 1).isAllOnesValue())
2620 return TyBits;
2621
2622 // If we are subtracting one from a positive number, there is no carry
2623 // out of the result.
2624 if (Known.isNonNegative())
2625 return Tmp;
2626 }
2627
2628 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2629 if (Tmp2 == 1) break;
2630 return std::min(Tmp, Tmp2) - 1;
2631
2632 case Instruction::Sub:
2633 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2634 if (Tmp2 == 1) break;
2635
2636 // Handle NEG.
2637 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2638 if (CLHS->isNullValue()) {
2639 KnownBits Known(TyBits);
2640 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2641 // If the input is known to be 0 or 1, the output is 0/-1, which is
2642 // all sign bits set.
2643 if ((Known.Zero | 1).isAllOnesValue())
2644 return TyBits;
2645
2646 // If the input is known to be positive (the sign bit is known clear),
2647 // the output of the NEG has the same number of sign bits as the
2648 // input.
2649 if (Known.isNonNegative())
2650 return Tmp2;
2651
2652 // Otherwise, we treat this like a SUB.
2653 }
2654
2655 // Sub can have at most one carry bit. Thus we know that the output
2656 // is, at worst, one more bit than the inputs.
2657 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2658 if (Tmp == 1) break;
2659 return std::min(Tmp, Tmp2) - 1;
2660
2661 case Instruction::Mul: {
2662 // The output of the Mul can be at most twice the valid bits in the
2663 // inputs.
2664 unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2665 if (SignBitsOp0 == 1) break;
2666 unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2667 if (SignBitsOp1 == 1) break;
2668 unsigned OutValidBits =
2669 (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
2670 return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
2671 }
2672
2673 case Instruction::PHI: {
2674 const PHINode *PN = cast<PHINode>(U);
2675 unsigned NumIncomingValues = PN->getNumIncomingValues();
2676 // Don't analyze large in-degree PHIs.
2677 if (NumIncomingValues > 4) break;
2678 // Unreachable blocks may have zero-operand PHI nodes.
2679 if (NumIncomingValues == 0) break;
2680
2681 // Take the minimum of all incoming values. This can't infinitely loop
2682 // because of our depth threshold.
2683 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2684 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2685 if (Tmp == 1) return Tmp;
2686 Tmp = std::min(
2687 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2688 }
2689 return Tmp;
2690 }
2691
2692 case Instruction::Trunc:
2693 // FIXME: it's tricky to do anything useful for this, but it is an
2694 // important case for targets like X86.
2695 break;
2696
2697 case Instruction::ExtractElement:
2698 // Look through extract element. At the moment we keep this simple and
2699 // skip tracking the specific element. But at least we might find
2700 // information valid for all elements of the vector (for example if vector
2701 // is sign extended, shifted, etc).
2702 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2703
2704 case Instruction::ShuffleVector: {
2705 // TODO: This is copied almost directly from the SelectionDAG version of
2706 // ComputeNumSignBits. It would be better if we could share common
2707 // code. If not, make sure that changes are translated to the DAG.
2708
2709 // Collect the minimum number of sign bits that are shared by every vector
2710 // element referenced by the shuffle.
2711 auto *Shuf = cast<ShuffleVectorInst>(U);
2712 int NumElts = Shuf->getOperand(0)->getType()->getVectorNumElements();
2713 int NumMaskElts = Shuf->getMask()->getType()->getVectorNumElements();
2714 APInt DemandedLHS(NumElts, 0), DemandedRHS(NumElts, 0);
2715 for (int i = 0; i != NumMaskElts; ++i) {
2716 int M = Shuf->getMaskValue(i);
2717 assert(M < NumElts * 2 && "Invalid shuffle mask constant")((M < NumElts * 2 && "Invalid shuffle mask constant"
) ? static_cast<void> (0) : __assert_fail ("M < NumElts * 2 && \"Invalid shuffle mask constant\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2717, __PRETTY_FUNCTION__))
;
2718 // For undef elements, we don't know anything about the common state of
2719 // the shuffle result.
2720 if (M == -1)
2721 return 1;
2722 if (M < NumElts)
2723 DemandedLHS.setBit(M % NumElts);
2724 else
2725 DemandedRHS.setBit(M % NumElts);
2726 }
2727 Tmp = std::numeric_limits<unsigned>::max();
2728 if (!!DemandedLHS)
2729 Tmp = ComputeNumSignBits(Shuf->getOperand(0), Depth + 1, Q);
2730 if (!!DemandedRHS) {
2731 Tmp2 = ComputeNumSignBits(Shuf->getOperand(1), Depth + 1, Q);
2732 Tmp = std::min(Tmp, Tmp2);
2733 }
2734 // If we don't know anything, early out and try computeKnownBits
2735 // fall-back.
2736 if (Tmp == 1)
2737 break;
2738 assert(Tmp <= V->getType()->getScalarSizeInBits() &&((Tmp <= V->getType()->getScalarSizeInBits() &&
"Failed to determine minimum sign bits") ? static_cast<void
> (0) : __assert_fail ("Tmp <= V->getType()->getScalarSizeInBits() && \"Failed to determine minimum sign bits\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2739, __PRETTY_FUNCTION__))
2739 "Failed to determine minimum sign bits")((Tmp <= V->getType()->getScalarSizeInBits() &&
"Failed to determine minimum sign bits") ? static_cast<void
> (0) : __assert_fail ("Tmp <= V->getType()->getScalarSizeInBits() && \"Failed to determine minimum sign bits\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2739, __PRETTY_FUNCTION__))
;
2740 return Tmp;
2741 }
2742 }
2743 }
2744
2745 // Finally, if we can prove that the top bits of the result are 0's or 1's,
2746 // use this information.
2747
2748 // If we can examine all elements of a vector constant successfully, we're
2749 // done (we can't do any better than that). If not, keep trying.
2750 if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2751 return VecSignBits;
2752
2753 KnownBits Known(TyBits);
2754 computeKnownBits(V, Known, Depth, Q);
2755
2756 // If we know that the sign bit is either zero or one, determine the number of
2757 // identical bits in the top of the input value.
2758 return std::max(FirstAnswer, Known.countMinSignBits());
2759}
2760
2761/// This function computes the integer multiple of Base that equals V.
2762/// If successful, it returns true and returns the multiple in
2763/// Multiple. If unsuccessful, it returns false. It looks
2764/// through SExt instructions only if LookThroughSExt is true.
2765bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2766 bool LookThroughSExt, unsigned Depth) {
2767 assert(V && "No Value?")((V && "No Value?") ? static_cast<void> (0) : __assert_fail
("V && \"No Value?\"", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2767, __PRETTY_FUNCTION__))
;
2768 assert(Depth <= MaxDepth && "Limit Search Depth")((Depth <= MaxDepth && "Limit Search Depth") ? static_cast
<void> (0) : __assert_fail ("Depth <= MaxDepth && \"Limit Search Depth\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2768, __PRETTY_FUNCTION__))
;
2769 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!")((V->getType()->isIntegerTy() && "Not integer or pointer type!"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntegerTy() && \"Not integer or pointer type!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 2769, __PRETTY_FUNCTION__))
;
2770
2771 Type *T = V->getType();
2772
2773 ConstantInt *CI = dyn_cast<ConstantInt>(V);
2774
2775 if (Base == 0)
2776 return false;
2777
2778 if (Base == 1) {
2779 Multiple = V;
2780 return true;
2781 }
2782
2783 ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2784 Constant *BaseVal = ConstantInt::get(T, Base);
2785 if (CO && CO == BaseVal) {
2786 // Multiple is 1.
2787 Multiple = ConstantInt::get(T, 1);
2788 return true;
2789 }
2790
2791 if (CI && CI->getZExtValue() % Base == 0) {
2792 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2793 return true;
2794 }
2795
2796 if (Depth == MaxDepth) return false; // Limit search depth.
2797
2798 Operator *I = dyn_cast<Operator>(V);
2799 if (!I) return false;
2800
2801 switch (I->getOpcode()) {
2802 default: break;
2803 case Instruction::SExt:
2804 if (!LookThroughSExt) return false;
2805 // otherwise fall through to ZExt
2806 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2807 case Instruction::ZExt:
2808 return ComputeMultiple(I->getOperand(0), Base, Multiple,
2809 LookThroughSExt, Depth+1);
2810 case Instruction::Shl:
2811 case Instruction::Mul: {
2812 Value *Op0 = I->getOperand(0);
2813 Value *Op1 = I->getOperand(1);
2814
2815 if (I->getOpcode() == Instruction::Shl) {
2816 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2817 if (!Op1CI) return false;
2818 // Turn Op0 << Op1 into Op0 * 2^Op1
2819 APInt Op1Int = Op1CI->getValue();
2820 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2821 APInt API(Op1Int.getBitWidth(), 0);
2822 API.setBit(BitToSet);
2823 Op1 = ConstantInt::get(V->getContext(), API);
2824 }
2825
2826 Value *Mul0 = nullptr;
2827 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2828 if (Constant *Op1C = dyn_cast<Constant>(Op1))
2829 if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2830 if (Op1C->getType()->getPrimitiveSizeInBits() <
2831 MulC->getType()->getPrimitiveSizeInBits())
2832 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2833 if (Op1C->getType()->getPrimitiveSizeInBits() >
2834 MulC->getType()->getPrimitiveSizeInBits())
2835 MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2836
2837 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2838 Multiple = ConstantExpr::getMul(MulC, Op1C);
2839 return true;
2840 }
2841
2842 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2843 if (Mul0CI->getValue() == 1) {
2844 // V == Base * Op1, so return Op1
2845 Multiple = Op1;
2846 return true;
2847 }
2848 }
2849
2850 Value *Mul1 = nullptr;
2851 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2852 if (Constant *Op0C = dyn_cast<Constant>(Op0))
2853 if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2854 if (Op0C->getType()->getPrimitiveSizeInBits() <
2855 MulC->getType()->getPrimitiveSizeInBits())
2856 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2857 if (Op0C->getType()->getPrimitiveSizeInBits() >
2858 MulC->getType()->getPrimitiveSizeInBits())
2859 MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2860
2861 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2862 Multiple = ConstantExpr::getMul(MulC, Op0C);
2863 return true;
2864 }
2865
2866 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2867 if (Mul1CI->getValue() == 1) {
2868 // V == Base * Op0, so return Op0
2869 Multiple = Op0;
2870 return true;
2871 }
2872 }
2873 }
2874 }
2875
2876 // We could not determine if V is a multiple of Base.
2877 return false;
2878}
2879
2880Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2881 const TargetLibraryInfo *TLI) {
2882 const Function *F = ICS.getCalledFunction();
2883 if (!F)
2884 return Intrinsic::not_intrinsic;
2885
2886 if (F->isIntrinsic())
2887 return F->getIntrinsicID();
2888
2889 if (!TLI)
2890 return Intrinsic::not_intrinsic;
2891
2892 LibFunc Func;
2893 // We're going to make assumptions on the semantics of the functions, check
2894 // that the target knows that it's available in this environment and it does
2895 // not have local linkage.
2896 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2897 return Intrinsic::not_intrinsic;
2898
2899 if (!ICS.onlyReadsMemory())
2900 return Intrinsic::not_intrinsic;
2901
2902 // Otherwise check if we have a call to a function that can be turned into a
2903 // vector intrinsic.
2904 switch (Func) {
2905 default:
2906 break;
2907 case LibFunc_sin:
2908 case LibFunc_sinf:
2909 case LibFunc_sinl:
2910 return Intrinsic::sin;
2911 case LibFunc_cos:
2912 case LibFunc_cosf:
2913 case LibFunc_cosl:
2914 return Intrinsic::cos;
2915 case LibFunc_exp:
2916 case LibFunc_expf:
2917 case LibFunc_expl:
2918 return Intrinsic::exp;
2919 case LibFunc_exp2:
2920 case LibFunc_exp2f:
2921 case LibFunc_exp2l:
2922 return Intrinsic::exp2;
2923 case LibFunc_log:
2924 case LibFunc_logf:
2925 case LibFunc_logl:
2926 return Intrinsic::log;
2927 case LibFunc_log10:
2928 case LibFunc_log10f:
2929 case LibFunc_log10l:
2930 return Intrinsic::log10;
2931 case LibFunc_log2:
2932 case LibFunc_log2f:
2933 case LibFunc_log2l:
2934 return Intrinsic::log2;
2935 case LibFunc_fabs:
2936 case LibFunc_fabsf:
2937 case LibFunc_fabsl:
2938 return Intrinsic::fabs;
2939 case LibFunc_fmin:
2940 case LibFunc_fminf:
2941 case LibFunc_fminl:
2942 return Intrinsic::minnum;
2943 case LibFunc_fmax:
2944 case LibFunc_fmaxf:
2945 case LibFunc_fmaxl:
2946 return Intrinsic::maxnum;
2947 case LibFunc_copysign:
2948 case LibFunc_copysignf:
2949 case LibFunc_copysignl:
2950 return Intrinsic::copysign;
2951 case LibFunc_floor:
2952 case LibFunc_floorf:
2953 case LibFunc_floorl:
2954 return Intrinsic::floor;
2955 case LibFunc_ceil:
2956 case LibFunc_ceilf:
2957 case LibFunc_ceill:
2958 return Intrinsic::ceil;
2959 case LibFunc_trunc:
2960 case LibFunc_truncf:
2961 case LibFunc_truncl:
2962 return Intrinsic::trunc;
2963 case LibFunc_rint:
2964 case LibFunc_rintf:
2965 case LibFunc_rintl:
2966 return Intrinsic::rint;
2967 case LibFunc_nearbyint:
2968 case LibFunc_nearbyintf:
2969 case LibFunc_nearbyintl:
2970 return Intrinsic::nearbyint;
2971 case LibFunc_round:
2972 case LibFunc_roundf:
2973 case LibFunc_roundl:
2974 return Intrinsic::round;
2975 case LibFunc_pow:
2976 case LibFunc_powf:
2977 case LibFunc_powl:
2978 return Intrinsic::pow;
2979 case LibFunc_sqrt:
2980 case LibFunc_sqrtf:
2981 case LibFunc_sqrtl:
2982 return Intrinsic::sqrt;
2983 }
2984
2985 return Intrinsic::not_intrinsic;
2986}
2987
2988/// Return true if we can prove that the specified FP value is never equal to
2989/// -0.0.
2990///
2991/// NOTE: this function will need to be revisited when we support non-default
2992/// rounding modes!
2993bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2994 unsigned Depth) {
2995 if (auto *CFP = dyn_cast<ConstantFP>(V))
2996 return !CFP->getValueAPF().isNegZero();
2997
2998 // Limit search depth.
2999 if (Depth == MaxDepth)
3000 return false;
3001
3002 auto *Op = dyn_cast<Operator>(V);
3003 if (!Op)
3004 return false;
3005
3006 // Check if the nsz fast-math flag is set.
3007 if (auto *FPO = dyn_cast<FPMathOperator>(Op))
3008 if (FPO->hasNoSignedZeros())
3009 return true;
3010
3011 // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
3012 if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
3013 return true;
3014
3015 // sitofp and uitofp turn into +0.0 for zero.
3016 if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
3017 return true;
3018
3019 if (auto *Call = dyn_cast<CallInst>(Op)) {
3020 Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
3021 switch (IID) {
3022 default:
3023 break;
3024 // sqrt(-0.0) = -0.0, no other negative results are possible.
3025 case Intrinsic::sqrt:
3026 case Intrinsic::canonicalize:
3027 return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
3028 // fabs(x) != -0.0
3029 case Intrinsic::fabs:
3030 return true;
3031 }
3032 }
3033
3034 return false;
3035}
3036
3037/// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
3038/// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
3039/// bit despite comparing equal.
3040static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
3041 const TargetLibraryInfo *TLI,
3042 bool SignBitOnly,
3043 unsigned Depth) {
3044 // TODO: This function does not do the right thing when SignBitOnly is true
3045 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
3046 // which flips the sign bits of NaNs. See
3047 // https://llvm.org/bugs/show_bug.cgi?id=31702.
3048
3049 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
3050 return !CFP->getValueAPF().isNegative() ||
3051 (!SignBitOnly && CFP->getValueAPF().isZero());
3052 }
3053
3054 // Handle vector of constants.
3055 if (auto *CV = dyn_cast<Constant>(V)) {
3056 if (CV->getType()->isVectorTy()) {
3057 unsigned NumElts = CV->getType()->getVectorNumElements();
3058 for (unsigned i = 0; i != NumElts; ++i) {
3059 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
3060 if (!CFP)
3061 return false;
3062 if (CFP->getValueAPF().isNegative() &&
3063 (SignBitOnly || !CFP->getValueAPF().isZero()))
3064 return false;
3065 }
3066
3067 // All non-negative ConstantFPs.
3068 return true;
3069 }
3070 }
3071
3072 if (Depth == MaxDepth)
3073 return false; // Limit search depth.
3074
3075 const Operator *I = dyn_cast<Operator>(V);
3076 if (!I)
3077 return false;
3078
3079 switch (I->getOpcode()) {
3080 default:
3081 break;
3082 // Unsigned integers are always nonnegative.
3083 case Instruction::UIToFP:
3084 return true;
3085 case Instruction::FMul:
3086 // x*x is always non-negative or a NaN.
3087 if (I->getOperand(0) == I->getOperand(1) &&
3088 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
3089 return true;
3090
3091 LLVM_FALLTHROUGH[[gnu::fallthrough]];
3092 case Instruction::FAdd:
3093 case Instruction::FDiv:
3094 case Instruction::FRem:
3095 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3096 Depth + 1) &&
3097 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3098 Depth + 1);
3099 case Instruction::Select:
3100 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3101 Depth + 1) &&
3102 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3103 Depth + 1);
3104 case Instruction::FPExt:
3105 case Instruction::FPTrunc:
3106 // Widening/narrowing never change sign.
3107 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3108 Depth + 1);
3109 case Instruction::ExtractElement:
3110 // Look through extract element. At the moment we keep this simple and skip
3111 // tracking the specific element. But at least we might find information
3112 // valid for all elements of the vector.
3113 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3114 Depth + 1);
3115 case Instruction::Call:
3116 const auto *CI = cast<CallInst>(I);
3117 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
3118 switch (IID) {
3119 default:
3120 break;
3121 case Intrinsic::maxnum:
3122 return (isKnownNeverNaN(I->getOperand(0), TLI) &&
3123 cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI,
3124 SignBitOnly, Depth + 1)) ||
3125 (isKnownNeverNaN(I->getOperand(1), TLI) &&
3126 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI,
3127 SignBitOnly, Depth + 1));
3128
3129 case Intrinsic::maximum:
3130 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3131 Depth + 1) ||
3132 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3133 Depth + 1);
3134 case Intrinsic::minnum:
3135 case Intrinsic::minimum:
3136 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3137 Depth + 1) &&
3138 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
3139 Depth + 1);
3140 case Intrinsic::exp:
3141 case Intrinsic::exp2:
3142 case Intrinsic::fabs:
3143 return true;
3144
3145 case Intrinsic::sqrt:
3146 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0.
3147 if (!SignBitOnly)
3148 return true;
3149 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
3150 CannotBeNegativeZero(CI->getOperand(0), TLI));
3151
3152 case Intrinsic::powi:
3153 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
3154 // powi(x,n) is non-negative if n is even.
3155 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
3156 return true;
3157 }
3158 // TODO: This is not correct. Given that exp is an integer, here are the
3159 // ways that pow can return a negative value:
3160 //
3161 // pow(x, exp) --> negative if exp is odd and x is negative.
3162 // pow(-0, exp) --> -inf if exp is negative odd.
3163 // pow(-0, exp) --> -0 if exp is positive odd.
3164 // pow(-inf, exp) --> -0 if exp is negative odd.
3165 // pow(-inf, exp) --> -inf if exp is positive odd.
3166 //
3167 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
3168 // but we must return false if x == -0. Unfortunately we do not currently
3169 // have a way of expressing this constraint. See details in
3170 // https://llvm.org/bugs/show_bug.cgi?id=31702.
3171 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
3172 Depth + 1);
3173
3174 case Intrinsic::fma:
3175 case Intrinsic::fmuladd:
3176 // x*x+y is non-negative if y is non-negative.
3177 return I->getOperand(0) == I->getOperand(1) &&
3178 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
3179 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
3180 Depth + 1);
3181 }
3182 break;
3183 }
3184 return false;
3185}
3186
3187bool llvm::CannotBeOrderedLessThanZero(const Value *V,
3188 const TargetLibraryInfo *TLI) {
3189 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
3190}
3191
3192bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
3193 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
3194}
3195
3196bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI,
3197 unsigned Depth) {
3198 assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type")((V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isFPOrFPVectorTy() && \"Querying for Inf on non-FP type\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3198, __PRETTY_FUNCTION__))
;
3199
3200 // If we're told that infinities won't happen, assume they won't.
3201 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3202 if (FPMathOp->hasNoInfs())
3203 return true;
3204
3205 // Handle scalar constants.
3206 if (auto *CFP = dyn_cast<ConstantFP>(V))
3207 return !CFP->isInfinity();
3208
3209 if (Depth == MaxDepth)
3210 return false;
3211
3212 if (auto *Inst = dyn_cast<Instruction>(V)) {
3213 switch (Inst->getOpcode()) {
3214 case Instruction::Select: {
3215 return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) &&
3216 isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1);
3217 }
3218 case Instruction::UIToFP:
3219 // If the input type fits into the floating type the result is finite.
3220 return ilogb(APFloat::getLargest(
3221 Inst->getType()->getScalarType()->getFltSemantics())) >=
3222 (int)Inst->getOperand(0)->getType()->getScalarSizeInBits();
3223 default:
3224 break;
3225 }
3226 }
3227
3228 // Bail out for constant expressions, but try to handle vector constants.
3229 if (!V->getType()->isVectorTy() || !isa<Constant>(V))
3230 return false;
3231
3232 // For vectors, verify that each element is not infinity.
3233 unsigned NumElts = V->getType()->getVectorNumElements();
3234 for (unsigned i = 0; i != NumElts; ++i) {
3235 Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3236 if (!Elt)
3237 return false;
3238 if (isa<UndefValue>(Elt))
3239 continue;
3240 auto *CElt = dyn_cast<ConstantFP>(Elt);
3241 if (!CElt || CElt->isInfinity())
3242 return false;
3243 }
3244 // All elements were confirmed non-infinity or undefined.
3245 return true;
3246}
3247
3248bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI,
3249 unsigned Depth) {
3250 assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type")((V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isFPOrFPVectorTy() && \"Querying for NaN on non-FP type\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3250, __PRETTY_FUNCTION__))
;
3251
3252 // If we're told that NaNs won't happen, assume they won't.
3253 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
3254 if (FPMathOp->hasNoNaNs())
3255 return true;
3256
3257 // Handle scalar constants.
3258 if (auto *CFP = dyn_cast<ConstantFP>(V))
3259 return !CFP->isNaN();
3260
3261 if (Depth == MaxDepth)
3262 return false;
3263
3264 if (auto *Inst = dyn_cast<Instruction>(V)) {
3265 switch (Inst->getOpcode()) {
3266 case Instruction::FAdd:
3267 case Instruction::FSub:
3268 // Adding positive and negative infinity produces NaN.
3269 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3270 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3271 (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) ||
3272 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1));
3273
3274 case Instruction::FMul:
3275 // Zero multiplied with infinity produces NaN.
3276 // FIXME: If neither side can be zero fmul never produces NaN.
3277 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) &&
3278 isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) &&
3279 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3280 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1);
3281
3282 case Instruction::FDiv:
3283 case Instruction::FRem:
3284 // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN.
3285 return false;
3286
3287 case Instruction::Select: {
3288 return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) &&
3289 isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1);
3290 }
3291 case Instruction::SIToFP:
3292 case Instruction::UIToFP:
3293 return true;
3294 case Instruction::FPTrunc:
3295 case Instruction::FPExt:
3296 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1);
3297 default:
3298 break;
3299 }
3300 }
3301
3302 if (const auto *II = dyn_cast<IntrinsicInst>(V)) {
3303 switch (II->getIntrinsicID()) {
3304 case Intrinsic::canonicalize:
3305 case Intrinsic::fabs:
3306 case Intrinsic::copysign:
3307 case Intrinsic::exp:
3308 case Intrinsic::exp2:
3309 case Intrinsic::floor:
3310 case Intrinsic::ceil:
3311 case Intrinsic::trunc:
3312 case Intrinsic::rint:
3313 case Intrinsic::nearbyint:
3314 case Intrinsic::round:
3315 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1);
3316 case Intrinsic::sqrt:
3317 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) &&
3318 CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI);
3319 case Intrinsic::minnum:
3320 case Intrinsic::maxnum:
3321 // If either operand is not NaN, the result is not NaN.
3322 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) ||
3323 isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1);
3324 default:
3325 return false;
3326 }
3327 }
3328
3329 // Bail out for constant expressions, but try to handle vector constants.
3330 if (!V->getType()->isVectorTy() || !isa<Constant>(V))
3331 return false;
3332
3333 // For vectors, verify that each element is not NaN.
3334 unsigned NumElts = V->getType()->getVectorNumElements();
3335 for (unsigned i = 0; i != NumElts; ++i) {
3336 Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
3337 if (!Elt)
3338 return false;
3339 if (isa<UndefValue>(Elt))
3340 continue;
3341 auto *CElt = dyn_cast<ConstantFP>(Elt);
3342 if (!CElt || CElt->isNaN())
3343 return false;
3344 }
3345 // All elements were confirmed not-NaN or undefined.
3346 return true;
3347}
3348
3349Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) {
3350
3351 // All byte-wide stores are splatable, even of arbitrary variables.
3352 if (V->getType()->isIntegerTy(8))
3353 return V;
3354
3355 LLVMContext &Ctx = V->getContext();
3356
3357 // Undef don't care.
3358 auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx));
3359 if (isa<UndefValue>(V))
3360 return UndefInt8;
3361
3362 const uint64_t Size = DL.getTypeStoreSize(V->getType());
3363 if (!Size)
3364 return UndefInt8;
3365
3366 Constant *C = dyn_cast<Constant>(V);
3367 if (!C) {
3368 // Conceptually, we could handle things like:
3369 // %a = zext i8 %X to i16
3370 // %b = shl i16 %a, 8
3371 // %c = or i16 %a, %b
3372 // but until there is an example that actually needs this, it doesn't seem
3373 // worth worrying about.
3374 return nullptr;
3375 }
3376
3377 // Handle 'null' ConstantArrayZero etc.
3378 if (C->isNullValue())
3379 return Constant::getNullValue(Type::getInt8Ty(Ctx));
3380
3381 // Constant floating-point values can be handled as integer values if the
3382 // corresponding integer value is "byteable". An important case is 0.0.
3383 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3384 Type *Ty = nullptr;
3385 if (CFP->getType()->isHalfTy())
3386 Ty = Type::getInt16Ty(Ctx);
3387 else if (CFP->getType()->isFloatTy())
3388 Ty = Type::getInt32Ty(Ctx);
3389 else if (CFP->getType()->isDoubleTy())
3390 Ty = Type::getInt64Ty(Ctx);
3391 // Don't handle long double formats, which have strange constraints.
3392 return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL)
3393 : nullptr;
3394 }
3395
3396 // We can handle constant integers that are multiple of 8 bits.
3397 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
3398 if (CI->getBitWidth() % 8 == 0) {
3399 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!")((CI->getBitWidth() > 8 && "8 bits should be handled above!"
) ? static_cast<void> (0) : __assert_fail ("CI->getBitWidth() > 8 && \"8 bits should be handled above!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3399, __PRETTY_FUNCTION__))
;
3400 if (!CI->getValue().isSplat(8))
3401 return nullptr;
3402 return ConstantInt::get(Ctx, CI->getValue().trunc(8));
3403 }
3404 }
3405
3406 if (auto *CE = dyn_cast<ConstantExpr>(C)) {
3407 if (CE->getOpcode() == Instruction::IntToPtr) {
3408 auto PS = DL.getPointerSizeInBits(
3409 cast<PointerType>(CE->getType())->getAddressSpace());
3410 return isBytewiseValue(
3411 ConstantExpr::getIntegerCast(CE->getOperand(0),
3412 Type::getIntNTy(Ctx, PS), false),
3413 DL);
3414 }
3415 }
3416
3417 auto Merge = [&](Value *LHS, Value *RHS) -> Value * {
3418 if (LHS == RHS)
3419 return LHS;
3420 if (!LHS || !RHS)
3421 return nullptr;
3422 if (LHS == UndefInt8)
3423 return RHS;
3424 if (RHS == UndefInt8)
3425 return LHS;
3426 return nullptr;
3427 };
3428
3429 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) {
3430 Value *Val = UndefInt8;
3431 for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I)
3432 if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL))))
3433 return nullptr;
3434 return Val;
3435 }
3436
3437 if (isa<ConstantAggregate>(C)) {
3438 Value *Val = UndefInt8;
3439 for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I)
3440 if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL))))
3441 return nullptr;
3442 return Val;
3443 }
3444
3445 // Don't try to handle the handful of other constants.
3446 return nullptr;
3447}
3448
3449// This is the recursive version of BuildSubAggregate. It takes a few different
3450// arguments. Idxs is the index within the nested struct From that we are
3451// looking at now (which is of type IndexedType). IdxSkip is the number of
3452// indices from Idxs that should be left out when inserting into the resulting
3453// struct. To is the result struct built so far, new insertvalue instructions
3454// build on that.
3455static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
3456 SmallVectorImpl<unsigned> &Idxs,
3457 unsigned IdxSkip,
3458 Instruction *InsertBefore) {
3459 StructType *STy = dyn_cast<StructType>(IndexedType);
1
Assuming 'IndexedType' is a 'StructType'
3460 if (STy
1.1
'STy' is non-null
1.1
'STy' is non-null
) {
2
Taking true branch
3461 // Save the original To argument so we can modify it
3462 Value *OrigTo = To;
3463 // General case, the type indexed by Idxs is a struct
3464 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3
Assuming 'i' is equal to 'e'
4
Loop condition is false. Execution continues on line 3483
3465 // Process each struct element recursively
3466 Idxs.push_back(i);
3467 Value *PrevTo = To;
3468 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3469 InsertBefore);
3470 Idxs.pop_back();
3471 if (!To) {
3472 // Couldn't find any inserted value for this index? Cleanup
3473 while (PrevTo != OrigTo) {
3474 InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3475 PrevTo = Del->getAggregateOperand();
3476 Del->eraseFromParent();
3477 }
3478 // Stop processing elements
3479 break;
3480 }
3481 }
3482 // If we successfully found a value for each of our subaggregates
3483 if (To)
5
Assuming 'To' is null
6
Taking false branch
3484 return To;
3485 }
3486 // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3487 // the struct's elements had a value that was inserted directly. In the latter
3488 // case, perhaps we can't determine each of the subelements individually, but
3489 // we might be able to find the complete struct somewhere.
3490
3491 // Find the value that is at that particular spot
3492 Value *V = FindInsertedValue(From, Idxs);
3493
3494 if (!V)
7
Assuming 'V' is non-null
8
Taking false branch
3495 return nullptr;
3496
3497 // Insert the value in the new (sub) aggregate
3498 return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
9
Passing null pointer value via 1st parameter 'Agg'
10
Calling 'InsertValueInst::Create'
3499 "tmp", InsertBefore);
3500}
3501
3502// This helper takes a nested struct and extracts a part of it (which is again a
3503// struct) into a new value. For example, given the struct:
3504// { a, { b, { c, d }, e } }
3505// and the indices "1, 1" this returns
3506// { c, d }.
3507//
3508// It does this by inserting an insertvalue for each element in the resulting
3509// struct, as opposed to just inserting a single struct. This will only work if
3510// each of the elements of the substruct are known (ie, inserted into From by an
3511// insertvalue instruction somewhere).
3512//
3513// All inserted insertvalue instructions are inserted before InsertBefore
3514static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
3515 Instruction *InsertBefore) {
3516 assert(InsertBefore && "Must have someplace to insert!")((InsertBefore && "Must have someplace to insert!") ?
static_cast<void> (0) : __assert_fail ("InsertBefore && \"Must have someplace to insert!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3516, __PRETTY_FUNCTION__))
;
3517 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3518 idx_range);
3519 Value *To = UndefValue::get(IndexedType);
3520 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3521 unsigned IdxSkip = Idxs.size();
3522
3523 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3524}
3525
3526/// Given an aggregate and a sequence of indices, see if the scalar value
3527/// indexed is already around as a register, for example if it was inserted
3528/// directly into the aggregate.
3529///
3530/// If InsertBefore is not null, this function will duplicate (modified)
3531/// insertvalues when a part of a nested struct is extracted.
3532Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
3533 Instruction *InsertBefore) {
3534 // Nothing to index? Just return V then (this is useful at the end of our
3535 // recursion).
3536 if (idx_range.empty())
3537 return V;
3538 // We have indices, so V should have an indexable type.
3539 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&(((V->getType()->isStructTy() || V->getType()->isArrayTy
()) && "Not looking at a struct or array?") ? static_cast
<void> (0) : __assert_fail ("(V->getType()->isStructTy() || V->getType()->isArrayTy()) && \"Not looking at a struct or array?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3540, __PRETTY_FUNCTION__))
3540 "Not looking at a struct or array?")(((V->getType()->isStructTy() || V->getType()->isArrayTy
()) && "Not looking at a struct or array?") ? static_cast
<void> (0) : __assert_fail ("(V->getType()->isStructTy() || V->getType()->isArrayTy()) && \"Not looking at a struct or array?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3540, __PRETTY_FUNCTION__))
;
3541 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&((ExtractValueInst::getIndexedType(V->getType(), idx_range
) && "Invalid indices for type?") ? static_cast<void
> (0) : __assert_fail ("ExtractValueInst::getIndexedType(V->getType(), idx_range) && \"Invalid indices for type?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3542, __PRETTY_FUNCTION__))
3542 "Invalid indices for type?")((ExtractValueInst::getIndexedType(V->getType(), idx_range
) && "Invalid indices for type?") ? static_cast<void
> (0) : __assert_fail ("ExtractValueInst::getIndexedType(V->getType(), idx_range) && \"Invalid indices for type?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3542, __PRETTY_FUNCTION__))
;
3543
3544 if (Constant *C = dyn_cast<Constant>(V)) {
3545 C = C->getAggregateElement(idx_range[0]);
3546 if (!C) return nullptr;
3547 return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3548 }
3549
3550 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3551 // Loop the indices for the insertvalue instruction in parallel with the
3552 // requested indices
3553 const unsigned *req_idx = idx_range.begin();
3554 for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3555 i != e; ++i, ++req_idx) {
3556 if (req_idx == idx_range.end()) {
3557 // We can't handle this without inserting insertvalues
3558 if (!InsertBefore)
3559 return nullptr;
3560
3561 // The requested index identifies a part of a nested aggregate. Handle
3562 // this specially. For example,
3563 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3564 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3565 // %C = extractvalue {i32, { i32, i32 } } %B, 1
3566 // This can be changed into
3567 // %A = insertvalue {i32, i32 } undef, i32 10, 0
3568 // %C = insertvalue {i32, i32 } %A, i32 11, 1
3569 // which allows the unused 0,0 element from the nested struct to be
3570 // removed.
3571 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3572 InsertBefore);
3573 }
3574
3575 // This insert value inserts something else than what we are looking for.
3576 // See if the (aggregate) value inserted into has the value we are
3577 // looking for, then.
3578 if (*req_idx != *i)
3579 return FindInsertedValue(I->getAggregateOperand(), idx_range,
3580 InsertBefore);
3581 }
3582 // If we end up here, the indices of the insertvalue match with those
3583 // requested (though possibly only partially). Now we recursively look at
3584 // the inserted value, passing any remaining indices.
3585 return FindInsertedValue(I->getInsertedValueOperand(),
3586 makeArrayRef(req_idx, idx_range.end()),
3587 InsertBefore);
3588 }
3589
3590 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3591 // If we're extracting a value from an aggregate that was extracted from
3592 // something else, we can extract from that something else directly instead.
3593 // However, we will need to chain I's indices with the requested indices.
3594
3595 // Calculate the number of indices required
3596 unsigned size = I->getNumIndices() + idx_range.size();
3597 // Allocate some space to put the new indices in
3598 SmallVector<unsigned, 5> Idxs;
3599 Idxs.reserve(size);
3600 // Add indices from the extract value instruction
3601 Idxs.append(I->idx_begin(), I->idx_end());
3602
3603 // Add requested indices
3604 Idxs.append(idx_range.begin(), idx_range.end());
3605
3606 assert(Idxs.size() == size((Idxs.size() == size && "Number of indices added not correct?"
) ? static_cast<void> (0) : __assert_fail ("Idxs.size() == size && \"Number of indices added not correct?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3607, __PRETTY_FUNCTION__))
3607 && "Number of indices added not correct?")((Idxs.size() == size && "Number of indices added not correct?"
) ? static_cast<void> (0) : __assert_fail ("Idxs.size() == size && \"Number of indices added not correct?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3607, __PRETTY_FUNCTION__))
;
3608
3609 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
3610 }
3611 // Otherwise, we don't know (such as, extracting from a function return value
3612 // or load instruction)
3613 return nullptr;
3614}
3615
3616bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP,
3617 unsigned CharSize) {
3618 // Make sure the GEP has exactly three arguments.
3619 if (GEP->getNumOperands() != 3)
3620 return false;
3621
3622 // Make sure the index-ee is a pointer to array of \p CharSize integers.
3623 // CharSize.
3624 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
3625 if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
3626 return false;
3627
3628 // Check to make sure that the first operand of the GEP is an integer and
3629 // has value 0 so that we are sure we're indexing into the initializer.
3630 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3631 if (!FirstIdx || !FirstIdx->isZero())
3632 return false;
3633
3634 return true;
3635}
3636
3637bool llvm::getConstantDataArrayInfo(const Value *V,
3638 ConstantDataArraySlice &Slice,
3639 unsigned ElementSize, uint64_t Offset) {
3640 assert(V)((V) ? static_cast<void> (0) : __assert_fail ("V", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3640, __PRETTY_FUNCTION__))
;
3641
3642 // Look through bitcast instructions and geps.
3643 V = V->stripPointerCasts();
3644
3645 // If the value is a GEP instruction or constant expression, treat it as an
3646 // offset.
3647 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3648 // The GEP operator should be based on a pointer to string constant, and is
3649 // indexing into the string constant.
3650 if (!isGEPBasedOnPointerToString(GEP, ElementSize))
3651 return false;
3652
3653 // If the second index isn't a ConstantInt, then this is a variable index
3654 // into the array. If this occurs, we can't say anything meaningful about
3655 // the string.
3656 uint64_t StartIdx = 0;
3657 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3658 StartIdx = CI->getZExtValue();
3659 else
3660 return false;
3661 return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3662 StartIdx + Offset);
3663 }
3664
3665 // The GEP instruction, constant or instruction, must reference a global
3666 // variable that is a constant and is initialized. The referenced constant
3667 // initializer is the array that we'll use for optimization.
3668 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3669 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3670 return false;
3671
3672 const ConstantDataArray *Array;
3673 ArrayType *ArrayTy;
3674 if (GV->getInitializer()->isNullValue()) {
3675 Type *GVTy = GV->getValueType();
3676 if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3677 // A zeroinitializer for the array; there is no ConstantDataArray.
3678 Array = nullptr;
3679 } else {
3680 const DataLayout &DL = GV->getParent()->getDataLayout();
3681 uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3682 uint64_t Length = SizeInBytes / (ElementSize / 8);
3683 if (Length <= Offset)
3684 return false;
3685
3686 Slice.Array = nullptr;
3687 Slice.Offset = 0;
3688 Slice.Length = Length - Offset;
3689 return true;
3690 }
3691 } else {
3692 // This must be a ConstantDataArray.
3693 Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3694 if (!Array)
3695 return false;
3696 ArrayTy = Array->getType();
3697 }
3698 if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3699 return false;
3700
3701 uint64_t NumElts = ArrayTy->getArrayNumElements();
3702 if (Offset > NumElts)
3703 return false;
3704
3705 Slice.Array = Array;
3706 Slice.Offset = Offset;
3707 Slice.Length = NumElts - Offset;
3708 return true;
3709}
3710
3711/// This function computes the length of a null-terminated C string pointed to
3712/// by V. If successful, it returns true and returns the string in Str.
3713/// If unsuccessful, it returns false.
3714bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
3715 uint64_t Offset, bool TrimAtNul) {
3716 ConstantDataArraySlice Slice;
3717 if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3718 return false;
3719
3720 if (Slice.Array == nullptr) {
3721 if (TrimAtNul) {
3722 Str = StringRef();
3723 return true;
3724 }
3725 if (Slice.Length == 1) {
3726 Str = StringRef("", 1);
3727 return true;
3728 }
3729 // We cannot instantiate a StringRef as we do not have an appropriate string
3730 // of 0s at hand.
3731 return false;
3732 }
3733
3734 // Start out with the entire array in the StringRef.
3735 Str = Slice.Array->getAsString();
3736 // Skip over 'offset' bytes.
3737 Str = Str.substr(Slice.Offset);
3738
3739 if (TrimAtNul) {
3740 // Trim off the \0 and anything after it. If the array is not nul
3741 // terminated, we just return the whole end of string. The client may know
3742 // some other way that the string is length-bound.
3743 Str = Str.substr(0, Str.find('\0'));
3744 }
3745 return true;
3746}
3747
3748// These next two are very similar to the above, but also look through PHI
3749// nodes.
3750// TODO: See if we can integrate these two together.
3751
3752/// If we can compute the length of the string pointed to by
3753/// the specified pointer, return 'len+1'. If we can't, return 0.
3754static uint64_t GetStringLengthH(const Value *V,
3755 SmallPtrSetImpl<const PHINode*> &PHIs,
3756 unsigned CharSize) {
3757 // Look through noop bitcast instructions.
3758 V = V->stripPointerCasts();
3759
3760 // If this is a PHI node, there are two cases: either we have already seen it
3761 // or we haven't.
3762 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3763 if (!PHIs.insert(PN).second)
3764 return ~0ULL; // already in the set.
3765
3766 // If it was new, see if all the input strings are the same length.
3767 uint64_t LenSoFar = ~0ULL;
3768 for (Value *IncValue : PN->incoming_values()) {
3769 uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3770 if (Len == 0) return 0; // Unknown length -> unknown.
3771
3772 if (Len == ~0ULL) continue;
3773
3774 if (Len != LenSoFar && LenSoFar != ~0ULL)
3775 return 0; // Disagree -> unknown.
3776 LenSoFar = Len;
3777 }
3778
3779 // Success, all agree.
3780 return LenSoFar;
3781 }
3782
3783 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3784 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3785 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3786 if (Len1 == 0) return 0;
3787 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3788 if (Len2 == 0) return 0;
3789 if (Len1 == ~0ULL) return Len2;
3790 if (Len2 == ~0ULL) return Len1;
3791 if (Len1 != Len2) return 0;
3792 return Len1;
3793 }
3794
3795 // Otherwise, see if we can read the string.
3796 ConstantDataArraySlice Slice;
3797 if (!getConstantDataArrayInfo(V, Slice, CharSize))
3798 return 0;
3799
3800 if (Slice.Array == nullptr)
3801 return 1;
3802
3803 // Search for nul characters
3804 unsigned NullIndex = 0;
3805 for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3806 if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3807 break;
3808 }
3809
3810 return NullIndex + 1;
3811}
3812
3813/// If we can compute the length of the string pointed to by
3814/// the specified pointer, return 'len+1'. If we can't, return 0.
3815uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3816 if (!V->getType()->isPointerTy())
3817 return 0;
3818
3819 SmallPtrSet<const PHINode*, 32> PHIs;
3820 uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3821 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3822 // an empty string as a length.
3823 return Len == ~0ULL ? 1 : Len;
3824}
3825
3826const Value *
3827llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call,
3828 bool MustPreserveNullness) {
3829 assert(Call &&((Call && "getArgumentAliasingToReturnedPointer only works on nonnull calls"
) ? static_cast<void> (0) : __assert_fail ("Call && \"getArgumentAliasingToReturnedPointer only works on nonnull calls\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3830, __PRETTY_FUNCTION__))
3830 "getArgumentAliasingToReturnedPointer only works on nonnull calls")((Call && "getArgumentAliasingToReturnedPointer only works on nonnull calls"
) ? static_cast<void> (0) : __assert_fail ("Call && \"getArgumentAliasingToReturnedPointer only works on nonnull calls\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3830, __PRETTY_FUNCTION__))
;
3831 if (const Value *RV = Call->getReturnedArgOperand())
3832 return RV;
3833 // This can be used only as a aliasing property.
3834 if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
3835 Call, MustPreserveNullness))
3836 return Call->getArgOperand(0);
3837 return nullptr;
3838}
3839
3840bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
3841 const CallBase *Call, bool MustPreserveNullness) {
3842 return Call->getIntrinsicID() == Intrinsic::launder_invariant_group ||
3843 Call->getIntrinsicID() == Intrinsic::strip_invariant_group ||
3844 Call->getIntrinsicID() == Intrinsic::aarch64_irg ||
3845 Call->getIntrinsicID() == Intrinsic::aarch64_tagp ||
3846 (!MustPreserveNullness &&
3847 Call->getIntrinsicID() == Intrinsic::ptrmask);
3848}
3849
3850/// \p PN defines a loop-variant pointer to an object. Check if the
3851/// previous iteration of the loop was referring to the same object as \p PN.
3852static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
3853 const LoopInfo *LI) {
3854 // Find the loop-defined value.
3855 Loop *L = LI->getLoopFor(PN->getParent());
3856 if (PN->getNumIncomingValues() != 2)
3857 return true;
3858
3859 // Find the value from previous iteration.
3860 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3861 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3862 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3863 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3864 return true;
3865
3866 // If a new pointer is loaded in the loop, the pointer references a different
3867 // object in every iteration. E.g.:
3868 // for (i)
3869 // int *p = a[i];
3870 // ...
3871 if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3872 if (!L->isLoopInvariant(Load->getPointerOperand()))
3873 return false;
3874 return true;
3875}
3876
3877Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3878 unsigned MaxLookup) {
3879 if (!V->getType()->isPointerTy())
3880 return V;
3881 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3882 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3883 V = GEP->getPointerOperand();
3884 } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3885 Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3886 V = cast<Operator>(V)->getOperand(0);
3887 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3888 if (GA->isInterposable())
3889 return V;
3890 V = GA->getAliasee();
3891 } else if (isa<AllocaInst>(V)) {
3892 // An alloca can't be further simplified.
3893 return V;
3894 } else {
3895 if (auto *Call = dyn_cast<CallBase>(V)) {
3896 // CaptureTracking can know about special capturing properties of some
3897 // intrinsics like launder.invariant.group, that can't be expressed with
3898 // the attributes, but have properties like returning aliasing pointer.
3899 // Because some analysis may assume that nocaptured pointer is not
3900 // returned from some special intrinsic (because function would have to
3901 // be marked with returns attribute), it is crucial to use this function
3902 // because it should be in sync with CaptureTracking. Not using it may
3903 // cause weird miscompilations where 2 aliasing pointers are assumed to
3904 // noalias.
3905 if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) {
3906 V = RP;
3907 continue;
3908 }
3909 }
3910
3911 // See if InstructionSimplify knows any relevant tricks.
3912 if (Instruction *I = dyn_cast<Instruction>(V))
3913 // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3914 if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3915 V = Simplified;
3916 continue;
3917 }
3918
3919 return V;
3920 }
3921 assert(V->getType()->isPointerTy() && "Unexpected operand type!")((V->getType()->isPointerTy() && "Unexpected operand type!"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isPointerTy() && \"Unexpected operand type!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3921, __PRETTY_FUNCTION__))
;
3922 }
3923 return V;
3924}
3925
3926void llvm::GetUnderlyingObjects(const Value *V,
3927 SmallVectorImpl<const Value *> &Objects,
3928 const DataLayout &DL, LoopInfo *LI,
3929 unsigned MaxLookup) {
3930 SmallPtrSet<const Value *, 4> Visited;
3931 SmallVector<const Value *, 4> Worklist;
3932 Worklist.push_back(V);
3933 do {
3934 const Value *P = Worklist.pop_back_val();
3935 P = GetUnderlyingObject(P, DL, MaxLookup);
3936
3937 if (!Visited.insert(P).second)
3938 continue;
3939
3940 if (auto *SI = dyn_cast<SelectInst>(P)) {
3941 Worklist.push_back(SI->getTrueValue());
3942 Worklist.push_back(SI->getFalseValue());
3943 continue;
3944 }
3945
3946 if (auto *PN = dyn_cast<PHINode>(P)) {
3947 // If this PHI changes the underlying object in every iteration of the
3948 // loop, don't look through it. Consider:
3949 // int **A;
3950 // for (i) {
3951 // Prev = Curr; // Prev = PHI (Prev_0, Curr)
3952 // Curr = A[i];
3953 // *Prev, *Curr;
3954 //
3955 // Prev is tracking Curr one iteration behind so they refer to different
3956 // underlying objects.
3957 if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3958 isSameUnderlyingObjectInLoop(PN, LI))
3959 for (Value *IncValue : PN->incoming_values())
3960 Worklist.push_back(IncValue);
3961 continue;
3962 }
3963
3964 Objects.push_back(P);
3965 } while (!Worklist.empty());
3966}
3967
3968/// This is the function that does the work of looking through basic
3969/// ptrtoint+arithmetic+inttoptr sequences.
3970static const Value *getUnderlyingObjectFromInt(const Value *V) {
3971 do {
3972 if (const Operator *U = dyn_cast<Operator>(V)) {
3973 // If we find a ptrtoint, we can transfer control back to the
3974 // regular getUnderlyingObjectFromInt.
3975 if (U->getOpcode() == Instruction::PtrToInt)
3976 return U->getOperand(0);
3977 // If we find an add of a constant, a multiplied value, or a phi, it's
3978 // likely that the other operand will lead us to the base
3979 // object. We don't have to worry about the case where the
3980 // object address is somehow being computed by the multiply,
3981 // because our callers only care when the result is an
3982 // identifiable object.
3983 if (U->getOpcode() != Instruction::Add ||
3984 (!isa<ConstantInt>(U->getOperand(1)) &&
3985 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
3986 !isa<PHINode>(U->getOperand(1))))
3987 return V;
3988 V = U->getOperand(0);
3989 } else {
3990 return V;
3991 }
3992 assert(V->getType()->isIntegerTy() && "Unexpected operand type!")((V->getType()->isIntegerTy() && "Unexpected operand type!"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntegerTy() && \"Unexpected operand type!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 3992, __PRETTY_FUNCTION__))
;
3993 } while (true);
3994}
3995
3996/// This is a wrapper around GetUnderlyingObjects and adds support for basic
3997/// ptrtoint+arithmetic+inttoptr sequences.
3998/// It returns false if unidentified object is found in GetUnderlyingObjects.
3999bool llvm::getUnderlyingObjectsForCodeGen(const Value *V,
4000 SmallVectorImpl<Value *> &Objects,
4001 const DataLayout &DL) {
4002 SmallPtrSet<const Value *, 16> Visited;
4003 SmallVector<const Value *, 4> Working(1, V);
4004 do {
4005 V = Working.pop_back_val();
4006
4007 SmallVector<const Value *, 4> Objs;
4008 GetUnderlyingObjects(V, Objs, DL);
4009
4010 for (const Value *V : Objs) {
4011 if (!Visited.insert(V).second)
4012 continue;
4013 if (Operator::getOpcode(V) == Instruction::IntToPtr) {
4014 const Value *O =
4015 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
4016 if (O->getType()->isPointerTy()) {
4017 Working.push_back(O);
4018 continue;
4019 }
4020 }
4021 // If GetUnderlyingObjects fails to find an identifiable object,
4022 // getUnderlyingObjectsForCodeGen also fails for safety.
4023 if (!isIdentifiedObject(V)) {
4024 Objects.clear();
4025 return false;
4026 }
4027 Objects.push_back(const_cast<Value *>(V));
4028 }
4029 } while (!Working.empty());
4030 return true;
4031}
4032
4033/// Return true if the only users of this pointer are lifetime markers.
4034bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
4035 for (const User *U : V->users()) {
4036 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
4037 if (!II) return false;
4038
4039 if (!II->isLifetimeStartOrEnd())
4040 return false;
4041 }
4042 return true;
4043}
4044
4045bool llvm::mustSuppressSpeculation(const LoadInst &LI) {
4046 if (!LI.isUnordered())
4047 return true;
4048 const Function &F = *LI.getFunction();
4049 // Speculative load may create a race that did not exist in the source.
4050 return F.hasFnAttribute(Attribute::SanitizeThread) ||
4051 // Speculative load may load data from dirty regions.
4052 F.hasFnAttribute(Attribute::SanitizeAddress) ||
4053 F.hasFnAttribute(Attribute::SanitizeHWAddress);
4054}
4055
4056
4057bool llvm::isSafeToSpeculativelyExecute(const Value *V,
4058 const Instruction *CtxI,
4059 const DominatorTree *DT) {
4060 const Operator *Inst = dyn_cast<Operator>(V);
4061 if (!Inst)
4062 return false;
4063
4064 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
4065 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
4066 if (C->canTrap())
4067 return false;
4068
4069 switch (Inst->getOpcode()) {
4070 default:
4071 return true;
4072 case Instruction::UDiv:
4073 case Instruction::URem: {
4074 // x / y is undefined if y == 0.
4075 const APInt *V;
4076 if (match(Inst->getOperand(1), m_APInt(V)))
4077 return *V != 0;
4078 return false;
4079 }
4080 case Instruction::SDiv:
4081 case Instruction::SRem: {
4082 // x / y is undefined if y == 0 or x == INT_MIN and y == -1
4083 const APInt *Numerator, *Denominator;
4084 if (!match(Inst->getOperand(1), m_APInt(Denominator)))
4085 return false;
4086 // We cannot hoist this division if the denominator is 0.
4087 if (*Denominator == 0)
4088 return false;
4089 // It's safe to hoist if the denominator is not 0 or -1.
4090 if (*Denominator != -1)
4091 return true;
4092 // At this point we know that the denominator is -1. It is safe to hoist as
4093 // long we know that the numerator is not INT_MIN.
4094 if (match(Inst->getOperand(0), m_APInt(Numerator)))
4095 return !Numerator->isMinSignedValue();
4096 // The numerator *might* be MinSignedValue.
4097 return false;
4098 }
4099 case Instruction::Load: {
4100 const LoadInst *LI = cast<LoadInst>(Inst);
4101 if (mustSuppressSpeculation(*LI))
4102 return false;
4103 const DataLayout &DL = LI->getModule()->getDataLayout();
4104 return isDereferenceableAndAlignedPointer(
4105 LI->getPointerOperand(), LI->getType(), MaybeAlign(LI->getAlignment()),
4106 DL, CtxI, DT);
4107 }
4108 case Instruction::Call: {
4109 auto *CI = cast<const CallInst>(Inst);
4110 const Function *Callee = CI->getCalledFunction();
4111
4112 // The called function could have undefined behavior or side-effects, even
4113 // if marked readnone nounwind.
4114 return Callee && Callee->isSpeculatable();
4115 }
4116 case Instruction::VAArg:
4117 case Instruction::Alloca:
4118 case Instruction::Invoke:
4119 case Instruction::CallBr:
4120 case Instruction::PHI:
4121 case Instruction::Store:
4122 case Instruction::Ret:
4123 case Instruction::Br:
4124 case Instruction::IndirectBr:
4125 case Instruction::Switch:
4126 case Instruction::Unreachable:
4127 case Instruction::Fence:
4128 case Instruction::AtomicRMW:
4129 case Instruction::AtomicCmpXchg:
4130 case Instruction::LandingPad:
4131 case Instruction::Resume:
4132 case Instruction::CatchSwitch:
4133 case Instruction::CatchPad:
4134 case Instruction::CatchRet:
4135 case Instruction::CleanupPad:
4136 case Instruction::CleanupRet:
4137 return false; // Misc instructions which have effects
4138 }
4139}
4140
4141bool llvm::mayBeMemoryDependent(const Instruction &I) {
4142 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
4143}
4144
4145/// Convert ConstantRange OverflowResult into ValueTracking OverflowResult.
4146static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) {
4147 switch (OR) {
4148 case ConstantRange::OverflowResult::MayOverflow:
4149 return OverflowResult::MayOverflow;
4150 case ConstantRange::OverflowResult::AlwaysOverflowsLow:
4151 return OverflowResult::AlwaysOverflowsLow;
4152 case ConstantRange::OverflowResult::AlwaysOverflowsHigh:
4153 return OverflowResult::AlwaysOverflowsHigh;
4154 case ConstantRange::OverflowResult::NeverOverflows:
4155 return OverflowResult::NeverOverflows;
4156 }
4157 llvm_unreachable("Unknown OverflowResult")::llvm::llvm_unreachable_internal("Unknown OverflowResult", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4157)
;
4158}
4159
4160/// Combine constant ranges from computeConstantRange() and computeKnownBits().
4161static ConstantRange computeConstantRangeIncludingKnownBits(
4162 const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth,
4163 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4164 OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) {
4165 KnownBits Known = computeKnownBits(
4166 V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo);
4167 ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned);
4168 ConstantRange CR2 = computeConstantRange(V, UseInstrInfo);
4169 ConstantRange::PreferredRangeType RangeType =
4170 ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned;
4171 return CR1.intersectWith(CR2, RangeType);
4172}
4173
4174OverflowResult llvm::computeOverflowForUnsignedMul(
4175 const Value *LHS, const Value *RHS, const DataLayout &DL,
4176 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4177 bool UseInstrInfo) {
4178 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4179 nullptr, UseInstrInfo);
4180 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4181 nullptr, UseInstrInfo);
4182 ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false);
4183 ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false);
4184 return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange));
4185}
4186
4187OverflowResult
4188llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS,
4189 const DataLayout &DL, AssumptionCache *AC,
4190 const Instruction *CxtI,
4191 const DominatorTree *DT, bool UseInstrInfo) {
4192 // Multiplying n * m significant bits yields a result of n + m significant
4193 // bits. If the total number of significant bits does not exceed the
4194 // result bit width (minus 1), there is no overflow.
4195 // This means if we have enough leading sign bits in the operands
4196 // we can guarantee that the result does not overflow.
4197 // Ref: "Hacker's Delight" by Henry Warren
4198 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
4199
4200 // Note that underestimating the number of sign bits gives a more
4201 // conservative answer.
4202 unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
4203 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
4204
4205 // First handle the easy case: if we have enough sign bits there's
4206 // definitely no overflow.
4207 if (SignBits > BitWidth + 1)
4208 return OverflowResult::NeverOverflows;
4209
4210 // There are two ambiguous cases where there can be no overflow:
4211 // SignBits == BitWidth + 1 and
4212 // SignBits == BitWidth
4213 // The second case is difficult to check, therefore we only handle the
4214 // first case.
4215 if (SignBits == BitWidth + 1) {
4216 // It overflows only when both arguments are negative and the true
4217 // product is exactly the minimum negative number.
4218 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
4219 // For simplicity we just check if at least one side is not negative.
4220 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT,
4221 nullptr, UseInstrInfo);
4222 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT,
4223 nullptr, UseInstrInfo);
4224 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
4225 return OverflowResult::NeverOverflows;
4226 }
4227 return OverflowResult::MayOverflow;
4228}
4229
4230OverflowResult llvm::computeOverflowForUnsignedAdd(
4231 const Value *LHS, const Value *RHS, const DataLayout &DL,
4232 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT,
4233 bool UseInstrInfo) {
4234 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
4235 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
4236 nullptr, UseInstrInfo);
4237 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
4238 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT,
4239 nullptr, UseInstrInfo);
4240 return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange));
4241}
4242
4243static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
4244 const Value *RHS,
4245 const AddOperator *Add,
4246 const DataLayout &DL,
4247 AssumptionCache *AC,
4248 const Instruction *CxtI,
4249 const DominatorTree *DT) {
4250 if (Add && Add->hasNoSignedWrap()) {
4251 return OverflowResult::NeverOverflows;
4252 }
4253
4254 // If LHS and RHS each have at least two sign bits, the addition will look
4255 // like
4256 //
4257 // XX..... +
4258 // YY.....
4259 //
4260 // If the carry into the most significant position is 0, X and Y can't both
4261 // be 1 and therefore the carry out of the addition is also 0.
4262 //
4263 // If the carry into the most significant position is 1, X and Y can't both
4264 // be 0 and therefore the carry out of the addition is also 1.
4265 //
4266 // Since the carry into the most significant position is always equal to
4267 // the carry out of the addition, there is no signed overflow.
4268 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4269 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4270 return OverflowResult::NeverOverflows;
4271
4272 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
4273 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4274 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
4275 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4276 OverflowResult OR =
4277 mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange));
4278 if (OR != OverflowResult::MayOverflow)
4279 return OR;
4280
4281 // The remaining code needs Add to be available. Early returns if not so.
4282 if (!Add)
4283 return OverflowResult::MayOverflow;
4284
4285 // If the sign of Add is the same as at least one of the operands, this add
4286 // CANNOT overflow. If this can be determined from the known bits of the
4287 // operands the above signedAddMayOverflow() check will have already done so.
4288 // The only other way to improve on the known bits is from an assumption, so
4289 // call computeKnownBitsFromAssume() directly.
4290 bool LHSOrRHSKnownNonNegative =
4291 (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative());
4292 bool LHSOrRHSKnownNegative =
4293 (LHSRange.isAllNegative() || RHSRange.isAllNegative());
4294 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
4295 KnownBits AddKnown(LHSRange.getBitWidth());
4296 computeKnownBitsFromAssume(
4297 Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true));
4298 if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
4299 (AddKnown.isNegative() && LHSOrRHSKnownNegative))
4300 return OverflowResult::NeverOverflows;
4301 }
4302
4303 return OverflowResult::MayOverflow;
4304}
4305
4306OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS,
4307 const Value *RHS,
4308 const DataLayout &DL,
4309 AssumptionCache *AC,
4310 const Instruction *CxtI,
4311 const DominatorTree *DT) {
4312 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
4313 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);
4314 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
4315 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT);
4316 return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange));
4317}
4318
4319OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS,
4320 const Value *RHS,
4321 const DataLayout &DL,
4322 AssumptionCache *AC,
4323 const Instruction *CxtI,
4324 const DominatorTree *DT) {
4325 // If LHS and RHS each have at least two sign bits, the subtraction
4326 // cannot overflow.
4327 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
4328 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
4329 return OverflowResult::NeverOverflows;
4330
4331 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits(
4332 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4333 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits(
4334 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT);
4335 return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange));
4336}
4337
4338bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO,
4339 const DominatorTree &DT) {
4340 SmallVector<const BranchInst *, 2> GuardingBranches;
4341 SmallVector<const ExtractValueInst *, 2> Results;
4342
4343 for (const User *U : WO->users()) {
4344 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
4345 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type")((EVI->getNumIndices() == 1 && "Obvious from CI's type"
) ? static_cast<void> (0) : __assert_fail ("EVI->getNumIndices() == 1 && \"Obvious from CI's type\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4345, __PRETTY_FUNCTION__))
;
4346
4347 if (EVI->getIndices()[0] == 0)
4348 Results.push_back(EVI);
4349 else {
4350 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type")((EVI->getIndices()[0] == 1 && "Obvious from CI's type"
) ? static_cast<void> (0) : __assert_fail ("EVI->getIndices()[0] == 1 && \"Obvious from CI's type\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4350, __PRETTY_FUNCTION__))
;
4351
4352 for (const auto *U : EVI->users())
4353 if (const auto *B = dyn_cast<BranchInst>(U)) {
4354 assert(B->isConditional() && "How else is it using an i1?")((B->isConditional() && "How else is it using an i1?"
) ? static_cast<void> (0) : __assert_fail ("B->isConditional() && \"How else is it using an i1?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4354, __PRETTY_FUNCTION__))
;
4355 GuardingBranches.push_back(B);
4356 }
4357 }
4358 } else {
4359 // We are using the aggregate directly in a way we don't want to analyze
4360 // here (storing it to a global, say).
4361 return false;
4362 }
4363 }
4364
4365 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4366 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4367 if (!NoWrapEdge.isSingleEdge())
4368 return false;
4369
4370 // Check if all users of the add are provably no-wrap.
4371 for (const auto *Result : Results) {
4372 // If the extractvalue itself is not executed on overflow, the we don't
4373 // need to check each use separately, since domination is transitive.
4374 if (DT.dominates(NoWrapEdge, Result->getParent()))
4375 continue;
4376
4377 for (auto &RU : Result->uses())
4378 if (!DT.dominates(NoWrapEdge, RU))
4379 return false;
4380 }
4381
4382 return true;
4383 };
4384
4385 return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4386}
4387
4388bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V) {
4389 // If the value is a freeze instruction, then it can never
4390 // be undef or poison.
4391 if (isa<FreezeInst>(V))
4392 return true;
4393 // TODO: Some instructions are guaranteed to return neither undef
4394 // nor poison if their arguments are not poison/undef.
4395
4396 // TODO: Deal with other Constant subclasses.
4397 if (isa<ConstantInt>(V) || isa<GlobalVariable>(V))
4398 return true;
4399
4400 return false;
4401}
4402
4403OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
4404 const DataLayout &DL,
4405 AssumptionCache *AC,
4406 const Instruction *CxtI,
4407 const DominatorTree *DT) {
4408 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
4409 Add, DL, AC, CxtI, DT);
4410}
4411
4412OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
4413 const Value *RHS,
4414 const DataLayout &DL,
4415 AssumptionCache *AC,
4416 const Instruction *CxtI,
4417 const DominatorTree *DT) {
4418 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
4419}
4420
4421bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
4422 // Note: An atomic operation isn't guaranteed to return in a reasonable amount
4423 // of time because it's possible for another thread to interfere with it for an
4424 // arbitrary length of time, but programs aren't allowed to rely on that.
4425
4426 // If there is no successor, then execution can't transfer to it.
4427 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
4428 return !CRI->unwindsToCaller();
4429 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
4430 return !CatchSwitch->unwindsToCaller();
4431 if (isa<ResumeInst>(I))
4432 return false;
4433 if (isa<ReturnInst>(I))
4434 return false;
4435 if (isa<UnreachableInst>(I))
4436 return false;
4437
4438 // Calls can throw, or contain an infinite loop, or kill the process.
4439 if (auto CS = ImmutableCallSite(I)) {
4440 // Call sites that throw have implicit non-local control flow.
4441 if (!CS.doesNotThrow())
4442 return false;
4443
4444 // A function which doens't throw and has "willreturn" attribute will
4445 // always return.
4446 if (CS.hasFnAttr(Attribute::WillReturn))
4447 return true;
4448
4449 // Non-throwing call sites can loop infinitely, call exit/pthread_exit
4450 // etc. and thus not return. However, LLVM already assumes that
4451 //
4452 // - Thread exiting actions are modeled as writes to memory invisible to
4453 // the program.
4454 //
4455 // - Loops that don't have side effects (side effects are volatile/atomic
4456 // stores and IO) always terminate (see http://llvm.org/PR965).
4457 // Furthermore IO itself is also modeled as writes to memory invisible to
4458 // the program.
4459 //
4460 // We rely on those assumptions here, and use the memory effects of the call
4461 // target as a proxy for checking that it always returns.
4462
4463 // FIXME: This isn't aggressive enough; a call which only writes to a global
4464 // is guaranteed to return.
4465 return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory();
4466 }
4467
4468 // Other instructions return normally.
4469 return true;
4470}
4471
4472bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) {
4473 // TODO: This is slightly conservative for invoke instruction since exiting
4474 // via an exception *is* normal control for them.
4475 for (auto I = BB->begin(), E = BB->end(); I != E; ++I)
4476 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
4477 return false;
4478 return true;
4479}
4480
4481bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
4482 const Loop *L) {
4483 // The loop header is guaranteed to be executed for every iteration.
4484 //
4485 // FIXME: Relax this constraint to cover all basic blocks that are
4486 // guaranteed to be executed at every iteration.
4487 if (I->getParent() != L->getHeader()) return false;
4488
4489 for (const Instruction &LI : *L->getHeader()) {
4490 if (&LI == I) return true;
4491 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
4492 }
4493 llvm_unreachable("Instruction not contained in its own parent basic block.")::llvm::llvm_unreachable_internal("Instruction not contained in its own parent basic block."
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4493)
;
4494}
4495
4496bool llvm::propagatesFullPoison(const Instruction *I) {
4497 // TODO: This should include all instructions apart from phis, selects and
4498 // call-like instructions.
4499 switch (I->getOpcode()) {
4500 case Instruction::Add:
4501 case Instruction::Sub:
4502 case Instruction::Xor:
4503 case Instruction::Trunc:
4504 case Instruction::BitCast:
4505 case Instruction::AddrSpaceCast:
4506 case Instruction::Mul:
4507 case Instruction::Shl:
4508 case Instruction::GetElementPtr:
4509 // These operations all propagate poison unconditionally. Note that poison
4510 // is not any particular value, so xor or subtraction of poison with
4511 // itself still yields poison, not zero.
4512 return true;
4513
4514 case Instruction::AShr:
4515 case Instruction::SExt:
4516 // For these operations, one bit of the input is replicated across
4517 // multiple output bits. A replicated poison bit is still poison.
4518 return true;
4519
4520 case Instruction::ICmp:
4521 // Comparing poison with any value yields poison. This is why, for
4522 // instance, x s< (x +nsw 1) can be folded to true.
4523 return true;
4524
4525 default:
4526 return false;
4527 }
4528}
4529
4530const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
4531 switch (I->getOpcode()) {
4532 case Instruction::Store:
4533 return cast<StoreInst>(I)->getPointerOperand();
4534
4535 case Instruction::Load:
4536 return cast<LoadInst>(I)->getPointerOperand();
4537
4538 case Instruction::AtomicCmpXchg:
4539 return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
4540
4541 case Instruction::AtomicRMW:
4542 return cast<AtomicRMWInst>(I)->getPointerOperand();
4543
4544 case Instruction::UDiv:
4545 case Instruction::SDiv:
4546 case Instruction::URem:
4547 case Instruction::SRem:
4548 return I->getOperand(1);
4549
4550 default:
4551 // Note: It's really tempting to think that a conditional branch or
4552 // switch should be listed here, but that's incorrect. It's not
4553 // branching off of poison which is UB, it is executing a side effecting
4554 // instruction which follows the branch.
4555 return nullptr;
4556 }
4557}
4558
4559bool llvm::mustTriggerUB(const Instruction *I,
4560 const SmallSet<const Value *, 16>& KnownPoison) {
4561 auto *NotPoison = getGuaranteedNonFullPoisonOp(I);
4562 return (NotPoison && KnownPoison.count(NotPoison));
4563}
4564
4565
4566bool llvm::programUndefinedIfFullPoison(const Instruction *PoisonI) {
4567 // We currently only look for uses of poison values within the same basic
4568 // block, as that makes it easier to guarantee that the uses will be
4569 // executed given that PoisonI is executed.
4570 //
4571 // FIXME: Expand this to consider uses beyond the same basic block. To do
4572 // this, look out for the distinction between post-dominance and strong
4573 // post-dominance.
4574 const BasicBlock *BB = PoisonI->getParent();
4575
4576 // Set of instructions that we have proved will yield poison if PoisonI
4577 // does.
4578 SmallSet<const Value *, 16> YieldsPoison;
4579 SmallSet<const BasicBlock *, 4> Visited;
4580 YieldsPoison.insert(PoisonI);
4581 Visited.insert(PoisonI->getParent());
4582
4583 BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
4584
4585 unsigned Iter = 0;
4586 while (Iter++ < MaxDepth) {
4587 for (auto &I : make_range(Begin, End)) {
4588 if (&I != PoisonI) {
4589 if (mustTriggerUB(&I, YieldsPoison))
4590 return true;
4591 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
4592 return false;
4593 }
4594
4595 // Mark poison that propagates from I through uses of I.
4596 if (YieldsPoison.count(&I)) {
4597 for (const User *User : I.users()) {
4598 const Instruction *UserI = cast<Instruction>(User);
4599 if (propagatesFullPoison(UserI))
4600 YieldsPoison.insert(User);
4601 }
4602 }
4603 }
4604
4605 if (auto *NextBB = BB->getSingleSuccessor()) {
4606 if (Visited.insert(NextBB).second) {
4607 BB = NextBB;
4608 Begin = BB->getFirstNonPHI()->getIterator();
4609 End = BB->end();
4610 continue;
4611 }
4612 }
4613
4614 break;
4615 }
4616 return false;
4617}
4618
4619static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
4620 if (FMF.noNaNs())
4621 return true;
4622
4623 if (auto *C = dyn_cast<ConstantFP>(V))
4624 return !C->isNaN();
4625
4626 if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4627 if (!C->getElementType()->isFloatingPointTy())
4628 return false;
4629 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4630 if (C->getElementAsAPFloat(I).isNaN())
4631 return false;
4632 }
4633 return true;
4634 }
4635
4636 return false;
4637}
4638
4639static bool isKnownNonZero(const Value *V) {
4640 if (auto *C = dyn_cast<ConstantFP>(V))
4641 return !C->isZero();
4642
4643 if (auto *C = dyn_cast<ConstantDataVector>(V)) {
4644 if (!C->getElementType()->isFloatingPointTy())
4645 return false;
4646 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) {
4647 if (C->getElementAsAPFloat(I).isZero())
4648 return false;
4649 }
4650 return true;
4651 }
4652
4653 return false;
4654}
4655
4656/// Match clamp pattern for float types without care about NaNs or signed zeros.
4657/// Given non-min/max outer cmp/select from the clamp pattern this
4658/// function recognizes if it can be substitued by a "canonical" min/max
4659/// pattern.
4660static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred,
4661 Value *CmpLHS, Value *CmpRHS,
4662 Value *TrueVal, Value *FalseVal,
4663 Value *&LHS, Value *&RHS) {
4664 // Try to match
4665 // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
4666 // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
4667 // and return description of the outer Max/Min.
4668
4669 // First, check if select has inverse order:
4670 if (CmpRHS == FalseVal) {
4671 std::swap(TrueVal, FalseVal);
4672 Pred = CmpInst::getInversePredicate(Pred);
4673 }
4674
4675 // Assume success now. If there's no match, callers should not use these anyway.
4676 LHS = TrueVal;
4677 RHS = FalseVal;
4678
4679 const APFloat *FC1;
4680 if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
4681 return {SPF_UNKNOWN, SPNB_NA, false};
4682
4683 const APFloat *FC2;
4684 switch (Pred) {
4685 case CmpInst::FCMP_OLT:
4686 case CmpInst::FCMP_OLE:
4687 case CmpInst::FCMP_ULT:
4688 case CmpInst::FCMP_ULE:
4689 if (match(FalseVal,
4690 m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
4691 m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4692 FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
4693 return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
4694 break;
4695 case CmpInst::FCMP_OGT:
4696 case CmpInst::FCMP_OGE:
4697 case CmpInst::FCMP_UGT:
4698 case CmpInst::FCMP_UGE:
4699 if (match(FalseVal,
4700 m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
4701 m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4702 FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
4703 return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
4704 break;
4705 default:
4706 break;
4707 }
4708
4709 return {SPF_UNKNOWN, SPNB_NA, false};
4710}
4711
4712/// Recognize variations of:
4713/// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
4714static SelectPatternResult matchClamp(CmpInst::Predicate Pred,
4715 Value *CmpLHS, Value *CmpRHS,
4716 Value *TrueVal, Value *FalseVal) {
4717 // Swap the select operands and predicate to match the patterns below.
4718 if (CmpRHS != TrueVal) {
4719 Pred = ICmpInst::getSwappedPredicate(Pred);
4720 std::swap(TrueVal, FalseVal);
4721 }
4722 const APInt *C1;
4723 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
4724 const APInt *C2;
4725 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
4726 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4727 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
4728 return {SPF_SMAX, SPNB_NA, false};
4729
4730 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
4731 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4732 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
4733 return {SPF_SMIN, SPNB_NA, false};
4734
4735 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
4736 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4737 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
4738 return {SPF_UMAX, SPNB_NA, false};
4739
4740 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
4741 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4742 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
4743 return {SPF_UMIN, SPNB_NA, false};
4744 }
4745 return {SPF_UNKNOWN, SPNB_NA, false};
4746}
4747
4748/// Recognize variations of:
4749/// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
4750static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred,
4751 Value *CmpLHS, Value *CmpRHS,
4752 Value *TVal, Value *FVal,
4753 unsigned Depth) {
4754 // TODO: Allow FP min/max with nnan/nsz.
4755 assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison")((CmpInst::isIntPredicate(Pred) && "Expected integer comparison"
) ? static_cast<void> (0) : __assert_fail ("CmpInst::isIntPredicate(Pred) && \"Expected integer comparison\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4755, __PRETTY_FUNCTION__))
;
4756
4757 Value *A = nullptr, *B = nullptr;
4758 SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
4759 if (!SelectPatternResult::isMinOrMax(L.Flavor))
4760 return {SPF_UNKNOWN, SPNB_NA, false};
4761
4762 Value *C = nullptr, *D = nullptr;
4763 SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
4764 if (L.Flavor != R.Flavor)
4765 return {SPF_UNKNOWN, SPNB_NA, false};
4766
4767 // We have something like: x Pred y ? min(a, b) : min(c, d).
4768 // Try to match the compare to the min/max operations of the select operands.
4769 // First, make sure we have the right compare predicate.
4770 switch (L.Flavor) {
4771 case SPF_SMIN:
4772 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
4773 Pred = ICmpInst::getSwappedPredicate(Pred);
4774 std::swap(CmpLHS, CmpRHS);
4775 }
4776 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
4777 break;
4778 return {SPF_UNKNOWN, SPNB_NA, false};
4779 case SPF_SMAX:
4780 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
4781 Pred = ICmpInst::getSwappedPredicate(Pred);
4782 std::swap(CmpLHS, CmpRHS);
4783 }
4784 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
4785 break;
4786 return {SPF_UNKNOWN, SPNB_NA, false};
4787 case SPF_UMIN:
4788 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
4789 Pred = ICmpInst::getSwappedPredicate(Pred);
4790 std::swap(CmpLHS, CmpRHS);
4791 }
4792 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
4793 break;
4794 return {SPF_UNKNOWN, SPNB_NA, false};
4795 case SPF_UMAX:
4796 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
4797 Pred = ICmpInst::getSwappedPredicate(Pred);
4798 std::swap(CmpLHS, CmpRHS);
4799 }
4800 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
4801 break;
4802 return {SPF_UNKNOWN, SPNB_NA, false};
4803 default:
4804 return {SPF_UNKNOWN, SPNB_NA, false};
4805 }
4806
4807 // If there is a common operand in the already matched min/max and the other
4808 // min/max operands match the compare operands (either directly or inverted),
4809 // then this is min/max of the same flavor.
4810
4811 // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4812 // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4813 if (D == B) {
4814 if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4815 match(A, m_Not(m_Specific(CmpRHS)))))
4816 return {L.Flavor, SPNB_NA, false};
4817 }
4818 // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4819 // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4820 if (C == B) {
4821 if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4822 match(A, m_Not(m_Specific(CmpRHS)))))
4823 return {L.Flavor, SPNB_NA, false};
4824 }
4825 // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4826 // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4827 if (D == A) {
4828 if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4829 match(B, m_Not(m_Specific(CmpRHS)))))
4830 return {L.Flavor, SPNB_NA, false};
4831 }
4832 // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4833 // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4834 if (C == A) {
4835 if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4836 match(B, m_Not(m_Specific(CmpRHS)))))
4837 return {L.Flavor, SPNB_NA, false};
4838 }
4839
4840 return {SPF_UNKNOWN, SPNB_NA, false};
4841}
4842
4843/// Match non-obvious integer minimum and maximum sequences.
4844static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
4845 Value *CmpLHS, Value *CmpRHS,
4846 Value *TrueVal, Value *FalseVal,
4847 Value *&LHS, Value *&RHS,
4848 unsigned Depth) {
4849 // Assume success. If there's no match, callers should not use these anyway.
4850 LHS = TrueVal;
4851 RHS = FalseVal;
4852
4853 SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
4854 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
4855 return SPR;
4856
4857 SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
4858 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
4859 return SPR;
4860
4861 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
4862 return {SPF_UNKNOWN, SPNB_NA, false};
4863
4864 // Z = X -nsw Y
4865 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
4866 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
4867 if (match(TrueVal, m_Zero()) &&
4868 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4869 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4870
4871 // Z = X -nsw Y
4872 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
4873 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
4874 if (match(FalseVal, m_Zero()) &&
4875 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4876 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4877
4878 const APInt *C1;
4879 if (!match(CmpRHS, m_APInt(C1)))
4880 return {SPF_UNKNOWN, SPNB_NA, false};
4881
4882 // An unsigned min/max can be written with a signed compare.
4883 const APInt *C2;
4884 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
4885 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
4886 // Is the sign bit set?
4887 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
4888 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
4889 if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
4890 C2->isMaxSignedValue())
4891 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4892
4893 // Is the sign bit clear?
4894 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
4895 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
4896 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
4897 C2->isMinSignedValue())
4898 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4899 }
4900
4901 // Look through 'not' ops to find disguised signed min/max.
4902 // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4903 // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4904 if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4905 match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4906 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4907
4908 // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4909 // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4910 if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4911 match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4912 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4913
4914 return {SPF_UNKNOWN, SPNB_NA, false};
4915}
4916
4917bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
4918 assert(X && Y && "Invalid operand")((X && Y && "Invalid operand") ? static_cast<
void> (0) : __assert_fail ("X && Y && \"Invalid operand\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 4918, __PRETTY_FUNCTION__))
;
4919
4920 // X = sub (0, Y) || X = sub nsw (0, Y)
4921 if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
4922 (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
4923 return true;
4924
4925 // Y = sub (0, X) || Y = sub nsw (0, X)
4926 if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
4927 (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
4928 return true;
4929
4930 // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
4931 Value *A, *B;
4932 return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
4933 match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
4934 (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
4935 match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
4936}
4937
4938static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
4939 FastMathFlags FMF,
4940 Value *CmpLHS, Value *CmpRHS,
4941 Value *TrueVal, Value *FalseVal,
4942 Value *&LHS, Value *&RHS,
4943 unsigned Depth) {
4944 if (CmpInst::isFPPredicate(Pred)) {
4945 // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one
4946 // 0.0 operand, set the compare's 0.0 operands to that same value for the
4947 // purpose of identifying min/max. Disregard vector constants with undefined
4948 // elements because those can not be back-propagated for analysis.
4949 Value *OutputZeroVal = nullptr;
4950 if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) &&
4951 !cast<Constant>(TrueVal)->containsUndefElement())
4952 OutputZeroVal = TrueVal;
4953 else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) &&
4954 !cast<Constant>(FalseVal)->containsUndefElement())
4955 OutputZeroVal = FalseVal;
4956
4957 if (OutputZeroVal) {
4958 if (match(CmpLHS, m_AnyZeroFP()))
4959 CmpLHS = OutputZeroVal;
4960 if (match(CmpRHS, m_AnyZeroFP()))
4961 CmpRHS = OutputZeroVal;
4962 }
4963 }
4964
4965 LHS = CmpLHS;
4966 RHS = CmpRHS;
4967
4968 // Signed zero may return inconsistent results between implementations.
4969 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4970 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4971 // Therefore, we behave conservatively and only proceed if at least one of the
4972 // operands is known to not be zero or if we don't care about signed zero.
4973 switch (Pred) {
4974 default: break;
4975 // FIXME: Include OGT/OLT/UGT/ULT.
4976 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
4977 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
4978 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4979 !isKnownNonZero(CmpRHS))
4980 return {SPF_UNKNOWN, SPNB_NA, false};
4981 }
4982
4983 SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4984 bool Ordered = false;
4985
4986 // When given one NaN and one non-NaN input:
4987 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4988 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4989 // ordered comparison fails), which could be NaN or non-NaN.
4990 // so here we discover exactly what NaN behavior is required/accepted.
4991 if (CmpInst::isFPPredicate(Pred)) {
4992 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4993 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4994
4995 if (LHSSafe && RHSSafe) {
4996 // Both operands are known non-NaN.
4997 NaNBehavior = SPNB_RETURNS_ANY;
4998 } else if (CmpInst::isOrdered(Pred)) {
4999 // An ordered comparison will return false when given a NaN, so it
5000 // returns the RHS.
5001 Ordered = true;
5002 if (LHSSafe)
5003 // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
5004 NaNBehavior = SPNB_RETURNS_NAN;
5005 else if (RHSSafe)
5006 NaNBehavior = SPNB_RETURNS_OTHER;
5007 else
5008 // Completely unsafe.
5009 return {SPF_UNKNOWN, SPNB_NA, false};
5010 } else {
5011 Ordered = false;
5012 // An unordered comparison will return true when given a NaN, so it
5013 // returns the LHS.
5014 if (LHSSafe)
5015 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
5016 NaNBehavior = SPNB_RETURNS_OTHER;
5017 else if (RHSSafe)
5018 NaNBehavior = SPNB_RETURNS_NAN;
5019 else
5020 // Completely unsafe.
5021 return {SPF_UNKNOWN, SPNB_NA, false};
5022 }
5023 }
5024
5025 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
5026 std::swap(CmpLHS, CmpRHS);
5027 Pred = CmpInst::getSwappedPredicate(Pred);
5028 if (NaNBehavior == SPNB_RETURNS_NAN)
5029 NaNBehavior = SPNB_RETURNS_OTHER;
5030 else if (NaNBehavior == SPNB_RETURNS_OTHER)
5031 NaNBehavior = SPNB_RETURNS_NAN;
5032 Ordered = !Ordered;
5033 }
5034
5035 // ([if]cmp X, Y) ? X : Y
5036 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
5037 switch (Pred) {
5038 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
5039 case ICmpInst::ICMP_UGT:
5040 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
5041 case ICmpInst::ICMP_SGT:
5042 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
5043 case ICmpInst::ICMP_ULT:
5044 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
5045 case ICmpInst::ICMP_SLT:
5046 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
5047 case FCmpInst::FCMP_UGT:
5048 case FCmpInst::FCMP_UGE:
5049 case FCmpInst::FCMP_OGT:
5050 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
5051 case FCmpInst::FCMP_ULT:
5052 case FCmpInst::FCMP_ULE:
5053 case FCmpInst::FCMP_OLT:
5054 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
5055 }
5056 }
5057
5058 if (isKnownNegation(TrueVal, FalseVal)) {
5059 // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
5060 // match against either LHS or sext(LHS).
5061 auto MaybeSExtCmpLHS =
5062 m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
5063 auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
5064 auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
5065 if (match(TrueVal, MaybeSExtCmpLHS)) {
5066 // Set the return values. If the compare uses the negated value (-X >s 0),
5067 // swap the return values because the negated value is always 'RHS'.
5068 LHS = TrueVal;
5069 RHS = FalseVal;
5070 if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
5071 std::swap(LHS, RHS);
5072
5073 // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
5074 // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
5075 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
5076 return {SPF_ABS, SPNB_NA, false};
5077
5078 // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X)
5079 if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne))
5080 return {SPF_ABS, SPNB_NA, false};
5081
5082 // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
5083 // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
5084 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
5085 return {SPF_NABS, SPNB_NA, false};
5086 }
5087 else if (match(FalseVal, MaybeSExtCmpLHS)) {
5088 // Set the return values. If the compare uses the negated value (-X >s 0),
5089 // swap the return values because the negated value is always 'RHS'.
5090 LHS = FalseVal;
5091 RHS = TrueVal;
5092 if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
5093 std::swap(LHS, RHS);
5094
5095 // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
5096 // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
5097 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
5098 return {SPF_NABS, SPNB_NA, false};
5099
5100 // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
5101 // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
5102 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
5103 return {SPF_ABS, SPNB_NA, false};
5104 }
5105 }
5106
5107 if (CmpInst::isIntPredicate(Pred))
5108 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
5109
5110 // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
5111 // may return either -0.0 or 0.0, so fcmp/select pair has stricter
5112 // semantics than minNum. Be conservative in such case.
5113 if (NaNBehavior != SPNB_RETURNS_ANY ||
5114 (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
5115 !isKnownNonZero(CmpRHS)))
5116 return {SPF_UNKNOWN, SPNB_NA, false};
5117
5118 return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
5119}
5120
5121/// Helps to match a select pattern in case of a type mismatch.
5122///
5123/// The function processes the case when type of true and false values of a
5124/// select instruction differs from type of the cmp instruction operands because
5125/// of a cast instruction. The function checks if it is legal to move the cast
5126/// operation after "select". If yes, it returns the new second value of
5127/// "select" (with the assumption that cast is moved):
5128/// 1. As operand of cast instruction when both values of "select" are same cast
5129/// instructions.
5130/// 2. As restored constant (by applying reverse cast operation) when the first
5131/// value of the "select" is a cast operation and the second value is a
5132/// constant.
5133/// NOTE: We return only the new second value because the first value could be
5134/// accessed as operand of cast instruction.
5135static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
5136 Instruction::CastOps *CastOp) {
5137 auto *Cast1 = dyn_cast<CastInst>(V1);
5138 if (!Cast1)
5139 return nullptr;
5140
5141 *CastOp = Cast1->getOpcode();
5142 Type *SrcTy = Cast1->getSrcTy();
5143 if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
5144 // If V1 and V2 are both the same cast from the same type, look through V1.
5145 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
5146 return Cast2->getOperand(0);
5147 return nullptr;
5148 }
5149
5150 auto *C = dyn_cast<Constant>(V2);
5151 if (!C)
5152 return nullptr;
5153
5154 Constant *CastedTo = nullptr;
5155 switch (*CastOp) {
5156 case Instruction::ZExt:
5157 if (CmpI->isUnsigned())
5158 CastedTo = ConstantExpr::getTrunc(C, SrcTy);
5159 break;
5160 case Instruction::SExt:
5161 if (CmpI->isSigned())
5162 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
5163 break;
5164 case Instruction::Trunc:
5165 Constant *CmpConst;
5166 if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
5167 CmpConst->getType() == SrcTy) {
5168 // Here we have the following case:
5169 //
5170 // %cond = cmp iN %x, CmpConst
5171 // %tr = trunc iN %x to iK
5172 // %narrowsel = select i1 %cond, iK %t, iK C
5173 //
5174 // We can always move trunc after select operation:
5175 //
5176 // %cond = cmp iN %x, CmpConst
5177 // %widesel = select i1 %cond, iN %x, iN CmpConst
5178 // %tr = trunc iN %widesel to iK
5179 //
5180 // Note that C could be extended in any way because we don't care about
5181 // upper bits after truncation. It can't be abs pattern, because it would
5182 // look like:
5183 //
5184 // select i1 %cond, x, -x.
5185 //
5186 // So only min/max pattern could be matched. Such match requires widened C
5187 // == CmpConst. That is why set widened C = CmpConst, condition trunc
5188 // CmpConst == C is checked below.
5189 CastedTo = CmpConst;
5190 } else {
5191 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
5192 }
5193 break;
5194 case Instruction::FPTrunc:
5195 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
5196 break;
5197 case Instruction::FPExt:
5198 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
5199 break;
5200 case Instruction::FPToUI:
5201 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
5202 break;
5203 case Instruction::FPToSI:
5204 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
5205 break;
5206 case Instruction::UIToFP:
5207 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
5208 break;
5209 case Instruction::SIToFP:
5210 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
5211 break;
5212 default:
5213 break;
5214 }
5215
5216 if (!CastedTo)
5217 return nullptr;
5218
5219 // Make sure the cast doesn't lose any information.
5220 Constant *CastedBack =
5221 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
5222 if (CastedBack != C)
5223 return nullptr;
5224
5225 return CastedTo;
5226}
5227
5228SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
5229 Instruction::CastOps *CastOp,
5230 unsigned Depth) {
5231 if (Depth >= MaxDepth)
5232 return {SPF_UNKNOWN, SPNB_NA, false};
5233
5234 SelectInst *SI = dyn_cast<SelectInst>(V);
5235 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
5236
5237 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
5238 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
5239
5240 Value *TrueVal = SI->getTrueValue();
5241 Value *FalseVal = SI->getFalseValue();
5242
5243 return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS,
5244 CastOp, Depth);
5245}
5246
5247SelectPatternResult llvm::matchDecomposedSelectPattern(
5248 CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS,
5249 Instruction::CastOps *CastOp, unsigned Depth) {
5250 CmpInst::Predicate Pred = CmpI->getPredicate();
5251 Value *CmpLHS = CmpI->getOperand(0);
5252 Value *CmpRHS = CmpI->getOperand(1);
5253 FastMathFlags FMF;
5254 if (isa<FPMathOperator>(CmpI))
5255 FMF = CmpI->getFastMathFlags();
5256
5257 // Bail out early.
5258 if (CmpI->isEquality())
5259 return {SPF_UNKNOWN, SPNB_NA, false};
5260
5261 // Deal with type mismatches.
5262 if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
5263 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
5264 // If this is a potential fmin/fmax with a cast to integer, then ignore
5265 // -0.0 because there is no corresponding integer value.
5266 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5267 FMF.setNoSignedZeros();
5268 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5269 cast<CastInst>(TrueVal)->getOperand(0), C,
5270 LHS, RHS, Depth);
5271 }
5272 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
5273 // If this is a potential fmin/fmax with a cast to integer, then ignore
5274 // -0.0 because there is no corresponding integer value.
5275 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
5276 FMF.setNoSignedZeros();
5277 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
5278 C, cast<CastInst>(FalseVal)->getOperand(0),
5279 LHS, RHS, Depth);
5280 }
5281 }
5282 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
5283 LHS, RHS, Depth);
5284}
5285
5286CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) {
5287 if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
5288 if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
5289 if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
5290 if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
5291 if (SPF == SPF_FMINNUM)
5292 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
5293 if (SPF == SPF_FMAXNUM)
5294 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
5295 llvm_unreachable("unhandled!")::llvm::llvm_unreachable_internal("unhandled!", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5295)
;
5296}
5297
5298SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) {
5299 if (SPF == SPF_SMIN) return SPF_SMAX;
5300 if (SPF == SPF_UMIN) return SPF_UMAX;
5301 if (SPF == SPF_SMAX) return SPF_SMIN;
5302 if (SPF == SPF_UMAX) return SPF_UMIN;
5303 llvm_unreachable("unhandled!")::llvm::llvm_unreachable_internal("unhandled!", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5303)
;
5304}
5305
5306CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) {
5307 return getMinMaxPred(getInverseMinMaxFlavor(SPF));
5308}
5309
5310/// Return true if "icmp Pred LHS RHS" is always true.
5311static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS,
5312 const Value *RHS, const DataLayout &DL,
5313 unsigned Depth) {
5314 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!")((!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!"
) ? static_cast<void> (0) : __assert_fail ("!LHS->getType()->isVectorTy() && \"TODO: extend to handle vectors!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5314, __PRETTY_FUNCTION__))
;
5315 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
5316 return true;
5317
5318 switch (Pred) {
5319 default:
5320 return false;
5321
5322 case CmpInst::ICMP_SLE: {
5323 const APInt *C;
5324
5325 // LHS s<= LHS +_{nsw} C if C >= 0
5326 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
5327 return !C->isNegative();
5328 return false;
5329 }
5330
5331 case CmpInst::ICMP_ULE: {
5332 const APInt *C;
5333
5334 // LHS u<= LHS +_{nuw} C for any C
5335 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
5336 return true;
5337
5338 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
5339 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
5340 const Value *&X,
5341 const APInt *&CA, const APInt *&CB) {
5342 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
5343 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
5344 return true;
5345
5346 // If X & C == 0 then (X | C) == X +_{nuw} C
5347 if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
5348 match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
5349 KnownBits Known(CA->getBitWidth());
5350 computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr,
5351 /*CxtI*/ nullptr, /*DT*/ nullptr);
5352 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
5353 return true;
5354 }
5355
5356 return false;
5357 };
5358
5359 const Value *X;
5360 const APInt *CLHS, *CRHS;
5361 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
5362 return CLHS->ule(*CRHS);
5363
5364 return false;
5365 }
5366 }
5367}
5368
5369/// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
5370/// ALHS ARHS" is true. Otherwise, return None.
5371static Optional<bool>
5372isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
5373 const Value *ARHS, const Value *BLHS, const Value *BRHS,
5374 const DataLayout &DL, unsigned Depth) {
5375 switch (Pred) {
5376 default:
5377 return None;
5378
5379 case CmpInst::ICMP_SLT:
5380 case CmpInst::ICMP_SLE:
5381 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) &&
5382 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth))
5383 return true;
5384 return None;
5385
5386 case CmpInst::ICMP_ULT:
5387 case CmpInst::ICMP_ULE:
5388 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) &&
5389 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth))
5390 return true;
5391 return None;
5392 }
5393}
5394
5395/// Return true if the operands of the two compares match. IsSwappedOps is true
5396/// when the operands match, but are swapped.
5397static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
5398 const Value *BLHS, const Value *BRHS,
5399 bool &IsSwappedOps) {
5400
5401 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
5402 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
5403 return IsMatchingOps || IsSwappedOps;
5404}
5405
5406/// Return true if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is true.
5407/// Return false if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is false.
5408/// Otherwise, return None if we can't infer anything.
5409static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
5410 CmpInst::Predicate BPred,
5411 bool AreSwappedOps) {
5412 // Canonicalize the predicate as if the operands were not commuted.
5413 if (AreSwappedOps)
5414 BPred = ICmpInst::getSwappedPredicate(BPred);
5415
5416 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
5417 return true;
5418 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
5419 return false;
5420
5421 return None;
5422}
5423
5424/// Return true if "icmp APred X, C1" implies "icmp BPred X, C2" is true.
5425/// Return false if "icmp APred X, C1" implies "icmp BPred X, C2" is false.
5426/// Otherwise, return None if we can't infer anything.
5427static Optional<bool>
5428isImpliedCondMatchingImmOperands(CmpInst::Predicate APred,
5429 const ConstantInt *C1,
5430 CmpInst::Predicate BPred,
5431 const ConstantInt *C2) {
5432 ConstantRange DomCR =
5433 ConstantRange::makeExactICmpRegion(APred, C1->getValue());
5434 ConstantRange CR =
5435 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
5436 ConstantRange Intersection = DomCR.intersectWith(CR);
5437 ConstantRange Difference = DomCR.difference(CR);
5438 if (Intersection.isEmptySet())
5439 return false;
5440 if (Difference.isEmptySet())
5441 return true;
5442 return None;
5443}
5444
5445/// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
5446/// false. Otherwise, return None if we can't infer anything.
5447static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
5448 const ICmpInst *RHS,
5449 const DataLayout &DL, bool LHSIsTrue,
5450 unsigned Depth) {
5451 Value *ALHS = LHS->getOperand(0);
5452 Value *ARHS = LHS->getOperand(1);
5453 // The rest of the logic assumes the LHS condition is true. If that's not the
5454 // case, invert the predicate to make it so.
5455 ICmpInst::Predicate APred =
5456 LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate();
5457
5458 Value *BLHS = RHS->getOperand(0);
5459 Value *BRHS = RHS->getOperand(1);
5460 ICmpInst::Predicate BPred = RHS->getPredicate();
5461
5462 // Can we infer anything when the two compares have matching operands?
5463 bool AreSwappedOps;
5464 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, AreSwappedOps)) {
5465 if (Optional<bool> Implication = isImpliedCondMatchingOperands(
5466 APred, BPred, AreSwappedOps))
5467 return Implication;
5468 // No amount of additional analysis will infer the second condition, so
5469 // early exit.
5470 return None;
5471 }
5472
5473 // Can we infer anything when the LHS operands match and the RHS operands are
5474 // constants (not necessarily matching)?
5475 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
5476 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
5477 APred, cast<ConstantInt>(ARHS), BPred, cast<ConstantInt>(BRHS)))
5478 return Implication;
5479 // No amount of additional analysis will infer the second condition, so
5480 // early exit.
5481 return None;
5482 }
5483
5484 if (APred == BPred)
5485 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth);
5486 return None;
5487}
5488
5489/// Return true if LHS implies RHS is true. Return false if LHS implies RHS is
5490/// false. Otherwise, return None if we can't infer anything. We expect the
5491/// RHS to be an icmp and the LHS to be an 'and' or an 'or' instruction.
5492static Optional<bool> isImpliedCondAndOr(const BinaryOperator *LHS,
5493 const ICmpInst *RHS,
5494 const DataLayout &DL, bool LHSIsTrue,
5495 unsigned Depth) {
5496 // The LHS must be an 'or' or an 'and' instruction.
5497 assert((LHS->getOpcode() == Instruction::And ||(((LHS->getOpcode() == Instruction::And || LHS->getOpcode
() == Instruction::Or) && "Expected LHS to be 'and' or 'or'."
) ? static_cast<void> (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or) && \"Expected LHS to be 'and' or 'or'.\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5499, __PRETTY_FUNCTION__))
5498 LHS->getOpcode() == Instruction::Or) &&(((LHS->getOpcode() == Instruction::And || LHS->getOpcode
() == Instruction::Or) && "Expected LHS to be 'and' or 'or'."
) ? static_cast<void> (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or) && \"Expected LHS to be 'and' or 'or'.\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5499, __PRETTY_FUNCTION__))
5499 "Expected LHS to be 'and' or 'or'.")(((LHS->getOpcode() == Instruction::And || LHS->getOpcode
() == Instruction::Or) && "Expected LHS to be 'and' or 'or'."
) ? static_cast<void> (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or) && \"Expected LHS to be 'and' or 'or'.\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5499, __PRETTY_FUNCTION__))
;
5500
5501 assert(Depth <= MaxDepth && "Hit recursion limit")((Depth <= MaxDepth && "Hit recursion limit") ? static_cast
<void> (0) : __assert_fail ("Depth <= MaxDepth && \"Hit recursion limit\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5501, __PRETTY_FUNCTION__))
;
5502
5503 // If the result of an 'or' is false, then we know both legs of the 'or' are
5504 // false. Similarly, if the result of an 'and' is true, then we know both
5505 // legs of the 'and' are true.
5506 Value *ALHS, *ARHS;
5507 if ((!LHSIsTrue && match(LHS, m_Or(m_Value(ALHS), m_Value(ARHS)))) ||
5508 (LHSIsTrue && match(LHS, m_And(m_Value(ALHS), m_Value(ARHS))))) {
5509 // FIXME: Make this non-recursion.
5510 if (Optional<bool> Implication =
5511 isImpliedCondition(ALHS, RHS, DL, LHSIsTrue, Depth + 1))
5512 return Implication;
5513 if (Optional<bool> Implication =
5514 isImpliedCondition(ARHS, RHS, DL, LHSIsTrue, Depth + 1))
5515 return Implication;
5516 return None;
5517 }
5518 return None;
5519}
5520
5521Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
5522 const DataLayout &DL, bool LHSIsTrue,
5523 unsigned Depth) {
5524 // Bail out when we hit the limit.
5525 if (Depth == MaxDepth)
5526 return None;
5527
5528 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for
5529 // example.
5530 if (LHS->getType() != RHS->getType())
5531 return None;
5532
5533 Type *OpTy = LHS->getType();
5534 assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!")((OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!"
) ? static_cast<void> (0) : __assert_fail ("OpTy->isIntOrIntVectorTy(1) && \"Expected integer type only!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5534, __PRETTY_FUNCTION__))
;
5535
5536 // LHS ==> RHS by definition
5537 if (LHS == RHS)
5538 return LHSIsTrue;
5539
5540 // FIXME: Extending the code below to handle vectors.
5541 if (OpTy->isVectorTy())
5542 return None;
5543
5544 assert(OpTy->isIntegerTy(1) && "implied by above")((OpTy->isIntegerTy(1) && "implied by above") ? static_cast
<void> (0) : __assert_fail ("OpTy->isIntegerTy(1) && \"implied by above\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5544, __PRETTY_FUNCTION__))
;
5545
5546 // Both LHS and RHS are icmps.
5547 const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS);
5548 const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS);
5549 if (LHSCmp && RHSCmp)
5550 return isImpliedCondICmps(LHSCmp, RHSCmp, DL, LHSIsTrue, Depth);
5551
5552 // The LHS should be an 'or' or an 'and' instruction. We expect the RHS to be
5553 // an icmp. FIXME: Add support for and/or on the RHS.
5554 const BinaryOperator *LHSBO = dyn_cast<BinaryOperator>(LHS);
5555 if (LHSBO && RHSCmp) {
5556 if ((LHSBO->getOpcode() == Instruction::And ||
5557 LHSBO->getOpcode() == Instruction::Or))
5558 return isImpliedCondAndOr(LHSBO, RHSCmp, DL, LHSIsTrue, Depth);
5559 }
5560 return None;
5561}
5562
5563Optional<bool> llvm::isImpliedByDomCondition(const Value *Cond,
5564 const Instruction *ContextI,
5565 const DataLayout &DL) {
5566 assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool")((Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool"
) ? static_cast<void> (0) : __assert_fail ("Cond->getType()->isIntOrIntVectorTy(1) && \"Condition must be bool\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5566, __PRETTY_FUNCTION__))
;
5567 if (!ContextI || !ContextI->getParent())
5568 return None;
5569
5570 // TODO: This is a poor/cheap way to determine dominance. Should we use a
5571 // dominator tree (eg, from a SimplifyQuery) instead?
5572 const BasicBlock *ContextBB = ContextI->getParent();
5573 const BasicBlock *PredBB = ContextBB->getSinglePredecessor();
5574 if (!PredBB)
5575 return None;
5576
5577 // We need a conditional branch in the predecessor.
5578 Value *PredCond;
5579 BasicBlock *TrueBB, *FalseBB;
5580 if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB)))
5581 return None;
5582
5583 // The branch should get simplified. Don't bother simplifying this condition.
5584 if (TrueBB == FalseBB)
5585 return None;
5586
5587 assert((TrueBB == ContextBB || FalseBB == ContextBB) &&(((TrueBB == ContextBB || FalseBB == ContextBB) && "Predecessor block does not point to successor?"
) ? static_cast<void> (0) : __assert_fail ("(TrueBB == ContextBB || FalseBB == ContextBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5588, __PRETTY_FUNCTION__))
5588 "Predecessor block does not point to successor?")(((TrueBB == ContextBB || FalseBB == ContextBB) && "Predecessor block does not point to successor?"
) ? static_cast<void> (0) : __assert_fail ("(TrueBB == ContextBB || FalseBB == ContextBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5588, __PRETTY_FUNCTION__))
;
5589
5590 // Is this condition implied by the predecessor condition?
5591 bool CondIsTrue = TrueBB == ContextBB;
5592 return isImpliedCondition(PredCond, Cond, DL, CondIsTrue);
5593}
5594
5595static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower,
5596 APInt &Upper, const InstrInfoQuery &IIQ) {
5597 unsigned Width = Lower.getBitWidth();
5598 const APInt *C;
5599 switch (BO.getOpcode()) {
5600 case Instruction::Add:
5601 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
5602 // FIXME: If we have both nuw and nsw, we should reduce the range further.
5603 if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
5604 // 'add nuw x, C' produces [C, UINT_MAX].
5605 Lower = *C;
5606 } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
5607 if (C->isNegative()) {
5608 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
5609 Lower = APInt::getSignedMinValue(Width);
5610 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
5611 } else {
5612 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
5613 Lower = APInt::getSignedMinValue(Width) + *C;
5614 Upper = APInt::getSignedMaxValue(Width) + 1;
5615 }
5616 }
5617 }
5618 break;
5619
5620 case Instruction::And:
5621 if (match(BO.getOperand(1), m_APInt(C)))
5622 // 'and x, C' produces [0, C].
5623 Upper = *C + 1;
5624 break;
5625
5626 case Instruction::Or:
5627 if (match(BO.getOperand(1), m_APInt(C)))
5628 // 'or x, C' produces [C, UINT_MAX].
5629 Lower = *C;
5630 break;
5631
5632 case Instruction::AShr:
5633 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
5634 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
5635 Lower = APInt::getSignedMinValue(Width).ashr(*C);
5636 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
5637 } else if (match(BO.getOperand(0), m_APInt(C))) {
5638 unsigned ShiftAmount = Width - 1;
5639 if (!C->isNullValue() && IIQ.isExact(&BO))
5640 ShiftAmount = C->countTrailingZeros();
5641 if (C->isNegative()) {
5642 // 'ashr C, x' produces [C, C >> (Width-1)]
5643 Lower = *C;
5644 Upper = C->ashr(ShiftAmount) + 1;
5645 } else {
5646 // 'ashr C, x' produces [C >> (Width-1), C]
5647 Lower = C->ashr(ShiftAmount);
5648 Upper = *C + 1;
5649 }
5650 }
5651 break;
5652
5653 case Instruction::LShr:
5654 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
5655 // 'lshr x, C' produces [0, UINT_MAX >> C].
5656 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
5657 } else if (match(BO.getOperand(0), m_APInt(C))) {
5658 // 'lshr C, x' produces [C >> (Width-1), C].
5659 unsigned ShiftAmount = Width - 1;
5660 if (!C->isNullValue() && IIQ.isExact(&BO))
5661 ShiftAmount = C->countTrailingZeros();
5662 Lower = C->lshr(ShiftAmount);
5663 Upper = *C + 1;
5664 }
5665 break;
5666
5667 case Instruction::Shl:
5668 if (match(BO.getOperand(0), m_APInt(C))) {
5669 if (IIQ.hasNoUnsignedWrap(&BO)) {
5670 // 'shl nuw C, x' produces [C, C << CLZ(C)]
5671 Lower = *C;
5672 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
5673 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
5674 if (C->isNegative()) {
5675 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
5676 unsigned ShiftAmount = C->countLeadingOnes() - 1;
5677 Lower = C->shl(ShiftAmount);
5678 Upper = *C + 1;
5679 } else {
5680 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
5681 unsigned ShiftAmount = C->countLeadingZeros() - 1;
5682 Lower = *C;
5683 Upper = C->shl(ShiftAmount) + 1;
5684 }
5685 }
5686 }
5687 break;
5688
5689 case Instruction::SDiv:
5690 if (match(BO.getOperand(1), m_APInt(C))) {
5691 APInt IntMin = APInt::getSignedMinValue(Width);
5692 APInt IntMax = APInt::getSignedMaxValue(Width);
5693 if (C->isAllOnesValue()) {
5694 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
5695 // where C != -1 and C != 0 and C != 1
5696 Lower = IntMin + 1;
5697 Upper = IntMax + 1;
5698 } else if (C->countLeadingZeros() < Width - 1) {
5699 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
5700 // where C != -1 and C != 0 and C != 1
5701 Lower = IntMin.sdiv(*C);
5702 Upper = IntMax.sdiv(*C);
5703 if (Lower.sgt(Upper))
5704 std::swap(Lower, Upper);
5705 Upper = Upper + 1;
5706 assert(Upper != Lower && "Upper part of range has wrapped!")((Upper != Lower && "Upper part of range has wrapped!"
) ? static_cast<void> (0) : __assert_fail ("Upper != Lower && \"Upper part of range has wrapped!\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5706, __PRETTY_FUNCTION__))
;
5707 }
5708 } else if (match(BO.getOperand(0), m_APInt(C))) {
5709 if (C->isMinSignedValue()) {
5710 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
5711 Lower = *C;
5712 Upper = Lower.lshr(1) + 1;
5713 } else {
5714 // 'sdiv C, x' produces [-|C|, |C|].
5715 Upper = C->abs() + 1;
5716 Lower = (-Upper) + 1;
5717 }
5718 }
5719 break;
5720
5721 case Instruction::UDiv:
5722 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
5723 // 'udiv x, C' produces [0, UINT_MAX / C].
5724 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
5725 } else if (match(BO.getOperand(0), m_APInt(C))) {
5726 // 'udiv C, x' produces [0, C].
5727 Upper = *C + 1;
5728 }
5729 break;
5730
5731 case Instruction::SRem:
5732 if (match(BO.getOperand(1), m_APInt(C))) {
5733 // 'srem x, C' produces (-|C|, |C|).
5734 Upper = C->abs();
5735 Lower = (-Upper) + 1;
5736 }
5737 break;
5738
5739 case Instruction::URem:
5740 if (match(BO.getOperand(1), m_APInt(C)))
5741 // 'urem x, C' produces [0, C).
5742 Upper = *C;
5743 break;
5744
5745 default:
5746 break;
5747 }
5748}
5749
5750static void setLimitsForIntrinsic(const IntrinsicInst &II, APInt &Lower,
5751 APInt &Upper) {
5752 unsigned Width = Lower.getBitWidth();
5753 const APInt *C;
5754 switch (II.getIntrinsicID()) {
5755 case Intrinsic::uadd_sat:
5756 // uadd.sat(x, C) produces [C, UINT_MAX].
5757 if (match(II.getOperand(0), m_APInt(C)) ||
5758 match(II.getOperand(1), m_APInt(C)))
5759 Lower = *C;
5760 break;
5761 case Intrinsic::sadd_sat:
5762 if (match(II.getOperand(0), m_APInt(C)) ||
5763 match(II.getOperand(1), m_APInt(C))) {
5764 if (C->isNegative()) {
5765 // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)].
5766 Lower = APInt::getSignedMinValue(Width);
5767 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
5768 } else {
5769 // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX].
5770 Lower = APInt::getSignedMinValue(Width) + *C;
5771 Upper = APInt::getSignedMaxValue(Width) + 1;
5772 }
5773 }
5774 break;
5775 case Intrinsic::usub_sat:
5776 // usub.sat(C, x) produces [0, C].
5777 if (match(II.getOperand(0), m_APInt(C)))
5778 Upper = *C + 1;
5779 // usub.sat(x, C) produces [0, UINT_MAX - C].
5780 else if (match(II.getOperand(1), m_APInt(C)))
5781 Upper = APInt::getMaxValue(Width) - *C + 1;
5782 break;
5783 case Intrinsic::ssub_sat:
5784 if (match(II.getOperand(0), m_APInt(C))) {
5785 if (C->isNegative()) {
5786 // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)].
5787 Lower = APInt::getSignedMinValue(Width);
5788 Upper = *C - APInt::getSignedMinValue(Width) + 1;
5789 } else {
5790 // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX].
5791 Lower = *C - APInt::getSignedMaxValue(Width);
5792 Upper = APInt::getSignedMaxValue(Width) + 1;
5793 }
5794 } else if (match(II.getOperand(1), m_APInt(C))) {
5795 if (C->isNegative()) {
5796 // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]:
5797 Lower = APInt::getSignedMinValue(Width) - *C;
5798 Upper = APInt::getSignedMaxValue(Width) + 1;
5799 } else {
5800 // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C].
5801 Lower = APInt::getSignedMinValue(Width);
5802 Upper = APInt::getSignedMaxValue(Width) - *C + 1;
5803 }
5804 }
5805 break;
5806 default:
5807 break;
5808 }
5809}
5810
5811static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower,
5812 APInt &Upper, const InstrInfoQuery &IIQ) {
5813 const Value *LHS = nullptr, *RHS = nullptr;
5814 SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS);
5815 if (R.Flavor == SPF_UNKNOWN)
5816 return;
5817
5818 unsigned BitWidth = SI.getType()->getScalarSizeInBits();
5819
5820 if (R.Flavor == SelectPatternFlavor::SPF_ABS) {
5821 // If the negation part of the abs (in RHS) has the NSW flag,
5822 // then the result of abs(X) is [0..SIGNED_MAX],
5823 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN.
5824 Lower = APInt::getNullValue(BitWidth);
5825 if (match(RHS, m_Neg(m_Specific(LHS))) &&
5826 IIQ.hasNoSignedWrap(cast<Instruction>(RHS)))
5827 Upper = APInt::getSignedMaxValue(BitWidth) + 1;
5828 else
5829 Upper = APInt::getSignedMinValue(BitWidth) + 1;
5830 return;
5831 }
5832
5833 if (R.Flavor == SelectPatternFlavor::SPF_NABS) {
5834 // The result of -abs(X) is <= 0.
5835 Lower = APInt::getSignedMinValue(BitWidth);
5836 Upper = APInt(BitWidth, 1);
5837 return;
5838 }
5839
5840 const APInt *C;
5841 if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C)))
5842 return;
5843
5844 switch (R.Flavor) {
5845 case SPF_UMIN:
5846 Upper = *C + 1;
5847 break;
5848 case SPF_UMAX:
5849 Lower = *C;
5850 break;
5851 case SPF_SMIN:
5852 Lower = APInt::getSignedMinValue(BitWidth);
5853 Upper = *C + 1;
5854 break;
5855 case SPF_SMAX:
5856 Lower = *C;
5857 Upper = APInt::getSignedMaxValue(BitWidth) + 1;
5858 break;
5859 default:
5860 break;
5861 }
5862}
5863
5864ConstantRange llvm::computeConstantRange(const Value *V, bool UseInstrInfo) {
5865 assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction")((V->getType()->isIntOrIntVectorTy() && "Expected integer instruction"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && \"Expected integer instruction\""
, "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/Analysis/ValueTracking.cpp"
, 5865, __PRETTY_FUN