Bug Summary

File:llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
Warning:line 3568, column 7
Called C++ object pointer is uninitialized

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name InstCombineCompares.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/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-12/lib/clang/12.0.0/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-12~++20201026111116+d3205bbca3e/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-10-27-053609-25509-1 -x c++ /build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

1//===- InstCombineCompares.cpp --------------------------------------------===//
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 implements the visitICmp and visitFCmp functions.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
14#include "llvm/ADT/APSInt.h"
15#include "llvm/ADT/SetVector.h"
16#include "llvm/ADT/Statistic.h"
17#include "llvm/Analysis/ConstantFolding.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/TargetLibraryInfo.h"
20#include "llvm/IR/ConstantRange.h"
21#include "llvm/IR/DataLayout.h"
22#include "llvm/IR/GetElementPtrTypeIterator.h"
23#include "llvm/IR/IntrinsicInst.h"
24#include "llvm/IR/PatternMatch.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/KnownBits.h"
27#include "llvm/Transforms/InstCombine/InstCombiner.h"
28
29using namespace llvm;
30using namespace PatternMatch;
31
32#define DEBUG_TYPE"instcombine" "instcombine"
33
34// How many times is a select replaced by one of its operands?
35STATISTIC(NumSel, "Number of select opts")static llvm::Statistic NumSel = {"instcombine", "NumSel", "Number of select opts"
}
;
36
37
38/// Compute Result = In1+In2, returning true if the result overflowed for this
39/// type.
40static bool addWithOverflow(APInt &Result, const APInt &In1,
41 const APInt &In2, bool IsSigned = false) {
42 bool Overflow;
43 if (IsSigned)
44 Result = In1.sadd_ov(In2, Overflow);
45 else
46 Result = In1.uadd_ov(In2, Overflow);
47
48 return Overflow;
49}
50
51/// Compute Result = In1-In2, returning true if the result overflowed for this
52/// type.
53static bool subWithOverflow(APInt &Result, const APInt &In1,
54 const APInt &In2, bool IsSigned = false) {
55 bool Overflow;
56 if (IsSigned)
57 Result = In1.ssub_ov(In2, Overflow);
58 else
59 Result = In1.usub_ov(In2, Overflow);
60
61 return Overflow;
62}
63
64/// Given an icmp instruction, return true if any use of this comparison is a
65/// branch on sign bit comparison.
66static bool hasBranchUse(ICmpInst &I) {
67 for (auto *U : I.users())
68 if (isa<BranchInst>(U))
69 return true;
70 return false;
71}
72
73/// Returns true if the exploded icmp can be expressed as a signed comparison
74/// to zero and updates the predicate accordingly.
75/// The signedness of the comparison is preserved.
76/// TODO: Refactor with decomposeBitTestICmp()?
77static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
78 if (!ICmpInst::isSigned(Pred))
79 return false;
80
81 if (C.isNullValue())
82 return ICmpInst::isRelational(Pred);
83
84 if (C.isOneValue()) {
85 if (Pred == ICmpInst::ICMP_SLT) {
86 Pred = ICmpInst::ICMP_SLE;
87 return true;
88 }
89 } else if (C.isAllOnesValue()) {
90 if (Pred == ICmpInst::ICMP_SGT) {
91 Pred = ICmpInst::ICMP_SGE;
92 return true;
93 }
94 }
95
96 return false;
97}
98
99/// Given a signed integer type and a set of known zero and one bits, compute
100/// the maximum and minimum values that could have the specified known zero and
101/// known one bits, returning them in Min/Max.
102/// TODO: Move to method on KnownBits struct?
103static void computeSignedMinMaxValuesFromKnownBits(const KnownBits &Known,
104 APInt &Min, APInt &Max) {
105 assert(Known.getBitWidth() == Min.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 107, __PRETTY_FUNCTION__))
106 Known.getBitWidth() == Max.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 107, __PRETTY_FUNCTION__))
107 "KnownZero, KnownOne and Min, Max must have equal bitwidth.")((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 107, __PRETTY_FUNCTION__))
;
108 APInt UnknownBits = ~(Known.Zero|Known.One);
109
110 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
111 // bit if it is unknown.
112 Min = Known.One;
113 Max = Known.One|UnknownBits;
114
115 if (UnknownBits.isNegative()) { // Sign bit is unknown
116 Min.setSignBit();
117 Max.clearSignBit();
118 }
119}
120
121/// Given an unsigned integer type and a set of known zero and one bits, compute
122/// the maximum and minimum values that could have the specified known zero and
123/// known one bits, returning them in Min/Max.
124/// TODO: Move to method on KnownBits struct?
125static void computeUnsignedMinMaxValuesFromKnownBits(const KnownBits &Known,
126 APInt &Min, APInt &Max) {
127 assert(Known.getBitWidth() == Min.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 129, __PRETTY_FUNCTION__))
128 Known.getBitWidth() == Max.getBitWidth() &&((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 129, __PRETTY_FUNCTION__))
129 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.")((Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth
() == Max.getBitWidth() && "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."
) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == Min.getBitWidth() && Known.getBitWidth() == Max.getBitWidth() && \"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 129, __PRETTY_FUNCTION__))
;
130 APInt UnknownBits = ~(Known.Zero|Known.One);
131
132 // The minimum value is when the unknown bits are all zeros.
133 Min = Known.One;
134 // The maximum value is when the unknown bits are all ones.
135 Max = Known.One|UnknownBits;
136}
137
138/// This is called when we see this pattern:
139/// cmp pred (load (gep GV, ...)), cmpcst
140/// where GV is a global variable with a constant initializer. Try to simplify
141/// this into some simple computation that does not need the load. For example
142/// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
143///
144/// If AndCst is non-null, then the loaded value is masked with that constant
145/// before doing the comparison. This handles cases like "A[i]&4 == 0".
146Instruction *
147InstCombinerImpl::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
148 GlobalVariable *GV, CmpInst &ICI,
149 ConstantInt *AndCst) {
150 Constant *Init = GV->getInitializer();
151 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
152 return nullptr;
153
154 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
155 // Don't blow up on huge arrays.
156 if (ArrayElementCount > MaxArraySizeForCombine)
157 return nullptr;
158
159 // There are many forms of this optimization we can handle, for now, just do
160 // the simple index into a single-dimensional array.
161 //
162 // Require: GEP GV, 0, i {{, constant indices}}
163 if (GEP->getNumOperands() < 3 ||
164 !isa<ConstantInt>(GEP->getOperand(1)) ||
165 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
166 isa<Constant>(GEP->getOperand(2)))
167 return nullptr;
168
169 // Check that indices after the variable are constants and in-range for the
170 // type they index. Collect the indices. This is typically for arrays of
171 // structs.
172 SmallVector<unsigned, 4> LaterIndices;
173
174 Type *EltTy = Init->getType()->getArrayElementType();
175 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
176 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
177 if (!Idx) return nullptr; // Variable index.
178
179 uint64_t IdxVal = Idx->getZExtValue();
180 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
181
182 if (StructType *STy = dyn_cast<StructType>(EltTy))
183 EltTy = STy->getElementType(IdxVal);
184 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
185 if (IdxVal >= ATy->getNumElements()) return nullptr;
186 EltTy = ATy->getElementType();
187 } else {
188 return nullptr; // Unknown type.
189 }
190
191 LaterIndices.push_back(IdxVal);
192 }
193
194 enum { Overdefined = -3, Undefined = -2 };
195
196 // Variables for our state machines.
197
198 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
199 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
200 // and 87 is the second (and last) index. FirstTrueElement is -2 when
201 // undefined, otherwise set to the first true element. SecondTrueElement is
202 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
203 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
204
205 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
206 // form "i != 47 & i != 87". Same state transitions as for true elements.
207 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
208
209 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
210 /// define a state machine that triggers for ranges of values that the index
211 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
212 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
213 /// index in the range (inclusive). We use -2 for undefined here because we
214 /// use relative comparisons and don't want 0-1 to match -1.
215 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
216
217 // MagicBitvector - This is a magic bitvector where we set a bit if the
218 // comparison is true for element 'i'. If there are 64 elements or less in
219 // the array, this will fully represent all the comparison results.
220 uint64_t MagicBitvector = 0;
221
222 // Scan the array and see if one of our patterns matches.
223 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
224 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
225 Constant *Elt = Init->getAggregateElement(i);
226 if (!Elt) return nullptr;
227
228 // If this is indexing an array of structures, get the structure element.
229 if (!LaterIndices.empty())
230 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
231
232 // If the element is masked, handle it.
233 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
234
235 // Find out if the comparison would be true or false for the i'th element.
236 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
237 CompareRHS, DL, &TLI);
238 // If the result is undef for this element, ignore it.
239 if (isa<UndefValue>(C)) {
240 // Extend range state machines to cover this element in case there is an
241 // undef in the middle of the range.
242 if (TrueRangeEnd == (int)i-1)
243 TrueRangeEnd = i;
244 if (FalseRangeEnd == (int)i-1)
245 FalseRangeEnd = i;
246 continue;
247 }
248
249 // If we can't compute the result for any of the elements, we have to give
250 // up evaluating the entire conditional.
251 if (!isa<ConstantInt>(C)) return nullptr;
252
253 // Otherwise, we know if the comparison is true or false for this element,
254 // update our state machines.
255 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
256
257 // State machine for single/double/range index comparison.
258 if (IsTrueForElt) {
259 // Update the TrueElement state machine.
260 if (FirstTrueElement == Undefined)
261 FirstTrueElement = TrueRangeEnd = i; // First true element.
262 else {
263 // Update double-compare state machine.
264 if (SecondTrueElement == Undefined)
265 SecondTrueElement = i;
266 else
267 SecondTrueElement = Overdefined;
268
269 // Update range state machine.
270 if (TrueRangeEnd == (int)i-1)
271 TrueRangeEnd = i;
272 else
273 TrueRangeEnd = Overdefined;
274 }
275 } else {
276 // Update the FalseElement state machine.
277 if (FirstFalseElement == Undefined)
278 FirstFalseElement = FalseRangeEnd = i; // First false element.
279 else {
280 // Update double-compare state machine.
281 if (SecondFalseElement == Undefined)
282 SecondFalseElement = i;
283 else
284 SecondFalseElement = Overdefined;
285
286 // Update range state machine.
287 if (FalseRangeEnd == (int)i-1)
288 FalseRangeEnd = i;
289 else
290 FalseRangeEnd = Overdefined;
291 }
292 }
293
294 // If this element is in range, update our magic bitvector.
295 if (i < 64 && IsTrueForElt)
296 MagicBitvector |= 1ULL << i;
297
298 // If all of our states become overdefined, bail out early. Since the
299 // predicate is expensive, only check it every 8 elements. This is only
300 // really useful for really huge arrays.
301 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
302 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
303 FalseRangeEnd == Overdefined)
304 return nullptr;
305 }
306
307 // Now that we've scanned the entire array, emit our new comparison(s). We
308 // order the state machines in complexity of the generated code.
309 Value *Idx = GEP->getOperand(2);
310
311 // If the index is larger than the pointer size of the target, truncate the
312 // index down like the GEP would do implicitly. We don't have to do this for
313 // an inbounds GEP because the index can't be out of range.
314 if (!GEP->isInBounds()) {
315 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
316 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
317 if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
318 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
319 }
320
321 // If the comparison is only true for one or two elements, emit direct
322 // comparisons.
323 if (SecondTrueElement != Overdefined) {
324 // None true -> false.
325 if (FirstTrueElement == Undefined)
326 return replaceInstUsesWith(ICI, Builder.getFalse());
327
328 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
329
330 // True for one element -> 'i == 47'.
331 if (SecondTrueElement == Undefined)
332 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
333
334 // True for two elements -> 'i == 47 | i == 72'.
335 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
336 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
337 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
338 return BinaryOperator::CreateOr(C1, C2);
339 }
340
341 // If the comparison is only false for one or two elements, emit direct
342 // comparisons.
343 if (SecondFalseElement != Overdefined) {
344 // None false -> true.
345 if (FirstFalseElement == Undefined)
346 return replaceInstUsesWith(ICI, Builder.getTrue());
347
348 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
349
350 // False for one element -> 'i != 47'.
351 if (SecondFalseElement == Undefined)
352 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
353
354 // False for two elements -> 'i != 47 & i != 72'.
355 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
356 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
357 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
358 return BinaryOperator::CreateAnd(C1, C2);
359 }
360
361 // If the comparison can be replaced with a range comparison for the elements
362 // where it is true, emit the range check.
363 if (TrueRangeEnd != Overdefined) {
364 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare")((TrueRangeEnd != FirstTrueElement && "Should emit single compare"
) ? static_cast<void> (0) : __assert_fail ("TrueRangeEnd != FirstTrueElement && \"Should emit single compare\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 364, __PRETTY_FUNCTION__))
;
365
366 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
367 if (FirstTrueElement) {
368 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
369 Idx = Builder.CreateAdd(Idx, Offs);
370 }
371
372 Value *End = ConstantInt::get(Idx->getType(),
373 TrueRangeEnd-FirstTrueElement+1);
374 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
375 }
376
377 // False range check.
378 if (FalseRangeEnd != Overdefined) {
379 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare")((FalseRangeEnd != FirstFalseElement && "Should emit single compare"
) ? static_cast<void> (0) : __assert_fail ("FalseRangeEnd != FirstFalseElement && \"Should emit single compare\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 379, __PRETTY_FUNCTION__))
;
380 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
381 if (FirstFalseElement) {
382 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
383 Idx = Builder.CreateAdd(Idx, Offs);
384 }
385
386 Value *End = ConstantInt::get(Idx->getType(),
387 FalseRangeEnd-FirstFalseElement);
388 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
389 }
390
391 // If a magic bitvector captures the entire comparison state
392 // of this load, replace it with computation that does:
393 // ((magic_cst >> i) & 1) != 0
394 {
395 Type *Ty = nullptr;
396
397 // Look for an appropriate type:
398 // - The type of Idx if the magic fits
399 // - The smallest fitting legal type
400 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
401 Ty = Idx->getType();
402 else
403 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
404
405 if (Ty) {
406 Value *V = Builder.CreateIntCast(Idx, Ty, false);
407 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
408 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
409 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
410 }
411 }
412
413 return nullptr;
414}
415
416/// Return a value that can be used to compare the *offset* implied by a GEP to
417/// zero. For example, if we have &A[i], we want to return 'i' for
418/// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
419/// are involved. The above expression would also be legal to codegen as
420/// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
421/// This latter form is less amenable to optimization though, and we are allowed
422/// to generate the first by knowing that pointer arithmetic doesn't overflow.
423///
424/// If we can't emit an optimized form for this expression, this returns null.
425///
426static Value *evaluateGEPOffsetExpression(User *GEP, InstCombinerImpl &IC,
427 const DataLayout &DL) {
428 gep_type_iterator GTI = gep_type_begin(GEP);
429
430 // Check to see if this gep only has a single variable index. If so, and if
431 // any constant indices are a multiple of its scale, then we can compute this
432 // in terms of the scale of the variable index. For example, if the GEP
433 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
434 // because the expression will cross zero at the same point.
435 unsigned i, e = GEP->getNumOperands();
436 int64_t Offset = 0;
437 for (i = 1; i != e; ++i, ++GTI) {
438 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
439 // Compute the aggregate offset of constant indices.
440 if (CI->isZero()) continue;
441
442 // Handle a struct index, which adds its field offset to the pointer.
443 if (StructType *STy = GTI.getStructTypeOrNull()) {
444 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
445 } else {
446 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
447 Offset += Size*CI->getSExtValue();
448 }
449 } else {
450 // Found our variable index.
451 break;
452 }
453 }
454
455 // If there are no variable indices, we must have a constant offset, just
456 // evaluate it the general way.
457 if (i == e) return nullptr;
458
459 Value *VariableIdx = GEP->getOperand(i);
460 // Determine the scale factor of the variable element. For example, this is
461 // 4 if the variable index is into an array of i32.
462 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
463
464 // Verify that there are no other variable indices. If so, emit the hard way.
465 for (++i, ++GTI; i != e; ++i, ++GTI) {
466 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
467 if (!CI) return nullptr;
468
469 // Compute the aggregate offset of constant indices.
470 if (CI->isZero()) continue;
471
472 // Handle a struct index, which adds its field offset to the pointer.
473 if (StructType *STy = GTI.getStructTypeOrNull()) {
474 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
475 } else {
476 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
477 Offset += Size*CI->getSExtValue();
478 }
479 }
480
481 // Okay, we know we have a single variable index, which must be a
482 // pointer/array/vector index. If there is no offset, life is simple, return
483 // the index.
484 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
485 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
486 if (Offset == 0) {
487 // Cast to intptrty in case a truncation occurs. If an extension is needed,
488 // we don't need to bother extending: the extension won't affect where the
489 // computation crosses zero.
490 if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
491 IntPtrWidth) {
492 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
493 }
494 return VariableIdx;
495 }
496
497 // Otherwise, there is an index. The computation we will do will be modulo
498 // the pointer size.
499 Offset = SignExtend64(Offset, IntPtrWidth);
500 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
501
502 // To do this transformation, any constant index must be a multiple of the
503 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
504 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
505 // multiple of the variable scale.
506 int64_t NewOffs = Offset / (int64_t)VariableScale;
507 if (Offset != NewOffs*(int64_t)VariableScale)
508 return nullptr;
509
510 // Okay, we can do this evaluation. Start by converting the index to intptr.
511 if (VariableIdx->getType() != IntPtrTy)
512 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
513 true /*Signed*/);
514 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
515 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
516}
517
518/// Returns true if we can rewrite Start as a GEP with pointer Base
519/// and some integer offset. The nodes that need to be re-written
520/// for this transformation will be added to Explored.
521static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
522 const DataLayout &DL,
523 SetVector<Value *> &Explored) {
524 SmallVector<Value *, 16> WorkList(1, Start);
525 Explored.insert(Base);
526
527 // The following traversal gives us an order which can be used
528 // when doing the final transformation. Since in the final
529 // transformation we create the PHI replacement instructions first,
530 // we don't have to get them in any particular order.
531 //
532 // However, for other instructions we will have to traverse the
533 // operands of an instruction first, which means that we have to
534 // do a post-order traversal.
535 while (!WorkList.empty()) {
536 SetVector<PHINode *> PHIs;
537
538 while (!WorkList.empty()) {
539 if (Explored.size() >= 100)
540 return false;
541
542 Value *V = WorkList.back();
543
544 if (Explored.count(V) != 0) {
545 WorkList.pop_back();
546 continue;
547 }
548
549 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
550 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
551 // We've found some value that we can't explore which is different from
552 // the base. Therefore we can't do this transformation.
553 return false;
554
555 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
556 auto *CI = cast<CastInst>(V);
557 if (!CI->isNoopCast(DL))
558 return false;
559
560 if (Explored.count(CI->getOperand(0)) == 0)
561 WorkList.push_back(CI->getOperand(0));
562 }
563
564 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
565 // We're limiting the GEP to having one index. This will preserve
566 // the original pointer type. We could handle more cases in the
567 // future.
568 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
569 GEP->getType() != Start->getType())
570 return false;
571
572 if (Explored.count(GEP->getOperand(0)) == 0)
573 WorkList.push_back(GEP->getOperand(0));
574 }
575
576 if (WorkList.back() == V) {
577 WorkList.pop_back();
578 // We've finished visiting this node, mark it as such.
579 Explored.insert(V);
580 }
581
582 if (auto *PN = dyn_cast<PHINode>(V)) {
583 // We cannot transform PHIs on unsplittable basic blocks.
584 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
585 return false;
586 Explored.insert(PN);
587 PHIs.insert(PN);
588 }
589 }
590
591 // Explore the PHI nodes further.
592 for (auto *PN : PHIs)
593 for (Value *Op : PN->incoming_values())
594 if (Explored.count(Op) == 0)
595 WorkList.push_back(Op);
596 }
597
598 // Make sure that we can do this. Since we can't insert GEPs in a basic
599 // block before a PHI node, we can't easily do this transformation if
600 // we have PHI node users of transformed instructions.
601 for (Value *Val : Explored) {
602 for (Value *Use : Val->uses()) {
603
604 auto *PHI = dyn_cast<PHINode>(Use);
605 auto *Inst = dyn_cast<Instruction>(Val);
606
607 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
608 Explored.count(PHI) == 0)
609 continue;
610
611 if (PHI->getParent() == Inst->getParent())
612 return false;
613 }
614 }
615 return true;
616}
617
618// Sets the appropriate insert point on Builder where we can add
619// a replacement Instruction for V (if that is possible).
620static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
621 bool Before = true) {
622 if (auto *PHI = dyn_cast<PHINode>(V)) {
623 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
624 return;
625 }
626 if (auto *I = dyn_cast<Instruction>(V)) {
627 if (!Before)
628 I = &*std::next(I->getIterator());
629 Builder.SetInsertPoint(I);
630 return;
631 }
632 if (auto *A = dyn_cast<Argument>(V)) {
633 // Set the insertion point in the entry block.
634 BasicBlock &Entry = A->getParent()->getEntryBlock();
635 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
636 return;
637 }
638 // Otherwise, this is a constant and we don't need to set a new
639 // insertion point.
640 assert(isa<Constant>(V) && "Setting insertion point for unknown value!")((isa<Constant>(V) && "Setting insertion point for unknown value!"
) ? static_cast<void> (0) : __assert_fail ("isa<Constant>(V) && \"Setting insertion point for unknown value!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 640, __PRETTY_FUNCTION__))
;
641}
642
643/// Returns a re-written value of Start as an indexed GEP using Base as a
644/// pointer.
645static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
646 const DataLayout &DL,
647 SetVector<Value *> &Explored) {
648 // Perform all the substitutions. This is a bit tricky because we can
649 // have cycles in our use-def chains.
650 // 1. Create the PHI nodes without any incoming values.
651 // 2. Create all the other values.
652 // 3. Add the edges for the PHI nodes.
653 // 4. Emit GEPs to get the original pointers.
654 // 5. Remove the original instructions.
655 Type *IndexType = IntegerType::get(
656 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
657
658 DenseMap<Value *, Value *> NewInsts;
659 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
660
661 // Create the new PHI nodes, without adding any incoming values.
662 for (Value *Val : Explored) {
663 if (Val == Base)
664 continue;
665 // Create empty phi nodes. This avoids cyclic dependencies when creating
666 // the remaining instructions.
667 if (auto *PHI = dyn_cast<PHINode>(Val))
668 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
669 PHI->getName() + ".idx", PHI);
670 }
671 IRBuilder<> Builder(Base->getContext());
672
673 // Create all the other instructions.
674 for (Value *Val : Explored) {
675
676 if (NewInsts.find(Val) != NewInsts.end())
677 continue;
678
679 if (auto *CI = dyn_cast<CastInst>(Val)) {
680 // Don't get rid of the intermediate variable here; the store can grow
681 // the map which will invalidate the reference to the input value.
682 Value *V = NewInsts[CI->getOperand(0)];
683 NewInsts[CI] = V;
684 continue;
685 }
686 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
687 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
688 : GEP->getOperand(1);
689 setInsertionPoint(Builder, GEP);
690 // Indices might need to be sign extended. GEPs will magically do
691 // this, but we need to do it ourselves here.
692 if (Index->getType()->getScalarSizeInBits() !=
693 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
694 Index = Builder.CreateSExtOrTrunc(
695 Index, NewInsts[GEP->getOperand(0)]->getType(),
696 GEP->getOperand(0)->getName() + ".sext");
697 }
698
699 auto *Op = NewInsts[GEP->getOperand(0)];
700 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
701 NewInsts[GEP] = Index;
702 else
703 NewInsts[GEP] = Builder.CreateNSWAdd(
704 Op, Index, GEP->getOperand(0)->getName() + ".add");
705 continue;
706 }
707 if (isa<PHINode>(Val))
708 continue;
709
710 llvm_unreachable("Unexpected instruction type")::llvm::llvm_unreachable_internal("Unexpected instruction type"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 710)
;
711 }
712
713 // Add the incoming values to the PHI nodes.
714 for (Value *Val : Explored) {
715 if (Val == Base)
716 continue;
717 // All the instructions have been created, we can now add edges to the
718 // phi nodes.
719 if (auto *PHI = dyn_cast<PHINode>(Val)) {
720 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
721 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
722 Value *NewIncoming = PHI->getIncomingValue(I);
723
724 if (NewInsts.find(NewIncoming) != NewInsts.end())
725 NewIncoming = NewInsts[NewIncoming];
726
727 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
728 }
729 }
730 }
731
732 for (Value *Val : Explored) {
733 if (Val == Base)
734 continue;
735
736 // Depending on the type, for external users we have to emit
737 // a GEP or a GEP + ptrtoint.
738 setInsertionPoint(Builder, Val, false);
739
740 // If required, create an inttoptr instruction for Base.
741 Value *NewBase = Base;
742 if (!Base->getType()->isPointerTy())
743 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
744 Start->getName() + "to.ptr");
745
746 Value *GEP = Builder.CreateInBoundsGEP(
747 Start->getType()->getPointerElementType(), NewBase,
748 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
749
750 if (!Val->getType()->isPointerTy()) {
751 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
752 Val->getName() + ".conv");
753 GEP = Cast;
754 }
755 Val->replaceAllUsesWith(GEP);
756 }
757
758 return NewInsts[Start];
759}
760
761/// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
762/// the input Value as a constant indexed GEP. Returns a pair containing
763/// the GEPs Pointer and Index.
764static std::pair<Value *, Value *>
765getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
766 Type *IndexType = IntegerType::get(V->getContext(),
767 DL.getIndexTypeSizeInBits(V->getType()));
768
769 Constant *Index = ConstantInt::getNullValue(IndexType);
770 while (true) {
771 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
772 // We accept only inbouds GEPs here to exclude the possibility of
773 // overflow.
774 if (!GEP->isInBounds())
775 break;
776 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
777 GEP->getType() == V->getType()) {
778 V = GEP->getOperand(0);
779 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
780 Index = ConstantExpr::getAdd(
781 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
782 continue;
783 }
784 break;
785 }
786 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
787 if (!CI->isNoopCast(DL))
788 break;
789 V = CI->getOperand(0);
790 continue;
791 }
792 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
793 if (!CI->isNoopCast(DL))
794 break;
795 V = CI->getOperand(0);
796 continue;
797 }
798 break;
799 }
800 return {V, Index};
801}
802
803/// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
804/// We can look through PHIs, GEPs and casts in order to determine a common base
805/// between GEPLHS and RHS.
806static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
807 ICmpInst::Predicate Cond,
808 const DataLayout &DL) {
809 // FIXME: Support vector of pointers.
810 if (GEPLHS->getType()->isVectorTy())
811 return nullptr;
812
813 if (!GEPLHS->hasAllConstantIndices())
814 return nullptr;
815
816 // Make sure the pointers have the same type.
817 if (GEPLHS->getType() != RHS->getType())
818 return nullptr;
819
820 Value *PtrBase, *Index;
821 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
822
823 // The set of nodes that will take part in this transformation.
824 SetVector<Value *> Nodes;
825
826 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
827 return nullptr;
828
829 // We know we can re-write this as
830 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
831 // Since we've only looked through inbouds GEPs we know that we
832 // can't have overflow on either side. We can therefore re-write
833 // this as:
834 // OFFSET1 cmp OFFSET2
835 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
836
837 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
838 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
839 // offset. Since Index is the offset of LHS to the base pointer, we will now
840 // compare the offsets instead of comparing the pointers.
841 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
842}
843
844/// Fold comparisons between a GEP instruction and something else. At this point
845/// we know that the GEP is on the LHS of the comparison.
846Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
847 ICmpInst::Predicate Cond,
848 Instruction &I) {
849 // Don't transform signed compares of GEPs into index compares. Even if the
850 // GEP is inbounds, the final add of the base pointer can have signed overflow
851 // and would change the result of the icmp.
852 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
853 // the maximum signed value for the pointer type.
854 if (ICmpInst::isSigned(Cond))
855 return nullptr;
856
857 // Look through bitcasts and addrspacecasts. We do not however want to remove
858 // 0 GEPs.
859 if (!isa<GetElementPtrInst>(RHS))
860 RHS = RHS->stripPointerCasts();
861
862 Value *PtrBase = GEPLHS->getOperand(0);
863 // FIXME: Support vector pointer GEPs.
864 if (PtrBase == RHS && GEPLHS->isInBounds() &&
865 !GEPLHS->getType()->isVectorTy()) {
866 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
867 // This transformation (ignoring the base and scales) is valid because we
868 // know pointers can't overflow since the gep is inbounds. See if we can
869 // output an optimized form.
870 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
871
872 // If not, synthesize the offset the hard way.
873 if (!Offset)
874 Offset = EmitGEPOffset(GEPLHS);
875 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
876 Constant::getNullValue(Offset->getType()));
877 }
878
879 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
880 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
881 !NullPointerIsDefined(I.getFunction(),
882 RHS->getType()->getPointerAddressSpace())) {
883 // For most address spaces, an allocation can't be placed at null, but null
884 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
885 // the only valid inbounds address derived from null, is null itself.
886 // Thus, we have four cases to consider:
887 // 1) Base == nullptr, Offset == 0 -> inbounds, null
888 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
889 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
890 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
891 //
892 // (Note if we're indexing a type of size 0, that simply collapses into one
893 // of the buckets above.)
894 //
895 // In general, we're allowed to make values less poison (i.e. remove
896 // sources of full UB), so in this case, we just select between the two
897 // non-poison cases (1 and 4 above).
898 //
899 // For vectors, we apply the same reasoning on a per-lane basis.
900 auto *Base = GEPLHS->getPointerOperand();
901 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
902 int NumElts = cast<FixedVectorType>(GEPLHS->getType())->getNumElements();
903 Base = Builder.CreateVectorSplat(NumElts, Base);
904 }
905 return new ICmpInst(Cond, Base,
906 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
907 cast<Constant>(RHS), Base->getType()));
908 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
909 // If the base pointers are different, but the indices are the same, just
910 // compare the base pointer.
911 if (PtrBase != GEPRHS->getOperand(0)) {
912 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
913 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
914 GEPRHS->getOperand(0)->getType();
915 if (IndicesTheSame)
916 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
917 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
918 IndicesTheSame = false;
919 break;
920 }
921
922 // If all indices are the same, just compare the base pointers.
923 Type *BaseType = GEPLHS->getOperand(0)->getType();
924 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
925 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
926
927 // If we're comparing GEPs with two base pointers that only differ in type
928 // and both GEPs have only constant indices or just one use, then fold
929 // the compare with the adjusted indices.
930 // FIXME: Support vector of pointers.
931 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
932 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
933 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
934 PtrBase->stripPointerCasts() ==
935 GEPRHS->getOperand(0)->stripPointerCasts() &&
936 !GEPLHS->getType()->isVectorTy()) {
937 Value *LOffset = EmitGEPOffset(GEPLHS);
938 Value *ROffset = EmitGEPOffset(GEPRHS);
939
940 // If we looked through an addrspacecast between different sized address
941 // spaces, the LHS and RHS pointers are different sized
942 // integers. Truncate to the smaller one.
943 Type *LHSIndexTy = LOffset->getType();
944 Type *RHSIndexTy = ROffset->getType();
945 if (LHSIndexTy != RHSIndexTy) {
946 if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
947 RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
948 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
949 } else
950 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
951 }
952
953 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
954 LOffset, ROffset);
955 return replaceInstUsesWith(I, Cmp);
956 }
957
958 // Otherwise, the base pointers are different and the indices are
959 // different. Try convert this to an indexed compare by looking through
960 // PHIs/casts.
961 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
962 }
963
964 // If one of the GEPs has all zero indices, recurse.
965 // FIXME: Handle vector of pointers.
966 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
967 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
968 ICmpInst::getSwappedPredicate(Cond), I);
969
970 // If the other GEP has all zero indices, recurse.
971 // FIXME: Handle vector of pointers.
972 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
973 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
974
975 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
976 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
977 // If the GEPs only differ by one index, compare it.
978 unsigned NumDifferences = 0; // Keep track of # differences.
979 unsigned DiffOperand = 0; // The operand that differs.
980 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
981 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
982 Type *LHSType = GEPLHS->getOperand(i)->getType();
983 Type *RHSType = GEPRHS->getOperand(i)->getType();
984 // FIXME: Better support for vector of pointers.
985 if (LHSType->getPrimitiveSizeInBits() !=
986 RHSType->getPrimitiveSizeInBits() ||
987 (GEPLHS->getType()->isVectorTy() &&
988 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
989 // Irreconcilable differences.
990 NumDifferences = 2;
991 break;
992 }
993
994 if (NumDifferences++) break;
995 DiffOperand = i;
996 }
997
998 if (NumDifferences == 0) // SAME GEP?
999 return replaceInstUsesWith(I, // No comparison is needed here.
1000 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
1001
1002 else if (NumDifferences == 1 && GEPsInBounds) {
1003 Value *LHSV = GEPLHS->getOperand(DiffOperand);
1004 Value *RHSV = GEPRHS->getOperand(DiffOperand);
1005 // Make sure we do a signed comparison here.
1006 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1007 }
1008 }
1009
1010 // Only lower this if the icmp is the only user of the GEP or if we expect
1011 // the result to fold to a constant!
1012 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1013 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1014 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1015 Value *L = EmitGEPOffset(GEPLHS);
1016 Value *R = EmitGEPOffset(GEPRHS);
1017 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1018 }
1019 }
1020
1021 // Try convert this to an indexed compare by looking through PHIs/casts as a
1022 // last resort.
1023 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1024}
1025
1026Instruction *InstCombinerImpl::foldAllocaCmp(ICmpInst &ICI,
1027 const AllocaInst *Alloca,
1028 const Value *Other) {
1029 assert(ICI.isEquality() && "Cannot fold non-equality comparison.")((ICI.isEquality() && "Cannot fold non-equality comparison."
) ? static_cast<void> (0) : __assert_fail ("ICI.isEquality() && \"Cannot fold non-equality comparison.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1029, __PRETTY_FUNCTION__))
;
1030
1031 // It would be tempting to fold away comparisons between allocas and any
1032 // pointer not based on that alloca (e.g. an argument). However, even
1033 // though such pointers cannot alias, they can still compare equal.
1034 //
1035 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1036 // doesn't escape we can argue that it's impossible to guess its value, and we
1037 // can therefore act as if any such guesses are wrong.
1038 //
1039 // The code below checks that the alloca doesn't escape, and that it's only
1040 // used in a comparison once (the current instruction). The
1041 // single-comparison-use condition ensures that we're trivially folding all
1042 // comparisons against the alloca consistently, and avoids the risk of
1043 // erroneously folding a comparison of the pointer with itself.
1044
1045 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1046
1047 SmallVector<const Use *, 32> Worklist;
1048 for (const Use &U : Alloca->uses()) {
1049 if (Worklist.size() >= MaxIter)
1050 return nullptr;
1051 Worklist.push_back(&U);
1052 }
1053
1054 unsigned NumCmps = 0;
1055 while (!Worklist.empty()) {
1056 assert(Worklist.size() <= MaxIter)((Worklist.size() <= MaxIter) ? static_cast<void> (0
) : __assert_fail ("Worklist.size() <= MaxIter", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1056, __PRETTY_FUNCTION__))
;
1057 const Use *U = Worklist.pop_back_val();
1058 const Value *V = U->getUser();
1059 --MaxIter;
1060
1061 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1062 isa<SelectInst>(V)) {
1063 // Track the uses.
1064 } else if (isa<LoadInst>(V)) {
1065 // Loading from the pointer doesn't escape it.
1066 continue;
1067 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1068 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1069 if (SI->getValueOperand() == U->get())
1070 return nullptr;
1071 continue;
1072 } else if (isa<ICmpInst>(V)) {
1073 if (NumCmps++)
1074 return nullptr; // Found more than one cmp.
1075 continue;
1076 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1077 switch (Intrin->getIntrinsicID()) {
1078 // These intrinsics don't escape or compare the pointer. Memset is safe
1079 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1080 // we don't allow stores, so src cannot point to V.
1081 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1082 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1083 continue;
1084 default:
1085 return nullptr;
1086 }
1087 } else {
1088 return nullptr;
1089 }
1090 for (const Use &U : V->uses()) {
1091 if (Worklist.size() >= MaxIter)
1092 return nullptr;
1093 Worklist.push_back(&U);
1094 }
1095 }
1096
1097 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1098 return replaceInstUsesWith(
1099 ICI,
1100 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1101}
1102
1103/// Fold "icmp pred (X+C), X".
1104Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
1105 ICmpInst::Predicate Pred) {
1106 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1107 // so the values can never be equal. Similarly for all other "or equals"
1108 // operators.
1109 assert(!!C && "C should not be zero!")((!!C && "C should not be zero!") ? static_cast<void
> (0) : __assert_fail ("!!C && \"C should not be zero!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1109, __PRETTY_FUNCTION__))
;
1110
1111 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1112 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1113 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1114 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1115 Constant *R = ConstantInt::get(X->getType(),
1116 APInt::getMaxValue(C.getBitWidth()) - C);
1117 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1118 }
1119
1120 // (X+1) >u X --> X <u (0-1) --> X != 255
1121 // (X+2) >u X --> X <u (0-2) --> X <u 254
1122 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1123 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1124 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1125 ConstantInt::get(X->getType(), -C));
1126
1127 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1128
1129 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1130 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1131 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1132 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1133 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1134 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1135 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1136 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1137 ConstantInt::get(X->getType(), SMax - C));
1138
1139 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1140 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1141 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1142 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1143 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1144 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1145
1146 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) ?
static_cast<void> (0) : __assert_fail ("Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1146, __PRETTY_FUNCTION__))
;
1147 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1148 ConstantInt::get(X->getType(), SMax - (C - 1)));
1149}
1150
1151/// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1152/// (icmp eq/ne A, Log2(AP2/AP1)) ->
1153/// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1154Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
1155 const APInt &AP1,
1156 const APInt &AP2) {
1157 assert(I.isEquality() && "Cannot fold icmp gt/lt")((I.isEquality() && "Cannot fold icmp gt/lt") ? static_cast
<void> (0) : __assert_fail ("I.isEquality() && \"Cannot fold icmp gt/lt\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1157, __PRETTY_FUNCTION__))
;
1158
1159 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1160 if (I.getPredicate() == I.ICMP_NE)
1161 Pred = CmpInst::getInversePredicate(Pred);
1162 return new ICmpInst(Pred, LHS, RHS);
1163 };
1164
1165 // Don't bother doing any work for cases which InstSimplify handles.
1166 if (AP2.isNullValue())
1167 return nullptr;
1168
1169 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1170 if (IsAShr) {
1171 if (AP2.isAllOnesValue())
1172 return nullptr;
1173 if (AP2.isNegative() != AP1.isNegative())
1174 return nullptr;
1175 if (AP2.sgt(AP1))
1176 return nullptr;
1177 }
1178
1179 if (!AP1)
1180 // 'A' must be large enough to shift out the highest set bit.
1181 return getICmp(I.ICMP_UGT, A,
1182 ConstantInt::get(A->getType(), AP2.logBase2()));
1183
1184 if (AP1 == AP2)
1185 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1186
1187 int Shift;
1188 if (IsAShr && AP1.isNegative())
1189 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1190 else
1191 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1192
1193 if (Shift > 0) {
1194 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1195 // There are multiple solutions if we are comparing against -1 and the LHS
1196 // of the ashr is not a power of two.
1197 if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1198 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1199 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1200 } else if (AP1 == AP2.lshr(Shift)) {
1201 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1202 }
1203 }
1204
1205 // Shifting const2 will never be equal to const1.
1206 // FIXME: This should always be handled by InstSimplify?
1207 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1208 return replaceInstUsesWith(I, TorF);
1209}
1210
1211/// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1212/// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1213Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1214 const APInt &AP1,
1215 const APInt &AP2) {
1216 assert(I.isEquality() && "Cannot fold icmp gt/lt")((I.isEquality() && "Cannot fold icmp gt/lt") ? static_cast
<void> (0) : __assert_fail ("I.isEquality() && \"Cannot fold icmp gt/lt\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1216, __PRETTY_FUNCTION__))
;
1217
1218 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1219 if (I.getPredicate() == I.ICMP_NE)
1220 Pred = CmpInst::getInversePredicate(Pred);
1221 return new ICmpInst(Pred, LHS, RHS);
1222 };
1223
1224 // Don't bother doing any work for cases which InstSimplify handles.
1225 if (AP2.isNullValue())
1226 return nullptr;
1227
1228 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1229
1230 if (!AP1 && AP2TrailingZeros != 0)
1231 return getICmp(
1232 I.ICMP_UGE, A,
1233 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1234
1235 if (AP1 == AP2)
1236 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1237
1238 // Get the distance between the lowest bits that are set.
1239 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1240
1241 if (Shift > 0 && AP2.shl(Shift) == AP1)
1242 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1243
1244 // Shifting const2 will never be equal to const1.
1245 // FIXME: This should always be handled by InstSimplify?
1246 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1247 return replaceInstUsesWith(I, TorF);
1248}
1249
1250/// The caller has matched a pattern of the form:
1251/// I = icmp ugt (add (add A, B), CI2), CI1
1252/// If this is of the form:
1253/// sum = a + b
1254/// if (sum+128 >u 255)
1255/// Then replace it with llvm.sadd.with.overflow.i8.
1256///
1257static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1258 ConstantInt *CI2, ConstantInt *CI1,
1259 InstCombinerImpl &IC) {
1260 // The transformation we're trying to do here is to transform this into an
1261 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1262 // with a narrower add, and discard the add-with-constant that is part of the
1263 // range check (if we can't eliminate it, this isn't profitable).
1264
1265 // In order to eliminate the add-with-constant, the compare can be its only
1266 // use.
1267 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1268 if (!AddWithCst->hasOneUse())
1269 return nullptr;
1270
1271 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1272 if (!CI2->getValue().isPowerOf2())
1273 return nullptr;
1274 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1275 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1276 return nullptr;
1277
1278 // The width of the new add formed is 1 more than the bias.
1279 ++NewWidth;
1280
1281 // Check to see that CI1 is an all-ones value with NewWidth bits.
1282 if (CI1->getBitWidth() == NewWidth ||
1283 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1284 return nullptr;
1285
1286 // This is only really a signed overflow check if the inputs have been
1287 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1288 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1289 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1290 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1291 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1292 return nullptr;
1293
1294 // In order to replace the original add with a narrower
1295 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1296 // and truncates that discard the high bits of the add. Verify that this is
1297 // the case.
1298 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1299 for (User *U : OrigAdd->users()) {
1300 if (U == AddWithCst)
1301 continue;
1302
1303 // Only accept truncates for now. We would really like a nice recursive
1304 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1305 // chain to see which bits of a value are actually demanded. If the
1306 // original add had another add which was then immediately truncated, we
1307 // could still do the transformation.
1308 TruncInst *TI = dyn_cast<TruncInst>(U);
1309 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1310 return nullptr;
1311 }
1312
1313 // If the pattern matches, truncate the inputs to the narrower type and
1314 // use the sadd_with_overflow intrinsic to efficiently compute both the
1315 // result and the overflow bit.
1316 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1317 Function *F = Intrinsic::getDeclaration(
1318 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1319
1320 InstCombiner::BuilderTy &Builder = IC.Builder;
1321
1322 // Put the new code above the original add, in case there are any uses of the
1323 // add between the add and the compare.
1324 Builder.SetInsertPoint(OrigAdd);
1325
1326 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1327 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1328 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1329 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1330 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1331
1332 // The inner add was the result of the narrow add, zero extended to the
1333 // wider type. Replace it with the result computed by the intrinsic.
1334 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1335 IC.eraseInstFromFunction(*OrigAdd);
1336
1337 // The original icmp gets replaced with the overflow value.
1338 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1339}
1340
1341/// If we have:
1342/// icmp eq/ne (urem/srem %x, %y), 0
1343/// iff %y is a power-of-two, we can replace this with a bit test:
1344/// icmp eq/ne (and %x, (add %y, -1)), 0
1345Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1346 // This fold is only valid for equality predicates.
1347 if (!I.isEquality())
1348 return nullptr;
1349 ICmpInst::Predicate Pred;
1350 Value *X, *Y, *Zero;
1351 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1352 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1353 return nullptr;
1354 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1355 return nullptr;
1356 // This may increase instruction count, we don't enforce that Y is a constant.
1357 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1358 Value *Masked = Builder.CreateAnd(X, Mask);
1359 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1360}
1361
1362/// Fold equality-comparison between zero and any (maybe truncated) right-shift
1363/// by one-less-than-bitwidth into a sign test on the original value.
1364Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1365 Instruction *Val;
1366 ICmpInst::Predicate Pred;
1367 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1368 return nullptr;
1369
1370 Value *X;
1371 Type *XTy;
1372
1373 Constant *C;
1374 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1375 XTy = X->getType();
1376 unsigned XBitWidth = XTy->getScalarSizeInBits();
1377 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1378 APInt(XBitWidth, XBitWidth - 1))))
1379 return nullptr;
1380 } else if (isa<BinaryOperator>(Val) &&
1381 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1382 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1383 /*AnalyzeForSignBitExtraction=*/true))) {
1384 XTy = X->getType();
1385 } else
1386 return nullptr;
1387
1388 return ICmpInst::Create(Instruction::ICmp,
1389 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1390 : ICmpInst::ICMP_SLT,
1391 X, ConstantInt::getNullValue(XTy));
1392}
1393
1394// Handle icmp pred X, 0
1395Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1396 CmpInst::Predicate Pred = Cmp.getPredicate();
1397 if (!match(Cmp.getOperand(1), m_Zero()))
1398 return nullptr;
1399
1400 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1401 if (Pred == ICmpInst::ICMP_SGT) {
1402 Value *A, *B;
1403 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1404 if (SPR.Flavor == SPF_SMIN) {
1405 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1406 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1407 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1408 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1409 }
1410 }
1411
1412 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1413 return New;
1414
1415 // Given:
1416 // icmp eq/ne (urem %x, %y), 0
1417 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1418 // icmp eq/ne %x, 0
1419 Value *X, *Y;
1420 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1421 ICmpInst::isEquality(Pred)) {
1422 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1423 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1424 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1425 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1426 }
1427
1428 return nullptr;
1429}
1430
1431/// Fold icmp Pred X, C.
1432/// TODO: This code structure does not make sense. The saturating add fold
1433/// should be moved to some other helper and extended as noted below (it is also
1434/// possible that code has been made unnecessary - do we canonicalize IR to
1435/// overflow/saturating intrinsics or not?).
1436Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1437 // Match the following pattern, which is a common idiom when writing
1438 // overflow-safe integer arithmetic functions. The source performs an addition
1439 // in wider type and explicitly checks for overflow using comparisons against
1440 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1441 //
1442 // TODO: This could probably be generalized to handle other overflow-safe
1443 // operations if we worked out the formulas to compute the appropriate magic
1444 // constants.
1445 //
1446 // sum = a + b
1447 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1448 CmpInst::Predicate Pred = Cmp.getPredicate();
1449 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1450 Value *A, *B;
1451 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1452 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1453 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1454 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1455 return Res;
1456
1457 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1458 Constant *C = dyn_cast<Constant>(Op1);
1459 if (!C)
1460 return nullptr;
1461
1462 if (auto *Phi = dyn_cast<PHINode>(Op0))
1463 if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1464 Type *Ty = Cmp.getType();
1465 Builder.SetInsertPoint(Phi);
1466 PHINode *NewPhi =
1467 Builder.CreatePHI(Ty, Phi->getNumOperands());
1468 for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1469 auto *Input =
1470 cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1471 auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1472 NewPhi->addIncoming(BoolInput, Predecessor);
1473 }
1474 NewPhi->takeName(&Cmp);
1475 return replaceInstUsesWith(Cmp, NewPhi);
1476 }
1477
1478 return nullptr;
1479}
1480
1481/// Canonicalize icmp instructions based on dominating conditions.
1482Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1483 // This is a cheap/incomplete check for dominance - just match a single
1484 // predecessor with a conditional branch.
1485 BasicBlock *CmpBB = Cmp.getParent();
1486 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1487 if (!DomBB)
1488 return nullptr;
1489
1490 Value *DomCond;
1491 BasicBlock *TrueBB, *FalseBB;
1492 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1493 return nullptr;
1494
1495 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&(((TrueBB == CmpBB || FalseBB == CmpBB) && "Predecessor block does not point to successor?"
) ? static_cast<void> (0) : __assert_fail ("(TrueBB == CmpBB || FalseBB == CmpBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1496, __PRETTY_FUNCTION__))
1496 "Predecessor block does not point to successor?")(((TrueBB == CmpBB || FalseBB == CmpBB) && "Predecessor block does not point to successor?"
) ? static_cast<void> (0) : __assert_fail ("(TrueBB == CmpBB || FalseBB == CmpBB) && \"Predecessor block does not point to successor?\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1496, __PRETTY_FUNCTION__))
;
1497
1498 // The branch should get simplified. Don't bother simplifying this condition.
1499 if (TrueBB == FalseBB)
1500 return nullptr;
1501
1502 // Try to simplify this compare to T/F based on the dominating condition.
1503 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1504 if (Imp)
1505 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1506
1507 CmpInst::Predicate Pred = Cmp.getPredicate();
1508 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1509 ICmpInst::Predicate DomPred;
1510 const APInt *C, *DomC;
1511 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1512 match(Y, m_APInt(C))) {
1513 // We have 2 compares of a variable with constants. Calculate the constant
1514 // ranges of those compares to see if we can transform the 2nd compare:
1515 // DomBB:
1516 // DomCond = icmp DomPred X, DomC
1517 // br DomCond, CmpBB, FalseBB
1518 // CmpBB:
1519 // Cmp = icmp Pred X, C
1520 ConstantRange CR = ConstantRange::makeAllowedICmpRegion(Pred, *C);
1521 ConstantRange DominatingCR =
1522 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1523 : ConstantRange::makeExactICmpRegion(
1524 CmpInst::getInversePredicate(DomPred), *DomC);
1525 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1526 ConstantRange Difference = DominatingCR.difference(CR);
1527 if (Intersection.isEmptySet())
1528 return replaceInstUsesWith(Cmp, Builder.getFalse());
1529 if (Difference.isEmptySet())
1530 return replaceInstUsesWith(Cmp, Builder.getTrue());
1531
1532 // Canonicalizing a sign bit comparison that gets used in a branch,
1533 // pessimizes codegen by generating branch on zero instruction instead
1534 // of a test and branch. So we avoid canonicalizing in such situations
1535 // because test and branch instruction has better branch displacement
1536 // than compare and branch instruction.
1537 bool UnusedBit;
1538 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1539 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1540 return nullptr;
1541
1542 if (const APInt *EqC = Intersection.getSingleElement())
1543 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1544 if (const APInt *NeC = Difference.getSingleElement())
1545 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1546 }
1547
1548 return nullptr;
1549}
1550
1551/// Fold icmp (trunc X, Y), C.
1552Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1553 TruncInst *Trunc,
1554 const APInt &C) {
1555 ICmpInst::Predicate Pred = Cmp.getPredicate();
1556 Value *X = Trunc->getOperand(0);
1557 if (C.isOneValue() && C.getBitWidth() > 1) {
1558 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1559 Value *V = nullptr;
1560 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1561 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1562 ConstantInt::get(V->getType(), 1));
1563 }
1564
1565 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1566 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1567 // of the high bits truncated out of x are known.
1568 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1569 SrcBits = X->getType()->getScalarSizeInBits();
1570 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1571
1572 // If all the high bits are known, we can do this xform.
1573 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1574 // Pull in the high bits from known-ones set.
1575 APInt NewRHS = C.zext(SrcBits);
1576 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1577 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1578 }
1579 }
1580
1581 return nullptr;
1582}
1583
1584/// Fold icmp (xor X, Y), C.
1585Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1586 BinaryOperator *Xor,
1587 const APInt &C) {
1588 Value *X = Xor->getOperand(0);
1589 Value *Y = Xor->getOperand(1);
1590 const APInt *XorC;
1591 if (!match(Y, m_APInt(XorC)))
1592 return nullptr;
1593
1594 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1595 // fold the xor.
1596 ICmpInst::Predicate Pred = Cmp.getPredicate();
1597 bool TrueIfSigned = false;
1598 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1599
1600 // If the sign bit of the XorCst is not set, there is no change to
1601 // the operation, just stop using the Xor.
1602 if (!XorC->isNegative())
1603 return replaceOperand(Cmp, 0, X);
1604
1605 // Emit the opposite comparison.
1606 if (TrueIfSigned)
1607 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1608 ConstantInt::getAllOnesValue(X->getType()));
1609 else
1610 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1611 ConstantInt::getNullValue(X->getType()));
1612 }
1613
1614 if (Xor->hasOneUse()) {
1615 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1616 if (!Cmp.isEquality() && XorC->isSignMask()) {
1617 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1618 : Cmp.getSignedPredicate();
1619 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1620 }
1621
1622 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1623 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1624 Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1625 : Cmp.getSignedPredicate();
1626 Pred = Cmp.getSwappedPredicate(Pred);
1627 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1628 }
1629 }
1630
1631 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1632 if (Pred == ICmpInst::ICMP_UGT) {
1633 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1634 if (*XorC == ~C && (C + 1).isPowerOf2())
1635 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1636 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1637 if (*XorC == C && (C + 1).isPowerOf2())
1638 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1639 }
1640 if (Pred == ICmpInst::ICMP_ULT) {
1641 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1642 if (*XorC == -C && C.isPowerOf2())
1643 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1644 ConstantInt::get(X->getType(), ~C));
1645 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1646 if (*XorC == C && (-C).isPowerOf2())
1647 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1648 ConstantInt::get(X->getType(), ~C));
1649 }
1650 return nullptr;
1651}
1652
1653/// Fold icmp (and (sh X, Y), C2), C1.
1654Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1655 BinaryOperator *And,
1656 const APInt &C1,
1657 const APInt &C2) {
1658 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1659 if (!Shift || !Shift->isShift())
1660 return nullptr;
1661
1662 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1663 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1664 // code produced by the clang front-end, for bitfield access.
1665 // This seemingly simple opportunity to fold away a shift turns out to be
1666 // rather complicated. See PR17827 for details.
1667 unsigned ShiftOpcode = Shift->getOpcode();
1668 bool IsShl = ShiftOpcode == Instruction::Shl;
1669 const APInt *C3;
1670 if (match(Shift->getOperand(1), m_APInt(C3))) {
1671 APInt NewAndCst, NewCmpCst;
1672 bool AnyCmpCstBitsShiftedOut;
1673 if (ShiftOpcode == Instruction::Shl) {
1674 // For a left shift, we can fold if the comparison is not signed. We can
1675 // also fold a signed comparison if the mask value and comparison value
1676 // are not negative. These constraints may not be obvious, but we can
1677 // prove that they are correct using an SMT solver.
1678 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1679 return nullptr;
1680
1681 NewCmpCst = C1.lshr(*C3);
1682 NewAndCst = C2.lshr(*C3);
1683 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1684 } else if (ShiftOpcode == Instruction::LShr) {
1685 // For a logical right shift, we can fold if the comparison is not signed.
1686 // We can also fold a signed comparison if the shifted mask value and the
1687 // shifted comparison value are not negative. These constraints may not be
1688 // obvious, but we can prove that they are correct using an SMT solver.
1689 NewCmpCst = C1.shl(*C3);
1690 NewAndCst = C2.shl(*C3);
1691 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1692 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1693 return nullptr;
1694 } else {
1695 // For an arithmetic shift, check that both constants don't use (in a
1696 // signed sense) the top bits being shifted out.
1697 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode")((ShiftOpcode == Instruction::AShr && "Unknown shift opcode"
) ? static_cast<void> (0) : __assert_fail ("ShiftOpcode == Instruction::AShr && \"Unknown shift opcode\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 1697, __PRETTY_FUNCTION__))
;
1698 NewCmpCst = C1.shl(*C3);
1699 NewAndCst = C2.shl(*C3);
1700 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1701 if (NewAndCst.ashr(*C3) != C2)
1702 return nullptr;
1703 }
1704
1705 if (AnyCmpCstBitsShiftedOut) {
1706 // If we shifted bits out, the fold is not going to work out. As a
1707 // special case, check to see if this means that the result is always
1708 // true or false now.
1709 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1710 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1711 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1712 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1713 } else {
1714 Value *NewAnd = Builder.CreateAnd(
1715 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1716 return new ICmpInst(Cmp.getPredicate(),
1717 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1718 }
1719 }
1720
1721 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1722 // preferable because it allows the C2 << Y expression to be hoisted out of a
1723 // loop if Y is invariant and X is not.
1724 if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1725 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1726 // Compute C2 << Y.
1727 Value *NewShift =
1728 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1729 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1730
1731 // Compute X & (C2 << Y).
1732 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1733 return replaceOperand(Cmp, 0, NewAnd);
1734 }
1735
1736 return nullptr;
1737}
1738
1739/// Fold icmp (and X, C2), C1.
1740Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1741 BinaryOperator *And,
1742 const APInt &C1) {
1743 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1744
1745 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1746 // TODO: We canonicalize to the longer form for scalars because we have
1747 // better analysis/folds for icmp, and codegen may be better with icmp.
1748 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1749 match(And->getOperand(1), m_One()))
1750 return new TruncInst(And->getOperand(0), Cmp.getType());
1751
1752 const APInt *C2;
1753 Value *X;
1754 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1755 return nullptr;
1756
1757 // Don't perform the following transforms if the AND has multiple uses
1758 if (!And->hasOneUse())
1759 return nullptr;
1760
1761 if (Cmp.isEquality() && C1.isNullValue()) {
1762 // Restrict this fold to single-use 'and' (PR10267).
1763 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1764 if (C2->isSignMask()) {
1765 Constant *Zero = Constant::getNullValue(X->getType());
1766 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1767 return new ICmpInst(NewPred, X, Zero);
1768 }
1769
1770 // Restrict this fold only for single-use 'and' (PR10267).
1771 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1772 if ((~(*C2) + 1).isPowerOf2()) {
1773 Constant *NegBOC =
1774 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1775 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1776 return new ICmpInst(NewPred, X, NegBOC);
1777 }
1778 }
1779
1780 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1781 // the input width without changing the value produced, eliminate the cast:
1782 //
1783 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1784 //
1785 // We can do this transformation if the constants do not have their sign bits
1786 // set or if it is an equality comparison. Extending a relational comparison
1787 // when we're checking the sign bit would not work.
1788 Value *W;
1789 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1790 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1791 // TODO: Is this a good transform for vectors? Wider types may reduce
1792 // throughput. Should this transform be limited (even for scalars) by using
1793 // shouldChangeType()?
1794 if (!Cmp.getType()->isVectorTy()) {
1795 Type *WideType = W->getType();
1796 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1797 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1798 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1799 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1800 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1801 }
1802 }
1803
1804 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1805 return I;
1806
1807 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1808 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1809 //
1810 // iff pred isn't signed
1811 if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1812 match(And->getOperand(1), m_One())) {
1813 Constant *One = cast<Constant>(And->getOperand(1));
1814 Value *Or = And->getOperand(0);
1815 Value *A, *B, *LShr;
1816 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1817 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1818 unsigned UsesRemoved = 0;
1819 if (And->hasOneUse())
1820 ++UsesRemoved;
1821 if (Or->hasOneUse())
1822 ++UsesRemoved;
1823 if (LShr->hasOneUse())
1824 ++UsesRemoved;
1825
1826 // Compute A & ((1 << B) | 1)
1827 Value *NewOr = nullptr;
1828 if (auto *C = dyn_cast<Constant>(B)) {
1829 if (UsesRemoved >= 1)
1830 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1831 } else {
1832 if (UsesRemoved >= 3)
1833 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1834 /*HasNUW=*/true),
1835 One, Or->getName());
1836 }
1837 if (NewOr) {
1838 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1839 return replaceOperand(Cmp, 0, NewAnd);
1840 }
1841 }
1842 }
1843
1844 return nullptr;
1845}
1846
1847/// Fold icmp (and X, Y), C.
1848Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1849 BinaryOperator *And,
1850 const APInt &C) {
1851 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1852 return I;
1853
1854 // TODO: These all require that Y is constant too, so refactor with the above.
1855
1856 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1857 Value *X = And->getOperand(0);
1858 Value *Y = And->getOperand(1);
1859 if (auto *LI = dyn_cast<LoadInst>(X))
1860 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1861 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1862 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1863 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1864 ConstantInt *C2 = cast<ConstantInt>(Y);
1865 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1866 return Res;
1867 }
1868
1869 if (!Cmp.isEquality())
1870 return nullptr;
1871
1872 // X & -C == -C -> X > u ~C
1873 // X & -C != -C -> X <= u ~C
1874 // iff C is a power of 2
1875 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1876 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1877 : CmpInst::ICMP_ULE;
1878 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1879 }
1880
1881 // (X & C2) == 0 -> (trunc X) >= 0
1882 // (X & C2) != 0 -> (trunc X) < 0
1883 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1884 const APInt *C2;
1885 if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1886 int32_t ExactLogBase2 = C2->exactLogBase2();
1887 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1888 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1889 if (auto *AndVTy = dyn_cast<VectorType>(And->getType()))
1890 NTy = FixedVectorType::get(
1891 NTy, cast<FixedVectorType>(AndVTy)->getNumElements());
1892 Value *Trunc = Builder.CreateTrunc(X, NTy);
1893 auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1894 : CmpInst::ICMP_SLT;
1895 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1896 }
1897 }
1898
1899 return nullptr;
1900}
1901
1902/// Fold icmp (or X, Y), C.
1903Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
1904 BinaryOperator *Or,
1905 const APInt &C) {
1906 ICmpInst::Predicate Pred = Cmp.getPredicate();
1907 if (C.isOneValue()) {
1908 // icmp slt signum(V) 1 --> icmp slt V, 1
1909 Value *V = nullptr;
1910 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1911 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1912 ConstantInt::get(V->getType(), 1));
1913 }
1914
1915 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1916 const APInt *MaskC;
1917 if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1918 if (*MaskC == C && (C + 1).isPowerOf2()) {
1919 // X | C == C --> X <=u C
1920 // X | C != C --> X >u C
1921 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1922 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1923 return new ICmpInst(Pred, OrOp0, OrOp1);
1924 }
1925
1926 // More general: canonicalize 'equality with set bits mask' to
1927 // 'equality with clear bits mask'.
1928 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1929 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1930 if (Or->hasOneUse()) {
1931 Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1932 Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1933 return new ICmpInst(Pred, And, NewC);
1934 }
1935 }
1936
1937 if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1938 return nullptr;
1939
1940 Value *P, *Q;
1941 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1942 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1943 // -> and (icmp eq P, null), (icmp eq Q, null).
1944 Value *CmpP =
1945 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1946 Value *CmpQ =
1947 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1948 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1949 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1950 }
1951
1952 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1953 // a shorter form that has more potential to be folded even further.
1954 Value *X1, *X2, *X3, *X4;
1955 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1956 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1957 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1958 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1959 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1960 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1961 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1962 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1963 }
1964
1965 return nullptr;
1966}
1967
1968/// Fold icmp (mul X, Y), C.
1969Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
1970 BinaryOperator *Mul,
1971 const APInt &C) {
1972 const APInt *MulC;
1973 if (!match(Mul->getOperand(1), m_APInt(MulC)))
1974 return nullptr;
1975
1976 // If this is a test of the sign bit and the multiply is sign-preserving with
1977 // a constant operand, use the multiply LHS operand instead.
1978 ICmpInst::Predicate Pred = Cmp.getPredicate();
1979 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1980 if (MulC->isNegative())
1981 Pred = ICmpInst::getSwappedPredicate(Pred);
1982 return new ICmpInst(Pred, Mul->getOperand(0),
1983 Constant::getNullValue(Mul->getType()));
1984 }
1985
1986 // If the multiply does not wrap, try to divide the compare constant by the
1987 // multiplication factor.
1988 if (Cmp.isEquality() && !MulC->isNullValue()) {
1989 // (mul nsw X, MulC) == C --> X == C /s MulC
1990 if (Mul->hasNoSignedWrap() && C.srem(*MulC).isNullValue()) {
1991 Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
1992 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1993 }
1994 // (mul nuw X, MulC) == C --> X == C /u MulC
1995 if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isNullValue()) {
1996 Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
1997 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1998 }
1999 }
2000
2001 return nullptr;
2002}
2003
2004/// Fold icmp (shl 1, Y), C.
2005static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
2006 const APInt &C) {
2007 Value *Y;
2008 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2009 return nullptr;
2010
2011 Type *ShiftType = Shl->getType();
2012 unsigned TypeBits = C.getBitWidth();
2013 bool CIsPowerOf2 = C.isPowerOf2();
2014 ICmpInst::Predicate Pred = Cmp.getPredicate();
2015 if (Cmp.isUnsigned()) {
2016 // (1 << Y) pred C -> Y pred Log2(C)
2017 if (!CIsPowerOf2) {
2018 // (1 << Y) < 30 -> Y <= 4
2019 // (1 << Y) <= 30 -> Y <= 4
2020 // (1 << Y) >= 30 -> Y > 4
2021 // (1 << Y) > 30 -> Y > 4
2022 if (Pred == ICmpInst::ICMP_ULT)
2023 Pred = ICmpInst::ICMP_ULE;
2024 else if (Pred == ICmpInst::ICMP_UGE)
2025 Pred = ICmpInst::ICMP_UGT;
2026 }
2027
2028 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2029 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2030 unsigned CLog2 = C.logBase2();
2031 if (CLog2 == TypeBits - 1) {
2032 if (Pred == ICmpInst::ICMP_UGE)
2033 Pred = ICmpInst::ICMP_EQ;
2034 else if (Pred == ICmpInst::ICMP_ULT)
2035 Pred = ICmpInst::ICMP_NE;
2036 }
2037 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2038 } else if (Cmp.isSigned()) {
2039 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2040 if (C.isAllOnesValue()) {
2041 // (1 << Y) <= -1 -> Y == 31
2042 if (Pred == ICmpInst::ICMP_SLE)
2043 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2044
2045 // (1 << Y) > -1 -> Y != 31
2046 if (Pred == ICmpInst::ICMP_SGT)
2047 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2048 } else if (!C) {
2049 // (1 << Y) < 0 -> Y == 31
2050 // (1 << Y) <= 0 -> Y == 31
2051 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2052 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2053
2054 // (1 << Y) >= 0 -> Y != 31
2055 // (1 << Y) > 0 -> Y != 31
2056 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2057 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2058 }
2059 } else if (Cmp.isEquality() && CIsPowerOf2) {
2060 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2061 }
2062
2063 return nullptr;
2064}
2065
2066/// Fold icmp (shl X, Y), C.
2067Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2068 BinaryOperator *Shl,
2069 const APInt &C) {
2070 const APInt *ShiftVal;
2071 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2072 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2073
2074 const APInt *ShiftAmt;
2075 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2076 return foldICmpShlOne(Cmp, Shl, C);
2077
2078 // Check that the shift amount is in range. If not, don't perform undefined
2079 // shifts. When the shift is visited, it will be simplified.
2080 unsigned TypeBits = C.getBitWidth();
2081 if (ShiftAmt->uge(TypeBits))
2082 return nullptr;
2083
2084 ICmpInst::Predicate Pred = Cmp.getPredicate();
2085 Value *X = Shl->getOperand(0);
2086 Type *ShType = Shl->getType();
2087
2088 // NSW guarantees that we are only shifting out sign bits from the high bits,
2089 // so we can ASHR the compare constant without needing a mask and eliminate
2090 // the shift.
2091 if (Shl->hasNoSignedWrap()) {
2092 if (Pred == ICmpInst::ICMP_SGT) {
2093 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2094 APInt ShiftedC = C.ashr(*ShiftAmt);
2095 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2096 }
2097 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2098 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2099 APInt ShiftedC = C.ashr(*ShiftAmt);
2100 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2101 }
2102 if (Pred == ICmpInst::ICMP_SLT) {
2103 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2104 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2105 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2106 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2107 assert(!C.isMinSignedValue() && "Unexpected icmp slt")((!C.isMinSignedValue() && "Unexpected icmp slt") ? static_cast
<void> (0) : __assert_fail ("!C.isMinSignedValue() && \"Unexpected icmp slt\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2107, __PRETTY_FUNCTION__))
;
2108 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2109 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2110 }
2111 // If this is a signed comparison to 0 and the shift is sign preserving,
2112 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2113 // do that if we're sure to not continue on in this function.
2114 if (isSignTest(Pred, C))
2115 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2116 }
2117
2118 // NUW guarantees that we are only shifting out zero bits from the high bits,
2119 // so we can LSHR the compare constant without needing a mask and eliminate
2120 // the shift.
2121 if (Shl->hasNoUnsignedWrap()) {
2122 if (Pred == ICmpInst::ICMP_UGT) {
2123 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2124 APInt ShiftedC = C.lshr(*ShiftAmt);
2125 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2126 }
2127 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2128 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2129 APInt ShiftedC = C.lshr(*ShiftAmt);
2130 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2131 }
2132 if (Pred == ICmpInst::ICMP_ULT) {
2133 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2134 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2135 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2136 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2137 assert(C.ugt(0) && "ult 0 should have been eliminated")((C.ugt(0) && "ult 0 should have been eliminated") ? static_cast
<void> (0) : __assert_fail ("C.ugt(0) && \"ult 0 should have been eliminated\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2137, __PRETTY_FUNCTION__))
;
2138 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2139 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2140 }
2141 }
2142
2143 if (Cmp.isEquality() && Shl->hasOneUse()) {
2144 // Strength-reduce the shift into an 'and'.
2145 Constant *Mask = ConstantInt::get(
2146 ShType,
2147 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2148 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2149 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2150 return new ICmpInst(Pred, And, LShrC);
2151 }
2152
2153 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2154 bool TrueIfSigned = false;
2155 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2156 // (X << 31) <s 0 --> (X & 1) != 0
2157 Constant *Mask = ConstantInt::get(
2158 ShType,
2159 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2160 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2161 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2162 And, Constant::getNullValue(ShType));
2163 }
2164
2165 // Simplify 'shl' inequality test into 'and' equality test.
2166 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2167 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2168 if ((C + 1).isPowerOf2() &&
2169 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2170 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2171 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2172 : ICmpInst::ICMP_NE,
2173 And, Constant::getNullValue(ShType));
2174 }
2175 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2176 if (C.isPowerOf2() &&
2177 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2178 Value *And =
2179 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2180 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2181 : ICmpInst::ICMP_NE,
2182 And, Constant::getNullValue(ShType));
2183 }
2184 }
2185
2186 // Transform (icmp pred iM (shl iM %v, N), C)
2187 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2188 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2189 // This enables us to get rid of the shift in favor of a trunc that may be
2190 // free on the target. It has the additional benefit of comparing to a
2191 // smaller constant that may be more target-friendly.
2192 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2193 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2194 DL.isLegalInteger(TypeBits - Amt)) {
2195 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2196 if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2197 TruncTy = FixedVectorType::get(
2198 TruncTy, cast<FixedVectorType>(ShVTy)->getNumElements());
2199 Constant *NewC =
2200 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2201 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2202 }
2203
2204 return nullptr;
2205}
2206
2207/// Fold icmp ({al}shr X, Y), C.
2208Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2209 BinaryOperator *Shr,
2210 const APInt &C) {
2211 // An exact shr only shifts out zero bits, so:
2212 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2213 Value *X = Shr->getOperand(0);
2214 CmpInst::Predicate Pred = Cmp.getPredicate();
2215 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2216 C.isNullValue())
2217 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2218
2219 const APInt *ShiftVal;
2220 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2221 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2222
2223 const APInt *ShiftAmt;
2224 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2225 return nullptr;
2226
2227 // Check that the shift amount is in range. If not, don't perform undefined
2228 // shifts. When the shift is visited it will be simplified.
2229 unsigned TypeBits = C.getBitWidth();
2230 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2231 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2232 return nullptr;
2233
2234 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2235 bool IsExact = Shr->isExact();
2236 Type *ShrTy = Shr->getType();
2237 // TODO: If we could guarantee that InstSimplify would handle all of the
2238 // constant-value-based preconditions in the folds below, then we could assert
2239 // those conditions rather than checking them. This is difficult because of
2240 // undef/poison (PR34838).
2241 if (IsAShr) {
2242 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2243 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2244 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2245 APInt ShiftedC = C.shl(ShAmtVal);
2246 if (ShiftedC.ashr(ShAmtVal) == C)
2247 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2248 }
2249 if (Pred == CmpInst::ICMP_SGT) {
2250 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2251 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2252 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2253 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2254 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2255 }
2256 } else {
2257 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2258 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2259 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2260 APInt ShiftedC = C.shl(ShAmtVal);
2261 if (ShiftedC.lshr(ShAmtVal) == C)
2262 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2263 }
2264 if (Pred == CmpInst::ICMP_UGT) {
2265 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2266 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2267 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2268 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2269 }
2270 }
2271
2272 if (!Cmp.isEquality())
2273 return nullptr;
2274
2275 // Handle equality comparisons of shift-by-constant.
2276
2277 // If the comparison constant changes with the shift, the comparison cannot
2278 // succeed (bits of the comparison constant cannot match the shifted value).
2279 // This should be known by InstSimplify and already be folded to true/false.
2280 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||((((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
(!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
"Expected icmp+shr simplify did not occur.") ? static_cast<
void> (0) : __assert_fail ("((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && \"Expected icmp+shr simplify did not occur.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2282, __PRETTY_FUNCTION__))
2281 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&((((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
(!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
"Expected icmp+shr simplify did not occur.") ? static_cast<
void> (0) : __assert_fail ("((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && \"Expected icmp+shr simplify did not occur.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2282, __PRETTY_FUNCTION__))
2282 "Expected icmp+shr simplify did not occur.")((((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
(!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
"Expected icmp+shr simplify did not occur.") ? static_cast<
void> (0) : __assert_fail ("((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && \"Expected icmp+shr simplify did not occur.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2282, __PRETTY_FUNCTION__))
;
2283
2284 // If the bits shifted out are known zero, compare the unshifted value:
2285 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2286 if (Shr->isExact())
2287 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2288
2289 if (Shr->hasOneUse()) {
2290 // Canonicalize the shift into an 'and':
2291 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2292 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2293 Constant *Mask = ConstantInt::get(ShrTy, Val);
2294 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2295 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2296 }
2297
2298 return nullptr;
2299}
2300
2301Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2302 BinaryOperator *SRem,
2303 const APInt &C) {
2304 // Match an 'is positive' or 'is negative' comparison of remainder by a
2305 // constant power-of-2 value:
2306 // (X % pow2C) sgt/slt 0
2307 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2308 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2309 return nullptr;
2310
2311 // TODO: The one-use check is standard because we do not typically want to
2312 // create longer instruction sequences, but this might be a special-case
2313 // because srem is not good for analysis or codegen.
2314 if (!SRem->hasOneUse())
2315 return nullptr;
2316
2317 const APInt *DivisorC;
2318 if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2319 return nullptr;
2320
2321 // Mask off the sign bit and the modulo bits (low-bits).
2322 Type *Ty = SRem->getType();
2323 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2324 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2325 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2326
2327 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2328 // bit is set. Example:
2329 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2330 if (Pred == ICmpInst::ICMP_SGT)
2331 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2332
2333 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2334 // bit is set. Example:
2335 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2336 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2337}
2338
2339/// Fold icmp (udiv X, Y), C.
2340Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2341 BinaryOperator *UDiv,
2342 const APInt &C) {
2343 const APInt *C2;
2344 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2345 return nullptr;
2346
2347 assert(*C2 != 0 && "udiv 0, X should have been simplified already.")((*C2 != 0 && "udiv 0, X should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("*C2 != 0 && \"udiv 0, X should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2347, __PRETTY_FUNCTION__))
;
2348
2349 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2350 Value *Y = UDiv->getOperand(1);
2351 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2352 assert(!C.isMaxValue() &&((!C.isMaxValue() && "icmp ugt X, UINT_MAX should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("!C.isMaxValue() && \"icmp ugt X, UINT_MAX should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2353, __PRETTY_FUNCTION__))
2353 "icmp ugt X, UINT_MAX should have been simplified already.")((!C.isMaxValue() && "icmp ugt X, UINT_MAX should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("!C.isMaxValue() && \"icmp ugt X, UINT_MAX should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2353, __PRETTY_FUNCTION__))
;
2354 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2355 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2356 }
2357
2358 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2359 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2360 assert(C != 0 && "icmp ult X, 0 should have been simplified already.")((C != 0 && "icmp ult X, 0 should have been simplified already."
) ? static_cast<void> (0) : __assert_fail ("C != 0 && \"icmp ult X, 0 should have been simplified already.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2360, __PRETTY_FUNCTION__))
;
2361 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2362 ConstantInt::get(Y->getType(), C2->udiv(C)));
2363 }
2364
2365 return nullptr;
2366}
2367
2368/// Fold icmp ({su}div X, Y), C.
2369Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2370 BinaryOperator *Div,
2371 const APInt &C) {
2372 // Fold: icmp pred ([us]div X, C2), C -> range test
2373 // Fold this div into the comparison, producing a range check.
2374 // Determine, based on the divide type, what the range is being
2375 // checked. If there is an overflow on the low or high side, remember
2376 // it, otherwise compute the range [low, hi) bounding the new value.
2377 // See: InsertRangeTest above for the kinds of replacements possible.
2378 const APInt *C2;
2379 if (!match(Div->getOperand(1), m_APInt(C2)))
2380 return nullptr;
2381
2382 // FIXME: If the operand types don't match the type of the divide
2383 // then don't attempt this transform. The code below doesn't have the
2384 // logic to deal with a signed divide and an unsigned compare (and
2385 // vice versa). This is because (x /s C2) <s C produces different
2386 // results than (x /s C2) <u C or (x /u C2) <s C or even
2387 // (x /u C2) <u C. Simply casting the operands and result won't
2388 // work. :( The if statement below tests that condition and bails
2389 // if it finds it.
2390 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2391 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2392 return nullptr;
2393
2394 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2395 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2396 // division-by-constant cases should be present, we can not assert that they
2397 // have happened before we reach this icmp instruction.
2398 if (C2->isNullValue() || C2->isOneValue() ||
2399 (DivIsSigned && C2->isAllOnesValue()))
2400 return nullptr;
2401
2402 // Compute Prod = C * C2. We are essentially solving an equation of
2403 // form X / C2 = C. We solve for X by multiplying C2 and C.
2404 // By solving for X, we can turn this into a range check instead of computing
2405 // a divide.
2406 APInt Prod = C * *C2;
2407
2408 // Determine if the product overflows by seeing if the product is not equal to
2409 // the divide. Make sure we do the same kind of divide as in the LHS
2410 // instruction that we're folding.
2411 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2412
2413 ICmpInst::Predicate Pred = Cmp.getPredicate();
2414
2415 // If the division is known to be exact, then there is no remainder from the
2416 // divide, so the covered range size is unit, otherwise it is the divisor.
2417 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2418
2419 // Figure out the interval that is being checked. For example, a comparison
2420 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2421 // Compute this interval based on the constants involved and the signedness of
2422 // the compare/divide. This computes a half-open interval, keeping track of
2423 // whether either value in the interval overflows. After analysis each
2424 // overflow variable is set to 0 if it's corresponding bound variable is valid
2425 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2426 int LoOverflow = 0, HiOverflow = 0;
2427 APInt LoBound, HiBound;
2428
2429 if (!DivIsSigned) { // udiv
2430 // e.g. X/5 op 3 --> [15, 20)
2431 LoBound = Prod;
2432 HiOverflow = LoOverflow = ProdOV;
2433 if (!HiOverflow) {
2434 // If this is not an exact divide, then many values in the range collapse
2435 // to the same result value.
2436 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2437 }
2438 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2439 if (C.isNullValue()) { // (X / pos) op 0
2440 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2441 LoBound = -(RangeSize - 1);
2442 HiBound = RangeSize;
2443 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2444 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2445 HiOverflow = LoOverflow = ProdOV;
2446 if (!HiOverflow)
2447 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2448 } else { // (X / pos) op neg
2449 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2450 HiBound = Prod + 1;
2451 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2452 if (!LoOverflow) {
2453 APInt DivNeg = -RangeSize;
2454 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2455 }
2456 }
2457 } else if (C2->isNegative()) { // Divisor is < 0.
2458 if (Div->isExact())
2459 RangeSize.negate();
2460 if (C.isNullValue()) { // (X / neg) op 0
2461 // e.g. X/-5 op 0 --> [-4, 5)
2462 LoBound = RangeSize + 1;
2463 HiBound = -RangeSize;
2464 if (HiBound == *C2) { // -INTMIN = INTMIN
2465 HiOverflow = 1; // [INTMIN+1, overflow)
2466 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2467 }
2468 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2469 // e.g. X/-5 op 3 --> [-19, -14)
2470 HiBound = Prod + 1;
2471 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2472 if (!LoOverflow)
2473 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2474 } else { // (X / neg) op neg
2475 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2476 LoOverflow = HiOverflow = ProdOV;
2477 if (!HiOverflow)
2478 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2479 }
2480
2481 // Dividing by a negative swaps the condition. LT <-> GT
2482 Pred = ICmpInst::getSwappedPredicate(Pred);
2483 }
2484
2485 Value *X = Div->getOperand(0);
2486 switch (Pred) {
2487 default: llvm_unreachable("Unhandled icmp opcode!")::llvm::llvm_unreachable_internal("Unhandled icmp opcode!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2487)
;
2488 case ICmpInst::ICMP_EQ:
2489 if (LoOverflow && HiOverflow)
2490 return replaceInstUsesWith(Cmp, Builder.getFalse());
2491 if (HiOverflow)
2492 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2493 ICmpInst::ICMP_UGE, X,
2494 ConstantInt::get(Div->getType(), LoBound));
2495 if (LoOverflow)
2496 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2497 ICmpInst::ICMP_ULT, X,
2498 ConstantInt::get(Div->getType(), HiBound));
2499 return replaceInstUsesWith(
2500 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2501 case ICmpInst::ICMP_NE:
2502 if (LoOverflow && HiOverflow)
2503 return replaceInstUsesWith(Cmp, Builder.getTrue());
2504 if (HiOverflow)
2505 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2506 ICmpInst::ICMP_ULT, X,
2507 ConstantInt::get(Div->getType(), LoBound));
2508 if (LoOverflow)
2509 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2510 ICmpInst::ICMP_UGE, X,
2511 ConstantInt::get(Div->getType(), HiBound));
2512 return replaceInstUsesWith(Cmp,
2513 insertRangeTest(X, LoBound, HiBound,
2514 DivIsSigned, false));
2515 case ICmpInst::ICMP_ULT:
2516 case ICmpInst::ICMP_SLT:
2517 if (LoOverflow == +1) // Low bound is greater than input range.
2518 return replaceInstUsesWith(Cmp, Builder.getTrue());
2519 if (LoOverflow == -1) // Low bound is less than input range.
2520 return replaceInstUsesWith(Cmp, Builder.getFalse());
2521 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2522 case ICmpInst::ICMP_UGT:
2523 case ICmpInst::ICMP_SGT:
2524 if (HiOverflow == +1) // High bound greater than input range.
2525 return replaceInstUsesWith(Cmp, Builder.getFalse());
2526 if (HiOverflow == -1) // High bound less than input range.
2527 return replaceInstUsesWith(Cmp, Builder.getTrue());
2528 if (Pred == ICmpInst::ICMP_UGT)
2529 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2530 ConstantInt::get(Div->getType(), HiBound));
2531 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2532 ConstantInt::get(Div->getType(), HiBound));
2533 }
2534
2535 return nullptr;
2536}
2537
2538/// Fold icmp (sub X, Y), C.
2539Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2540 BinaryOperator *Sub,
2541 const APInt &C) {
2542 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2543 ICmpInst::Predicate Pred = Cmp.getPredicate();
2544 const APInt *C2;
2545 APInt SubResult;
2546
2547 // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2548 if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2549 return new ICmpInst(Cmp.getPredicate(), Y,
2550 ConstantInt::get(Y->getType(), 0));
2551
2552 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2553 if (match(X, m_APInt(C2)) &&
2554 ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2555 (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2556 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2557 return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2558 ConstantInt::get(Y->getType(), SubResult));
2559
2560 // The following transforms are only worth it if the only user of the subtract
2561 // is the icmp.
2562 if (!Sub->hasOneUse())
2563 return nullptr;
2564
2565 if (Sub->hasNoSignedWrap()) {
2566 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2567 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2568 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2569
2570 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2571 if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2572 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2573
2574 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2575 if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2576 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2577
2578 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2579 if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2580 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2581 }
2582
2583 if (!match(X, m_APInt(C2)))
2584 return nullptr;
2585
2586 // C2 - Y <u C -> (Y | (C - 1)) == C2
2587 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2588 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2589 (*C2 & (C - 1)) == (C - 1))
2590 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2591
2592 // C2 - Y >u C -> (Y | C) != C2
2593 // iff C2 & C == C and C + 1 is a power of 2
2594 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2595 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2596
2597 return nullptr;
2598}
2599
2600/// Fold icmp (add X, Y), C.
2601Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
2602 BinaryOperator *Add,
2603 const APInt &C) {
2604 Value *Y = Add->getOperand(1);
2605 const APInt *C2;
2606 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2607 return nullptr;
2608
2609 // Fold icmp pred (add X, C2), C.
2610 Value *X = Add->getOperand(0);
2611 Type *Ty = Add->getType();
2612 CmpInst::Predicate Pred = Cmp.getPredicate();
2613
2614 // If the add does not wrap, we can always adjust the compare by subtracting
2615 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2616 // are canonicalized to SGT/SLT/UGT/ULT.
2617 if ((Add->hasNoSignedWrap() &&
2618 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2619 (Add->hasNoUnsignedWrap() &&
2620 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2621 bool Overflow;
2622 APInt NewC =
2623 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2624 // If there is overflow, the result must be true or false.
2625 // TODO: Can we assert there is no overflow because InstSimplify always
2626 // handles those cases?
2627 if (!Overflow)
2628 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2629 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2630 }
2631
2632 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2633 const APInt &Upper = CR.getUpper();
2634 const APInt &Lower = CR.getLower();
2635 if (Cmp.isSigned()) {
2636 if (Lower.isSignMask())
2637 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2638 if (Upper.isSignMask())
2639 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2640 } else {
2641 if (Lower.isMinValue())
2642 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2643 if (Upper.isMinValue())
2644 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2645 }
2646
2647 if (!Add->hasOneUse())
2648 return nullptr;
2649
2650 // X+C <u C2 -> (X & -C2) == C
2651 // iff C & (C2-1) == 0
2652 // C2 is a power of 2
2653 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2654 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2655 ConstantExpr::getNeg(cast<Constant>(Y)));
2656
2657 // X+C >u C2 -> (X & ~C2) != C
2658 // iff C & C2 == 0
2659 // C2+1 is a power of 2
2660 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2661 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2662 ConstantExpr::getNeg(cast<Constant>(Y)));
2663
2664 return nullptr;
2665}
2666
2667bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2668 Value *&RHS, ConstantInt *&Less,
2669 ConstantInt *&Equal,
2670 ConstantInt *&Greater) {
2671 // TODO: Generalize this to work with other comparison idioms or ensure
2672 // they get canonicalized into this form.
2673
2674 // select i1 (a == b),
2675 // i32 Equal,
2676 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2677 // where Equal, Less and Greater are placeholders for any three constants.
2678 ICmpInst::Predicate PredA;
2679 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2680 !ICmpInst::isEquality(PredA))
2681 return false;
2682 Value *EqualVal = SI->getTrueValue();
2683 Value *UnequalVal = SI->getFalseValue();
2684 // We still can get non-canonical predicate here, so canonicalize.
2685 if (PredA == ICmpInst::ICMP_NE)
2686 std::swap(EqualVal, UnequalVal);
2687 if (!match(EqualVal, m_ConstantInt(Equal)))
2688 return false;
2689 ICmpInst::Predicate PredB;
2690 Value *LHS2, *RHS2;
2691 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2692 m_ConstantInt(Less), m_ConstantInt(Greater))))
2693 return false;
2694 // We can get predicate mismatch here, so canonicalize if possible:
2695 // First, ensure that 'LHS' match.
2696 if (LHS2 != LHS) {
2697 // x sgt y <--> y slt x
2698 std::swap(LHS2, RHS2);
2699 PredB = ICmpInst::getSwappedPredicate(PredB);
2700 }
2701 if (LHS2 != LHS)
2702 return false;
2703 // We also need to canonicalize 'RHS'.
2704 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2705 // x sgt C-1 <--> x sge C <--> not(x slt C)
2706 auto FlippedStrictness =
2707 InstCombiner::getFlippedStrictnessPredicateAndConstant(
2708 PredB, cast<Constant>(RHS2));
2709 if (!FlippedStrictness)
2710 return false;
2711 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check")((FlippedStrictness->first == ICmpInst::ICMP_SGE &&
"Sanity check") ? static_cast<void> (0) : __assert_fail
("FlippedStrictness->first == ICmpInst::ICMP_SGE && \"Sanity check\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2711, __PRETTY_FUNCTION__))
;
2712 RHS2 = FlippedStrictness->second;
2713 // And kind-of perform the result swap.
2714 std::swap(Less, Greater);
2715 PredB = ICmpInst::ICMP_SLT;
2716 }
2717 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2718}
2719
2720Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
2721 SelectInst *Select,
2722 ConstantInt *C) {
2723
2724 assert(C && "Cmp RHS should be a constant int!")((C && "Cmp RHS should be a constant int!") ? static_cast
<void> (0) : __assert_fail ("C && \"Cmp RHS should be a constant int!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2724, __PRETTY_FUNCTION__))
;
2725 // If we're testing a constant value against the result of a three way
2726 // comparison, the result can be expressed directly in terms of the
2727 // original values being compared. Note: We could possibly be more
2728 // aggressive here and remove the hasOneUse test. The original select is
2729 // really likely to simplify or sink when we remove a test of the result.
2730 Value *OrigLHS, *OrigRHS;
2731 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2732 if (Cmp.hasOneUse() &&
2733 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2734 C3GreaterThan)) {
2735 assert(C1LessThan && C2Equal && C3GreaterThan)((C1LessThan && C2Equal && C3GreaterThan) ? static_cast
<void> (0) : __assert_fail ("C1LessThan && C2Equal && C3GreaterThan"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 2735, __PRETTY_FUNCTION__))
;
2736
2737 bool TrueWhenLessThan =
2738 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2739 ->isAllOnesValue();
2740 bool TrueWhenEqual =
2741 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2742 ->isAllOnesValue();
2743 bool TrueWhenGreaterThan =
2744 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2745 ->isAllOnesValue();
2746
2747 // This generates the new instruction that will replace the original Cmp
2748 // Instruction. Instead of enumerating the various combinations when
2749 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2750 // false, we rely on chaining of ORs and future passes of InstCombine to
2751 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2752
2753 // When none of the three constants satisfy the predicate for the RHS (C),
2754 // the entire original Cmp can be simplified to a false.
2755 Value *Cond = Builder.getFalse();
2756 if (TrueWhenLessThan)
2757 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2758 OrigLHS, OrigRHS));
2759 if (TrueWhenEqual)
2760 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2761 OrigLHS, OrigRHS));
2762 if (TrueWhenGreaterThan)
2763 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2764 OrigLHS, OrigRHS));
2765
2766 return replaceInstUsesWith(Cmp, Cond);
2767 }
2768 return nullptr;
2769}
2770
2771static Instruction *foldICmpBitCast(ICmpInst &Cmp,
2772 InstCombiner::BuilderTy &Builder) {
2773 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2774 if (!Bitcast)
2775 return nullptr;
2776
2777 ICmpInst::Predicate Pred = Cmp.getPredicate();
2778 Value *Op1 = Cmp.getOperand(1);
2779 Value *BCSrcOp = Bitcast->getOperand(0);
2780
2781 // Make sure the bitcast doesn't change the number of vector elements.
2782 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2783 Bitcast->getDestTy()->getScalarSizeInBits()) {
2784 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2785 Value *X;
2786 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2787 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2788 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2789 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2790 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2791 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2792 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2793 match(Op1, m_Zero()))
2794 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2795
2796 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2797 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2798 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2799
2800 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2801 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2802 return new ICmpInst(Pred, X,
2803 ConstantInt::getAllOnesValue(X->getType()));
2804 }
2805
2806 // Zero-equality checks are preserved through unsigned floating-point casts:
2807 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2808 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2809 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2810 if (Cmp.isEquality() && match(Op1, m_Zero()))
2811 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2812
2813 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2814 // the FP extend/truncate because that cast does not change the sign-bit.
2815 // This is true for all standard IEEE-754 types and the X86 80-bit type.
2816 // The sign-bit is always the most significant bit in those types.
2817 const APInt *C;
2818 bool TrueIfSigned;
2819 if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2820 InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2821 if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2822 match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2823 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2824 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2825 Type *XType = X->getType();
2826
2827 // We can't currently handle Power style floating point operations here.
2828 if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2829
2830 Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2831 if (auto *XVTy = dyn_cast<VectorType>(XType))
2832 NewType = FixedVectorType::get(
2833 NewType, cast<FixedVectorType>(XVTy)->getNumElements());
2834 Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2835 if (TrueIfSigned)
2836 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2837 ConstantInt::getNullValue(NewType));
2838 else
2839 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2840 ConstantInt::getAllOnesValue(NewType));
2841 }
2842 }
2843 }
2844 }
2845
2846 // Test to see if the operands of the icmp are casted versions of other
2847 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2848 if (Bitcast->getType()->isPointerTy() &&
2849 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2850 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2851 // so eliminate it as well.
2852 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2853 Op1 = BC2->getOperand(0);
2854
2855 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2856 return new ICmpInst(Pred, BCSrcOp, Op1);
2857 }
2858
2859 // Folding: icmp <pred> iN X, C
2860 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2861 // and C is a splat of a K-bit pattern
2862 // and SC is a constant vector = <C', C', C', ..., C'>
2863 // Into:
2864 // %E = extractelement <M x iK> %vec, i32 C'
2865 // icmp <pred> iK %E, trunc(C)
2866 const APInt *C;
2867 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2868 !Bitcast->getType()->isIntegerTy() ||
2869 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2870 return nullptr;
2871
2872 Value *Vec;
2873 ArrayRef<int> Mask;
2874 if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2875 // Check whether every element of Mask is the same constant
2876 if (is_splat(Mask)) {
2877 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2878 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2879 if (C->isSplat(EltTy->getBitWidth())) {
2880 // Fold the icmp based on the value of C
2881 // If C is M copies of an iK sized bit pattern,
2882 // then:
2883 // => %E = extractelement <N x iK> %vec, i32 Elem
2884 // icmp <pred> iK %SplatVal, <pattern>
2885 Value *Elem = Builder.getInt32(Mask[0]);
2886 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2887 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2888 return new ICmpInst(Pred, Extract, NewC);
2889 }
2890 }
2891 }
2892 return nullptr;
2893}
2894
2895/// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2896/// where X is some kind of instruction.
2897Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
2898 const APInt *C;
2899 if (!match(Cmp.getOperand(1), m_APInt(C)))
2900 return nullptr;
2901
2902 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2903 switch (BO->getOpcode()) {
2904 case Instruction::Xor:
2905 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2906 return I;
2907 break;
2908 case Instruction::And:
2909 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2910 return I;
2911 break;
2912 case Instruction::Or:
2913 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2914 return I;
2915 break;
2916 case Instruction::Mul:
2917 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2918 return I;
2919 break;
2920 case Instruction::Shl:
2921 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2922 return I;
2923 break;
2924 case Instruction::LShr:
2925 case Instruction::AShr:
2926 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2927 return I;
2928 break;
2929 case Instruction::SRem:
2930 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2931 return I;
2932 break;
2933 case Instruction::UDiv:
2934 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2935 return I;
2936 LLVM_FALLTHROUGH[[gnu::fallthrough]];
2937 case Instruction::SDiv:
2938 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2939 return I;
2940 break;
2941 case Instruction::Sub:
2942 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2943 return I;
2944 break;
2945 case Instruction::Add:
2946 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2947 return I;
2948 break;
2949 default:
2950 break;
2951 }
2952 // TODO: These folds could be refactored to be part of the above calls.
2953 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2954 return I;
2955 }
2956
2957 // Match against CmpInst LHS being instructions other than binary operators.
2958
2959 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2960 // For now, we only support constant integers while folding the
2961 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2962 // similar to the cases handled by binary ops above.
2963 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2964 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2965 return I;
2966 }
2967
2968 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2969 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2970 return I;
2971 }
2972
2973 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2974 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2975 return I;
2976
2977 return nullptr;
2978}
2979
2980/// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2981/// icmp eq/ne BO, C.
2982Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
2983 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
2984 // TODO: Some of these folds could work with arbitrary constants, but this
2985 // function is limited to scalar and vector splat constants.
2986 if (!Cmp.isEquality())
2987 return nullptr;
2988
2989 ICmpInst::Predicate Pred = Cmp.getPredicate();
2990 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2991 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2992 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2993
2994 switch (BO->getOpcode()) {
2995 case Instruction::SRem:
2996 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2997 if (C.isNullValue() && BO->hasOneUse()) {
2998 const APInt *BOC;
2999 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3000 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3001 return new ICmpInst(Pred, NewRem,
3002 Constant::getNullValue(BO->getType()));
3003 }
3004 }
3005 break;
3006 case Instruction::Add: {
3007 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3008 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3009 if (BO->hasOneUse())
3010 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3011 } else if (C.isNullValue()) {
3012 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3013 // efficiently invertible, or if the add has just this one use.
3014 if (Value *NegVal = dyn_castNegVal(BOp1))
3015 return new ICmpInst(Pred, BOp0, NegVal);
3016 if (Value *NegVal = dyn_castNegVal(BOp0))
3017 return new ICmpInst(Pred, NegVal, BOp1);
3018 if (BO->hasOneUse()) {
3019 Value *Neg = Builder.CreateNeg(BOp1);
3020 Neg->takeName(BO);
3021 return new ICmpInst(Pred, BOp0, Neg);
3022 }
3023 }
3024 break;
3025 }
3026 case Instruction::Xor:
3027 if (BO->hasOneUse()) {
3028 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3029 // For the xor case, we can xor two constants together, eliminating
3030 // the explicit xor.
3031 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3032 } else if (C.isNullValue()) {
3033 // Replace ((xor A, B) != 0) with (A != B)
3034 return new ICmpInst(Pred, BOp0, BOp1);
3035 }
3036 }
3037 break;
3038 case Instruction::Sub:
3039 if (BO->hasOneUse()) {
3040 // Only check for constant LHS here, as constant RHS will be canonicalized
3041 // to add and use the fold above.
3042 if (Constant *BOC = dyn_cast<Constant>(BOp0)) {
3043 // Replace ((sub BOC, B) != C) with (B != BOC-C).
3044 return new ICmpInst(Pred, BOp1, ConstantExpr::getSub(BOC, RHS));
3045 } else if (C.isNullValue()) {
3046 // Replace ((sub A, B) != 0) with (A != B).
3047 return new ICmpInst(Pred, BOp0, BOp1);
3048 }
3049 }
3050 break;
3051 case Instruction::Or: {
3052 const APInt *BOC;
3053 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3054 // Comparing if all bits outside of a constant mask are set?
3055 // Replace (X | C) == -1 with (X & ~C) == ~C.
3056 // This removes the -1 constant.
3057 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3058 Value *And = Builder.CreateAnd(BOp0, NotBOC);
3059 return new ICmpInst(Pred, And, NotBOC);
3060 }
3061 break;
3062 }
3063 case Instruction::And: {
3064 const APInt *BOC;
3065 if (match(BOp1, m_APInt(BOC))) {
3066 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3067 if (C == *BOC && C.isPowerOf2())
3068 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3069 BO, Constant::getNullValue(RHS->getType()));
3070 }
3071 break;
3072 }
3073 case Instruction::UDiv:
3074 if (C.isNullValue()) {
3075 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3076 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3077 return new ICmpInst(NewPred, BOp1, BOp0);
3078 }
3079 break;
3080 default:
3081 break;
3082 }
3083 return nullptr;
3084}
3085
3086/// Fold an equality icmp with LLVM intrinsic and constant operand.
3087Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3088 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3089 Type *Ty = II->getType();
3090 unsigned BitWidth = C.getBitWidth();
3091 switch (II->getIntrinsicID()) {
3092 case Intrinsic::abs:
3093 // abs(A) == 0 -> A == 0
3094 // abs(A) == INT_MIN -> A == INT_MIN
3095 if (C.isNullValue() || C.isMinSignedValue())
3096 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3097 ConstantInt::get(Ty, C));
3098 break;
3099
3100 case Intrinsic::bswap:
3101 // bswap(A) == C -> A == bswap(C)
3102 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3103 ConstantInt::get(Ty, C.byteSwap()));
3104
3105 case Intrinsic::ctlz:
3106 case Intrinsic::cttz: {
3107 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3108 if (C == BitWidth)
3109 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3110 ConstantInt::getNullValue(Ty));
3111
3112 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3113 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3114 // Limit to one use to ensure we don't increase instruction count.
3115 unsigned Num = C.getLimitedValue(BitWidth);
3116 if (Num != BitWidth && II->hasOneUse()) {
3117 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3118 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3119 : APInt::getHighBitsSet(BitWidth, Num + 1);
3120 APInt Mask2 = IsTrailing
3121 ? APInt::getOneBitSet(BitWidth, Num)
3122 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3123 return new ICmpInst(Cmp.getPredicate(),
3124 Builder.CreateAnd(II->getArgOperand(0), Mask1),
3125 ConstantInt::get(Ty, Mask2));
3126 }
3127 break;
3128 }
3129
3130 case Intrinsic::ctpop: {
3131 // popcount(A) == 0 -> A == 0 and likewise for !=
3132 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3133 bool IsZero = C.isNullValue();
3134 if (IsZero || C == BitWidth)
3135 return new ICmpInst(Cmp.getPredicate(), II->getArgOperand(0),
3136 IsZero ? Constant::getNullValue(Ty) : Constant::getAllOnesValue(Ty));
3137
3138 break;
3139 }
3140
3141 case Intrinsic::uadd_sat: {
3142 // uadd.sat(a, b) == 0 -> (a | b) == 0
3143 if (C.isNullValue()) {
3144 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3145 return new ICmpInst(Cmp.getPredicate(), Or, Constant::getNullValue(Ty));
3146 }
3147 break;
3148 }
3149
3150 case Intrinsic::usub_sat: {
3151 // usub.sat(a, b) == 0 -> a <= b
3152 if (C.isNullValue()) {
3153 ICmpInst::Predicate NewPred = Cmp.getPredicate() == ICmpInst::ICMP_EQ
3154 ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3155 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3156 }
3157 break;
3158 }
3159 default:
3160 break;
3161 }
3162
3163 return nullptr;
3164}
3165
3166/// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3167Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3168 IntrinsicInst *II,
3169 const APInt &C) {
3170 if (Cmp.isEquality())
3171 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3172
3173 Type *Ty = II->getType();
3174 unsigned BitWidth = C.getBitWidth();
3175 switch (II->getIntrinsicID()) {
3176 case Intrinsic::ctlz: {
3177 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3178 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3179 unsigned Num = C.getLimitedValue();
3180 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3181 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3182 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3183 }
3184
3185 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3186 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3187 C.uge(1) && C.ule(BitWidth)) {
3188 unsigned Num = C.getLimitedValue();
3189 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3190 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3191 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3192 }
3193 break;
3194 }
3195 case Intrinsic::cttz: {
3196 // Limit to one use to ensure we don't increase instruction count.
3197 if (!II->hasOneUse())
3198 return nullptr;
3199
3200 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3201 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3202 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3203 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3204 Builder.CreateAnd(II->getArgOperand(0), Mask),
3205 ConstantInt::getNullValue(Ty));
3206 }
3207
3208 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3209 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3210 C.uge(1) && C.ule(BitWidth)) {
3211 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3212 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3213 Builder.CreateAnd(II->getArgOperand(0), Mask),
3214 ConstantInt::getNullValue(Ty));
3215 }
3216 break;
3217 }
3218 default:
3219 break;
3220 }
3221
3222 return nullptr;
3223}
3224
3225/// Handle icmp with constant (but not simple integer constant) RHS.
3226Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3227 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3228 Constant *RHSC = dyn_cast<Constant>(Op1);
3229 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3230 if (!RHSC || !LHSI)
3231 return nullptr;
3232
3233 switch (LHSI->getOpcode()) {
3234 case Instruction::GetElementPtr:
3235 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3236 if (RHSC->isNullValue() &&
3237 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3238 return new ICmpInst(
3239 I.getPredicate(), LHSI->getOperand(0),
3240 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3241 break;
3242 case Instruction::PHI:
3243 // Only fold icmp into the PHI if the phi and icmp are in the same
3244 // block. If in the same block, we're encouraging jump threading. If
3245 // not, we are just pessimizing the code by making an i1 phi.
3246 if (LHSI->getParent() == I.getParent())
3247 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3248 return NV;
3249 break;
3250 case Instruction::Select: {
3251 // If either operand of the select is a constant, we can fold the
3252 // comparison into the select arms, which will cause one to be
3253 // constant folded and the select turned into a bitwise or.
3254 Value *Op1 = nullptr, *Op2 = nullptr;
3255 ConstantInt *CI = nullptr;
3256 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3257 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3258 CI = dyn_cast<ConstantInt>(Op1);
3259 }
3260 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3261 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3262 CI = dyn_cast<ConstantInt>(Op2);
3263 }
3264
3265 // We only want to perform this transformation if it will not lead to
3266 // additional code. This is true if either both sides of the select
3267 // fold to a constant (in which case the icmp is replaced with a select
3268 // which will usually simplify) or this is the only user of the
3269 // select (in which case we are trading a select+icmp for a simpler
3270 // select+icmp) or all uses of the select can be replaced based on
3271 // dominance information ("Global cases").
3272 bool Transform = false;
3273 if (Op1 && Op2)
3274 Transform = true;
3275 else if (Op1 || Op2) {
3276 // Local case
3277 if (LHSI->hasOneUse())
3278 Transform = true;
3279 // Global cases
3280 else if (CI && !CI->isZero())
3281 // When Op1 is constant try replacing select with second operand.
3282 // Otherwise Op2 is constant and try replacing select with first
3283 // operand.
3284 Transform =
3285 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3286 }
3287 if (Transform) {
3288 if (!Op1)
3289 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3290 I.getName());
3291 if (!Op2)
3292 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3293 I.getName());
3294 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3295 }
3296 break;
3297 }
3298 case Instruction::IntToPtr:
3299 // icmp pred inttoptr(X), null -> icmp pred X, 0
3300 if (RHSC->isNullValue() &&
3301 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3302 return new ICmpInst(
3303 I.getPredicate(), LHSI->getOperand(0),
3304 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3305 break;
3306
3307 case Instruction::Load:
3308 // Try to optimize things like "A[i] > 4" to index computations.
3309 if (GetElementPtrInst *GEP =
3310 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3311 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3312 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3313 !cast<LoadInst>(LHSI)->isVolatile())
3314 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3315 return Res;
3316 }
3317 break;
3318 }
3319
3320 return nullptr;
3321}
3322
3323/// Some comparisons can be simplified.
3324/// In this case, we are looking for comparisons that look like
3325/// a check for a lossy truncation.
3326/// Folds:
3327/// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3328/// Where Mask is some pattern that produces all-ones in low bits:
3329/// (-1 >> y)
3330/// ((-1 << y) >> y) <- non-canonical, has extra uses
3331/// ~(-1 << y)
3332/// ((1 << y) + (-1)) <- non-canonical, has extra uses
3333/// The Mask can be a constant, too.
3334/// For some predicates, the operands are commutative.
3335/// For others, x can only be on a specific side.
3336static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3337 InstCombiner::BuilderTy &Builder) {
3338 ICmpInst::Predicate SrcPred;
3339 Value *X, *M, *Y;
3340 auto m_VariableMask = m_CombineOr(
3341 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3342 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3343 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3344 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3345 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3346 if (!match(&I, m_c_ICmp(SrcPred,
3347 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3348 m_Deferred(X))))
3349 return nullptr;
3350
3351 ICmpInst::Predicate DstPred;
3352 switch (SrcPred) {
3353 case ICmpInst::Predicate::ICMP_EQ:
3354 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3355 DstPred = ICmpInst::Predicate::ICMP_ULE;
3356 break;
3357 case ICmpInst::Predicate::ICMP_NE:
3358 // x & (-1 >> y) != x -> x u> (-1 >> y)
3359 DstPred = ICmpInst::Predicate::ICMP_UGT;
3360 break;
3361 case ICmpInst::Predicate::ICMP_ULT:
3362 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3363 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3364 DstPred = ICmpInst::Predicate::ICMP_UGT;
3365 break;
3366 case ICmpInst::Predicate::ICMP_UGE:
3367 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3368 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3369 DstPred = ICmpInst::Predicate::ICMP_ULE;
3370 break;
3371 case ICmpInst::Predicate::ICMP_SLT:
3372 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3373 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3374 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3375 return nullptr;
3376 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3377 return nullptr;
3378 DstPred = ICmpInst::Predicate::ICMP_SGT;
3379 break;
3380 case ICmpInst::Predicate::ICMP_SGE:
3381 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3382 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3383 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3384 return nullptr;
3385 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3386 return nullptr;
3387 DstPred = ICmpInst::Predicate::ICMP_SLE;
3388 break;
3389 case ICmpInst::Predicate::ICMP_SGT:
3390 case ICmpInst::Predicate::ICMP_SLE:
3391 return nullptr;
3392 case ICmpInst::Predicate::ICMP_UGT:
3393 case ICmpInst::Predicate::ICMP_ULE:
3394 llvm_unreachable("Instsimplify took care of commut. variant")::llvm::llvm_unreachable_internal("Instsimplify took care of commut. variant"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3394)
;
3395 break;
3396 default:
3397 llvm_unreachable("All possible folds are handled.")::llvm::llvm_unreachable_internal("All possible folds are handled."
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3397)
;
3398 }
3399
3400 // The mask value may be a vector constant that has undefined elements. But it
3401 // may not be safe to propagate those undefs into the new compare, so replace
3402 // those elements by copying an existing, defined, and safe scalar constant.
3403 Type *OpTy = M->getType();
3404 auto *VecC = dyn_cast<Constant>(M);
3405 if (OpTy->isVectorTy() && VecC && VecC->containsUndefElement()) {
3406 auto *OpVTy = cast<FixedVectorType>(OpTy);
3407 Constant *SafeReplacementConstant = nullptr;
3408 for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3409 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3410 SafeReplacementConstant = VecC->getAggregateElement(i);
3411 break;
3412 }
3413 }
3414 assert(SafeReplacementConstant && "Failed to find undef replacement")((SafeReplacementConstant && "Failed to find undef replacement"
) ? static_cast<void> (0) : __assert_fail ("SafeReplacementConstant && \"Failed to find undef replacement\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3414, __PRETTY_FUNCTION__))
;
3415 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3416 }
3417
3418 return Builder.CreateICmp(DstPred, X, M);
3419}
3420
3421/// Some comparisons can be simplified.
3422/// In this case, we are looking for comparisons that look like
3423/// a check for a lossy signed truncation.
3424/// Folds: (MaskedBits is a constant.)
3425/// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3426/// Into:
3427/// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3428/// Where KeptBits = bitwidth(%x) - MaskedBits
3429static Value *
3430foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3431 InstCombiner::BuilderTy &Builder) {
3432 ICmpInst::Predicate SrcPred;
3433 Value *X;
3434 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3435 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3436 if (!match(&I, m_c_ICmp(SrcPred,
3437 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3438 m_APInt(C1))),
3439 m_Deferred(X))))
3440 return nullptr;
3441
3442 // Potential handling of non-splats: for each element:
3443 // * if both are undef, replace with constant 0.
3444 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3445 // * if both are not undef, and are different, bailout.
3446 // * else, only one is undef, then pick the non-undef one.
3447
3448 // The shift amount must be equal.
3449 if (*C0 != *C1)
3450 return nullptr;
3451 const APInt &MaskedBits = *C0;
3452 assert(MaskedBits != 0 && "shift by zero should be folded away already.")((MaskedBits != 0 && "shift by zero should be folded away already."
) ? static_cast<void> (0) : __assert_fail ("MaskedBits != 0 && \"shift by zero should be folded away already.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3452, __PRETTY_FUNCTION__))
;
3453
3454 ICmpInst::Predicate DstPred;
3455 switch (SrcPred) {
3456 case ICmpInst::Predicate::ICMP_EQ:
3457 // ((%x << MaskedBits) a>> MaskedBits) == %x
3458 // =>
3459 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3460 DstPred = ICmpInst::Predicate::ICMP_ULT;
3461 break;
3462 case ICmpInst::Predicate::ICMP_NE:
3463 // ((%x << MaskedBits) a>> MaskedBits) != %x
3464 // =>
3465 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3466 DstPred = ICmpInst::Predicate::ICMP_UGE;
3467 break;
3468 // FIXME: are more folds possible?
3469 default:
3470 return nullptr;
3471 }
3472
3473 auto *XType = X->getType();
3474 const unsigned XBitWidth = XType->getScalarSizeInBits();
3475 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3476 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched")((BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched"
) ? static_cast<void> (0) : __assert_fail ("BitWidth.ugt(MaskedBits) && \"shifts should leave some bits untouched\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3476, __PRETTY_FUNCTION__))
;
3477
3478 // KeptBits = bitwidth(%x) - MaskedBits
3479 const APInt KeptBits = BitWidth - MaskedBits;
3480 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable")((KeptBits.ugt(0) && KeptBits.ult(BitWidth) &&
"unreachable") ? static_cast<void> (0) : __assert_fail
("KeptBits.ugt(0) && KeptBits.ult(BitWidth) && \"unreachable\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3480, __PRETTY_FUNCTION__))
;
3481 // ICmpCst = (1 << KeptBits)
3482 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3483 assert(ICmpCst.isPowerOf2())((ICmpCst.isPowerOf2()) ? static_cast<void> (0) : __assert_fail
("ICmpCst.isPowerOf2()", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3483, __PRETTY_FUNCTION__))
;
3484 // AddCst = (1 << (KeptBits-1))
3485 const APInt AddCst = ICmpCst.lshr(1);
3486 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2())((AddCst.ult(ICmpCst) && AddCst.isPowerOf2()) ? static_cast
<void> (0) : __assert_fail ("AddCst.ult(ICmpCst) && AddCst.isPowerOf2()"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3486, __PRETTY_FUNCTION__))
;
3487
3488 // T0 = add %x, AddCst
3489 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3490 // T1 = T0 DstPred ICmpCst
3491 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3492
3493 return T1;
3494}
3495
3496// Given pattern:
3497// icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3498// we should move shifts to the same hand of 'and', i.e. rewrite as
3499// icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3500// We are only interested in opposite logical shifts here.
3501// One of the shifts can be truncated.
3502// If we can, we want to end up creating 'lshr' shift.
3503static Value *
3504foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3505 InstCombiner::BuilderTy &Builder) {
3506 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
32
Assuming the condition is false
34
Taking false branch
3507 !I.getOperand(0)->hasOneUse())
33
Assuming the condition is false
3508 return nullptr;
3509
3510 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3511
3512 // Look for an 'and' of two logical shifts, one of which may be truncated.
3513 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3514 Instruction *XShift, *MaybeTruncation, *YShift;
3515 if (!match(
35
Calling 'match<llvm::Value, llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 28, true>>'
43
Returning from 'match<llvm::Value, llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinOpPred_match<llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::class_match<llvm::Value>, llvm::PatternMatch::is_logical_shift_op>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 28, true>>'
44
Taking false branch
3516 I.getOperand(0),
3517 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3518 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3519 m_AnyLogicalShift, m_Instruction(YShift))),
3520 m_Instruction(MaybeTruncation)))))
3521 return nullptr;
3522
3523 // We potentially looked past 'trunc', but only when matching YShift,
3524 // therefore YShift must have the widest type.
3525 Instruction *WidestShift = YShift;
3526 // Therefore XShift must have the shallowest type.
3527 // Or they both have identical types if there was no truncation.
3528 Instruction *NarrowestShift = XShift;
3529
3530 Type *WidestTy = WidestShift->getType();
3531 Type *NarrowestTy = NarrowestShift->getType();
3532 assert(NarrowestTy == I.getOperand(0)->getType() &&((NarrowestTy == I.getOperand(0)->getType() && "We did not look past any shifts while matching XShift though."
) ? static_cast<void> (0) : __assert_fail ("NarrowestTy == I.getOperand(0)->getType() && \"We did not look past any shifts while matching XShift though.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3533, __PRETTY_FUNCTION__))
45
Assuming the condition is true
46
'?' condition is true
3533 "We did not look past any shifts while matching XShift though.")((NarrowestTy == I.getOperand(0)->getType() && "We did not look past any shifts while matching XShift though."
) ? static_cast<void> (0) : __assert_fail ("NarrowestTy == I.getOperand(0)->getType() && \"We did not look past any shifts while matching XShift though.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3533, __PRETTY_FUNCTION__))
;
3534 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
47
Assuming the condition is false
3535
3536 // If YShift is a 'lshr', swap the shifts around.
3537 if (match(YShift, m_LShr(m_Value(), m_Value())))
48
Taking false branch
3538 std::swap(XShift, YShift);
3539
3540 // The shifts must be in opposite directions.
3541 auto XShiftOpcode = XShift->getOpcode();
3542 if (XShiftOpcode == YShift->getOpcode())
49
Assuming the condition is false
50
Taking false branch
3543 return nullptr; // Do not care about same-direction shifts here.
3544
3545 Value *X, *XShAmt, *Y, *YShAmt;
51
'XShAmt' declared without an initial value
3546 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
52
Calling 'm_Value'
56
Returning from 'm_Value'
57
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
65
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'
66
Calling 'm_BinOp<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'
68
Returning from 'm_BinOp<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'
3547 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3548
3549 // If one of the values being shifted is a constant, then we will end with
3550 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3551 // however, we will need to ensure that we won't increase instruction count.
3552 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
69
Assuming 'X' is a 'Constant'
3553 // At least one of the hands of the 'and' should be one-use shift.
3554 if (!match(I.getOperand(0),
3555 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3556 return nullptr;
3557 if (HadTrunc) {
3558 // Due to the 'trunc', we will need to widen X. For that either the old
3559 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3560 if (!MaybeTruncation->hasOneUse() &&
3561 !NarrowestShift->getOperand(1)->hasOneUse())
3562 return nullptr;
3563 }
3564 }
3565
3566 // We have two shift amounts from two different shifts. The types of those
3567 // shift amounts may not match. If that's the case let's bailout now.
3568 if (XShAmt->getType() != YShAmt->getType())
70
Called C++ object pointer is uninitialized
3569 return nullptr;
3570
3571 // As input, we have the following pattern:
3572 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3573 // We want to rewrite that as:
3574 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3575 // While we know that originally (Q+K) would not overflow
3576 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3577 // shift amounts. so it may now overflow in smaller bitwidth.
3578 // To ensure that does not happen, we need to ensure that the total maximal
3579 // shift amount is still representable in that smaller bit width.
3580 unsigned MaximalPossibleTotalShiftAmount =
3581 (WidestTy->getScalarSizeInBits() - 1) +
3582 (NarrowestTy->getScalarSizeInBits() - 1);
3583 APInt MaximalRepresentableShiftAmount =
3584 APInt::getAllOnesValue(XShAmt->getType()->getScalarSizeInBits());
3585 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3586 return nullptr;
3587
3588 // Can we fold (XShAmt+YShAmt) ?
3589 auto *NewShAmt = dyn_cast_or_null<Constant>(
3590 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3591 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3592 if (!NewShAmt)
3593 return nullptr;
3594 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3595 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3596
3597 // Is the new shift amount smaller than the bit width?
3598 // FIXME: could also rely on ConstantRange.
3599 if (!match(NewShAmt,
3600 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3601 APInt(WidestBitWidth, WidestBitWidth))))
3602 return nullptr;
3603
3604 // An extra legality check is needed if we had trunc-of-lshr.
3605 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3606 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3607 WidestShift]() {
3608 // It isn't obvious whether it's worth it to analyze non-constants here.
3609 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3610 // If *any* of these preconditions matches we can perform the fold.
3611 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3612 ? NewShAmt->getSplatValue()
3613 : NewShAmt;
3614 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3615 if (NewShAmtSplat &&
3616 (NewShAmtSplat->isNullValue() ||
3617 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3618 return true;
3619 // We consider *min* leading zeros so a single outlier
3620 // blocks the transform as opposed to allowing it.
3621 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3622 KnownBits Known = computeKnownBits(C, SQ.DL);
3623 unsigned MinLeadZero = Known.countMinLeadingZeros();
3624 // If the value being shifted has at most lowest bit set we can fold.
3625 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3626 if (MaxActiveBits <= 1)
3627 return true;
3628 // Precondition: NewShAmt u<= countLeadingZeros(C)
3629 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3630 return true;
3631 }
3632 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3633 KnownBits Known = computeKnownBits(C, SQ.DL);
3634 unsigned MinLeadZero = Known.countMinLeadingZeros();
3635 // If the value being shifted has at most lowest bit set we can fold.
3636 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3637 if (MaxActiveBits <= 1)
3638 return true;
3639 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3640 if (NewShAmtSplat) {
3641 APInt AdjNewShAmt =
3642 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3643 if (AdjNewShAmt.ule(MinLeadZero))
3644 return true;
3645 }
3646 }
3647 return false; // Can't tell if it's ok.
3648 };
3649 if (!CanFold())
3650 return nullptr;
3651 }
3652
3653 // All good, we can do this fold.
3654 X = Builder.CreateZExt(X, WidestTy);
3655 Y = Builder.CreateZExt(Y, WidestTy);
3656 // The shift is the same that was for X.
3657 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3658 ? Builder.CreateLShr(X, NewShAmt)
3659 : Builder.CreateShl(X, NewShAmt);
3660 Value *T1 = Builder.CreateAnd(T0, Y);
3661 return Builder.CreateICmp(I.getPredicate(), T1,
3662 Constant::getNullValue(WidestTy));
3663}
3664
3665/// Fold
3666/// (-1 u/ x) u< y
3667/// ((x * y) u/ x) != y
3668/// to
3669/// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3670/// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3671/// will mean that we are looking for the opposite answer.
3672Value *InstCombinerImpl::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3673 ICmpInst::Predicate Pred;
3674 Value *X, *Y;
3675 Instruction *Mul;
3676 bool NeedNegation;
3677 // Look for: (-1 u/ x) u</u>= y
3678 if (!I.isEquality() &&
3679 match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3680 m_Value(Y)))) {
3681 Mul = nullptr;
3682
3683 // Are we checking that overflow does not happen, or does happen?
3684 switch (Pred) {
3685 case ICmpInst::Predicate::ICMP_ULT:
3686 NeedNegation = false;
3687 break; // OK
3688 case ICmpInst::Predicate::ICMP_UGE:
3689 NeedNegation = true;
3690 break; // OK
3691 default:
3692 return nullptr; // Wrong predicate.
3693 }
3694 } else // Look for: ((x * y) u/ x) !=/== y
3695 if (I.isEquality() &&
3696 match(&I, m_c_ICmp(Pred, m_Value(Y),
3697 m_OneUse(m_UDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3698 m_Value(X)),
3699 m_Instruction(Mul)),
3700 m_Deferred(X)))))) {
3701 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3702 } else
3703 return nullptr;
3704
3705 BuilderTy::InsertPointGuard Guard(Builder);
3706 // If the pattern included (x * y), we'll want to insert new instructions
3707 // right before that original multiplication so that we can replace it.
3708 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3709 if (MulHadOtherUses)
3710 Builder.SetInsertPoint(Mul);
3711
3712 Function *F = Intrinsic::getDeclaration(
3713 I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3714 CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3715
3716 // If the multiplication was used elsewhere, to ensure that we don't leave
3717 // "duplicate" instructions, replace uses of that original multiplication
3718 // with the multiplication result from the with.overflow intrinsic.
3719 if (MulHadOtherUses)
3720 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3721
3722 Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3723 if (NeedNegation) // This technically increases instruction count.
3724 Res = Builder.CreateNot(Res, "umul.not.ov");
3725
3726 // If we replaced the mul, erase it. Do this after all uses of Builder,
3727 // as the mul is used as insertion point.
3728 if (MulHadOtherUses)
3729 eraseInstFromFunction(*Mul);
3730
3731 return Res;
3732}
3733
3734static Instruction *foldICmpXNegX(ICmpInst &I) {
3735 CmpInst::Predicate Pred;
3736 Value *X;
3737 if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3738 return nullptr;
3739
3740 if (ICmpInst::isSigned(Pred))
3741 Pred = ICmpInst::getSwappedPredicate(Pred);
3742 else if (ICmpInst::isUnsigned(Pred))
3743 Pred = ICmpInst::getSignedPredicate(Pred);
3744 // else for equality-comparisons just keep the predicate.
3745
3746 return ICmpInst::Create(Instruction::ICmp, Pred, X,
3747 Constant::getNullValue(X->getType()), I.getName());
3748}
3749
3750/// Try to fold icmp (binop), X or icmp X, (binop).
3751/// TODO: A large part of this logic is duplicated in InstSimplify's
3752/// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3753/// duplication.
3754Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
3755 const SimplifyQuery &SQ) {
3756 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3757 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3758
3759 // Special logic for binary operators.
3760 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
1
Assuming 'Op0' is a 'BinaryOperator'
3761 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2
Assuming 'Op1' is not a 'BinaryOperator'
3762 if (!BO0
2.1
'BO0' is non-null
2.1
'BO0' is non-null
&& !BO1)
3763 return nullptr;
3764
3765 if (Instruction *NewICmp
2.2
'NewICmp' is null
2.2
'NewICmp' is null
= foldICmpXNegX(I))
3
Taking false branch
3766 return NewICmp;
3767
3768 const CmpInst::Predicate Pred = I.getPredicate();
3769 Value *X;
3770
3771 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3772 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3773 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
4
Taking false branch
3774 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3775 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3776 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3777 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
5
Taking false branch
3778 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3779 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3780
3781 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3782 if (BO0
5.1
'BO0' is non-null
5.1
'BO0' is non-null
&& isa<OverflowingBinaryOperator>(BO0))
6
Assuming 'BO0' is not a 'OverflowingBinaryOperator'
7
Taking false branch
3783 NoOp0WrapProblem =
3784 ICmpInst::isEquality(Pred) ||
3785 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3786 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3787 if (BO1
7.1
'BO1' is null
7.1
'BO1' is null
&& isa<OverflowingBinaryOperator>(BO1))
3788 NoOp1WrapProblem =
3789 ICmpInst::isEquality(Pred) ||
3790 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3791 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3792
3793 // Analyze the case when either Op0 or Op1 is an add instruction.
3794 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3795 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3796 if (BO0
7.2
'BO0' is non-null
7.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Add) {
8
Assuming the condition is false
9
Taking false branch
3797 A = BO0->getOperand(0);
3798 B = BO0->getOperand(1);
3799 }
3800 if (BO1
9.1
'BO1' is null
9.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Add) {
3801 C = BO1->getOperand(0);
3802 D = BO1->getOperand(1);
3803 }
3804
3805 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3806 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3807 if ((A
9.2
'A' is not equal to 'Op1'
9.2
'A' is not equal to 'Op1'
== Op1 || B
9.3
'B' is not equal to 'Op1'
9.3
'B' is not equal to 'Op1'
== Op1) && NoOp0WrapProblem)
3808 return new ICmpInst(Pred, A == Op1 ? B : A,
3809 Constant::getNullValue(Op1->getType()));
3810
3811 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3812 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3813 if ((C
9.4
'C' is not equal to 'Op0'
9.4
'C' is not equal to 'Op0'
== Op0 || D
9.5
'D' is not equal to 'Op0'
9.5
'D' is not equal to 'Op0'
== Op0) && NoOp1WrapProblem)
3814 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3815 C == Op0 ? D : C);
3816
3817 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3818 if (A
9.6
'A' is null
9.6
'A' is null
&& C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3819 NoOp1WrapProblem) {
3820 // Determine Y and Z in the form icmp (X+Y), (X+Z).
3821 Value *Y, *Z;
3822 if (A == C) {
3823 // C + B == C + D -> B == D
3824 Y = B;
3825 Z = D;
3826 } else if (A == D) {
3827 // D + B == C + D -> B == C
3828 Y = B;
3829 Z = C;
3830 } else if (B == C) {
3831 // A + C == C + D -> A == D
3832 Y = A;
3833 Z = D;
3834 } else {
3835 assert(B == D)((B == D) ? static_cast<void> (0) : __assert_fail ("B == D"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 3835, __PRETTY_FUNCTION__))
;
3836 // A + D == C + D -> A == C
3837 Y = A;
3838 Z = C;
3839 }
3840 return new ICmpInst(Pred, Y, Z);
3841 }
3842
3843 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
3844 if (A
9.7
'A' is null
9.7
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
10
Taking false branch
3845 match(B, m_AllOnes()))
3846 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3847
3848 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
3849 if (A
10.1
'A' is null
10.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
11
Taking false branch
3850 match(B, m_AllOnes()))
3851 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3852
3853 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
3854 if (A
11.1
'A' is null
11.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
12
Taking false branch
3855 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3856
3857 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
3858 if (A
12.1
'A' is null
12.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
13
Taking false branch
3859 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3860
3861 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
3862 if (C
13.1
'C' is null
13.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
14
Taking false branch
3863 match(D, m_AllOnes()))
3864 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3865
3866 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
3867 if (C
14.1
'C' is null
14.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
15
Taking false branch
3868 match(D, m_AllOnes()))
3869 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3870
3871 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
3872 if (C
15.1
'C' is null
15.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
16
Taking false branch
3873 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3874
3875 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
3876 if (C
16.1
'C' is null
16.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
17
Taking false branch
3877 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3878
3879 // TODO: The subtraction-related identities shown below also hold, but
3880 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3881 // wouldn't happen even if they were implemented.
3882 //
3883 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
3884 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
3885 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
3886 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
3887
3888 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
3889 if (A
17.1
'A' is null
17.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
18
Taking false branch
3890 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3891
3892 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
3893 if (A
18.1
'A' is null
18.1
'A' is null
&& NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
19
Taking false branch
3894 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3895
3896 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
3897 if (C
19.1
'C' is null
19.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
20
Taking false branch
3898 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3899
3900 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
3901 if (C
20.1
'C' is null
20.1
'C' is null
&& NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
21
Taking false branch
3902 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3903
3904 // if C1 has greater magnitude than C2:
3905 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
3906 // s.t. C3 = C1 - C2
3907 //
3908 // if C2 has greater magnitude than C1:
3909 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
3910 // s.t. C3 = C2 - C1
3911 if (A
21.1
'A' is null
21.1
'A' is null
&& C && NoOp0WrapProblem && NoOp1WrapProblem &&
3912 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3913 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3914 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3915 const APInt &AP1 = C1->getValue();
3916 const APInt &AP2 = C2->getValue();
3917 if (AP1.isNegative() == AP2.isNegative()) {
3918 APInt AP1Abs = C1->getValue().abs();
3919 APInt AP2Abs = C2->getValue().abs();
3920 if (AP1Abs.uge(AP2Abs)) {
3921 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3922 Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3923 return new ICmpInst(Pred, NewAdd, C);
3924 } else {
3925 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3926 Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3927 return new ICmpInst(Pred, A, NewAdd);
3928 }
3929 }
3930 }
3931
3932 // Analyze the case when either Op0 or Op1 is a sub instruction.
3933 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3934 A = nullptr;
3935 B = nullptr;
3936 C = nullptr;
3937 D = nullptr;
3938 if (BO0
21.2
'BO0' is non-null
21.2
'BO0' is non-null
&& BO0->getOpcode() == Instruction::Sub) {
22
Assuming the condition is false
23
Taking false branch
3939 A = BO0->getOperand(0);
3940 B = BO0->getOperand(1);
3941 }
3942 if (BO1
23.1
'BO1' is null
23.1
'BO1' is null
&& BO1->getOpcode() == Instruction::Sub) {
3943 C = BO1->getOperand(0);
3944 D = BO1->getOperand(1);
3945 }
3946
3947 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
3948 if (A
23.2
'A' is not equal to 'Op1'
23.2
'A' is not equal to 'Op1'
== Op1 && NoOp0WrapProblem)
3949 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3950 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
3951 if (C
23.3
'C' is not equal to 'Op0'
23.3
'C' is not equal to 'Op0'
== Op0 && NoOp1WrapProblem)
3952 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3953
3954 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3955 // (A - B) u>/u<= A --> B u>/u<= A
3956 if (A
23.4
'A' is not equal to 'Op1'
23.4
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3957 return new ICmpInst(Pred, B, A);
3958 // C u</u>= (C - D) --> C u</u>= D
3959 if (C
23.5
'C' is not equal to 'Op0'
23.5
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3960 return new ICmpInst(Pred, C, D);
3961 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
3962 if (A
23.6
'A' is not equal to 'Op1'
23.6
'A' is not equal to 'Op1'
== Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
3963 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3964 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
3965 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
3966 if (C
23.7
'C' is not equal to 'Op0'
23.7
'C' is not equal to 'Op0'
== Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
3967 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
3968 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
3969
3970 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
3971 if (B
23.8
'B' is null
23.8
'B' is null
&& D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
3972 return new ICmpInst(Pred, A, C);
3973
3974 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
3975 if (A
23.9
'A' is null
23.9
'A' is null
&& C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
3976 return new ICmpInst(Pred, D, B);
3977
3978 // icmp (0-X) < cst --> x > -cst
3979 if (NoOp0WrapProblem
23.10
'NoOp0WrapProblem' is false
23.10
'NoOp0WrapProblem' is false
&& ICmpInst::isSigned(Pred)) {
3980 Value *X;
3981 if (match(BO0, m_Neg(m_Value(X))))
3982 if (Constant *RHSC = dyn_cast<Constant>(Op1))
3983 if (RHSC->isNotMinSignedValue())
3984 return new ICmpInst(I.getSwappedPredicate(), X,
3985 ConstantExpr::getNeg(RHSC));
3986 }
3987
3988 {
3989 // Try to remove shared constant multiplier from equality comparison:
3990 // X * C == Y * C (with no overflowing/aliasing) --> X == Y
3991 Value *X, *Y;
3992 const APInt *C;
3993 if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
24
Taking false branch
3994 match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
3995 if (!C->countTrailingZeros() ||
3996 (BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
3997 (BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
3998 return new ICmpInst(Pred, X, Y);
3999 }
4000
4001 BinaryOperator *SRem = nullptr;
4002 // icmp (srem X, Y), Y
4003 if (BO0
24.1
'BO0' is non-null
24.1
'BO0' is non-null
&& BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
25
Assuming the condition is false
4004 SRem = BO0;
4005 // icmp Y, (srem X, Y)
4006 else if (BO1
25.1
'BO1' is null
25.1
'BO1' is null
&& BO1->getOpcode() == Instruction::SRem &&
4007 Op0 == BO1->getOperand(1))
4008 SRem = BO1;
4009 if (SRem
25.2
'SRem' is null
25.2
'SRem' is null
) {
26
Taking false branch
4010 // We don't check hasOneUse to avoid increasing register pressure because
4011 // the value we use is the same value this instruction was already using.
4012 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4013 default:
4014 break;
4015 case ICmpInst::ICMP_EQ:
4016 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4017 case ICmpInst::ICMP_NE:
4018 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4019 case ICmpInst::ICMP_SGT:
4020 case ICmpInst::ICMP_SGE:
4021 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4022 Constant::getAllOnesValue(SRem->getType()));
4023 case ICmpInst::ICMP_SLT:
4024 case ICmpInst::ICMP_SLE:
4025 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4026 Constant::getNullValue(SRem->getType()));
4027 }
4028 }
4029
4030 if (BO0
26.1
'BO0' is non-null
26.1
'BO0' is non-null
&& BO1
26.2
'BO1' is null
26.2
'BO1' is null
&& BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4031 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4032 switch (BO0->getOpcode()) {
4033 default:
4034 break;
4035 case Instruction::Add:
4036 case Instruction::Sub:
4037 case Instruction::Xor: {
4038 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4039 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4040
4041 const APInt *C;
4042 if (match(BO0->getOperand(1), m_APInt(C))) {
4043 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4044 if (C->isSignMask()) {
4045 ICmpInst::Predicate NewPred =
4046 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
4047 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4048 }
4049
4050 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4051 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4052 ICmpInst::Predicate NewPred =
4053 I.isSigned() ? I.getUnsignedPredicate() : I.getSignedPredicate();
4054 NewPred = I.getSwappedPredicate(NewPred);
4055 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4056 }
4057 }
4058 break;
4059 }
4060 case Instruction::Mul: {
4061 if (!I.isEquality())
4062 break;
4063
4064 const APInt *C;
4065 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
4066 !C->isOneValue()) {
4067 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4068 // Mask = -1 >> count-trailing-zeros(C).
4069 if (unsigned TZs = C->countTrailingZeros()) {
4070 Constant *Mask = ConstantInt::get(
4071 BO0->getType(),
4072 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4073 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4074 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4075 return new ICmpInst(Pred, And1, And2);
4076 }
4077 }
4078 break;
4079 }
4080 case Instruction::UDiv:
4081 case Instruction::LShr:
4082 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4083 break;
4084 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4085
4086 case Instruction::SDiv:
4087 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4088 break;
4089 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4090
4091 case Instruction::AShr:
4092 if (!BO0->isExact() || !BO1->isExact())
4093 break;
4094 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4095
4096 case Instruction::Shl: {
4097 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4098 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4099 if (!NUW && !NSW)
4100 break;
4101 if (!NSW && I.isSigned())
4102 break;
4103 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4104 }
4105 }
4106 }
4107
4108 if (BO0
26.3
'BO0' is non-null
26.3
'BO0' is non-null
) {
27
Taking true branch
4109 // Transform A & (L - 1) `ult` L --> L != 0
4110 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4111 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4112
4113 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4114 auto *Zero = Constant::getNullValue(BO0->getType());
4115 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4116 }
4117 }
4118
4119 if (Value *V
27.1
'V' is null
27.1
'V' is null
= foldUnsignedMultiplicationOverflowCheck(I))
28
Taking false branch
4120 return replaceInstUsesWith(I, V);
4121
4122 if (Value *V
28.1
'V' is null
28.1
'V' is null
= foldICmpWithLowBitMaskedVal(I, Builder))
29
Taking false branch
4123 return replaceInstUsesWith(I, V);
4124
4125 if (Value *V
29.1
'V' is null
29.1
'V' is null
= foldICmpWithTruncSignExtendedVal(I, Builder))
30
Taking false branch
4126 return replaceInstUsesWith(I, V);
4127
4128 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
31
Calling 'foldShiftIntoShiftInAnotherHandOfAndInICmp'
4129 return replaceInstUsesWith(I, V);
4130
4131 return nullptr;
4132}
4133
4134/// Fold icmp Pred min|max(X, Y), X.
4135static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4136 ICmpInst::Predicate Pred = Cmp.getPredicate();
4137 Value *Op0 = Cmp.getOperand(0);
4138 Value *X = Cmp.getOperand(1);
4139
4140 // Canonicalize minimum or maximum operand to LHS of the icmp.
4141 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4142 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4143 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4144 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4145 std::swap(Op0, X);
4146 Pred = Cmp.getSwappedPredicate();
4147 }
4148
4149 Value *Y;
4150 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4151 // smin(X, Y) == X --> X s<= Y
4152 // smin(X, Y) s>= X --> X s<= Y
4153 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4154 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4155
4156 // smin(X, Y) != X --> X s> Y
4157 // smin(X, Y) s< X --> X s> Y
4158 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4159 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4160
4161 // These cases should be handled in InstSimplify:
4162 // smin(X, Y) s<= X --> true
4163 // smin(X, Y) s> X --> false
4164 return nullptr;
4165 }
4166
4167 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4168 // smax(X, Y) == X --> X s>= Y
4169 // smax(X, Y) s<= X --> X s>= Y
4170 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4171 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4172
4173 // smax(X, Y) != X --> X s< Y
4174 // smax(X, Y) s> X --> X s< Y
4175 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4176 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4177
4178 // These cases should be handled in InstSimplify:
4179 // smax(X, Y) s>= X --> true
4180 // smax(X, Y) s< X --> false
4181 return nullptr;
4182 }
4183
4184 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4185 // umin(X, Y) == X --> X u<= Y
4186 // umin(X, Y) u>= X --> X u<= Y
4187 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4188 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4189
4190 // umin(X, Y) != X --> X u> Y
4191 // umin(X, Y) u< X --> X u> Y
4192 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4193 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4194
4195 // These cases should be handled in InstSimplify:
4196 // umin(X, Y) u<= X --> true
4197 // umin(X, Y) u> X --> false
4198 return nullptr;
4199 }
4200
4201 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4202 // umax(X, Y) == X --> X u>= Y
4203 // umax(X, Y) u<= X --> X u>= Y
4204 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4205 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4206
4207 // umax(X, Y) != X --> X u< Y
4208 // umax(X, Y) u> X --> X u< Y
4209 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4210 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4211
4212 // These cases should be handled in InstSimplify:
4213 // umax(X, Y) u>= X --> true
4214 // umax(X, Y) u< X --> false
4215 return nullptr;
4216 }
4217
4218 return nullptr;
4219}
4220
4221Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
4222 if (!I.isEquality())
4223 return nullptr;
4224
4225 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4226 const CmpInst::Predicate Pred = I.getPredicate();
4227 Value *A, *B, *C, *D;
4228 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4229 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4230 Value *OtherVal = A == Op1 ? B : A;
4231 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4232 }
4233
4234 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4235 // A^c1 == C^c2 --> A == C^(c1^c2)
4236 ConstantInt *C1, *C2;
4237 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4238 Op1->hasOneUse()) {
4239 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4240 Value *Xor = Builder.CreateXor(C, NC);
4241 return new ICmpInst(Pred, A, Xor);
4242 }
4243
4244 // A^B == A^D -> B == D
4245 if (A == C)
4246 return new ICmpInst(Pred, B, D);
4247 if (A == D)
4248 return new ICmpInst(Pred, B, C);
4249 if (B == C)
4250 return new ICmpInst(Pred, A, D);
4251 if (B == D)
4252 return new ICmpInst(Pred, A, C);
4253 }
4254 }
4255
4256 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4257 // A == (A^B) -> B == 0
4258 Value *OtherVal = A == Op0 ? B : A;
4259 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4260 }
4261
4262 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4263 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4264 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4265 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4266
4267 if (A == C) {
4268 X = B;
4269 Y = D;
4270 Z = A;
4271 } else if (A == D) {
4272 X = B;
4273 Y = C;
4274 Z = A;
4275 } else if (B == C) {
4276 X = A;
4277 Y = D;
4278 Z = B;
4279 } else if (B == D) {
4280 X = A;
4281 Y = C;
4282 Z = B;
4283 }
4284
4285 if (X) { // Build (X^Y) & Z
4286 Op1 = Builder.CreateXor(X, Y);
4287 Op1 = Builder.CreateAnd(Op1, Z);
4288 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4289 }
4290 }
4291
4292 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4293 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4294 ConstantInt *Cst1;
4295 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4296 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4297 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4298 match(Op1, m_ZExt(m_Value(A))))) {
4299 APInt Pow2 = Cst1->getValue() + 1;
4300 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4301 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4302 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4303 }
4304
4305 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4306 // For lshr and ashr pairs.
4307 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4308 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4309 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4310 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4311 unsigned TypeBits = Cst1->getBitWidth();
4312 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4313 if (ShAmt < TypeBits && ShAmt != 0) {
4314 ICmpInst::Predicate NewPred =
4315 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4316 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4317 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4318 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4319 }
4320 }
4321
4322 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4323 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4324 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4325 unsigned TypeBits = Cst1->getBitWidth();
4326 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4327 if (ShAmt < TypeBits && ShAmt != 0) {
4328 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4329 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4330 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4331 I.getName() + ".mask");
4332 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4333 }
4334 }
4335
4336 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4337 // "icmp (and X, mask), cst"
4338 uint64_t ShAmt = 0;
4339 if (Op0->hasOneUse() &&
4340 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4341 match(Op1, m_ConstantInt(Cst1)) &&
4342 // Only do this when A has multiple uses. This is most important to do
4343 // when it exposes other optimizations.
4344 !A->hasOneUse()) {
4345 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4346
4347 if (ShAmt < ASize) {
4348 APInt MaskV =
4349 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4350 MaskV <<= ShAmt;
4351
4352 APInt CmpV = Cst1->getValue().zext(ASize);
4353 CmpV <<= ShAmt;
4354
4355 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4356 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4357 }
4358 }
4359
4360 // If both operands are byte-swapped or bit-reversed, just compare the
4361 // original values.
4362 // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4363 // and handle more intrinsics.
4364 if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4365 (match(Op0, m_BitReverse(m_Value(A))) &&
4366 match(Op1, m_BitReverse(m_Value(B)))))
4367 return new ICmpInst(Pred, A, B);
4368
4369 // Canonicalize checking for a power-of-2-or-zero value:
4370 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4371 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4372 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4373 m_Deferred(A)))) ||
4374 !match(Op1, m_ZeroInt()))
4375 A = nullptr;
4376
4377 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4378 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4379 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4380 A = Op1;
4381 else if (match(Op1,
4382 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4383 A = Op0;
4384
4385 if (A) {
4386 Type *Ty = A->getType();
4387 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4388 return Pred == ICmpInst::ICMP_EQ
4389 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4390 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4391 }
4392
4393 return nullptr;
4394}
4395
4396static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4397 InstCombiner::BuilderTy &Builder) {
4398 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0")((isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0"
) ? static_cast<void> (0) : __assert_fail ("isa<CastInst>(ICmp.getOperand(0)) && \"Expected cast for operand 0\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4398, __PRETTY_FUNCTION__))
;
4399 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4400 Value *X;
4401 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4402 return nullptr;
4403
4404 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4405 bool IsSignedCmp = ICmp.isSigned();
4406 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4407 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4408 // and the other is a zext), then we can't handle this.
4409 // TODO: This is too strict. We can handle some predicates (equality?).
4410 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4411 return nullptr;
4412
4413 // Not an extension from the same type?
4414 Value *Y = CastOp1->getOperand(0);
4415 Type *XTy = X->getType(), *YTy = Y->getType();
4416 if (XTy != YTy) {
4417 // One of the casts must have one use because we are creating a new cast.
4418 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4419 return nullptr;
4420 // Extend the narrower operand to the type of the wider operand.
4421 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4422 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4423 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4424 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4425 else
4426 return nullptr;
4427 }
4428
4429 // (zext X) == (zext Y) --> X == Y
4430 // (sext X) == (sext Y) --> X == Y
4431 if (ICmp.isEquality())
4432 return new ICmpInst(ICmp.getPredicate(), X, Y);
4433
4434 // A signed comparison of sign extended values simplifies into a
4435 // signed comparison.
4436 if (IsSignedCmp && IsSignedExt)
4437 return new ICmpInst(ICmp.getPredicate(), X, Y);
4438
4439 // The other three cases all fold into an unsigned comparison.
4440 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4441 }
4442
4443 // Below here, we are only folding a compare with constant.
4444 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4445 if (!C)
4446 return nullptr;
4447
4448 // Compute the constant that would happen if we truncated to SrcTy then
4449 // re-extended to DestTy.
4450 Type *SrcTy = CastOp0->getSrcTy();
4451 Type *DestTy = CastOp0->getDestTy();
4452 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4453 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4454
4455 // If the re-extended constant didn't change...
4456 if (Res2 == C) {
4457 if (ICmp.isEquality())
4458 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4459
4460 // A signed comparison of sign extended values simplifies into a
4461 // signed comparison.
4462 if (IsSignedExt && IsSignedCmp)
4463 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4464
4465 // The other three cases all fold into an unsigned comparison.
4466 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4467 }
4468
4469 // The re-extended constant changed, partly changed (in the case of a vector),
4470 // or could not be determined to be equal (in the case of a constant
4471 // expression), so the constant cannot be represented in the shorter type.
4472 // All the cases that fold to true or false will have already been handled
4473 // by SimplifyICmpInst, so only deal with the tricky case.
4474 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4475 return nullptr;
4476
4477 // Is source op positive?
4478 // icmp ult (sext X), C --> icmp sgt X, -1
4479 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4480 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4481
4482 // Is source op negative?
4483 // icmp ugt (sext X), C --> icmp slt X, 0
4484 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!")((ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"
) ? static_cast<void> (0) : __assert_fail ("ICmp.getPredicate() == ICmpInst::ICMP_UGT && \"ICmp should be folded!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4484, __PRETTY_FUNCTION__))
;
4485 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4486}
4487
4488/// Handle icmp (cast x), (cast or constant).
4489Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
4490 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4491 if (!CastOp0)
4492 return nullptr;
4493 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4494 return nullptr;
4495
4496 Value *Op0Src = CastOp0->getOperand(0);
4497 Type *SrcTy = CastOp0->getSrcTy();
4498 Type *DestTy = CastOp0->getDestTy();
4499
4500 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4501 // integer type is the same size as the pointer type.
4502 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4503 if (isa<VectorType>(SrcTy)) {
4504 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4505 DestTy = cast<VectorType>(DestTy)->getElementType();
4506 }
4507 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4508 };
4509 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4510 CompatibleSizes(SrcTy, DestTy)) {
4511 Value *NewOp1 = nullptr;
4512 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4513 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4514 if (PtrSrc->getType()->getPointerAddressSpace() ==
4515 Op0Src->getType()->getPointerAddressSpace()) {
4516 NewOp1 = PtrToIntOp1->getOperand(0);
4517 // If the pointer types don't match, insert a bitcast.
4518 if (Op0Src->getType() != NewOp1->getType())
4519 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4520 }
4521 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4522 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4523 }
4524
4525 if (NewOp1)
4526 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4527 }
4528
4529 return foldICmpWithZextOrSext(ICmp, Builder);
4530}
4531
4532static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4533 switch (BinaryOp) {
4534 default:
4535 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4535)
;
4536 case Instruction::Add:
4537 case Instruction::Sub:
4538 return match(RHS, m_Zero());
4539 case Instruction::Mul:
4540 return match(RHS, m_One());
4541 }
4542}
4543
4544OverflowResult
4545InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
4546 bool IsSigned, Value *LHS, Value *RHS,
4547 Instruction *CxtI) const {
4548 switch (BinaryOp) {
4549 default:
4550 llvm_unreachable("Unsupported binary op")::llvm::llvm_unreachable_internal("Unsupported binary op", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4550)
;
4551 case Instruction::Add:
4552 if (IsSigned)
4553 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4554 else
4555 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4556 case Instruction::Sub:
4557 if (IsSigned)
4558 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4559 else
4560 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4561 case Instruction::Mul:
4562 if (IsSigned)
4563 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4564 else
4565 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4566 }
4567}
4568
4569bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
4570 bool IsSigned, Value *LHS,
4571 Value *RHS, Instruction &OrigI,
4572 Value *&Result,
4573 Constant *&Overflow) {
4574 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4575 std::swap(LHS, RHS);
4576
4577 // If the overflow check was an add followed by a compare, the insertion point
4578 // may be pointing to the compare. We want to insert the new instructions
4579 // before the add in case there are uses of the add between the add and the
4580 // compare.
4581 Builder.SetInsertPoint(&OrigI);
4582
4583 if (isNeutralValue(BinaryOp, RHS)) {
4584 Result = LHS;
4585 Overflow = Builder.getFalse();
4586 return true;
4587 }
4588
4589 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4590 case OverflowResult::MayOverflow:
4591 return false;
4592 case OverflowResult::AlwaysOverflowsLow:
4593 case OverflowResult::AlwaysOverflowsHigh:
4594 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4595 Result->takeName(&OrigI);
4596 Overflow = Builder.getTrue();
4597 return true;
4598 case OverflowResult::NeverOverflows:
4599 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4600 Result->takeName(&OrigI);
4601 Overflow = Builder.getFalse();
4602 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4603 if (IsSigned)
4604 Inst->setHasNoSignedWrap();
4605 else
4606 Inst->setHasNoUnsignedWrap();
4607 }
4608 return true;
4609 }
4610
4611 llvm_unreachable("Unexpected overflow result")::llvm::llvm_unreachable_internal("Unexpected overflow result"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4611)
;
4612}
4613
4614/// Recognize and process idiom involving test for multiplication
4615/// overflow.
4616///
4617/// The caller has matched a pattern of the form:
4618/// I = cmp u (mul(zext A, zext B), V
4619/// The function checks if this is a test for overflow and if so replaces
4620/// multiplication with call to 'mul.with.overflow' intrinsic.
4621///
4622/// \param I Compare instruction.
4623/// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4624/// the compare instruction. Must be of integer type.
4625/// \param OtherVal The other argument of compare instruction.
4626/// \returns Instruction which must replace the compare instruction, NULL if no
4627/// replacement required.
4628static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4629 Value *OtherVal,
4630 InstCombinerImpl &IC) {
4631 // Don't bother doing this transformation for pointers, don't do it for
4632 // vectors.
4633 if (!isa<IntegerType>(MulVal->getType()))
4634 return nullptr;
4635
4636 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal)((I.getOperand(0) == MulVal || I.getOperand(1) == MulVal) ? static_cast
<void> (0) : __assert_fail ("I.getOperand(0) == MulVal || I.getOperand(1) == MulVal"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4636, __PRETTY_FUNCTION__))
;
4637 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal)((I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal)
? static_cast<void> (0) : __assert_fail ("I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4637, __PRETTY_FUNCTION__))
;
4638 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4639 if (!MulInstr)
4640 return nullptr;
4641 assert(MulInstr->getOpcode() == Instruction::Mul)((MulInstr->getOpcode() == Instruction::Mul) ? static_cast
<void> (0) : __assert_fail ("MulInstr->getOpcode() == Instruction::Mul"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4641, __PRETTY_FUNCTION__))
;
4642
4643 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4644 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4645 assert(LHS->getOpcode() == Instruction::ZExt)((LHS->getOpcode() == Instruction::ZExt) ? static_cast<
void> (0) : __assert_fail ("LHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4645, __PRETTY_FUNCTION__))
;
4646 assert(RHS->getOpcode() == Instruction::ZExt)((RHS->getOpcode() == Instruction::ZExt) ? static_cast<
void> (0) : __assert_fail ("RHS->getOpcode() == Instruction::ZExt"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4646, __PRETTY_FUNCTION__))
;
4647 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4648
4649 // Calculate type and width of the result produced by mul.with.overflow.
4650 Type *TyA = A->getType(), *TyB = B->getType();
4651 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4652 WidthB = TyB->getPrimitiveSizeInBits();
4653 unsigned MulWidth;
4654 Type *MulType;
4655 if (WidthB > WidthA) {
4656 MulWidth = WidthB;
4657 MulType = TyB;
4658 } else {
4659 MulWidth = WidthA;
4660 MulType = TyA;
4661 }
4662
4663 // In order to replace the original mul with a narrower mul.with.overflow,
4664 // all uses must ignore upper bits of the product. The number of used low
4665 // bits must be not greater than the width of mul.with.overflow.
4666 if (MulVal->hasNUsesOrMore(2))
4667 for (User *U : MulVal->users()) {
4668 if (U == &I)
4669 continue;
4670 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4671 // Check if truncation ignores bits above MulWidth.
4672 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4673 if (TruncWidth > MulWidth)
4674 return nullptr;
4675 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4676 // Check if AND ignores bits above MulWidth.
4677 if (BO->getOpcode() != Instruction::And)
4678 return nullptr;
4679 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4680 const APInt &CVal = CI->getValue();
4681 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4682 return nullptr;
4683 } else {
4684 // In this case we could have the operand of the binary operation
4685 // being defined in another block, and performing the replacement
4686 // could break the dominance relation.
4687 return nullptr;
4688 }
4689 } else {
4690 // Other uses prohibit this transformation.
4691 return nullptr;
4692 }
4693 }
4694
4695 // Recognize patterns
4696 switch (I.getPredicate()) {
4697 case ICmpInst::ICMP_EQ:
4698 case ICmpInst::ICMP_NE:
4699 // Recognize pattern:
4700 // mulval = mul(zext A, zext B)
4701 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4702 ConstantInt *CI;
4703 Value *ValToMask;
4704 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4705 if (ValToMask != MulVal)
4706 return nullptr;
4707 const APInt &CVal = CI->getValue() + 1;
4708 if (CVal.isPowerOf2()) {
4709 unsigned MaskWidth = CVal.logBase2();
4710 if (MaskWidth == MulWidth)
4711 break; // Recognized
4712 }
4713 }
4714 return nullptr;
4715
4716 case ICmpInst::ICMP_UGT:
4717 // Recognize pattern:
4718 // mulval = mul(zext A, zext B)
4719 // cmp ugt mulval, max
4720 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4721 APInt MaxVal = APInt::getMaxValue(MulWidth);
4722 MaxVal = MaxVal.zext(CI->getBitWidth());
4723 if (MaxVal.eq(CI->getValue()))
4724 break; // Recognized
4725 }
4726 return nullptr;
4727
4728 case ICmpInst::ICMP_UGE:
4729 // Recognize pattern:
4730 // mulval = mul(zext A, zext B)
4731 // cmp uge mulval, max+1
4732 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4733 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4734 if (MaxVal.eq(CI->getValue()))
4735 break; // Recognized
4736 }
4737 return nullptr;
4738
4739 case ICmpInst::ICMP_ULE:
4740 // Recognize pattern:
4741 // mulval = mul(zext A, zext B)
4742 // cmp ule mulval, max
4743 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4744 APInt MaxVal = APInt::getMaxValue(MulWidth);
4745 MaxVal = MaxVal.zext(CI->getBitWidth());
4746 if (MaxVal.eq(CI->getValue()))
4747 break; // Recognized
4748 }
4749 return nullptr;
4750
4751 case ICmpInst::ICMP_ULT:
4752 // Recognize pattern:
4753 // mulval = mul(zext A, zext B)
4754 // cmp ule mulval, max + 1
4755 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4756 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4757 if (MaxVal.eq(CI->getValue()))
4758 break; // Recognized
4759 }
4760 return nullptr;
4761
4762 default:
4763 return nullptr;
4764 }
4765
4766 InstCombiner::BuilderTy &Builder = IC.Builder;
4767 Builder.SetInsertPoint(MulInstr);
4768
4769 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4770 Value *MulA = A, *MulB = B;
4771 if (WidthA < MulWidth)
4772 MulA = Builder.CreateZExt(A, MulType);
4773 if (WidthB < MulWidth)
4774 MulB = Builder.CreateZExt(B, MulType);
4775 Function *F = Intrinsic::getDeclaration(
4776 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4777 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4778 IC.addToWorklist(MulInstr);
4779
4780 // If there are uses of mul result other than the comparison, we know that
4781 // they are truncation or binary AND. Change them to use result of
4782 // mul.with.overflow and adjust properly mask/size.
4783 if (MulVal->hasNUsesOrMore(2)) {
4784 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4785 for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4786 User *U = *UI++;
4787 if (U == &I || U == OtherVal)
4788 continue;
4789 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4790 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4791 IC.replaceInstUsesWith(*TI, Mul);
4792 else
4793 TI->setOperand(0, Mul);
4794 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4795 assert(BO->getOpcode() == Instruction::And)((BO->getOpcode() == Instruction::And) ? static_cast<void
> (0) : __assert_fail ("BO->getOpcode() == Instruction::And"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4795, __PRETTY_FUNCTION__))
;
4796 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4797 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4798 APInt ShortMask = CI->getValue().trunc(MulWidth);
4799 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4800 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
4801 IC.replaceInstUsesWith(*BO, Zext);
4802 } else {
4803 llvm_unreachable("Unexpected Binary operation")::llvm::llvm_unreachable_internal("Unexpected Binary operation"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4803)
;
4804 }
4805 IC.addToWorklist(cast<Instruction>(U));
4806 }
4807 }
4808 if (isa<Instruction>(OtherVal))
4809 IC.addToWorklist(cast<Instruction>(OtherVal));
4810
4811 // The original icmp gets replaced with the overflow value, maybe inverted
4812 // depending on predicate.
4813 bool Inverse = false;
4814 switch (I.getPredicate()) {
4815 case ICmpInst::ICMP_NE:
4816 break;
4817 case ICmpInst::ICMP_EQ:
4818 Inverse = true;
4819 break;
4820 case ICmpInst::ICMP_UGT:
4821 case ICmpInst::ICMP_UGE:
4822 if (I.getOperand(0) == MulVal)
4823 break;
4824 Inverse = true;
4825 break;
4826 case ICmpInst::ICMP_ULT:
4827 case ICmpInst::ICMP_ULE:
4828 if (I.getOperand(1) == MulVal)
4829 break;
4830 Inverse = true;
4831 break;
4832 default:
4833 llvm_unreachable("Unexpected predicate")::llvm::llvm_unreachable_internal("Unexpected predicate", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4833)
;
4834 }
4835 if (Inverse) {
4836 Value *Res = Builder.CreateExtractValue(Call, 1);
4837 return BinaryOperator::CreateNot(Res);
4838 }
4839
4840 return ExtractValueInst::Create(Call, 1);
4841}
4842
4843/// When performing a comparison against a constant, it is possible that not all
4844/// the bits in the LHS are demanded. This helper method computes the mask that
4845/// IS demanded.
4846static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4847 const APInt *RHS;
4848 if (!match(I.getOperand(1), m_APInt(RHS)))
4849 return APInt::getAllOnesValue(BitWidth);
4850
4851 // If this is a normal comparison, it demands all bits. If it is a sign bit
4852 // comparison, it only demands the sign bit.
4853 bool UnusedBit;
4854 if (InstCombiner::isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4855 return APInt::getSignMask(BitWidth);
4856
4857 switch (I.getPredicate()) {
4858 // For a UGT comparison, we don't care about any bits that
4859 // correspond to the trailing ones of the comparand. The value of these
4860 // bits doesn't impact the outcome of the comparison, because any value
4861 // greater than the RHS must differ in a bit higher than these due to carry.
4862 case ICmpInst::ICMP_UGT:
4863 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4864
4865 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4866 // Any value less than the RHS must differ in a higher bit because of carries.
4867 case ICmpInst::ICMP_ULT:
4868 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4869
4870 default:
4871 return APInt::getAllOnesValue(BitWidth);
4872 }
4873}
4874
4875/// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4876/// should be swapped.
4877/// The decision is based on how many times these two operands are reused
4878/// as subtract operands and their positions in those instructions.
4879/// The rationale is that several architectures use the same instruction for
4880/// both subtract and cmp. Thus, it is better if the order of those operands
4881/// match.
4882/// \return true if Op0 and Op1 should be swapped.
4883static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4884 // Filter out pointer values as those cannot appear directly in subtract.
4885 // FIXME: we may want to go through inttoptrs or bitcasts.
4886 if (Op0->getType()->isPointerTy())
4887 return false;
4888 // If a subtract already has the same operands as a compare, swapping would be
4889 // bad. If a subtract has the same operands as a compare but in reverse order,
4890 // then swapping is good.
4891 int GoodToSwap = 0;
4892 for (const User *U : Op0->users()) {
4893 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4894 GoodToSwap++;
4895 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4896 GoodToSwap--;
4897 }
4898 return GoodToSwap > 0;
4899}
4900
4901/// Check that one use is in the same block as the definition and all
4902/// other uses are in blocks dominated by a given block.
4903///
4904/// \param DI Definition
4905/// \param UI Use
4906/// \param DB Block that must dominate all uses of \p DI outside
4907/// the parent block
4908/// \return true when \p UI is the only use of \p DI in the parent block
4909/// and all other uses of \p DI are in blocks dominated by \p DB.
4910///
4911bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
4912 const Instruction *UI,
4913 const BasicBlock *DB) const {
4914 assert(DI && UI && "Instruction not defined\n")((DI && UI && "Instruction not defined\n") ? static_cast
<void> (0) : __assert_fail ("DI && UI && \"Instruction not defined\\n\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4914, __PRETTY_FUNCTION__))
;
4915 // Ignore incomplete definitions.
4916 if (!DI->getParent())
4917 return false;
4918 // DI and UI must be in the same block.
4919 if (DI->getParent() != UI->getParent())
4920 return false;
4921 // Protect from self-referencing blocks.
4922 if (DI->getParent() == DB)
4923 return false;
4924 for (const User *U : DI->users()) {
4925 auto *Usr = cast<Instruction>(U);
4926 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4927 return false;
4928 }
4929 return true;
4930}
4931
4932/// Return true when the instruction sequence within a block is select-cmp-br.
4933static bool isChainSelectCmpBranch(const SelectInst *SI) {
4934 const BasicBlock *BB = SI->getParent();
4935 if (!BB)
4936 return false;
4937 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4938 if (!BI || BI->getNumSuccessors() != 2)
4939 return false;
4940 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4941 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4942 return false;
4943 return true;
4944}
4945
4946/// True when a select result is replaced by one of its operands
4947/// in select-icmp sequence. This will eventually result in the elimination
4948/// of the select.
4949///
4950/// \param SI Select instruction
4951/// \param Icmp Compare instruction
4952/// \param SIOpd Operand that replaces the select
4953///
4954/// Notes:
4955/// - The replacement is global and requires dominator information
4956/// - The caller is responsible for the actual replacement
4957///
4958/// Example:
4959///
4960/// entry:
4961/// %4 = select i1 %3, %C* %0, %C* null
4962/// %5 = icmp eq %C* %4, null
4963/// br i1 %5, label %9, label %7
4964/// ...
4965/// ; <label>:7 ; preds = %entry
4966/// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4967/// ...
4968///
4969/// can be transformed to
4970///
4971/// %5 = icmp eq %C* %0, null
4972/// %6 = select i1 %3, i1 %5, i1 true
4973/// br i1 %6, label %9, label %7
4974/// ...
4975/// ; <label>:7 ; preds = %entry
4976/// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4977///
4978/// Similar when the first operand of the select is a constant or/and
4979/// the compare is for not equal rather than equal.
4980///
4981/// NOTE: The function is only called when the select and compare constants
4982/// are equal, the optimization can work only for EQ predicates. This is not a
4983/// major restriction since a NE compare should be 'normalized' to an equal
4984/// compare, which usually happens in the combiner and test case
4985/// select-cmp-br.ll checks for it.
4986bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
4987 const ICmpInst *Icmp,
4988 const unsigned SIOpd) {
4989 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!")(((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!"
) ? static_cast<void> (0) : __assert_fail ("(SIOpd == 1 || SIOpd == 2) && \"Invalid select operand!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 4989, __PRETTY_FUNCTION__))
;
4990 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4991 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4992 // The check for the single predecessor is not the best that can be
4993 // done. But it protects efficiently against cases like when SI's
4994 // home block has two successors, Succ and Succ1, and Succ1 predecessor
4995 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4996 // replaced can be reached on either path. So the uniqueness check
4997 // guarantees that the path all uses of SI (outside SI's parent) are on
4998 // is disjoint from all other paths out of SI. But that information
4999 // is more expensive to compute, and the trade-off here is in favor
5000 // of compile-time. It should also be noticed that we check for a single
5001 // predecessor and not only uniqueness. This to handle the situation when
5002 // Succ and Succ1 points to the same basic block.
5003 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
5004 NumSel++;
5005 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
5006 return true;
5007 }
5008 }
5009 return false;
5010}
5011
5012/// Try to fold the comparison based on range information we can get by checking
5013/// whether bits are known to be zero or one in the inputs.
5014Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
5015 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5016 Type *Ty = Op0->getType();
5017 ICmpInst::Predicate Pred = I.getPredicate();
5018
5019 // Get scalar or pointer size.
5020 unsigned BitWidth = Ty->isIntOrIntVectorTy()
5021 ? Ty->getScalarSizeInBits()
5022 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
5023
5024 if (!BitWidth)
5025 return nullptr;
5026
5027 KnownBits Op0Known(BitWidth);
5028 KnownBits Op1Known(BitWidth);
5029
5030 if (SimplifyDemandedBits(&I, 0,
5031 getDemandedBitsLHSMask(I, BitWidth),
5032 Op0Known, 0))
5033 return &I;
5034
5035 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
5036 Op1Known, 0))
5037 return &I;
5038
5039 // Given the known and unknown bits, compute a range that the LHS could be
5040 // in. Compute the Min, Max and RHS values based on the known bits. For the
5041 // EQ and NE we use unsigned values.
5042 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
5043 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
5044 if (I.isSigned()) {
5045 computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
5046 computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
5047 } else {
5048 computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
5049 computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
5050 }
5051
5052 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
5053 // out that the LHS or RHS is a constant. Constant fold this now, so that
5054 // code below can assume that Min != Max.
5055 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
5056 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
5057 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
5058 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
5059
5060 // Based on the range information we know about the LHS, see if we can
5061 // simplify this comparison. For example, (x&4) < 8 is always true.
5062 switch (Pred) {
5063 default:
5064 llvm_unreachable("Unknown icmp opcode!")::llvm::llvm_unreachable_internal("Unknown icmp opcode!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5064)
;
5065 case ICmpInst::ICMP_EQ:
5066 case ICmpInst::ICMP_NE: {
5067 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
5068 return Pred == CmpInst::ICMP_EQ
5069 ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
5070 : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5071 }
5072
5073 // If all bits are known zero except for one, then we know at most one bit
5074 // is set. If the comparison is against zero, then this is a check to see if
5075 // *that* bit is set.
5076 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
5077 if (Op1Known.isZero()) {
5078 // If the LHS is an AND with the same constant, look through it.
5079 Value *LHS = nullptr;
5080 const APInt *LHSC;
5081 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
5082 *LHSC != Op0KnownZeroInverted)
5083 LHS = Op0;
5084
5085 Value *X;
5086 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
5087 APInt ValToCheck = Op0KnownZeroInverted;
5088 Type *XTy = X->getType();
5089 if (ValToCheck.isPowerOf2()) {
5090 // ((1 << X) & 8) == 0 -> X != 3
5091 // ((1 << X) & 8) != 0 -> X == 3
5092 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5093 auto NewPred = ICmpInst::getInversePredicate(Pred);
5094 return new ICmpInst(NewPred, X, CmpC);
5095 } else if ((++ValToCheck).isPowerOf2()) {
5096 // ((1 << X) & 7) == 0 -> X >= 3
5097 // ((1 << X) & 7) != 0 -> X < 3
5098 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5099 auto NewPred =
5100 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
5101 return new ICmpInst(NewPred, X, CmpC);
5102 }
5103 }
5104
5105 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5106 const APInt *CI;
5107 if (Op0KnownZeroInverted.isOneValue() &&
5108 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5109 // ((8 >>u X) & 1) == 0 -> X != 3
5110 // ((8 >>u X) & 1) != 0 -> X == 3
5111 unsigned CmpVal = CI->countTrailingZeros();
5112 auto NewPred = ICmpInst::getInversePredicate(Pred);
5113 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5114 }
5115 }
5116 break;
5117 }
5118 case ICmpInst::ICMP_ULT: {
5119 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5120 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5121 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5122 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5123 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5124 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5125
5126 const APInt *CmpC;
5127 if (match(Op1, m_APInt(CmpC))) {
5128 // A <u C -> A == C-1 if min(A)+1 == C
5129 if (*CmpC == Op0Min + 1)
5130 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5131 ConstantInt::get(Op1->getType(), *CmpC - 1));
5132 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5133 // exceeds the log2 of C.
5134 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5135 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5136 Constant::getNullValue(Op1->getType()));
5137 }
5138 break;
5139 }
5140 case ICmpInst::ICMP_UGT: {
5141 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5142 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5143 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5144 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5145 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5146 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5147
5148 const APInt *CmpC;
5149 if (match(Op1, m_APInt(CmpC))) {
5150 // A >u C -> A == C+1 if max(a)-1 == C
5151 if (*CmpC == Op0Max - 1)
5152 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5153 ConstantInt::get(Op1->getType(), *CmpC + 1));
5154 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5155 // exceeds the log2 of C.
5156 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5157 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5158 Constant::getNullValue(Op1->getType()));
5159 }
5160 break;
5161 }
5162 case ICmpInst::ICMP_SLT: {
5163 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5164 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5165 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5166 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5167 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5168 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5169 const APInt *CmpC;
5170 if (match(Op1, m_APInt(CmpC))) {
5171 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5172 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5173 ConstantInt::get(Op1->getType(), *CmpC - 1));
5174 }
5175 break;
5176 }
5177 case ICmpInst::ICMP_SGT: {
5178 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5179 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5180 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5181 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5182 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5183 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5184 const APInt *CmpC;
5185 if (match(Op1, m_APInt(CmpC))) {
5186 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5187 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5188 ConstantInt::get(Op1->getType(), *CmpC + 1));
5189 }
5190 break;
5191 }
5192 case ICmpInst::ICMP_SGE:
5193 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_SGE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5193, __PRETTY_FUNCTION__))
;
5194 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5195 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5196 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5197 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5198 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5199 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5200 break;
5201 case ICmpInst::ICMP_SLE:
5202 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_SLE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5202, __PRETTY_FUNCTION__))
;
5203 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5204 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5205 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5206 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5207 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5208 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5209 break;
5210 case ICmpInst::ICMP_UGE:
5211 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_UGE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5211, __PRETTY_FUNCTION__))
;
5212 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5213 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5214 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5215 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5216 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5217 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5218 break;
5219 case ICmpInst::ICMP_ULE:
5220 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!")((!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!"
) ? static_cast<void> (0) : __assert_fail ("!isa<ConstantInt>(Op1) && \"ICMP_ULE with ConstantInt not folded!\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5220, __PRETTY_FUNCTION__))
;
5221 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5222 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5223 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5224 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5225 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5226 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5227 break;
5228 }
5229
5230 // Turn a signed comparison into an unsigned one if both operands are known to
5231 // have the same sign.
5232 if (I.isSigned() &&
5233 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5234 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5235 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5236
5237 return nullptr;
5238}
5239
5240llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
5241InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5242 Constant *C) {
5243 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&((ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate
(Pred) && "Only for relational integer predicates.") ?
static_cast<void> (0) : __assert_fail ("ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && \"Only for relational integer predicates.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5244, __PRETTY_FUNCTION__))
5244 "Only for relational integer predicates.")((ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate
(Pred) && "Only for relational integer predicates.") ?
static_cast<void> (0) : __assert_fail ("ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && \"Only for relational integer predicates.\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5244, __PRETTY_FUNCTION__))
;
5245
5246 Type *Type = C->getType();
5247 bool IsSigned = ICmpInst::isSigned(Pred);
5248
5249 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5250 bool WillIncrement =
5251 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5252
5253 // Check if the constant operand can be safely incremented/decremented
5254 // without overflowing/underflowing.
5255 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5256 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5257 };
5258
5259 Constant *SafeReplacementConstant = nullptr;
5260 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5261 // Bail out if the constant can't be safely incremented/decremented.
5262 if (!ConstantIsOk(CI))
5263 return llvm::None;
5264 } else if (auto *FVTy = dyn_cast<FixedVectorType>(Type)) {
5265 unsigned NumElts = FVTy->getNumElements();
5266 for (unsigned i = 0; i != NumElts; ++i) {
5267 Constant *Elt = C->getAggregateElement(i);
5268 if (!Elt)
5269 return llvm::None;
5270
5271 if (isa<UndefValue>(Elt))
5272 continue;
5273
5274 // Bail out if we can't determine if this constant is min/max or if we
5275 // know that this constant is min/max.
5276 auto *CI = dyn_cast<ConstantInt>(Elt);
5277 if (!CI || !ConstantIsOk(CI))
5278 return llvm::None;
5279
5280 if (!SafeReplacementConstant)
5281 SafeReplacementConstant = CI;
5282 }
5283 } else {
5284 // ConstantExpr?
5285 return llvm::None;
5286 }
5287
5288 // It may not be safe to change a compare predicate in the presence of
5289 // undefined elements, so replace those elements with the first safe constant
5290 // that we found.
5291 if (C->containsUndefElement()) {
5292 assert(SafeReplacementConstant && "Replacement constant not set")((SafeReplacementConstant && "Replacement constant not set"
) ? static_cast<void> (0) : __assert_fail ("SafeReplacementConstant && \"Replacement constant not set\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5292, __PRETTY_FUNCTION__))
;
5293 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5294 }
5295
5296 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5297
5298 // Increment or decrement the constant.
5299 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5300 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5301
5302 return std::make_pair(NewPred, NewC);
5303}
5304
5305/// If we have an icmp le or icmp ge instruction with a constant operand, turn
5306/// it into the appropriate icmp lt or icmp gt instruction. This transform
5307/// allows them to be folded in visitICmpInst.
5308static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5309 ICmpInst::Predicate Pred = I.getPredicate();
5310 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5311 InstCombiner::isCanonicalPredicate(Pred))
5312 return nullptr;
5313
5314 Value *Op0 = I.getOperand(0);
5315 Value *Op1 = I.getOperand(1);
5316 auto *Op1C = dyn_cast<Constant>(Op1);
5317 if (!Op1C)
5318 return nullptr;
5319
5320 auto FlippedStrictness =
5321 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5322 if (!FlippedStrictness)
5323 return nullptr;
5324
5325 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5326}
5327
5328/// If we have a comparison with a non-canonical predicate, if we can update
5329/// all the users, invert the predicate and adjust all the users.
5330CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
5331 // Is the predicate already canonical?
5332 CmpInst::Predicate Pred = I.getPredicate();
5333 if (InstCombiner::isCanonicalPredicate(Pred))
5334 return nullptr;
5335
5336 // Can all users be adjusted to predicate inversion?
5337 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
5338 return nullptr;
5339
5340 // Ok, we can canonicalize comparison!
5341 // Let's first invert the comparison's predicate.
5342 I.setPredicate(CmpInst::getInversePredicate(Pred));
5343 I.setName(I.getName() + ".not");
5344
5345 // And now let's adjust every user.
5346 for (User *U : I.users()) {
5347 switch (cast<Instruction>(U)->getOpcode()) {
5348 case Instruction::Select: {
5349 auto *SI = cast<SelectInst>(U);
5350 SI->swapValues();
5351 SI->swapProfMetadata();
5352 break;
5353 }
5354 case Instruction::Br:
5355 cast<BranchInst>(U)->swapSuccessors(); // swaps prof metadata too
5356 break;
5357 case Instruction::Xor:
5358 replaceInstUsesWith(cast<Instruction>(*U), &I);
5359 break;
5360 default:
5361 llvm_unreachable("Got unexpected user - out of sync with "::llvm::llvm_unreachable_internal("Got unexpected user - out of sync with "
"canFreelyInvertAllUsersOf() ?", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5362)
5362 "canFreelyInvertAllUsersOf() ?")::llvm::llvm_unreachable_internal("Got unexpected user - out of sync with "
"canFreelyInvertAllUsersOf() ?", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5362)
;
5363 }
5364 }
5365
5366 return &I;
5367}
5368
5369/// Integer compare with boolean values can always be turned into bitwise ops.
5370static Instruction *canonicalizeICmpBool(ICmpInst &I,
5371 InstCombiner::BuilderTy &Builder) {
5372 Value *A = I.getOperand(0), *B = I.getOperand(1);
5373 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only")((A->getType()->isIntOrIntVectorTy(1) && "Bools only"
) ? static_cast<void> (0) : __assert_fail ("A->getType()->isIntOrIntVectorTy(1) && \"Bools only\""
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5373, __PRETTY_FUNCTION__))
;
5374
5375 // A boolean compared to true/false can be simplified to Op0/true/false in
5376 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5377 // Cases not handled by InstSimplify are always 'not' of Op0.
5378 if (match(B, m_Zero())) {
5379 switch (I.getPredicate()) {
5380 case CmpInst::ICMP_EQ: // A == 0 -> !A
5381 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5382 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5383 return BinaryOperator::CreateNot(A);
5384 default:
5385 llvm_unreachable("ICmp i1 X, C not simplified as expected.")::llvm::llvm_unreachable_internal("ICmp i1 X, C not simplified as expected."
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5385)
;
5386 }
5387 } else if (match(B, m_One())) {
5388 switch (I.getPredicate()) {
5389 case CmpInst::ICMP_NE: // A != 1 -> !A
5390 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5391 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5392 return BinaryOperator::CreateNot(A);
5393 default:
5394 llvm_unreachable("ICmp i1 X, C not simplified as expected.")::llvm::llvm_unreachable_internal("ICmp i1 X, C not simplified as expected."
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5394)
;
5395 }
5396 }
5397
5398 switch (I.getPredicate()) {
5399 default:
5400 llvm_unreachable("Invalid icmp instruction!")::llvm::llvm_unreachable_internal("Invalid icmp instruction!"
, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp"
, 5400)
;
5401 case ICmpInst::ICMP_EQ:
5402 // icmp eq i1 A, B -> ~(A ^ B)
5403 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5404
5405 case ICmpInst::ICMP_NE:
5406 // icmp ne i1 A, B -> A ^ B
5407 return BinaryOperator::CreateXor(A, B);
5408
5409 case ICmpInst::ICMP_UGT:
5410 // icmp ugt -> icmp ult
5411 std::swap(A, B);
5412 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5413 case ICmpInst::ICMP_ULT:
5414 // icmp ult i1 A, B -> ~A & B
5415 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5416
5417 case ICmpInst::ICMP_SGT:
5418 // icmp sgt -> icmp slt
5419 std::swap(A, B);
5420 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5421 case ICmpInst::ICMP_SLT:
5422 // icmp slt i1 A, B -> A & ~B
5423 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5424
5425 case ICmpInst::ICMP_UGE:
5426 // icmp uge -> icmp ule
5427 std::swap(A, B);
5428 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5429 case ICmpInst::ICMP_ULE:
5430 // icmp ule i1 A, B -> ~A | B
5431 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5432
5433 case ICmpInst::ICMP_SGE:
5434 // icmp sge -> icmp sle
5435 std::swap(A, B);
5436 LLVM_FALLTHROUGH[[gnu::fallthrough]];
5437 case ICmpInst::ICMP_SLE:
5438 // icmp sle i1 A, B -> A | ~B
5439 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5440 }
5441}
5442
5443// Transform pattern like:
5444// (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5445// (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5446// Into:
5447// (X l>> Y) != 0
5448// (X l>> Y) == 0
5449static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5450 InstCombiner::BuilderTy &Builder) {
5451 ICmpInst::Predicate Pred, NewPred;
5452 Value *X, *Y;
5453 if (match(&Cmp,
5454 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5455 switch (Pred) {
5456 case ICmpInst::ICMP_ULE:
5457 NewPred = ICmpInst::ICMP_NE;
5458 break;
5459 case ICmpInst::ICMP_UGT:
5460 NewPred = ICmpInst::ICMP_EQ;
5461 break;
5462 default:
5463 return nullptr;
5464 }
5465 } else if (match(&Cmp, m_c_ICmp(Pred,
5466 m_OneUse(m_CombineOr(
5467 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5468 m_Add(m_Shl(m_One(), m_Value(Y)),
5469 m_AllOnes()))),
5470 m_Value(X)))) {
5471 // The variant with 'add' is not canonical, (the variant with 'not' is)
5472 // we only get it because it has extra uses, and can't be canonicalized,
5473
5474 switch (Pred) {
5475 case ICmpInst::ICMP_ULT:
5476 NewPred = ICmpInst::ICMP_NE;
5477 break;
5478 case ICmpInst::ICMP_UGE:
5479 NewPred = ICmpInst::ICMP_EQ;
5480 break;
5481 default:
5482 return nullptr;
5483 }
5484 } else
5485 return nullptr;
5486
5487 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5488 Constant *Zero = Constant::getNullValue(NewX->getType());
5489 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5490}
5491
5492static Instruction *foldVectorCmp(CmpInst &Cmp,
5493 InstCombiner::BuilderTy &Builder) {
5494 const CmpInst::Predicate Pred = Cmp.getPredicate();
5495 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5496 Value *V1, *V2;
5497 ArrayRef<int> M;
5498 if (!match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(M))))
5499 return nullptr;
5500
5501 // If both arguments of the cmp are shuffles that use the same mask and
5502 // shuffle within a single vector, move the shuffle after the cmp:
5503 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5504 Type *V1Ty = V1->getType();
5505 if (match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(M))) &&
5506 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
5507 Value *NewCmp = Builder.CreateCmp(Pred, V1, V2);
5508 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
5509 }
5510
5511 // Try to canonicalize compare with splatted operand and splat constant.
5512 // TODO: We could generalize this for more than splats. See/use the code in
5513 // InstCombiner::foldVectorBinop().
5514 Constant *C;
5515 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
5516 return nullptr;
5517
5518 // Length-changing splats are ok, so adjust the constants as needed:
5519 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
5520 Constant *ScalarC = C->getSplatValue(/* AllowUndefs */ true);
5521 int MaskSplatIndex;
5522 if (ScalarC && match(M, m_SplatOrUndefMask(MaskSplatIndex))) {
5523 // We allow undefs in matching, but this transform removes those for safety.
5524 // Demanded elements analysis should be able to recover some/all of that.
5525 C = ConstantVector::getSplat(cast<VectorType>(V1Ty)->getElementCount(),
5526 ScalarC);
5527 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
5528 Value *NewCmp = Builder.CreateCmp(Pred, V1, C);
5529 return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()),
5530 NewM);
5531 }
5532
5533 return nullptr;
5534}
5535
5536// extract(uadd.with.overflow(A, B), 0) ult A
5537// -> extract(uadd.with.overflow(A, B), 1)
5538static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5539 CmpInst::Predicate Pred = I.getPredicate();
5540 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5541
5542 Value *UAddOv;
5543 Value *A, *B;
5544 auto UAddOvResultPat = m_ExtractValue<0>(
5545 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5546 if (match(Op0, UAddOvResultPat) &&
5547 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5548 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5549 (match(A, m_One()) || match(B, m_One()))) ||
5550 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5551 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5552 // extract(uadd.with.overflow(A, B), 0) < A
5553 // extract(uadd.with.overflow(A, 1), 0) == 0
5554 // extract(uadd.with.overflow(A, -1), 0) != -1
5555 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5556 else if (match(Op1, UAddOvResultPat) &&
5557 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5558 // A > extract(uadd.with.overflow(A, B), 0)
5559 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5560 else
5561 return nullptr;
5562
5563 return ExtractValueInst::Create(UAddOv, 1);
5564}
5565
5566Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
5567 bool Changed = false;
5568 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5569 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5570 unsigned Op0Cplxity = getComplexity(Op0);
5571 unsigned Op1Cplxity = getComplexity(Op1);
5572
5573 /// Orders the operands of the compare so that they are listed from most
5574 /// complex to least complex. This puts constants before unary operators,
5575 /// before binary operators.
5576 if (Op0Cplxity < Op1Cplxity ||
5577 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5578 I.swapOperands();
5579 std::swap(Op0, Op1);
5580 Changed = true;
5581 }
5582
5583 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, Q))
5584 return replaceInstUsesWith(I, V);
5585
5586 // Comparing -val or val with non-zero is the same as just comparing val
5587 // ie, abs(val) != 0 -> val != 0
5588 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5589 Value *Cond, *SelectTrue, *SelectFalse;
5590 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5591 m_Value(SelectFalse)))) {
5592 if (Value *V = dyn_castNegVal(SelectTrue)) {
5593 if (V == SelectFalse)
5594 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5595 }
5596 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5597 if (V == SelectTrue)
5598