1 //===- InstCombineCompares.cpp --------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitICmp and visitFCmp functions.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "InstCombine.h"
15 #include "llvm/IntrinsicInst.h"
16 #include "llvm/Analysis/ConstantFolding.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/Target/TargetData.h"
20 #include "llvm/Support/ConstantRange.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/PatternMatch.h"
23 using namespace llvm;
24 using namespace PatternMatch;
25
getOne(Constant * C)26 static ConstantInt *getOne(Constant *C) {
27 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
28 }
29
30 /// AddOne - Add one to a ConstantInt
AddOne(Constant * C)31 static Constant *AddOne(Constant *C) {
32 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
33 }
34 /// SubOne - Subtract one from a ConstantInt
SubOne(Constant * C)35 static Constant *SubOne(Constant *C) {
36 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
37 }
38
ExtractElement(Constant * V,Constant * Idx)39 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
40 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
41 }
42
HasAddOverflow(ConstantInt * Result,ConstantInt * In1,ConstantInt * In2,bool IsSigned)43 static bool HasAddOverflow(ConstantInt *Result,
44 ConstantInt *In1, ConstantInt *In2,
45 bool IsSigned) {
46 if (!IsSigned)
47 return Result->getValue().ult(In1->getValue());
48
49 if (In2->isNegative())
50 return Result->getValue().sgt(In1->getValue());
51 return Result->getValue().slt(In1->getValue());
52 }
53
54 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
55 /// overflowed for this type.
AddWithOverflow(Constant * & Result,Constant * In1,Constant * In2,bool IsSigned=false)56 static bool AddWithOverflow(Constant *&Result, Constant *In1,
57 Constant *In2, bool IsSigned = false) {
58 Result = ConstantExpr::getAdd(In1, In2);
59
60 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
63 if (HasAddOverflow(ExtractElement(Result, Idx),
64 ExtractElement(In1, Idx),
65 ExtractElement(In2, Idx),
66 IsSigned))
67 return true;
68 }
69 return false;
70 }
71
72 return HasAddOverflow(cast<ConstantInt>(Result),
73 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
74 IsSigned);
75 }
76
HasSubOverflow(ConstantInt * Result,ConstantInt * In1,ConstantInt * In2,bool IsSigned)77 static bool HasSubOverflow(ConstantInt *Result,
78 ConstantInt *In1, ConstantInt *In2,
79 bool IsSigned) {
80 if (!IsSigned)
81 return Result->getValue().ugt(In1->getValue());
82
83 if (In2->isNegative())
84 return Result->getValue().slt(In1->getValue());
85
86 return Result->getValue().sgt(In1->getValue());
87 }
88
89 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
90 /// overflowed for this type.
SubWithOverflow(Constant * & Result,Constant * In1,Constant * In2,bool IsSigned=false)91 static bool SubWithOverflow(Constant *&Result, Constant *In1,
92 Constant *In2, bool IsSigned = false) {
93 Result = ConstantExpr::getSub(In1, In2);
94
95 if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
98 if (HasSubOverflow(ExtractElement(Result, Idx),
99 ExtractElement(In1, Idx),
100 ExtractElement(In2, Idx),
101 IsSigned))
102 return true;
103 }
104 return false;
105 }
106
107 return HasSubOverflow(cast<ConstantInt>(Result),
108 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
109 IsSigned);
110 }
111
112 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
113 /// comparison only checks the sign bit. If it only checks the sign bit, set
114 /// TrueIfSigned if the result of the comparison is true when the input value is
115 /// signed.
isSignBitCheck(ICmpInst::Predicate pred,ConstantInt * RHS,bool & TrueIfSigned)116 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
117 bool &TrueIfSigned) {
118 switch (pred) {
119 case ICmpInst::ICMP_SLT: // True if LHS s< 0
120 TrueIfSigned = true;
121 return RHS->isZero();
122 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
123 TrueIfSigned = true;
124 return RHS->isAllOnesValue();
125 case ICmpInst::ICMP_SGT: // True if LHS s> -1
126 TrueIfSigned = false;
127 return RHS->isAllOnesValue();
128 case ICmpInst::ICMP_UGT:
129 // True if LHS u> RHS and RHS == high-bit-mask - 1
130 TrueIfSigned = true;
131 return RHS->isMaxValue(true);
132 case ICmpInst::ICMP_UGE:
133 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
134 TrueIfSigned = true;
135 return RHS->getValue().isSignBit();
136 default:
137 return false;
138 }
139 }
140
141 // isHighOnes - Return true if the constant is of the form 1+0+.
142 // This is the same as lowones(~X).
isHighOnes(const ConstantInt * CI)143 static bool isHighOnes(const ConstantInt *CI) {
144 return (~CI->getValue() + 1).isPowerOf2();
145 }
146
147 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
148 /// set of known zero and one bits, compute the maximum and minimum values that
149 /// could have the specified known zero and known one bits, returning them in
150 /// min/max.
ComputeSignedMinMaxValuesFromKnownBits(const APInt & KnownZero,const APInt & KnownOne,APInt & Min,APInt & Max)151 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
152 const APInt& KnownOne,
153 APInt& Min, APInt& Max) {
154 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
155 KnownZero.getBitWidth() == Min.getBitWidth() &&
156 KnownZero.getBitWidth() == Max.getBitWidth() &&
157 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
158 APInt UnknownBits = ~(KnownZero|KnownOne);
159
160 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
161 // bit if it is unknown.
162 Min = KnownOne;
163 Max = KnownOne|UnknownBits;
164
165 if (UnknownBits.isNegative()) { // Sign bit is unknown
166 Min.setBit(Min.getBitWidth()-1);
167 Max.clearBit(Max.getBitWidth()-1);
168 }
169 }
170
171 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
172 // a set of known zero and one bits, compute the maximum and minimum values that
173 // could have the specified known zero and known one bits, returning them in
174 // min/max.
ComputeUnsignedMinMaxValuesFromKnownBits(const APInt & KnownZero,const APInt & KnownOne,APInt & Min,APInt & Max)175 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
176 const APInt &KnownOne,
177 APInt &Min, APInt &Max) {
178 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
179 KnownZero.getBitWidth() == Min.getBitWidth() &&
180 KnownZero.getBitWidth() == Max.getBitWidth() &&
181 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
182 APInt UnknownBits = ~(KnownZero|KnownOne);
183
184 // The minimum value is when the unknown bits are all zeros.
185 Min = KnownOne;
186 // The maximum value is when the unknown bits are all ones.
187 Max = KnownOne|UnknownBits;
188 }
189
190
191
192 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
193 /// cmp pred (load (gep GV, ...)), cmpcst
194 /// where GV is a global variable with a constant initializer. Try to simplify
195 /// this into some simple computation that does not need the load. For example
196 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
197 ///
198 /// If AndCst is non-null, then the loaded value is masked with that constant
199 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
200 Instruction *InstCombiner::
FoldCmpLoadFromIndexedGlobal(GetElementPtrInst * GEP,GlobalVariable * GV,CmpInst & ICI,ConstantInt * AndCst)201 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
202 CmpInst &ICI, ConstantInt *AndCst) {
203 // We need TD information to know the pointer size unless this is inbounds.
204 if (!GEP->isInBounds() && TD == 0) return 0;
205
206 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
207 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
208
209 // There are many forms of this optimization we can handle, for now, just do
210 // the simple index into a single-dimensional array.
211 //
212 // Require: GEP GV, 0, i {{, constant indices}}
213 if (GEP->getNumOperands() < 3 ||
214 !isa<ConstantInt>(GEP->getOperand(1)) ||
215 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
216 isa<Constant>(GEP->getOperand(2)))
217 return 0;
218
219 // Check that indices after the variable are constants and in-range for the
220 // type they index. Collect the indices. This is typically for arrays of
221 // structs.
222 SmallVector<unsigned, 4> LaterIndices;
223
224 Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
225 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
226 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
227 if (Idx == 0) return 0; // Variable index.
228
229 uint64_t IdxVal = Idx->getZExtValue();
230 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
231
232 if (StructType *STy = dyn_cast<StructType>(EltTy))
233 EltTy = STy->getElementType(IdxVal);
234 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
235 if (IdxVal >= ATy->getNumElements()) return 0;
236 EltTy = ATy->getElementType();
237 } else {
238 return 0; // Unknown type.
239 }
240
241 LaterIndices.push_back(IdxVal);
242 }
243
244 enum { Overdefined = -3, Undefined = -2 };
245
246 // Variables for our state machines.
247
248 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
249 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
250 // and 87 is the second (and last) index. FirstTrueElement is -2 when
251 // undefined, otherwise set to the first true element. SecondTrueElement is
252 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
253 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
254
255 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
256 // form "i != 47 & i != 87". Same state transitions as for true elements.
257 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
258
259 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
260 /// define a state machine that triggers for ranges of values that the index
261 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
262 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
263 /// index in the range (inclusive). We use -2 for undefined here because we
264 /// use relative comparisons and don't want 0-1 to match -1.
265 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
266
267 // MagicBitvector - This is a magic bitvector where we set a bit if the
268 // comparison is true for element 'i'. If there are 64 elements or less in
269 // the array, this will fully represent all the comparison results.
270 uint64_t MagicBitvector = 0;
271
272
273 // Scan the array and see if one of our patterns matches.
274 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
275 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
276 Constant *Elt = Init->getOperand(i);
277
278 // If this is indexing an array of structures, get the structure element.
279 if (!LaterIndices.empty())
280 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
281
282 // If the element is masked, handle it.
283 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
284
285 // Find out if the comparison would be true or false for the i'th element.
286 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
287 CompareRHS, TD);
288 // If the result is undef for this element, ignore it.
289 if (isa<UndefValue>(C)) {
290 // Extend range state machines to cover this element in case there is an
291 // undef in the middle of the range.
292 if (TrueRangeEnd == (int)i-1)
293 TrueRangeEnd = i;
294 if (FalseRangeEnd == (int)i-1)
295 FalseRangeEnd = i;
296 continue;
297 }
298
299 // If we can't compute the result for any of the elements, we have to give
300 // up evaluating the entire conditional.
301 if (!isa<ConstantInt>(C)) return 0;
302
303 // Otherwise, we know if the comparison is true or false for this element,
304 // update our state machines.
305 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
306
307 // State machine for single/double/range index comparison.
308 if (IsTrueForElt) {
309 // Update the TrueElement state machine.
310 if (FirstTrueElement == Undefined)
311 FirstTrueElement = TrueRangeEnd = i; // First true element.
312 else {
313 // Update double-compare state machine.
314 if (SecondTrueElement == Undefined)
315 SecondTrueElement = i;
316 else
317 SecondTrueElement = Overdefined;
318
319 // Update range state machine.
320 if (TrueRangeEnd == (int)i-1)
321 TrueRangeEnd = i;
322 else
323 TrueRangeEnd = Overdefined;
324 }
325 } else {
326 // Update the FalseElement state machine.
327 if (FirstFalseElement == Undefined)
328 FirstFalseElement = FalseRangeEnd = i; // First false element.
329 else {
330 // Update double-compare state machine.
331 if (SecondFalseElement == Undefined)
332 SecondFalseElement = i;
333 else
334 SecondFalseElement = Overdefined;
335
336 // Update range state machine.
337 if (FalseRangeEnd == (int)i-1)
338 FalseRangeEnd = i;
339 else
340 FalseRangeEnd = Overdefined;
341 }
342 }
343
344
345 // If this element is in range, update our magic bitvector.
346 if (i < 64 && IsTrueForElt)
347 MagicBitvector |= 1ULL << i;
348
349 // If all of our states become overdefined, bail out early. Since the
350 // predicate is expensive, only check it every 8 elements. This is only
351 // really useful for really huge arrays.
352 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
353 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
354 FalseRangeEnd == Overdefined)
355 return 0;
356 }
357
358 // Now that we've scanned the entire array, emit our new comparison(s). We
359 // order the state machines in complexity of the generated code.
360 Value *Idx = GEP->getOperand(2);
361
362 // If the index is larger than the pointer size of the target, truncate the
363 // index down like the GEP would do implicitly. We don't have to do this for
364 // an inbounds GEP because the index can't be out of range.
365 if (!GEP->isInBounds() &&
366 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
367 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
368
369 // If the comparison is only true for one or two elements, emit direct
370 // comparisons.
371 if (SecondTrueElement != Overdefined) {
372 // None true -> false.
373 if (FirstTrueElement == Undefined)
374 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
375
376 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
377
378 // True for one element -> 'i == 47'.
379 if (SecondTrueElement == Undefined)
380 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
381
382 // True for two elements -> 'i == 47 | i == 72'.
383 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
384 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
385 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
386 return BinaryOperator::CreateOr(C1, C2);
387 }
388
389 // If the comparison is only false for one or two elements, emit direct
390 // comparisons.
391 if (SecondFalseElement != Overdefined) {
392 // None false -> true.
393 if (FirstFalseElement == Undefined)
394 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
395
396 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
397
398 // False for one element -> 'i != 47'.
399 if (SecondFalseElement == Undefined)
400 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
401
402 // False for two elements -> 'i != 47 & i != 72'.
403 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
404 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
405 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
406 return BinaryOperator::CreateAnd(C1, C2);
407 }
408
409 // If the comparison can be replaced with a range comparison for the elements
410 // where it is true, emit the range check.
411 if (TrueRangeEnd != Overdefined) {
412 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
413
414 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
415 if (FirstTrueElement) {
416 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
417 Idx = Builder->CreateAdd(Idx, Offs);
418 }
419
420 Value *End = ConstantInt::get(Idx->getType(),
421 TrueRangeEnd-FirstTrueElement+1);
422 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
423 }
424
425 // False range check.
426 if (FalseRangeEnd != Overdefined) {
427 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
428 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
429 if (FirstFalseElement) {
430 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
431 Idx = Builder->CreateAdd(Idx, Offs);
432 }
433
434 Value *End = ConstantInt::get(Idx->getType(),
435 FalseRangeEnd-FirstFalseElement);
436 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
437 }
438
439
440 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
441 // of this load, replace it with computation that does:
442 // ((magic_cst >> i) & 1) != 0
443 if (Init->getNumOperands() <= 32 ||
444 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
445 Type *Ty;
446 if (Init->getNumOperands() <= 32)
447 Ty = Type::getInt32Ty(Init->getContext());
448 else
449 Ty = Type::getInt64Ty(Init->getContext());
450 Value *V = Builder->CreateIntCast(Idx, Ty, false);
451 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
452 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
453 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
454 }
455
456 return 0;
457 }
458
459
460 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
461 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
462 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
463 /// be complex, and scales are involved. The above expression would also be
464 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
465 /// This later form is less amenable to optimization though, and we are allowed
466 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
467 ///
468 /// If we can't emit an optimized form for this expression, this returns null.
469 ///
EvaluateGEPOffsetExpression(User * GEP,InstCombiner & IC)470 static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
471 TargetData &TD = *IC.getTargetData();
472 gep_type_iterator GTI = gep_type_begin(GEP);
473
474 // Check to see if this gep only has a single variable index. If so, and if
475 // any constant indices are a multiple of its scale, then we can compute this
476 // in terms of the scale of the variable index. For example, if the GEP
477 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
478 // because the expression will cross zero at the same point.
479 unsigned i, e = GEP->getNumOperands();
480 int64_t Offset = 0;
481 for (i = 1; i != e; ++i, ++GTI) {
482 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
483 // Compute the aggregate offset of constant indices.
484 if (CI->isZero()) continue;
485
486 // Handle a struct index, which adds its field offset to the pointer.
487 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
488 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
489 } else {
490 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
491 Offset += Size*CI->getSExtValue();
492 }
493 } else {
494 // Found our variable index.
495 break;
496 }
497 }
498
499 // If there are no variable indices, we must have a constant offset, just
500 // evaluate it the general way.
501 if (i == e) return 0;
502
503 Value *VariableIdx = GEP->getOperand(i);
504 // Determine the scale factor of the variable element. For example, this is
505 // 4 if the variable index is into an array of i32.
506 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
507
508 // Verify that there are no other variable indices. If so, emit the hard way.
509 for (++i, ++GTI; i != e; ++i, ++GTI) {
510 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
511 if (!CI) return 0;
512
513 // Compute the aggregate offset of constant indices.
514 if (CI->isZero()) continue;
515
516 // Handle a struct index, which adds its field offset to the pointer.
517 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
518 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
519 } else {
520 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
521 Offset += Size*CI->getSExtValue();
522 }
523 }
524
525 // Okay, we know we have a single variable index, which must be a
526 // pointer/array/vector index. If there is no offset, life is simple, return
527 // the index.
528 unsigned IntPtrWidth = TD.getPointerSizeInBits();
529 if (Offset == 0) {
530 // Cast to intptrty in case a truncation occurs. If an extension is needed,
531 // we don't need to bother extending: the extension won't affect where the
532 // computation crosses zero.
533 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
534 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
535 VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
536 }
537 return VariableIdx;
538 }
539
540 // Otherwise, there is an index. The computation we will do will be modulo
541 // the pointer size, so get it.
542 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
543
544 Offset &= PtrSizeMask;
545 VariableScale &= PtrSizeMask;
546
547 // To do this transformation, any constant index must be a multiple of the
548 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
549 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
550 // multiple of the variable scale.
551 int64_t NewOffs = Offset / (int64_t)VariableScale;
552 if (Offset != NewOffs*(int64_t)VariableScale)
553 return 0;
554
555 // Okay, we can do this evaluation. Start by converting the index to intptr.
556 Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
557 if (VariableIdx->getType() != IntPtrTy)
558 VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
559 true /*Signed*/);
560 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
561 return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
562 }
563
564 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
565 /// else. At this point we know that the GEP is on the LHS of the comparison.
FoldGEPICmp(GEPOperator * GEPLHS,Value * RHS,ICmpInst::Predicate Cond,Instruction & I)566 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
567 ICmpInst::Predicate Cond,
568 Instruction &I) {
569 // Look through bitcasts.
570 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
571 RHS = BCI->getOperand(0);
572
573 Value *PtrBase = GEPLHS->getOperand(0);
574 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
575 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
576 // This transformation (ignoring the base and scales) is valid because we
577 // know pointers can't overflow since the gep is inbounds. See if we can
578 // output an optimized form.
579 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
580
581 // If not, synthesize the offset the hard way.
582 if (Offset == 0)
583 Offset = EmitGEPOffset(GEPLHS);
584 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
585 Constant::getNullValue(Offset->getType()));
586 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
587 // If the base pointers are different, but the indices are the same, just
588 // compare the base pointer.
589 if (PtrBase != GEPRHS->getOperand(0)) {
590 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
591 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
592 GEPRHS->getOperand(0)->getType();
593 if (IndicesTheSame)
594 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
595 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
596 IndicesTheSame = false;
597 break;
598 }
599
600 // If all indices are the same, just compare the base pointers.
601 if (IndicesTheSame)
602 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
603 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
604
605 // Otherwise, the base pointers are different and the indices are
606 // different, bail out.
607 return 0;
608 }
609
610 // If one of the GEPs has all zero indices, recurse.
611 bool AllZeros = true;
612 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
613 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
614 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
615 AllZeros = false;
616 break;
617 }
618 if (AllZeros)
619 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
620 ICmpInst::getSwappedPredicate(Cond), I);
621
622 // If the other GEP has all zero indices, recurse.
623 AllZeros = true;
624 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
625 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
626 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
627 AllZeros = false;
628 break;
629 }
630 if (AllZeros)
631 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
632
633 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
634 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
635 // If the GEPs only differ by one index, compare it.
636 unsigned NumDifferences = 0; // Keep track of # differences.
637 unsigned DiffOperand = 0; // The operand that differs.
638 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
639 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
640 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
641 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
642 // Irreconcilable differences.
643 NumDifferences = 2;
644 break;
645 } else {
646 if (NumDifferences++) break;
647 DiffOperand = i;
648 }
649 }
650
651 if (NumDifferences == 0) // SAME GEP?
652 return ReplaceInstUsesWith(I, // No comparison is needed here.
653 ConstantInt::get(Type::getInt1Ty(I.getContext()),
654 ICmpInst::isTrueWhenEqual(Cond)));
655
656 else if (NumDifferences == 1 && GEPsInBounds) {
657 Value *LHSV = GEPLHS->getOperand(DiffOperand);
658 Value *RHSV = GEPRHS->getOperand(DiffOperand);
659 // Make sure we do a signed comparison here.
660 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
661 }
662 }
663
664 // Only lower this if the icmp is the only user of the GEP or if we expect
665 // the result to fold to a constant!
666 if (TD &&
667 GEPsInBounds &&
668 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
669 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
670 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
671 Value *L = EmitGEPOffset(GEPLHS);
672 Value *R = EmitGEPOffset(GEPRHS);
673 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
674 }
675 }
676 return 0;
677 }
678
679 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
FoldICmpAddOpCst(ICmpInst & ICI,Value * X,ConstantInt * CI,ICmpInst::Predicate Pred,Value * TheAdd)680 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
681 Value *X, ConstantInt *CI,
682 ICmpInst::Predicate Pred,
683 Value *TheAdd) {
684 // If we have X+0, exit early (simplifying logic below) and let it get folded
685 // elsewhere. icmp X+0, X -> icmp X, X
686 if (CI->isZero()) {
687 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
688 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
689 }
690
691 // (X+4) == X -> false.
692 if (Pred == ICmpInst::ICMP_EQ)
693 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
694
695 // (X+4) != X -> true.
696 if (Pred == ICmpInst::ICMP_NE)
697 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
698
699 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
700 // so the values can never be equal. Similarly for all other "or equals"
701 // operators.
702
703 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
704 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
705 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
706 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
707 Value *R =
708 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
709 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
710 }
711
712 // (X+1) >u X --> X <u (0-1) --> X != 255
713 // (X+2) >u X --> X <u (0-2) --> X <u 254
714 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
715 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
716 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
717
718 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
719 ConstantInt *SMax = ConstantInt::get(X->getContext(),
720 APInt::getSignedMaxValue(BitWidth));
721
722 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
723 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
724 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
725 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
726 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
727 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
728 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
729 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
730
731 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
732 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
733 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
734 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
735 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
736 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
737
738 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
739 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
740 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
741 }
742
743 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
744 /// and CmpRHS are both known to be integer constants.
FoldICmpDivCst(ICmpInst & ICI,BinaryOperator * DivI,ConstantInt * DivRHS)745 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
746 ConstantInt *DivRHS) {
747 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
748 const APInt &CmpRHSV = CmpRHS->getValue();
749
750 // FIXME: If the operand types don't match the type of the divide
751 // then don't attempt this transform. The code below doesn't have the
752 // logic to deal with a signed divide and an unsigned compare (and
753 // vice versa). This is because (x /s C1) <s C2 produces different
754 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
755 // (x /u C1) <u C2. Simply casting the operands and result won't
756 // work. :( The if statement below tests that condition and bails
757 // if it finds it.
758 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
759 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
760 return 0;
761 if (DivRHS->isZero())
762 return 0; // The ProdOV computation fails on divide by zero.
763 if (DivIsSigned && DivRHS->isAllOnesValue())
764 return 0; // The overflow computation also screws up here
765 if (DivRHS->isOne()) {
766 // This eliminates some funny cases with INT_MIN.
767 ICI.setOperand(0, DivI->getOperand(0)); // X/1 == X.
768 return &ICI;
769 }
770
771 // Compute Prod = CI * DivRHS. We are essentially solving an equation
772 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
773 // C2 (CI). By solving for X we can turn this into a range check
774 // instead of computing a divide.
775 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
776
777 // Determine if the product overflows by seeing if the product is
778 // not equal to the divide. Make sure we do the same kind of divide
779 // as in the LHS instruction that we're folding.
780 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
781 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
782
783 // Get the ICmp opcode
784 ICmpInst::Predicate Pred = ICI.getPredicate();
785
786 /// If the division is known to be exact, then there is no remainder from the
787 /// divide, so the covered range size is unit, otherwise it is the divisor.
788 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
789
790 // Figure out the interval that is being checked. For example, a comparison
791 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
792 // Compute this interval based on the constants involved and the signedness of
793 // the compare/divide. This computes a half-open interval, keeping track of
794 // whether either value in the interval overflows. After analysis each
795 // overflow variable is set to 0 if it's corresponding bound variable is valid
796 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
797 int LoOverflow = 0, HiOverflow = 0;
798 Constant *LoBound = 0, *HiBound = 0;
799
800 if (!DivIsSigned) { // udiv
801 // e.g. X/5 op 3 --> [15, 20)
802 LoBound = Prod;
803 HiOverflow = LoOverflow = ProdOV;
804 if (!HiOverflow) {
805 // If this is not an exact divide, then many values in the range collapse
806 // to the same result value.
807 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
808 }
809
810 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
811 if (CmpRHSV == 0) { // (X / pos) op 0
812 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
813 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
814 HiBound = RangeSize;
815 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
816 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
817 HiOverflow = LoOverflow = ProdOV;
818 if (!HiOverflow)
819 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
820 } else { // (X / pos) op neg
821 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
822 HiBound = AddOne(Prod);
823 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
824 if (!LoOverflow) {
825 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
826 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
827 }
828 }
829 } else if (DivRHS->isNegative()) { // Divisor is < 0.
830 if (DivI->isExact())
831 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
832 if (CmpRHSV == 0) { // (X / neg) op 0
833 // e.g. X/-5 op 0 --> [-4, 5)
834 LoBound = AddOne(RangeSize);
835 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
836 if (HiBound == DivRHS) { // -INTMIN = INTMIN
837 HiOverflow = 1; // [INTMIN+1, overflow)
838 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
839 }
840 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
841 // e.g. X/-5 op 3 --> [-19, -14)
842 HiBound = AddOne(Prod);
843 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
844 if (!LoOverflow)
845 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
846 } else { // (X / neg) op neg
847 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
848 LoOverflow = HiOverflow = ProdOV;
849 if (!HiOverflow)
850 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
851 }
852
853 // Dividing by a negative swaps the condition. LT <-> GT
854 Pred = ICmpInst::getSwappedPredicate(Pred);
855 }
856
857 Value *X = DivI->getOperand(0);
858 switch (Pred) {
859 default: llvm_unreachable("Unhandled icmp opcode!");
860 case ICmpInst::ICMP_EQ:
861 if (LoOverflow && HiOverflow)
862 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
863 if (HiOverflow)
864 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
865 ICmpInst::ICMP_UGE, X, LoBound);
866 if (LoOverflow)
867 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
868 ICmpInst::ICMP_ULT, X, HiBound);
869 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
870 DivIsSigned, true));
871 case ICmpInst::ICMP_NE:
872 if (LoOverflow && HiOverflow)
873 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
874 if (HiOverflow)
875 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
876 ICmpInst::ICMP_ULT, X, LoBound);
877 if (LoOverflow)
878 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
879 ICmpInst::ICMP_UGE, X, HiBound);
880 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
881 DivIsSigned, false));
882 case ICmpInst::ICMP_ULT:
883 case ICmpInst::ICMP_SLT:
884 if (LoOverflow == +1) // Low bound is greater than input range.
885 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
886 if (LoOverflow == -1) // Low bound is less than input range.
887 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
888 return new ICmpInst(Pred, X, LoBound);
889 case ICmpInst::ICMP_UGT:
890 case ICmpInst::ICMP_SGT:
891 if (HiOverflow == +1) // High bound greater than input range.
892 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
893 if (HiOverflow == -1) // High bound less than input range.
894 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
895 if (Pred == ICmpInst::ICMP_UGT)
896 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
897 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
898 }
899 }
900
901 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
FoldICmpShrCst(ICmpInst & ICI,BinaryOperator * Shr,ConstantInt * ShAmt)902 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
903 ConstantInt *ShAmt) {
904 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
905
906 // Check that the shift amount is in range. If not, don't perform
907 // undefined shifts. When the shift is visited it will be
908 // simplified.
909 uint32_t TypeBits = CmpRHSV.getBitWidth();
910 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
911 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
912 return 0;
913
914 if (!ICI.isEquality()) {
915 // If we have an unsigned comparison and an ashr, we can't simplify this.
916 // Similarly for signed comparisons with lshr.
917 if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
918 return 0;
919
920 // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
921 // by a power of 2. Since we already have logic to simplify these,
922 // transform to div and then simplify the resultant comparison.
923 if (Shr->getOpcode() == Instruction::AShr &&
924 (!Shr->isExact() || ShAmtVal == TypeBits - 1))
925 return 0;
926
927 // Revisit the shift (to delete it).
928 Worklist.Add(Shr);
929
930 Constant *DivCst =
931 ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
932
933 Value *Tmp =
934 Shr->getOpcode() == Instruction::AShr ?
935 Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
936 Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
937
938 ICI.setOperand(0, Tmp);
939
940 // If the builder folded the binop, just return it.
941 BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
942 if (TheDiv == 0)
943 return &ICI;
944
945 // Otherwise, fold this div/compare.
946 assert(TheDiv->getOpcode() == Instruction::SDiv ||
947 TheDiv->getOpcode() == Instruction::UDiv);
948
949 Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
950 assert(Res && "This div/cst should have folded!");
951 return Res;
952 }
953
954
955 // If we are comparing against bits always shifted out, the
956 // comparison cannot succeed.
957 APInt Comp = CmpRHSV << ShAmtVal;
958 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
959 if (Shr->getOpcode() == Instruction::LShr)
960 Comp = Comp.lshr(ShAmtVal);
961 else
962 Comp = Comp.ashr(ShAmtVal);
963
964 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
965 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
966 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
967 IsICMP_NE);
968 return ReplaceInstUsesWith(ICI, Cst);
969 }
970
971 // Otherwise, check to see if the bits shifted out are known to be zero.
972 // If so, we can compare against the unshifted value:
973 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
974 if (Shr->hasOneUse() && Shr->isExact())
975 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
976
977 if (Shr->hasOneUse()) {
978 // Otherwise strength reduce the shift into an and.
979 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
980 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
981
982 Value *And = Builder->CreateAnd(Shr->getOperand(0),
983 Mask, Shr->getName()+".mask");
984 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
985 }
986 return 0;
987 }
988
989
990 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
991 ///
visitICmpInstWithInstAndIntCst(ICmpInst & ICI,Instruction * LHSI,ConstantInt * RHS)992 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
993 Instruction *LHSI,
994 ConstantInt *RHS) {
995 const APInt &RHSV = RHS->getValue();
996
997 switch (LHSI->getOpcode()) {
998 case Instruction::Trunc:
999 if (ICI.isEquality() && LHSI->hasOneUse()) {
1000 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1001 // of the high bits truncated out of x are known.
1002 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
1003 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
1004 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
1005 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
1006 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
1007
1008 // If all the high bits are known, we can do this xform.
1009 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1010 // Pull in the high bits from known-ones set.
1011 APInt NewRHS = RHS->getValue().zext(SrcBits);
1012 NewRHS |= KnownOne;
1013 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1014 ConstantInt::get(ICI.getContext(), NewRHS));
1015 }
1016 }
1017 break;
1018
1019 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1020 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1021 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1022 // fold the xor.
1023 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1024 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1025 Value *CompareVal = LHSI->getOperand(0);
1026
1027 // If the sign bit of the XorCST is not set, there is no change to
1028 // the operation, just stop using the Xor.
1029 if (!XorCST->isNegative()) {
1030 ICI.setOperand(0, CompareVal);
1031 Worklist.Add(LHSI);
1032 return &ICI;
1033 }
1034
1035 // Was the old condition true if the operand is positive?
1036 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1037
1038 // If so, the new one isn't.
1039 isTrueIfPositive ^= true;
1040
1041 if (isTrueIfPositive)
1042 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1043 SubOne(RHS));
1044 else
1045 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1046 AddOne(RHS));
1047 }
1048
1049 if (LHSI->hasOneUse()) {
1050 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1051 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1052 const APInt &SignBit = XorCST->getValue();
1053 ICmpInst::Predicate Pred = ICI.isSigned()
1054 ? ICI.getUnsignedPredicate()
1055 : ICI.getSignedPredicate();
1056 return new ICmpInst(Pred, LHSI->getOperand(0),
1057 ConstantInt::get(ICI.getContext(),
1058 RHSV ^ SignBit));
1059 }
1060
1061 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1062 if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
1063 const APInt &NotSignBit = XorCST->getValue();
1064 ICmpInst::Predicate Pred = ICI.isSigned()
1065 ? ICI.getUnsignedPredicate()
1066 : ICI.getSignedPredicate();
1067 Pred = ICI.getSwappedPredicate(Pred);
1068 return new ICmpInst(Pred, LHSI->getOperand(0),
1069 ConstantInt::get(ICI.getContext(),
1070 RHSV ^ NotSignBit));
1071 }
1072 }
1073 }
1074 break;
1075 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1076 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1077 LHSI->getOperand(0)->hasOneUse()) {
1078 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1079
1080 // If the LHS is an AND of a truncating cast, we can widen the
1081 // and/compare to be the input width without changing the value
1082 // produced, eliminating a cast.
1083 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1084 // We can do this transformation if either the AND constant does not
1085 // have its sign bit set or if it is an equality comparison.
1086 // Extending a relational comparison when we're checking the sign
1087 // bit would not work.
1088 if (ICI.isEquality() ||
1089 (!AndCST->isNegative() && RHSV.isNonNegative())) {
1090 Value *NewAnd =
1091 Builder->CreateAnd(Cast->getOperand(0),
1092 ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
1093 NewAnd->takeName(LHSI);
1094 return new ICmpInst(ICI.getPredicate(), NewAnd,
1095 ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
1096 }
1097 }
1098
1099 // If the LHS is an AND of a zext, and we have an equality compare, we can
1100 // shrink the and/compare to the smaller type, eliminating the cast.
1101 if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
1102 IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
1103 // Make sure we don't compare the upper bits, SimplifyDemandedBits
1104 // should fold the icmp to true/false in that case.
1105 if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
1106 Value *NewAnd =
1107 Builder->CreateAnd(Cast->getOperand(0),
1108 ConstantExpr::getTrunc(AndCST, Ty));
1109 NewAnd->takeName(LHSI);
1110 return new ICmpInst(ICI.getPredicate(), NewAnd,
1111 ConstantExpr::getTrunc(RHS, Ty));
1112 }
1113 }
1114
1115 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1116 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1117 // happens a LOT in code produced by the C front-end, for bitfield
1118 // access.
1119 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1120 if (Shift && !Shift->isShift())
1121 Shift = 0;
1122
1123 ConstantInt *ShAmt;
1124 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1125 Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1126 Type *AndTy = AndCST->getType(); // Type of the and.
1127
1128 // We can fold this as long as we can't shift unknown bits
1129 // into the mask. This can only happen with signed shift
1130 // rights, as they sign-extend.
1131 if (ShAmt) {
1132 bool CanFold = Shift->isLogicalShift();
1133 if (!CanFold) {
1134 // To test for the bad case of the signed shr, see if any
1135 // of the bits shifted in could be tested after the mask.
1136 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1137 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1138
1139 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1140 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1141 AndCST->getValue()) == 0)
1142 CanFold = true;
1143 }
1144
1145 if (CanFold) {
1146 Constant *NewCst;
1147 if (Shift->getOpcode() == Instruction::Shl)
1148 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1149 else
1150 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1151
1152 // Check to see if we are shifting out any of the bits being
1153 // compared.
1154 if (ConstantExpr::get(Shift->getOpcode(),
1155 NewCst, ShAmt) != RHS) {
1156 // If we shifted bits out, the fold is not going to work out.
1157 // As a special case, check to see if this means that the
1158 // result is always true or false now.
1159 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1160 return ReplaceInstUsesWith(ICI,
1161 ConstantInt::getFalse(ICI.getContext()));
1162 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1163 return ReplaceInstUsesWith(ICI,
1164 ConstantInt::getTrue(ICI.getContext()));
1165 } else {
1166 ICI.setOperand(1, NewCst);
1167 Constant *NewAndCST;
1168 if (Shift->getOpcode() == Instruction::Shl)
1169 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1170 else
1171 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1172 LHSI->setOperand(1, NewAndCST);
1173 LHSI->setOperand(0, Shift->getOperand(0));
1174 Worklist.Add(Shift); // Shift is dead.
1175 return &ICI;
1176 }
1177 }
1178 }
1179
1180 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1181 // preferable because it allows the C<<Y expression to be hoisted out
1182 // of a loop if Y is invariant and X is not.
1183 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1184 ICI.isEquality() && !Shift->isArithmeticShift() &&
1185 !isa<Constant>(Shift->getOperand(0))) {
1186 // Compute C << Y.
1187 Value *NS;
1188 if (Shift->getOpcode() == Instruction::LShr) {
1189 NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
1190 } else {
1191 // Insert a logical shift.
1192 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
1193 }
1194
1195 // Compute X & (C << Y).
1196 Value *NewAnd =
1197 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1198
1199 ICI.setOperand(0, NewAnd);
1200 return &ICI;
1201 }
1202 }
1203
1204 // Try to optimize things like "A[i]&42 == 0" to index computations.
1205 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1206 if (GetElementPtrInst *GEP =
1207 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1208 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1209 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1210 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1211 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1212 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1213 return Res;
1214 }
1215 }
1216 break;
1217
1218 case Instruction::Or: {
1219 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1220 break;
1221 Value *P, *Q;
1222 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1223 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1224 // -> and (icmp eq P, null), (icmp eq Q, null).
1225 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1226 Constant::getNullValue(P->getType()));
1227 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1228 Constant::getNullValue(Q->getType()));
1229 Instruction *Op;
1230 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1231 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1232 else
1233 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1234 return Op;
1235 }
1236 break;
1237 }
1238
1239 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1240 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1241 if (!ShAmt) break;
1242
1243 uint32_t TypeBits = RHSV.getBitWidth();
1244
1245 // Check that the shift amount is in range. If not, don't perform
1246 // undefined shifts. When the shift is visited it will be
1247 // simplified.
1248 if (ShAmt->uge(TypeBits))
1249 break;
1250
1251 if (ICI.isEquality()) {
1252 // If we are comparing against bits always shifted out, the
1253 // comparison cannot succeed.
1254 Constant *Comp =
1255 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1256 ShAmt);
1257 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1258 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1259 Constant *Cst =
1260 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1261 return ReplaceInstUsesWith(ICI, Cst);
1262 }
1263
1264 // If the shift is NUW, then it is just shifting out zeros, no need for an
1265 // AND.
1266 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1267 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1268 ConstantExpr::getLShr(RHS, ShAmt));
1269
1270 if (LHSI->hasOneUse()) {
1271 // Otherwise strength reduce the shift into an and.
1272 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1273 Constant *Mask =
1274 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1275 TypeBits-ShAmtVal));
1276
1277 Value *And =
1278 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1279 return new ICmpInst(ICI.getPredicate(), And,
1280 ConstantExpr::getLShr(RHS, ShAmt));
1281 }
1282 }
1283
1284 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1285 bool TrueIfSigned = false;
1286 if (LHSI->hasOneUse() &&
1287 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1288 // (X << 31) <s 0 --> (X&1) != 0
1289 Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
1290 APInt::getOneBitSet(TypeBits,
1291 TypeBits-ShAmt->getZExtValue()-1));
1292 Value *And =
1293 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1294 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1295 And, Constant::getNullValue(And->getType()));
1296 }
1297 break;
1298 }
1299
1300 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1301 case Instruction::AShr: {
1302 // Handle equality comparisons of shift-by-constant.
1303 BinaryOperator *BO = cast<BinaryOperator>(LHSI);
1304 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1305 if (Instruction *Res = FoldICmpShrCst(ICI, BO, ShAmt))
1306 return Res;
1307 }
1308
1309 // Handle exact shr's.
1310 if (ICI.isEquality() && BO->isExact() && BO->hasOneUse()) {
1311 if (RHSV.isMinValue())
1312 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0), RHS);
1313 }
1314 break;
1315 }
1316
1317 case Instruction::SDiv:
1318 case Instruction::UDiv:
1319 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1320 // Fold this div into the comparison, producing a range check.
1321 // Determine, based on the divide type, what the range is being
1322 // checked. If there is an overflow on the low or high side, remember
1323 // it, otherwise compute the range [low, hi) bounding the new value.
1324 // See: InsertRangeTest above for the kinds of replacements possible.
1325 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1326 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1327 DivRHS))
1328 return R;
1329 break;
1330
1331 case Instruction::Add:
1332 // Fold: icmp pred (add X, C1), C2
1333 if (!ICI.isEquality()) {
1334 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1335 if (!LHSC) break;
1336 const APInt &LHSV = LHSC->getValue();
1337
1338 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1339 .subtract(LHSV);
1340
1341 if (ICI.isSigned()) {
1342 if (CR.getLower().isSignBit()) {
1343 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1344 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1345 } else if (CR.getUpper().isSignBit()) {
1346 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1347 ConstantInt::get(ICI.getContext(),CR.getLower()));
1348 }
1349 } else {
1350 if (CR.getLower().isMinValue()) {
1351 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1352 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1353 } else if (CR.getUpper().isMinValue()) {
1354 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1355 ConstantInt::get(ICI.getContext(),CR.getLower()));
1356 }
1357 }
1358 }
1359 break;
1360 }
1361
1362 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1363 if (ICI.isEquality()) {
1364 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1365
1366 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1367 // the second operand is a constant, simplify a bit.
1368 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1369 switch (BO->getOpcode()) {
1370 case Instruction::SRem:
1371 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1372 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1373 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1374 if (V.sgt(1) && V.isPowerOf2()) {
1375 Value *NewRem =
1376 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1377 BO->getName());
1378 return new ICmpInst(ICI.getPredicate(), NewRem,
1379 Constant::getNullValue(BO->getType()));
1380 }
1381 }
1382 break;
1383 case Instruction::Add:
1384 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1385 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1386 if (BO->hasOneUse())
1387 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1388 ConstantExpr::getSub(RHS, BOp1C));
1389 } else if (RHSV == 0) {
1390 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1391 // efficiently invertible, or if the add has just this one use.
1392 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1393
1394 if (Value *NegVal = dyn_castNegVal(BOp1))
1395 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1396 if (Value *NegVal = dyn_castNegVal(BOp0))
1397 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1398 if (BO->hasOneUse()) {
1399 Value *Neg = Builder->CreateNeg(BOp1);
1400 Neg->takeName(BO);
1401 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1402 }
1403 }
1404 break;
1405 case Instruction::Xor:
1406 // For the xor case, we can xor two constants together, eliminating
1407 // the explicit xor.
1408 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
1409 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1410 ConstantExpr::getXor(RHS, BOC));
1411 } else if (RHSV == 0) {
1412 // Replace ((xor A, B) != 0) with (A != B)
1413 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1414 BO->getOperand(1));
1415 }
1416 break;
1417 case Instruction::Sub:
1418 // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
1419 if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
1420 if (BO->hasOneUse())
1421 return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
1422 ConstantExpr::getSub(BOp0C, RHS));
1423 } else if (RHSV == 0) {
1424 // Replace ((sub A, B) != 0) with (A != B)
1425 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1426 BO->getOperand(1));
1427 }
1428 break;
1429 case Instruction::Or:
1430 // If bits are being or'd in that are not present in the constant we
1431 // are comparing against, then the comparison could never succeed!
1432 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1433 Constant *NotCI = ConstantExpr::getNot(RHS);
1434 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1435 return ReplaceInstUsesWith(ICI,
1436 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1437 isICMP_NE));
1438 }
1439 break;
1440
1441 case Instruction::And:
1442 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1443 // If bits are being compared against that are and'd out, then the
1444 // comparison can never succeed!
1445 if ((RHSV & ~BOC->getValue()) != 0)
1446 return ReplaceInstUsesWith(ICI,
1447 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1448 isICMP_NE));
1449
1450 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1451 if (RHS == BOC && RHSV.isPowerOf2())
1452 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1453 ICmpInst::ICMP_NE, LHSI,
1454 Constant::getNullValue(RHS->getType()));
1455
1456 // Don't perform the following transforms if the AND has multiple uses
1457 if (!BO->hasOneUse())
1458 break;
1459
1460 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1461 if (BOC->getValue().isSignBit()) {
1462 Value *X = BO->getOperand(0);
1463 Constant *Zero = Constant::getNullValue(X->getType());
1464 ICmpInst::Predicate pred = isICMP_NE ?
1465 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1466 return new ICmpInst(pred, X, Zero);
1467 }
1468
1469 // ((X & ~7) == 0) --> X < 8
1470 if (RHSV == 0 && isHighOnes(BOC)) {
1471 Value *X = BO->getOperand(0);
1472 Constant *NegX = ConstantExpr::getNeg(BOC);
1473 ICmpInst::Predicate pred = isICMP_NE ?
1474 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1475 return new ICmpInst(pred, X, NegX);
1476 }
1477 }
1478 default: break;
1479 }
1480 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1481 // Handle icmp {eq|ne} <intrinsic>, intcst.
1482 switch (II->getIntrinsicID()) {
1483 case Intrinsic::bswap:
1484 Worklist.Add(II);
1485 ICI.setOperand(0, II->getArgOperand(0));
1486 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1487 return &ICI;
1488 case Intrinsic::ctlz:
1489 case Intrinsic::cttz:
1490 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1491 if (RHSV == RHS->getType()->getBitWidth()) {
1492 Worklist.Add(II);
1493 ICI.setOperand(0, II->getArgOperand(0));
1494 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1495 return &ICI;
1496 }
1497 break;
1498 case Intrinsic::ctpop:
1499 // popcount(A) == 0 -> A == 0 and likewise for !=
1500 if (RHS->isZero()) {
1501 Worklist.Add(II);
1502 ICI.setOperand(0, II->getArgOperand(0));
1503 ICI.setOperand(1, RHS);
1504 return &ICI;
1505 }
1506 break;
1507 default:
1508 break;
1509 }
1510 }
1511 }
1512 return 0;
1513 }
1514
1515 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1516 /// We only handle extending casts so far.
1517 ///
visitICmpInstWithCastAndCast(ICmpInst & ICI)1518 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1519 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1520 Value *LHSCIOp = LHSCI->getOperand(0);
1521 Type *SrcTy = LHSCIOp->getType();
1522 Type *DestTy = LHSCI->getType();
1523 Value *RHSCIOp;
1524
1525 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1526 // integer type is the same size as the pointer type.
1527 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1528 TD->getPointerSizeInBits() ==
1529 cast<IntegerType>(DestTy)->getBitWidth()) {
1530 Value *RHSOp = 0;
1531 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1532 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1533 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1534 RHSOp = RHSC->getOperand(0);
1535 // If the pointer types don't match, insert a bitcast.
1536 if (LHSCIOp->getType() != RHSOp->getType())
1537 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1538 }
1539
1540 if (RHSOp)
1541 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1542 }
1543
1544 // The code below only handles extension cast instructions, so far.
1545 // Enforce this.
1546 if (LHSCI->getOpcode() != Instruction::ZExt &&
1547 LHSCI->getOpcode() != Instruction::SExt)
1548 return 0;
1549
1550 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1551 bool isSignedCmp = ICI.isSigned();
1552
1553 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1554 // Not an extension from the same type?
1555 RHSCIOp = CI->getOperand(0);
1556 if (RHSCIOp->getType() != LHSCIOp->getType())
1557 return 0;
1558
1559 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1560 // and the other is a zext), then we can't handle this.
1561 if (CI->getOpcode() != LHSCI->getOpcode())
1562 return 0;
1563
1564 // Deal with equality cases early.
1565 if (ICI.isEquality())
1566 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1567
1568 // A signed comparison of sign extended values simplifies into a
1569 // signed comparison.
1570 if (isSignedCmp && isSignedExt)
1571 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1572
1573 // The other three cases all fold into an unsigned comparison.
1574 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1575 }
1576
1577 // If we aren't dealing with a constant on the RHS, exit early
1578 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1579 if (!CI)
1580 return 0;
1581
1582 // Compute the constant that would happen if we truncated to SrcTy then
1583 // reextended to DestTy.
1584 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1585 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1586 Res1, DestTy);
1587
1588 // If the re-extended constant didn't change...
1589 if (Res2 == CI) {
1590 // Deal with equality cases early.
1591 if (ICI.isEquality())
1592 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1593
1594 // A signed comparison of sign extended values simplifies into a
1595 // signed comparison.
1596 if (isSignedExt && isSignedCmp)
1597 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1598
1599 // The other three cases all fold into an unsigned comparison.
1600 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1601 }
1602
1603 // The re-extended constant changed so the constant cannot be represented
1604 // in the shorter type. Consequently, we cannot emit a simple comparison.
1605 // All the cases that fold to true or false will have already been handled
1606 // by SimplifyICmpInst, so only deal with the tricky case.
1607
1608 if (isSignedCmp || !isSignedExt)
1609 return 0;
1610
1611 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1612 // should have been folded away previously and not enter in here.
1613
1614 // We're performing an unsigned comp with a sign extended value.
1615 // This is true if the input is >= 0. [aka >s -1]
1616 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1617 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1618
1619 // Finally, return the value computed.
1620 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1621 return ReplaceInstUsesWith(ICI, Result);
1622
1623 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1624 return BinaryOperator::CreateNot(Result);
1625 }
1626
1627 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1628 /// I = icmp ugt (add (add A, B), CI2), CI1
1629 /// If this is of the form:
1630 /// sum = a + b
1631 /// if (sum+128 >u 255)
1632 /// Then replace it with llvm.sadd.with.overflow.i8.
1633 ///
ProcessUGT_ADDCST_ADD(ICmpInst & I,Value * A,Value * B,ConstantInt * CI2,ConstantInt * CI1,InstCombiner & IC)1634 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1635 ConstantInt *CI2, ConstantInt *CI1,
1636 InstCombiner &IC) {
1637 // The transformation we're trying to do here is to transform this into an
1638 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1639 // with a narrower add, and discard the add-with-constant that is part of the
1640 // range check (if we can't eliminate it, this isn't profitable).
1641
1642 // In order to eliminate the add-with-constant, the compare can be its only
1643 // use.
1644 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1645 if (!AddWithCst->hasOneUse()) return 0;
1646
1647 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1648 if (!CI2->getValue().isPowerOf2()) return 0;
1649 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1650 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1651
1652 // The width of the new add formed is 1 more than the bias.
1653 ++NewWidth;
1654
1655 // Check to see that CI1 is an all-ones value with NewWidth bits.
1656 if (CI1->getBitWidth() == NewWidth ||
1657 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1658 return 0;
1659
1660 // In order to replace the original add with a narrower
1661 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1662 // and truncates that discard the high bits of the add. Verify that this is
1663 // the case.
1664 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1665 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1666 UI != E; ++UI) {
1667 if (*UI == AddWithCst) continue;
1668
1669 // Only accept truncates for now. We would really like a nice recursive
1670 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1671 // chain to see which bits of a value are actually demanded. If the
1672 // original add had another add which was then immediately truncated, we
1673 // could still do the transformation.
1674 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1675 if (TI == 0 ||
1676 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1677 }
1678
1679 // If the pattern matches, truncate the inputs to the narrower type and
1680 // use the sadd_with_overflow intrinsic to efficiently compute both the
1681 // result and the overflow bit.
1682 Module *M = I.getParent()->getParent()->getParent();
1683
1684 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1685 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1686 NewType);
1687
1688 InstCombiner::BuilderTy *Builder = IC.Builder;
1689
1690 // Put the new code above the original add, in case there are any uses of the
1691 // add between the add and the compare.
1692 Builder->SetInsertPoint(OrigAdd);
1693
1694 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1695 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1696 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1697 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1698 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1699
1700 // The inner add was the result of the narrow add, zero extended to the
1701 // wider type. Replace it with the result computed by the intrinsic.
1702 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1703
1704 // The original icmp gets replaced with the overflow value.
1705 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1706 }
1707
ProcessUAddIdiom(Instruction & I,Value * OrigAddV,InstCombiner & IC)1708 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1709 InstCombiner &IC) {
1710 // Don't bother doing this transformation for pointers, don't do it for
1711 // vectors.
1712 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1713
1714 // If the add is a constant expr, then we don't bother transforming it.
1715 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1716 if (OrigAdd == 0) return 0;
1717
1718 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1719
1720 // Put the new code above the original add, in case there are any uses of the
1721 // add between the add and the compare.
1722 InstCombiner::BuilderTy *Builder = IC.Builder;
1723 Builder->SetInsertPoint(OrigAdd);
1724
1725 Module *M = I.getParent()->getParent()->getParent();
1726 Type *Ty = LHS->getType();
1727 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1728 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1729 Value *Add = Builder->CreateExtractValue(Call, 0);
1730
1731 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1732
1733 // The original icmp gets replaced with the overflow value.
1734 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1735 }
1736
1737 // DemandedBitsLHSMask - When performing a comparison against a constant,
1738 // it is possible that not all the bits in the LHS are demanded. This helper
1739 // method computes the mask that IS demanded.
DemandedBitsLHSMask(ICmpInst & I,unsigned BitWidth,bool isSignCheck)1740 static APInt DemandedBitsLHSMask(ICmpInst &I,
1741 unsigned BitWidth, bool isSignCheck) {
1742 if (isSignCheck)
1743 return APInt::getSignBit(BitWidth);
1744
1745 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1746 if (!CI) return APInt::getAllOnesValue(BitWidth);
1747 const APInt &RHS = CI->getValue();
1748
1749 switch (I.getPredicate()) {
1750 // For a UGT comparison, we don't care about any bits that
1751 // correspond to the trailing ones of the comparand. The value of these
1752 // bits doesn't impact the outcome of the comparison, because any value
1753 // greater than the RHS must differ in a bit higher than these due to carry.
1754 case ICmpInst::ICMP_UGT: {
1755 unsigned trailingOnes = RHS.countTrailingOnes();
1756 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1757 return ~lowBitsSet;
1758 }
1759
1760 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1761 // Any value less than the RHS must differ in a higher bit because of carries.
1762 case ICmpInst::ICMP_ULT: {
1763 unsigned trailingZeros = RHS.countTrailingZeros();
1764 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1765 return ~lowBitsSet;
1766 }
1767
1768 default:
1769 return APInt::getAllOnesValue(BitWidth);
1770 }
1771
1772 }
1773
visitICmpInst(ICmpInst & I)1774 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1775 bool Changed = false;
1776 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1777
1778 /// Orders the operands of the compare so that they are listed from most
1779 /// complex to least complex. This puts constants before unary operators,
1780 /// before binary operators.
1781 if (getComplexity(Op0) < getComplexity(Op1)) {
1782 I.swapOperands();
1783 std::swap(Op0, Op1);
1784 Changed = true;
1785 }
1786
1787 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1788 return ReplaceInstUsesWith(I, V);
1789
1790 Type *Ty = Op0->getType();
1791
1792 // icmp's with boolean values can always be turned into bitwise operations
1793 if (Ty->isIntegerTy(1)) {
1794 switch (I.getPredicate()) {
1795 default: llvm_unreachable("Invalid icmp instruction!");
1796 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1797 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1798 return BinaryOperator::CreateNot(Xor);
1799 }
1800 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1801 return BinaryOperator::CreateXor(Op0, Op1);
1802
1803 case ICmpInst::ICMP_UGT:
1804 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1805 // FALL THROUGH
1806 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1807 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1808 return BinaryOperator::CreateAnd(Not, Op1);
1809 }
1810 case ICmpInst::ICMP_SGT:
1811 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1812 // FALL THROUGH
1813 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1814 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1815 return BinaryOperator::CreateAnd(Not, Op0);
1816 }
1817 case ICmpInst::ICMP_UGE:
1818 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1819 // FALL THROUGH
1820 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1821 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1822 return BinaryOperator::CreateOr(Not, Op1);
1823 }
1824 case ICmpInst::ICMP_SGE:
1825 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1826 // FALL THROUGH
1827 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1828 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1829 return BinaryOperator::CreateOr(Not, Op0);
1830 }
1831 }
1832 }
1833
1834 unsigned BitWidth = 0;
1835 if (Ty->isIntOrIntVectorTy())
1836 BitWidth = Ty->getScalarSizeInBits();
1837 else if (TD) // Pointers require TD info to get their size.
1838 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1839
1840 bool isSignBit = false;
1841
1842 // See if we are doing a comparison with a constant.
1843 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1844 Value *A = 0, *B = 0;
1845
1846 // Match the following pattern, which is a common idiom when writing
1847 // overflow-safe integer arithmetic function. The source performs an
1848 // addition in wider type, and explicitly checks for overflow using
1849 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1850 // sadd_with_overflow intrinsic.
1851 //
1852 // TODO: This could probably be generalized to handle other overflow-safe
1853 // operations if we worked out the formulas to compute the appropriate
1854 // magic constants.
1855 //
1856 // sum = a + b
1857 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1858 {
1859 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1860 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1861 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1862 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1863 return Res;
1864 }
1865
1866 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1867 if (I.isEquality() && CI->isZero() &&
1868 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1869 // (icmp cond A B) if cond is equality
1870 return new ICmpInst(I.getPredicate(), A, B);
1871 }
1872
1873 // If we have an icmp le or icmp ge instruction, turn it into the
1874 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1875 // them being folded in the code below. The SimplifyICmpInst code has
1876 // already handled the edge cases for us, so we just assert on them.
1877 switch (I.getPredicate()) {
1878 default: break;
1879 case ICmpInst::ICMP_ULE:
1880 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1881 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1882 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1883 case ICmpInst::ICMP_SLE:
1884 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1885 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1886 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1887 case ICmpInst::ICMP_UGE:
1888 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1889 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1890 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1891 case ICmpInst::ICMP_SGE:
1892 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1893 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1894 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1895 }
1896
1897 // If this comparison is a normal comparison, it demands all
1898 // bits, if it is a sign bit comparison, it only demands the sign bit.
1899 bool UnusedBit;
1900 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1901 }
1902
1903 // See if we can fold the comparison based on range information we can get
1904 // by checking whether bits are known to be zero or one in the input.
1905 if (BitWidth != 0) {
1906 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1907 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1908
1909 if (SimplifyDemandedBits(I.getOperandUse(0),
1910 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1911 Op0KnownZero, Op0KnownOne, 0))
1912 return &I;
1913 if (SimplifyDemandedBits(I.getOperandUse(1),
1914 APInt::getAllOnesValue(BitWidth),
1915 Op1KnownZero, Op1KnownOne, 0))
1916 return &I;
1917
1918 // Given the known and unknown bits, compute a range that the LHS could be
1919 // in. Compute the Min, Max and RHS values based on the known bits. For the
1920 // EQ and NE we use unsigned values.
1921 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1922 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1923 if (I.isSigned()) {
1924 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1925 Op0Min, Op0Max);
1926 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1927 Op1Min, Op1Max);
1928 } else {
1929 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1930 Op0Min, Op0Max);
1931 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1932 Op1Min, Op1Max);
1933 }
1934
1935 // If Min and Max are known to be the same, then SimplifyDemandedBits
1936 // figured out that the LHS is a constant. Just constant fold this now so
1937 // that code below can assume that Min != Max.
1938 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1939 return new ICmpInst(I.getPredicate(),
1940 ConstantInt::get(Op0->getType(), Op0Min), Op1);
1941 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1942 return new ICmpInst(I.getPredicate(), Op0,
1943 ConstantInt::get(Op1->getType(), Op1Min));
1944
1945 // Based on the range information we know about the LHS, see if we can
1946 // simplify this comparison. For example, (x&4) < 8 is always true.
1947 switch (I.getPredicate()) {
1948 default: llvm_unreachable("Unknown icmp opcode!");
1949 case ICmpInst::ICMP_EQ: {
1950 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1951 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
1952
1953 // If all bits are known zero except for one, then we know at most one
1954 // bit is set. If the comparison is against zero, then this is a check
1955 // to see if *that* bit is set.
1956 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1957 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1958 // If the LHS is an AND with the same constant, look through it.
1959 Value *LHS = 0;
1960 ConstantInt *LHSC = 0;
1961 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1962 LHSC->getValue() != Op0KnownZeroInverted)
1963 LHS = Op0;
1964
1965 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1966 // then turn "((1 << x)&8) == 0" into "x != 3".
1967 Value *X = 0;
1968 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1969 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1970 return new ICmpInst(ICmpInst::ICMP_NE, X,
1971 ConstantInt::get(X->getType(), CmpVal));
1972 }
1973
1974 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1975 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1976 const APInt *CI;
1977 if (Op0KnownZeroInverted == 1 &&
1978 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1979 return new ICmpInst(ICmpInst::ICMP_NE, X,
1980 ConstantInt::get(X->getType(),
1981 CI->countTrailingZeros()));
1982 }
1983
1984 break;
1985 }
1986 case ICmpInst::ICMP_NE: {
1987 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1988 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
1989
1990 // If all bits are known zero except for one, then we know at most one
1991 // bit is set. If the comparison is against zero, then this is a check
1992 // to see if *that* bit is set.
1993 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1994 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1995 // If the LHS is an AND with the same constant, look through it.
1996 Value *LHS = 0;
1997 ConstantInt *LHSC = 0;
1998 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1999 LHSC->getValue() != Op0KnownZeroInverted)
2000 LHS = Op0;
2001
2002 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
2003 // then turn "((1 << x)&8) != 0" into "x == 3".
2004 Value *X = 0;
2005 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
2006 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
2007 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2008 ConstantInt::get(X->getType(), CmpVal));
2009 }
2010
2011 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
2012 // then turn "((8 >>u x)&1) != 0" into "x == 3".
2013 const APInt *CI;
2014 if (Op0KnownZeroInverted == 1 &&
2015 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
2016 return new ICmpInst(ICmpInst::ICMP_EQ, X,
2017 ConstantInt::get(X->getType(),
2018 CI->countTrailingZeros()));
2019 }
2020
2021 break;
2022 }
2023 case ICmpInst::ICMP_ULT:
2024 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
2025 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2026 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
2027 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2028 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
2029 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2030 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2031 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
2032 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2033 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2034
2035 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
2036 if (CI->isMinValue(true))
2037 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
2038 Constant::getAllOnesValue(Op0->getType()));
2039 }
2040 break;
2041 case ICmpInst::ICMP_UGT:
2042 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2043 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2044 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2045 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2046
2047 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2048 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2049 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2050 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2051 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2052 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2053
2054 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2055 if (CI->isMaxValue(true))
2056 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2057 Constant::getNullValue(Op0->getType()));
2058 }
2059 break;
2060 case ICmpInst::ICMP_SLT:
2061 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2062 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2063 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2064 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2065 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2066 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2067 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2068 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2069 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2070 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2071 }
2072 break;
2073 case ICmpInst::ICMP_SGT:
2074 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2075 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2076 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2077 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2078
2079 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2080 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2081 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2082 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2083 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2084 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2085 }
2086 break;
2087 case ICmpInst::ICMP_SGE:
2088 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2089 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2090 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2091 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2092 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2093 break;
2094 case ICmpInst::ICMP_SLE:
2095 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2096 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2097 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2098 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2099 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2100 break;
2101 case ICmpInst::ICMP_UGE:
2102 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2103 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2104 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2105 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2106 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2107 break;
2108 case ICmpInst::ICMP_ULE:
2109 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2110 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2111 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2112 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2113 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2114 break;
2115 }
2116
2117 // Turn a signed comparison into an unsigned one if both operands
2118 // are known to have the same sign.
2119 if (I.isSigned() &&
2120 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2121 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2122 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2123 }
2124
2125 // Test if the ICmpInst instruction is used exclusively by a select as
2126 // part of a minimum or maximum operation. If so, refrain from doing
2127 // any other folding. This helps out other analyses which understand
2128 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2129 // and CodeGen. And in this case, at least one of the comparison
2130 // operands has at least one user besides the compare (the select),
2131 // which would often largely negate the benefit of folding anyway.
2132 if (I.hasOneUse())
2133 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2134 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2135 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2136 return 0;
2137
2138 // See if we are doing a comparison between a constant and an instruction that
2139 // can be folded into the comparison.
2140 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2141 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2142 // instruction, see if that instruction also has constants so that the
2143 // instruction can be folded into the icmp
2144 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2145 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2146 return Res;
2147 }
2148
2149 // Handle icmp with constant (but not simple integer constant) RHS
2150 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2151 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2152 switch (LHSI->getOpcode()) {
2153 case Instruction::GetElementPtr:
2154 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2155 if (RHSC->isNullValue() &&
2156 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2157 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2158 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2159 break;
2160 case Instruction::PHI:
2161 // Only fold icmp into the PHI if the phi and icmp are in the same
2162 // block. If in the same block, we're encouraging jump threading. If
2163 // not, we are just pessimizing the code by making an i1 phi.
2164 if (LHSI->getParent() == I.getParent())
2165 if (Instruction *NV = FoldOpIntoPhi(I))
2166 return NV;
2167 break;
2168 case Instruction::Select: {
2169 // If either operand of the select is a constant, we can fold the
2170 // comparison into the select arms, which will cause one to be
2171 // constant folded and the select turned into a bitwise or.
2172 Value *Op1 = 0, *Op2 = 0;
2173 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2174 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2175 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2176 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2177
2178 // We only want to perform this transformation if it will not lead to
2179 // additional code. This is true if either both sides of the select
2180 // fold to a constant (in which case the icmp is replaced with a select
2181 // which will usually simplify) or this is the only user of the
2182 // select (in which case we are trading a select+icmp for a simpler
2183 // select+icmp).
2184 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2185 if (!Op1)
2186 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2187 RHSC, I.getName());
2188 if (!Op2)
2189 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2190 RHSC, I.getName());
2191 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2192 }
2193 break;
2194 }
2195 case Instruction::IntToPtr:
2196 // icmp pred inttoptr(X), null -> icmp pred X, 0
2197 if (RHSC->isNullValue() && TD &&
2198 TD->getIntPtrType(RHSC->getContext()) ==
2199 LHSI->getOperand(0)->getType())
2200 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2201 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2202 break;
2203
2204 case Instruction::Load:
2205 // Try to optimize things like "A[i] > 4" to index computations.
2206 if (GetElementPtrInst *GEP =
2207 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2208 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2209 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2210 !cast<LoadInst>(LHSI)->isVolatile())
2211 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2212 return Res;
2213 }
2214 break;
2215 }
2216 }
2217
2218 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2219 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2220 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2221 return NI;
2222 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2223 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2224 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2225 return NI;
2226
2227 // Test to see if the operands of the icmp are casted versions of other
2228 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2229 // now.
2230 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2231 if (Op0->getType()->isPointerTy() &&
2232 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2233 // We keep moving the cast from the left operand over to the right
2234 // operand, where it can often be eliminated completely.
2235 Op0 = CI->getOperand(0);
2236
2237 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2238 // so eliminate it as well.
2239 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2240 Op1 = CI2->getOperand(0);
2241
2242 // If Op1 is a constant, we can fold the cast into the constant.
2243 if (Op0->getType() != Op1->getType()) {
2244 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2245 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2246 } else {
2247 // Otherwise, cast the RHS right before the icmp
2248 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2249 }
2250 }
2251 return new ICmpInst(I.getPredicate(), Op0, Op1);
2252 }
2253 }
2254
2255 if (isa<CastInst>(Op0)) {
2256 // Handle the special case of: icmp (cast bool to X), <cst>
2257 // This comes up when you have code like
2258 // int X = A < B;
2259 // if (X) ...
2260 // For generality, we handle any zero-extension of any operand comparison
2261 // with a constant or another cast from the same type.
2262 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2263 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2264 return R;
2265 }
2266
2267 // Special logic for binary operators.
2268 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
2269 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
2270 if (BO0 || BO1) {
2271 CmpInst::Predicate Pred = I.getPredicate();
2272 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
2273 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
2274 NoOp0WrapProblem = ICmpInst::isEquality(Pred) ||
2275 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
2276 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
2277 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
2278 NoOp1WrapProblem = ICmpInst::isEquality(Pred) ||
2279 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
2280 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
2281
2282 // Analyze the case when either Op0 or Op1 is an add instruction.
2283 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
2284 Value *A = 0, *B = 0, *C = 0, *D = 0;
2285 if (BO0 && BO0->getOpcode() == Instruction::Add)
2286 A = BO0->getOperand(0), B = BO0->getOperand(1);
2287 if (BO1 && BO1->getOpcode() == Instruction::Add)
2288 C = BO1->getOperand(0), D = BO1->getOperand(1);
2289
2290 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2291 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
2292 return new ICmpInst(Pred, A == Op1 ? B : A,
2293 Constant::getNullValue(Op1->getType()));
2294
2295 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2296 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
2297 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
2298 C == Op0 ? D : C);
2299
2300 // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
2301 if (A && C && (A == C || A == D || B == C || B == D) &&
2302 NoOp0WrapProblem && NoOp1WrapProblem &&
2303 // Try not to increase register pressure.
2304 BO0->hasOneUse() && BO1->hasOneUse()) {
2305 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2306 Value *Y = (A == C || A == D) ? B : A;
2307 Value *Z = (C == A || C == B) ? D : C;
2308 return new ICmpInst(Pred, Y, Z);
2309 }
2310
2311 // Analyze the case when either Op0 or Op1 is a sub instruction.
2312 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
2313 A = 0; B = 0; C = 0; D = 0;
2314 if (BO0 && BO0->getOpcode() == Instruction::Sub)
2315 A = BO0->getOperand(0), B = BO0->getOperand(1);
2316 if (BO1 && BO1->getOpcode() == Instruction::Sub)
2317 C = BO1->getOperand(0), D = BO1->getOperand(1);
2318
2319 // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
2320 if (A == Op1 && NoOp0WrapProblem)
2321 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
2322
2323 // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
2324 if (C == Op0 && NoOp1WrapProblem)
2325 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
2326
2327 // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
2328 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
2329 // Try not to increase register pressure.
2330 BO0->hasOneUse() && BO1->hasOneUse())
2331 return new ICmpInst(Pred, A, C);
2332
2333 // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
2334 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
2335 // Try not to increase register pressure.
2336 BO0->hasOneUse() && BO1->hasOneUse())
2337 return new ICmpInst(Pred, D, B);
2338
2339 BinaryOperator *SRem = NULL;
2340 // icmp (srem X, Y), Y
2341 if (BO0 && BO0->getOpcode() == Instruction::SRem &&
2342 Op1 == BO0->getOperand(1))
2343 SRem = BO0;
2344 // icmp Y, (srem X, Y)
2345 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
2346 Op0 == BO1->getOperand(1))
2347 SRem = BO1;
2348 if (SRem) {
2349 // We don't check hasOneUse to avoid increasing register pressure because
2350 // the value we use is the same value this instruction was already using.
2351 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
2352 default: break;
2353 case ICmpInst::ICMP_EQ:
2354 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
2355 case ICmpInst::ICMP_NE:
2356 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
2357 case ICmpInst::ICMP_SGT:
2358 case ICmpInst::ICMP_SGE:
2359 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
2360 Constant::getAllOnesValue(SRem->getType()));
2361 case ICmpInst::ICMP_SLT:
2362 case ICmpInst::ICMP_SLE:
2363 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
2364 Constant::getNullValue(SRem->getType()));
2365 }
2366 }
2367
2368 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
2369 BO0->hasOneUse() && BO1->hasOneUse() &&
2370 BO0->getOperand(1) == BO1->getOperand(1)) {
2371 switch (BO0->getOpcode()) {
2372 default: break;
2373 case Instruction::Add:
2374 case Instruction::Sub:
2375 case Instruction::Xor:
2376 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2377 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2378 BO1->getOperand(0));
2379 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2380 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2381 if (CI->getValue().isSignBit()) {
2382 ICmpInst::Predicate Pred = I.isSigned()
2383 ? I.getUnsignedPredicate()
2384 : I.getSignedPredicate();
2385 return new ICmpInst(Pred, BO0->getOperand(0),
2386 BO1->getOperand(0));
2387 }
2388
2389 if (CI->isMaxValue(true)) {
2390 ICmpInst::Predicate Pred = I.isSigned()
2391 ? I.getUnsignedPredicate()
2392 : I.getSignedPredicate();
2393 Pred = I.getSwappedPredicate(Pred);
2394 return new ICmpInst(Pred, BO0->getOperand(0),
2395 BO1->getOperand(0));
2396 }
2397 }
2398 break;
2399 case Instruction::Mul:
2400 if (!I.isEquality())
2401 break;
2402
2403 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO0->getOperand(1))) {
2404 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2405 // Mask = -1 >> count-trailing-zeros(Cst).
2406 if (!CI->isZero() && !CI->isOne()) {
2407 const APInt &AP = CI->getValue();
2408 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2409 APInt::getLowBitsSet(AP.getBitWidth(),
2410 AP.getBitWidth() -
2411 AP.countTrailingZeros()));
2412 Value *And1 = Builder->CreateAnd(BO0->getOperand(0), Mask);
2413 Value *And2 = Builder->CreateAnd(BO1->getOperand(0), Mask);
2414 return new ICmpInst(I.getPredicate(), And1, And2);
2415 }
2416 }
2417 break;
2418 case Instruction::UDiv:
2419 case Instruction::LShr:
2420 if (I.isSigned())
2421 break;
2422 // fall-through
2423 case Instruction::SDiv:
2424 case Instruction::AShr:
2425 if (!BO0->isExact() || !BO1->isExact())
2426 break;
2427 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2428 BO1->getOperand(0));
2429 case Instruction::Shl: {
2430 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
2431 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
2432 if (!NUW && !NSW)
2433 break;
2434 if (!NSW && I.isSigned())
2435 break;
2436 return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
2437 BO1->getOperand(0));
2438 }
2439 }
2440 }
2441 }
2442
2443 { Value *A, *B;
2444 // ~x < ~y --> y < x
2445 // ~x < cst --> ~cst < x
2446 if (match(Op0, m_Not(m_Value(A)))) {
2447 if (match(Op1, m_Not(m_Value(B))))
2448 return new ICmpInst(I.getPredicate(), B, A);
2449 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2450 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2451 }
2452
2453 // (a+b) <u a --> llvm.uadd.with.overflow.
2454 // (a+b) <u b --> llvm.uadd.with.overflow.
2455 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2456 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2457 (Op1 == A || Op1 == B))
2458 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2459 return R;
2460
2461 // a >u (a+b) --> llvm.uadd.with.overflow.
2462 // b >u (a+b) --> llvm.uadd.with.overflow.
2463 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2464 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2465 (Op0 == A || Op0 == B))
2466 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2467 return R;
2468 }
2469
2470 if (I.isEquality()) {
2471 Value *A, *B, *C, *D;
2472
2473 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2474 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2475 Value *OtherVal = A == Op1 ? B : A;
2476 return new ICmpInst(I.getPredicate(), OtherVal,
2477 Constant::getNullValue(A->getType()));
2478 }
2479
2480 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2481 // A^c1 == C^c2 --> A == C^(c1^c2)
2482 ConstantInt *C1, *C2;
2483 if (match(B, m_ConstantInt(C1)) &&
2484 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2485 Constant *NC = ConstantInt::get(I.getContext(),
2486 C1->getValue() ^ C2->getValue());
2487 Value *Xor = Builder->CreateXor(C, NC);
2488 return new ICmpInst(I.getPredicate(), A, Xor);
2489 }
2490
2491 // A^B == A^D -> B == D
2492 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2493 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2494 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2495 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2496 }
2497 }
2498
2499 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2500 (A == Op0 || B == Op0)) {
2501 // A == (A^B) -> B == 0
2502 Value *OtherVal = A == Op0 ? B : A;
2503 return new ICmpInst(I.getPredicate(), OtherVal,
2504 Constant::getNullValue(A->getType()));
2505 }
2506
2507 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2508 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
2509 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
2510 Value *X = 0, *Y = 0, *Z = 0;
2511
2512 if (A == C) {
2513 X = B; Y = D; Z = A;
2514 } else if (A == D) {
2515 X = B; Y = C; Z = A;
2516 } else if (B == C) {
2517 X = A; Y = D; Z = B;
2518 } else if (B == D) {
2519 X = A; Y = C; Z = B;
2520 }
2521
2522 if (X) { // Build (X^Y) & Z
2523 Op1 = Builder->CreateXor(X, Y);
2524 Op1 = Builder->CreateAnd(Op1, Z);
2525 I.setOperand(0, Op1);
2526 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2527 return &I;
2528 }
2529 }
2530
2531 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
2532 // "icmp (and X, mask), cst"
2533 uint64_t ShAmt = 0;
2534 ConstantInt *Cst1;
2535 if (Op0->hasOneUse() &&
2536 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
2537 m_ConstantInt(ShAmt))))) &&
2538 match(Op1, m_ConstantInt(Cst1)) &&
2539 // Only do this when A has multiple uses. This is most important to do
2540 // when it exposes other optimizations.
2541 !A->hasOneUse()) {
2542 unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
2543
2544 if (ShAmt < ASize) {
2545 APInt MaskV =
2546 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
2547 MaskV <<= ShAmt;
2548
2549 APInt CmpV = Cst1->getValue().zext(ASize);
2550 CmpV <<= ShAmt;
2551
2552 Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
2553 return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
2554 }
2555 }
2556 }
2557
2558 {
2559 Value *X; ConstantInt *Cst;
2560 // icmp X+Cst, X
2561 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2562 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2563
2564 // icmp X, X+Cst
2565 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2566 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2567 }
2568 return Changed ? &I : 0;
2569 }
2570
2571
2572
2573
2574
2575
2576 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2577 ///
FoldFCmp_IntToFP_Cst(FCmpInst & I,Instruction * LHSI,Constant * RHSC)2578 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2579 Instruction *LHSI,
2580 Constant *RHSC) {
2581 if (!isa<ConstantFP>(RHSC)) return 0;
2582 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2583
2584 // Get the width of the mantissa. We don't want to hack on conversions that
2585 // might lose information from the integer, e.g. "i64 -> float"
2586 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2587 if (MantissaWidth == -1) return 0; // Unknown.
2588
2589 // Check to see that the input is converted from an integer type that is small
2590 // enough that preserves all bits. TODO: check here for "known" sign bits.
2591 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2592 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2593
2594 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2595 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2596 if (LHSUnsigned)
2597 ++InputSize;
2598
2599 // If the conversion would lose info, don't hack on this.
2600 if ((int)InputSize > MantissaWidth)
2601 return 0;
2602
2603 // Otherwise, we can potentially simplify the comparison. We know that it
2604 // will always come through as an integer value and we know the constant is
2605 // not a NAN (it would have been previously simplified).
2606 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2607
2608 ICmpInst::Predicate Pred;
2609 switch (I.getPredicate()) {
2610 default: llvm_unreachable("Unexpected predicate!");
2611 case FCmpInst::FCMP_UEQ:
2612 case FCmpInst::FCMP_OEQ:
2613 Pred = ICmpInst::ICMP_EQ;
2614 break;
2615 case FCmpInst::FCMP_UGT:
2616 case FCmpInst::FCMP_OGT:
2617 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2618 break;
2619 case FCmpInst::FCMP_UGE:
2620 case FCmpInst::FCMP_OGE:
2621 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2622 break;
2623 case FCmpInst::FCMP_ULT:
2624 case FCmpInst::FCMP_OLT:
2625 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2626 break;
2627 case FCmpInst::FCMP_ULE:
2628 case FCmpInst::FCMP_OLE:
2629 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2630 break;
2631 case FCmpInst::FCMP_UNE:
2632 case FCmpInst::FCMP_ONE:
2633 Pred = ICmpInst::ICMP_NE;
2634 break;
2635 case FCmpInst::FCMP_ORD:
2636 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2637 case FCmpInst::FCMP_UNO:
2638 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2639 }
2640
2641 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2642
2643 // Now we know that the APFloat is a normal number, zero or inf.
2644
2645 // See if the FP constant is too large for the integer. For example,
2646 // comparing an i8 to 300.0.
2647 unsigned IntWidth = IntTy->getScalarSizeInBits();
2648
2649 if (!LHSUnsigned) {
2650 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2651 // and large values.
2652 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2653 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2654 APFloat::rmNearestTiesToEven);
2655 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2656 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2657 Pred == ICmpInst::ICMP_SLE)
2658 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2659 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2660 }
2661 } else {
2662 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2663 // +INF and large values.
2664 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2665 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2666 APFloat::rmNearestTiesToEven);
2667 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2668 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2669 Pred == ICmpInst::ICMP_ULE)
2670 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2671 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2672 }
2673 }
2674
2675 if (!LHSUnsigned) {
2676 // See if the RHS value is < SignedMin.
2677 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2678 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2679 APFloat::rmNearestTiesToEven);
2680 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2681 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2682 Pred == ICmpInst::ICMP_SGE)
2683 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2684 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2685 }
2686 }
2687
2688 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2689 // [0, UMAX], but it may still be fractional. See if it is fractional by
2690 // casting the FP value to the integer value and back, checking for equality.
2691 // Don't do this for zero, because -0.0 is not fractional.
2692 Constant *RHSInt = LHSUnsigned
2693 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2694 : ConstantExpr::getFPToSI(RHSC, IntTy);
2695 if (!RHS.isZero()) {
2696 bool Equal = LHSUnsigned
2697 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2698 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2699 if (!Equal) {
2700 // If we had a comparison against a fractional value, we have to adjust
2701 // the compare predicate and sometimes the value. RHSC is rounded towards
2702 // zero at this point.
2703 switch (Pred) {
2704 default: llvm_unreachable("Unexpected integer comparison!");
2705 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2706 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2707 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2708 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2709 case ICmpInst::ICMP_ULE:
2710 // (float)int <= 4.4 --> int <= 4
2711 // (float)int <= -4.4 --> false
2712 if (RHS.isNegative())
2713 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2714 break;
2715 case ICmpInst::ICMP_SLE:
2716 // (float)int <= 4.4 --> int <= 4
2717 // (float)int <= -4.4 --> int < -4
2718 if (RHS.isNegative())
2719 Pred = ICmpInst::ICMP_SLT;
2720 break;
2721 case ICmpInst::ICMP_ULT:
2722 // (float)int < -4.4 --> false
2723 // (float)int < 4.4 --> int <= 4
2724 if (RHS.isNegative())
2725 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2726 Pred = ICmpInst::ICMP_ULE;
2727 break;
2728 case ICmpInst::ICMP_SLT:
2729 // (float)int < -4.4 --> int < -4
2730 // (float)int < 4.4 --> int <= 4
2731 if (!RHS.isNegative())
2732 Pred = ICmpInst::ICMP_SLE;
2733 break;
2734 case ICmpInst::ICMP_UGT:
2735 // (float)int > 4.4 --> int > 4
2736 // (float)int > -4.4 --> true
2737 if (RHS.isNegative())
2738 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2739 break;
2740 case ICmpInst::ICMP_SGT:
2741 // (float)int > 4.4 --> int > 4
2742 // (float)int > -4.4 --> int >= -4
2743 if (RHS.isNegative())
2744 Pred = ICmpInst::ICMP_SGE;
2745 break;
2746 case ICmpInst::ICMP_UGE:
2747 // (float)int >= -4.4 --> true
2748 // (float)int >= 4.4 --> int > 4
2749 if (!RHS.isNegative())
2750 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2751 Pred = ICmpInst::ICMP_UGT;
2752 break;
2753 case ICmpInst::ICMP_SGE:
2754 // (float)int >= -4.4 --> int >= -4
2755 // (float)int >= 4.4 --> int > 4
2756 if (!RHS.isNegative())
2757 Pred = ICmpInst::ICMP_SGT;
2758 break;
2759 }
2760 }
2761 }
2762
2763 // Lower this FP comparison into an appropriate integer version of the
2764 // comparison.
2765 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2766 }
2767
visitFCmpInst(FCmpInst & I)2768 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2769 bool Changed = false;
2770
2771 /// Orders the operands of the compare so that they are listed from most
2772 /// complex to least complex. This puts constants before unary operators,
2773 /// before binary operators.
2774 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2775 I.swapOperands();
2776 Changed = true;
2777 }
2778
2779 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2780
2781 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2782 return ReplaceInstUsesWith(I, V);
2783
2784 // Simplify 'fcmp pred X, X'
2785 if (Op0 == Op1) {
2786 switch (I.getPredicate()) {
2787 default: llvm_unreachable("Unknown predicate!");
2788 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2789 case FCmpInst::FCMP_ULT: // True if unordered or less than
2790 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2791 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2792 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2793 I.setPredicate(FCmpInst::FCMP_UNO);
2794 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2795 return &I;
2796
2797 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2798 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2799 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2800 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2801 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2802 I.setPredicate(FCmpInst::FCMP_ORD);
2803 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2804 return &I;
2805 }
2806 }
2807
2808 // Handle fcmp with constant RHS
2809 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2810 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2811 switch (LHSI->getOpcode()) {
2812 case Instruction::FPExt: {
2813 // fcmp (fpext x), C -> fcmp x, (fptrunc C) if fptrunc is lossless
2814 FPExtInst *LHSExt = cast<FPExtInst>(LHSI);
2815 ConstantFP *RHSF = dyn_cast<ConstantFP>(RHSC);
2816 if (!RHSF)
2817 break;
2818
2819 // We can't convert a PPC double double.
2820 if (RHSF->getType()->isPPC_FP128Ty())
2821 break;
2822
2823 const fltSemantics *Sem;
2824 // FIXME: This shouldn't be here.
2825 if (LHSExt->getSrcTy()->isFloatTy())
2826 Sem = &APFloat::IEEEsingle;
2827 else if (LHSExt->getSrcTy()->isDoubleTy())
2828 Sem = &APFloat::IEEEdouble;
2829 else if (LHSExt->getSrcTy()->isFP128Ty())
2830 Sem = &APFloat::IEEEquad;
2831 else if (LHSExt->getSrcTy()->isX86_FP80Ty())
2832 Sem = &APFloat::x87DoubleExtended;
2833 else
2834 break;
2835
2836 bool Lossy;
2837 APFloat F = RHSF->getValueAPF();
2838 F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
2839
2840 // Avoid lossy conversions and denormals. Zero is a special case
2841 // that's OK to convert.
2842 APFloat Fabs = F;
2843 Fabs.clearSign();
2844 if (!Lossy &&
2845 ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
2846 APFloat::cmpLessThan) || Fabs.isZero()))
2847
2848 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2849 ConstantFP::get(RHSC->getContext(), F));
2850 break;
2851 }
2852 case Instruction::PHI:
2853 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2854 // block. If in the same block, we're encouraging jump threading. If
2855 // not, we are just pessimizing the code by making an i1 phi.
2856 if (LHSI->getParent() == I.getParent())
2857 if (Instruction *NV = FoldOpIntoPhi(I))
2858 return NV;
2859 break;
2860 case Instruction::SIToFP:
2861 case Instruction::UIToFP:
2862 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2863 return NV;
2864 break;
2865 case Instruction::Select: {
2866 // If either operand of the select is a constant, we can fold the
2867 // comparison into the select arms, which will cause one to be
2868 // constant folded and the select turned into a bitwise or.
2869 Value *Op1 = 0, *Op2 = 0;
2870 if (LHSI->hasOneUse()) {
2871 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2872 // Fold the known value into the constant operand.
2873 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2874 // Insert a new FCmp of the other select operand.
2875 Op2 = Builder->CreateFCmp(I.getPredicate(),
2876 LHSI->getOperand(2), RHSC, I.getName());
2877 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2878 // Fold the known value into the constant operand.
2879 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2880 // Insert a new FCmp of the other select operand.
2881 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2882 RHSC, I.getName());
2883 }
2884 }
2885
2886 if (Op1)
2887 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2888 break;
2889 }
2890 case Instruction::FSub: {
2891 // fcmp pred (fneg x), C -> fcmp swap(pred) x, -C
2892 Value *Op;
2893 if (match(LHSI, m_FNeg(m_Value(Op))))
2894 return new FCmpInst(I.getSwappedPredicate(), Op,
2895 ConstantExpr::getFNeg(RHSC));
2896 break;
2897 }
2898 case Instruction::Load:
2899 if (GetElementPtrInst *GEP =
2900 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2901 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2902 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2903 !cast<LoadInst>(LHSI)->isVolatile())
2904 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2905 return Res;
2906 }
2907 break;
2908 }
2909 }
2910
2911 // fcmp pred (fneg x), (fneg y) -> fcmp swap(pred) x, y
2912 Value *X, *Y;
2913 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
2914 return new FCmpInst(I.getSwappedPredicate(), X, Y);
2915
2916 // fcmp (fpext x), (fpext y) -> fcmp x, y
2917 if (FPExtInst *LHSExt = dyn_cast<FPExtInst>(Op0))
2918 if (FPExtInst *RHSExt = dyn_cast<FPExtInst>(Op1))
2919 if (LHSExt->getSrcTy() == RHSExt->getSrcTy())
2920 return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
2921 RHSExt->getOperand(0));
2922
2923 return Changed ? &I : 0;
2924 }
2925