1 //===- InstCombineCasts.cpp -----------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visit functions for cast operations.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/SetVector.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/IR/DataLayout.h"
17 #include "llvm/IR/DebugInfo.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Support/KnownBits.h"
20 #include "llvm/Transforms/InstCombine/InstCombiner.h"
21 #include <optional>
22
23 using namespace llvm;
24 using namespace PatternMatch;
25
26 #define DEBUG_TYPE "instcombine"
27
28 /// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns
29 /// true for, actually insert the code to evaluate the expression.
EvaluateInDifferentType(Value * V,Type * Ty,bool isSigned)30 Value *InstCombinerImpl::EvaluateInDifferentType(Value *V, Type *Ty,
31 bool isSigned) {
32 if (Constant *C = dyn_cast<Constant>(V))
33 return ConstantFoldIntegerCast(C, Ty, isSigned, DL);
34
35 // Otherwise, it must be an instruction.
36 Instruction *I = cast<Instruction>(V);
37 Instruction *Res = nullptr;
38 unsigned Opc = I->getOpcode();
39 switch (Opc) {
40 case Instruction::Add:
41 case Instruction::Sub:
42 case Instruction::Mul:
43 case Instruction::And:
44 case Instruction::Or:
45 case Instruction::Xor:
46 case Instruction::AShr:
47 case Instruction::LShr:
48 case Instruction::Shl:
49 case Instruction::UDiv:
50 case Instruction::URem: {
51 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
52 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
53 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
54 break;
55 }
56 case Instruction::Trunc:
57 case Instruction::ZExt:
58 case Instruction::SExt:
59 // If the source type of the cast is the type we're trying for then we can
60 // just return the source. There's no need to insert it because it is not
61 // new.
62 if (I->getOperand(0)->getType() == Ty)
63 return I->getOperand(0);
64
65 // Otherwise, must be the same type of cast, so just reinsert a new one.
66 // This also handles the case of zext(trunc(x)) -> zext(x).
67 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,
68 Opc == Instruction::SExt);
69 break;
70 case Instruction::Select: {
71 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
72 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
73 Res = SelectInst::Create(I->getOperand(0), True, False);
74 break;
75 }
76 case Instruction::PHI: {
77 PHINode *OPN = cast<PHINode>(I);
78 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());
79 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
80 Value *V =
81 EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
82 NPN->addIncoming(V, OPN->getIncomingBlock(i));
83 }
84 Res = NPN;
85 break;
86 }
87 case Instruction::FPToUI:
88 case Instruction::FPToSI:
89 Res = CastInst::Create(
90 static_cast<Instruction::CastOps>(Opc), I->getOperand(0), Ty);
91 break;
92 case Instruction::Call:
93 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
94 switch (II->getIntrinsicID()) {
95 default:
96 llvm_unreachable("Unsupported call!");
97 case Intrinsic::vscale: {
98 Function *Fn =
99 Intrinsic::getDeclaration(I->getModule(), Intrinsic::vscale, {Ty});
100 Res = CallInst::Create(Fn->getFunctionType(), Fn);
101 break;
102 }
103 }
104 }
105 break;
106 case Instruction::ShuffleVector: {
107 auto *ScalarTy = cast<VectorType>(Ty)->getElementType();
108 auto *VTy = cast<VectorType>(I->getOperand(0)->getType());
109 auto *FixedTy = VectorType::get(ScalarTy, VTy->getElementCount());
110 Value *Op0 = EvaluateInDifferentType(I->getOperand(0), FixedTy, isSigned);
111 Value *Op1 = EvaluateInDifferentType(I->getOperand(1), FixedTy, isSigned);
112 Res = new ShuffleVectorInst(Op0, Op1,
113 cast<ShuffleVectorInst>(I)->getShuffleMask());
114 break;
115 }
116 default:
117 // TODO: Can handle more cases here.
118 llvm_unreachable("Unreachable!");
119 }
120
121 Res->takeName(I);
122 return InsertNewInstWith(Res, I->getIterator());
123 }
124
125 Instruction::CastOps
isEliminableCastPair(const CastInst * CI1,const CastInst * CI2)126 InstCombinerImpl::isEliminableCastPair(const CastInst *CI1,
127 const CastInst *CI2) {
128 Type *SrcTy = CI1->getSrcTy();
129 Type *MidTy = CI1->getDestTy();
130 Type *DstTy = CI2->getDestTy();
131
132 Instruction::CastOps firstOp = CI1->getOpcode();
133 Instruction::CastOps secondOp = CI2->getOpcode();
134 Type *SrcIntPtrTy =
135 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;
136 Type *MidIntPtrTy =
137 MidTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(MidTy) : nullptr;
138 Type *DstIntPtrTy =
139 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;
140 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
141 DstTy, SrcIntPtrTy, MidIntPtrTy,
142 DstIntPtrTy);
143
144 // We don't want to form an inttoptr or ptrtoint that converts to an integer
145 // type that differs from the pointer size.
146 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||
147 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))
148 Res = 0;
149
150 return Instruction::CastOps(Res);
151 }
152
153 /// Implement the transforms common to all CastInst visitors.
commonCastTransforms(CastInst & CI)154 Instruction *InstCombinerImpl::commonCastTransforms(CastInst &CI) {
155 Value *Src = CI.getOperand(0);
156 Type *Ty = CI.getType();
157
158 if (auto *SrcC = dyn_cast<Constant>(Src))
159 if (Constant *Res = ConstantFoldCastOperand(CI.getOpcode(), SrcC, Ty, DL))
160 return replaceInstUsesWith(CI, Res);
161
162 // Try to eliminate a cast of a cast.
163 if (auto *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
164 if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {
165 // The first cast (CSrc) is eliminable so we need to fix up or replace
166 // the second cast (CI). CSrc will then have a good chance of being dead.
167 auto *Res = CastInst::Create(NewOpc, CSrc->getOperand(0), Ty);
168 // Point debug users of the dying cast to the new one.
169 if (CSrc->hasOneUse())
170 replaceAllDbgUsesWith(*CSrc, *Res, CI, DT);
171 return Res;
172 }
173 }
174
175 if (auto *Sel = dyn_cast<SelectInst>(Src)) {
176 // We are casting a select. Try to fold the cast into the select if the
177 // select does not have a compare instruction with matching operand types
178 // or the select is likely better done in a narrow type.
179 // Creating a select with operands that are different sizes than its
180 // condition may inhibit other folds and lead to worse codegen.
181 auto *Cmp = dyn_cast<CmpInst>(Sel->getCondition());
182 if (!Cmp || Cmp->getOperand(0)->getType() != Sel->getType() ||
183 (CI.getOpcode() == Instruction::Trunc &&
184 shouldChangeType(CI.getSrcTy(), CI.getType()))) {
185 if (Instruction *NV = FoldOpIntoSelect(CI, Sel)) {
186 replaceAllDbgUsesWith(*Sel, *NV, CI, DT);
187 return NV;
188 }
189 }
190 }
191
192 // If we are casting a PHI, then fold the cast into the PHI.
193 if (auto *PN = dyn_cast<PHINode>(Src)) {
194 // Don't do this if it would create a PHI node with an illegal type from a
195 // legal type.
196 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
197 shouldChangeType(CI.getSrcTy(), CI.getType()))
198 if (Instruction *NV = foldOpIntoPhi(CI, PN))
199 return NV;
200 }
201
202 // Canonicalize a unary shuffle after the cast if neither operation changes
203 // the size or element size of the input vector.
204 // TODO: We could allow size-changing ops if that doesn't harm codegen.
205 // cast (shuffle X, Mask) --> shuffle (cast X), Mask
206 Value *X;
207 ArrayRef<int> Mask;
208 if (match(Src, m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(Mask))))) {
209 // TODO: Allow scalable vectors?
210 auto *SrcTy = dyn_cast<FixedVectorType>(X->getType());
211 auto *DestTy = dyn_cast<FixedVectorType>(Ty);
212 if (SrcTy && DestTy &&
213 SrcTy->getNumElements() == DestTy->getNumElements() &&
214 SrcTy->getPrimitiveSizeInBits() == DestTy->getPrimitiveSizeInBits()) {
215 Value *CastX = Builder.CreateCast(CI.getOpcode(), X, DestTy);
216 return new ShuffleVectorInst(CastX, Mask);
217 }
218 }
219
220 return nullptr;
221 }
222
223 /// Constants and extensions/truncates from the destination type are always
224 /// free to be evaluated in that type. This is a helper for canEvaluate*.
canAlwaysEvaluateInType(Value * V,Type * Ty)225 static bool canAlwaysEvaluateInType(Value *V, Type *Ty) {
226 if (isa<Constant>(V))
227 return match(V, m_ImmConstant());
228
229 Value *X;
230 if ((match(V, m_ZExtOrSExt(m_Value(X))) || match(V, m_Trunc(m_Value(X)))) &&
231 X->getType() == Ty)
232 return true;
233
234 return false;
235 }
236
237 /// Filter out values that we can not evaluate in the destination type for free.
238 /// This is a helper for canEvaluate*.
canNotEvaluateInType(Value * V,Type * Ty)239 static bool canNotEvaluateInType(Value *V, Type *Ty) {
240 if (!isa<Instruction>(V))
241 return true;
242 // We don't extend or shrink something that has multiple uses -- doing so
243 // would require duplicating the instruction which isn't profitable.
244 if (!V->hasOneUse())
245 return true;
246
247 return false;
248 }
249
250 /// Return true if we can evaluate the specified expression tree as type Ty
251 /// instead of its larger type, and arrive with the same value.
252 /// This is used by code that tries to eliminate truncates.
253 ///
254 /// Ty will always be a type smaller than V. We should return true if trunc(V)
255 /// can be computed by computing V in the smaller type. If V is an instruction,
256 /// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only
257 /// makes sense if x and y can be efficiently truncated.
258 ///
259 /// This function works on both vectors and scalars.
260 ///
canEvaluateTruncated(Value * V,Type * Ty,InstCombinerImpl & IC,Instruction * CxtI)261 static bool canEvaluateTruncated(Value *V, Type *Ty, InstCombinerImpl &IC,
262 Instruction *CxtI) {
263 if (canAlwaysEvaluateInType(V, Ty))
264 return true;
265 if (canNotEvaluateInType(V, Ty))
266 return false;
267
268 auto *I = cast<Instruction>(V);
269 Type *OrigTy = V->getType();
270 switch (I->getOpcode()) {
271 case Instruction::Add:
272 case Instruction::Sub:
273 case Instruction::Mul:
274 case Instruction::And:
275 case Instruction::Or:
276 case Instruction::Xor:
277 // These operators can all arbitrarily be extended or truncated.
278 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
279 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
280
281 case Instruction::UDiv:
282 case Instruction::URem: {
283 // UDiv and URem can be truncated if all the truncated bits are zero.
284 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
285 uint32_t BitWidth = Ty->getScalarSizeInBits();
286 assert(BitWidth < OrigBitWidth && "Unexpected bitwidths!");
287 APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
288 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, 0, CxtI) &&
289 IC.MaskedValueIsZero(I->getOperand(1), Mask, 0, CxtI)) {
290 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
291 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
292 }
293 break;
294 }
295 case Instruction::Shl: {
296 // If we are truncating the result of this SHL, and if it's a shift of an
297 // inrange amount, we can always perform a SHL in a smaller type.
298 uint32_t BitWidth = Ty->getScalarSizeInBits();
299 KnownBits AmtKnownBits =
300 llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
301 if (AmtKnownBits.getMaxValue().ult(BitWidth))
302 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
303 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
304 break;
305 }
306 case Instruction::LShr: {
307 // If this is a truncate of a logical shr, we can truncate it to a smaller
308 // lshr iff we know that the bits we would otherwise be shifting in are
309 // already zeros.
310 // TODO: It is enough to check that the bits we would be shifting in are
311 // zero - use AmtKnownBits.getMaxValue().
312 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
313 uint32_t BitWidth = Ty->getScalarSizeInBits();
314 KnownBits AmtKnownBits =
315 llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
316 APInt ShiftedBits = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
317 if (AmtKnownBits.getMaxValue().ult(BitWidth) &&
318 IC.MaskedValueIsZero(I->getOperand(0), ShiftedBits, 0, CxtI)) {
319 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
320 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
321 }
322 break;
323 }
324 case Instruction::AShr: {
325 // If this is a truncate of an arithmetic shr, we can truncate it to a
326 // smaller ashr iff we know that all the bits from the sign bit of the
327 // original type and the sign bit of the truncate type are similar.
328 // TODO: It is enough to check that the bits we would be shifting in are
329 // similar to sign bit of the truncate type.
330 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
331 uint32_t BitWidth = Ty->getScalarSizeInBits();
332 KnownBits AmtKnownBits =
333 llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());
334 unsigned ShiftedBits = OrigBitWidth - BitWidth;
335 if (AmtKnownBits.getMaxValue().ult(BitWidth) &&
336 ShiftedBits < IC.ComputeNumSignBits(I->getOperand(0), 0, CxtI))
337 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
338 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
339 break;
340 }
341 case Instruction::Trunc:
342 // trunc(trunc(x)) -> trunc(x)
343 return true;
344 case Instruction::ZExt:
345 case Instruction::SExt:
346 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
347 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest
348 return true;
349 case Instruction::Select: {
350 SelectInst *SI = cast<SelectInst>(I);
351 return canEvaluateTruncated(SI->getTrueValue(), Ty, IC, CxtI) &&
352 canEvaluateTruncated(SI->getFalseValue(), Ty, IC, CxtI);
353 }
354 case Instruction::PHI: {
355 // We can change a phi if we can change all operands. Note that we never
356 // get into trouble with cyclic PHIs here because we only consider
357 // instructions with a single use.
358 PHINode *PN = cast<PHINode>(I);
359 for (Value *IncValue : PN->incoming_values())
360 if (!canEvaluateTruncated(IncValue, Ty, IC, CxtI))
361 return false;
362 return true;
363 }
364 case Instruction::FPToUI:
365 case Instruction::FPToSI: {
366 // If the integer type can hold the max FP value, it is safe to cast
367 // directly to that type. Otherwise, we may create poison via overflow
368 // that did not exist in the original code.
369 Type *InputTy = I->getOperand(0)->getType()->getScalarType();
370 const fltSemantics &Semantics = InputTy->getFltSemantics();
371 uint32_t MinBitWidth =
372 APFloatBase::semanticsIntSizeInBits(Semantics,
373 I->getOpcode() == Instruction::FPToSI);
374 return Ty->getScalarSizeInBits() >= MinBitWidth;
375 }
376 case Instruction::ShuffleVector:
377 return canEvaluateTruncated(I->getOperand(0), Ty, IC, CxtI) &&
378 canEvaluateTruncated(I->getOperand(1), Ty, IC, CxtI);
379 default:
380 // TODO: Can handle more cases here.
381 break;
382 }
383
384 return false;
385 }
386
387 /// Given a vector that is bitcast to an integer, optionally logically
388 /// right-shifted, and truncated, convert it to an extractelement.
389 /// Example (big endian):
390 /// trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32
391 /// --->
392 /// extractelement <4 x i32> %X, 1
foldVecTruncToExtElt(TruncInst & Trunc,InstCombinerImpl & IC)393 static Instruction *foldVecTruncToExtElt(TruncInst &Trunc,
394 InstCombinerImpl &IC) {
395 Value *TruncOp = Trunc.getOperand(0);
396 Type *DestType = Trunc.getType();
397 if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))
398 return nullptr;
399
400 Value *VecInput = nullptr;
401 ConstantInt *ShiftVal = nullptr;
402 if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),
403 m_LShr(m_BitCast(m_Value(VecInput)),
404 m_ConstantInt(ShiftVal)))) ||
405 !isa<VectorType>(VecInput->getType()))
406 return nullptr;
407
408 VectorType *VecType = cast<VectorType>(VecInput->getType());
409 unsigned VecWidth = VecType->getPrimitiveSizeInBits();
410 unsigned DestWidth = DestType->getPrimitiveSizeInBits();
411 unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;
412
413 if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))
414 return nullptr;
415
416 // If the element type of the vector doesn't match the result type,
417 // bitcast it to a vector type that we can extract from.
418 unsigned NumVecElts = VecWidth / DestWidth;
419 if (VecType->getElementType() != DestType) {
420 VecType = FixedVectorType::get(DestType, NumVecElts);
421 VecInput = IC.Builder.CreateBitCast(VecInput, VecType, "bc");
422 }
423
424 unsigned Elt = ShiftAmount / DestWidth;
425 if (IC.getDataLayout().isBigEndian())
426 Elt = NumVecElts - 1 - Elt;
427
428 return ExtractElementInst::Create(VecInput, IC.Builder.getInt32(Elt));
429 }
430
431 /// Funnel/Rotate left/right may occur in a wider type than necessary because of
432 /// type promotion rules. Try to narrow the inputs and convert to funnel shift.
narrowFunnelShift(TruncInst & Trunc)433 Instruction *InstCombinerImpl::narrowFunnelShift(TruncInst &Trunc) {
434 assert((isa<VectorType>(Trunc.getSrcTy()) ||
435 shouldChangeType(Trunc.getSrcTy(), Trunc.getType())) &&
436 "Don't narrow to an illegal scalar type");
437
438 // Bail out on strange types. It is possible to handle some of these patterns
439 // even with non-power-of-2 sizes, but it is not a likely scenario.
440 Type *DestTy = Trunc.getType();
441 unsigned NarrowWidth = DestTy->getScalarSizeInBits();
442 unsigned WideWidth = Trunc.getSrcTy()->getScalarSizeInBits();
443 if (!isPowerOf2_32(NarrowWidth))
444 return nullptr;
445
446 // First, find an or'd pair of opposite shifts:
447 // trunc (or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1))
448 BinaryOperator *Or0, *Or1;
449 if (!match(Trunc.getOperand(0), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1)))))
450 return nullptr;
451
452 Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;
453 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) ||
454 !match(Or1, m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) ||
455 Or0->getOpcode() == Or1->getOpcode())
456 return nullptr;
457
458 // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).
459 if (Or0->getOpcode() == BinaryOperator::LShr) {
460 std::swap(Or0, Or1);
461 std::swap(ShVal0, ShVal1);
462 std::swap(ShAmt0, ShAmt1);
463 }
464 assert(Or0->getOpcode() == BinaryOperator::Shl &&
465 Or1->getOpcode() == BinaryOperator::LShr &&
466 "Illegal or(shift,shift) pair");
467
468 // Match the shift amount operands for a funnel/rotate pattern. This always
469 // matches a subtraction on the R operand.
470 auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {
471 // The shift amounts may add up to the narrow bit width:
472 // (shl ShVal0, L) | (lshr ShVal1, Width - L)
473 // If this is a funnel shift (different operands are shifted), then the
474 // shift amount can not over-shift (create poison) in the narrow type.
475 unsigned MaxShiftAmountWidth = Log2_32(NarrowWidth);
476 APInt HiBitMask = ~APInt::getLowBitsSet(WideWidth, MaxShiftAmountWidth);
477 if (ShVal0 == ShVal1 || MaskedValueIsZero(L, HiBitMask))
478 if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L)))))
479 return L;
480
481 // The following patterns currently only work for rotation patterns.
482 // TODO: Add more general funnel-shift compatible patterns.
483 if (ShVal0 != ShVal1)
484 return nullptr;
485
486 // The shift amount may be masked with negation:
487 // (shl ShVal0, (X & (Width - 1))) | (lshr ShVal1, ((-X) & (Width - 1)))
488 Value *X;
489 unsigned Mask = Width - 1;
490 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
491 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
492 return X;
493
494 // Same as above, but the shift amount may be extended after masking:
495 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
496 match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))))
497 return X;
498
499 return nullptr;
500 };
501
502 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, NarrowWidth);
503 bool IsFshl = true; // Sub on LSHR.
504 if (!ShAmt) {
505 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, NarrowWidth);
506 IsFshl = false; // Sub on SHL.
507 }
508 if (!ShAmt)
509 return nullptr;
510
511 // The right-shifted value must have high zeros in the wide type (for example
512 // from 'zext', 'and' or 'shift'). High bits of the left-shifted value are
513 // truncated, so those do not matter.
514 APInt HiBitMask = APInt::getHighBitsSet(WideWidth, WideWidth - NarrowWidth);
515 if (!MaskedValueIsZero(ShVal1, HiBitMask, 0, &Trunc))
516 return nullptr;
517
518 // Adjust the width of ShAmt for narrowed funnel shift operation:
519 // - Zero-extend if ShAmt is narrower than the destination type.
520 // - Truncate if ShAmt is wider, discarding non-significant high-order bits.
521 // This prepares ShAmt for llvm.fshl.i8(trunc(ShVal), trunc(ShVal),
522 // zext/trunc(ShAmt)).
523 Value *NarrowShAmt = Builder.CreateZExtOrTrunc(ShAmt, DestTy);
524
525 Value *X, *Y;
526 X = Y = Builder.CreateTrunc(ShVal0, DestTy);
527 if (ShVal0 != ShVal1)
528 Y = Builder.CreateTrunc(ShVal1, DestTy);
529 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
530 Function *F = Intrinsic::getDeclaration(Trunc.getModule(), IID, DestTy);
531 return CallInst::Create(F, {X, Y, NarrowShAmt});
532 }
533
534 /// Try to narrow the width of math or bitwise logic instructions by pulling a
535 /// truncate ahead of binary operators.
narrowBinOp(TruncInst & Trunc)536 Instruction *InstCombinerImpl::narrowBinOp(TruncInst &Trunc) {
537 Type *SrcTy = Trunc.getSrcTy();
538 Type *DestTy = Trunc.getType();
539 unsigned SrcWidth = SrcTy->getScalarSizeInBits();
540 unsigned DestWidth = DestTy->getScalarSizeInBits();
541
542 if (!isa<VectorType>(SrcTy) && !shouldChangeType(SrcTy, DestTy))
543 return nullptr;
544
545 BinaryOperator *BinOp;
546 if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(BinOp))))
547 return nullptr;
548
549 Value *BinOp0 = BinOp->getOperand(0);
550 Value *BinOp1 = BinOp->getOperand(1);
551 switch (BinOp->getOpcode()) {
552 case Instruction::And:
553 case Instruction::Or:
554 case Instruction::Xor:
555 case Instruction::Add:
556 case Instruction::Sub:
557 case Instruction::Mul: {
558 Constant *C;
559 if (match(BinOp0, m_Constant(C))) {
560 // trunc (binop C, X) --> binop (trunc C', X)
561 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
562 Value *TruncX = Builder.CreateTrunc(BinOp1, DestTy);
563 return BinaryOperator::Create(BinOp->getOpcode(), NarrowC, TruncX);
564 }
565 if (match(BinOp1, m_Constant(C))) {
566 // trunc (binop X, C) --> binop (trunc X, C')
567 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);
568 Value *TruncX = Builder.CreateTrunc(BinOp0, DestTy);
569 return BinaryOperator::Create(BinOp->getOpcode(), TruncX, NarrowC);
570 }
571 Value *X;
572 if (match(BinOp0, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {
573 // trunc (binop (ext X), Y) --> binop X, (trunc Y)
574 Value *NarrowOp1 = Builder.CreateTrunc(BinOp1, DestTy);
575 return BinaryOperator::Create(BinOp->getOpcode(), X, NarrowOp1);
576 }
577 if (match(BinOp1, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {
578 // trunc (binop Y, (ext X)) --> binop (trunc Y), X
579 Value *NarrowOp0 = Builder.CreateTrunc(BinOp0, DestTy);
580 return BinaryOperator::Create(BinOp->getOpcode(), NarrowOp0, X);
581 }
582 break;
583 }
584 case Instruction::LShr:
585 case Instruction::AShr: {
586 // trunc (*shr (trunc A), C) --> trunc(*shr A, C)
587 Value *A;
588 Constant *C;
589 if (match(BinOp0, m_Trunc(m_Value(A))) && match(BinOp1, m_Constant(C))) {
590 unsigned MaxShiftAmt = SrcWidth - DestWidth;
591 // If the shift is small enough, all zero/sign bits created by the shift
592 // are removed by the trunc.
593 if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,
594 APInt(SrcWidth, MaxShiftAmt)))) {
595 auto *OldShift = cast<Instruction>(Trunc.getOperand(0));
596 bool IsExact = OldShift->isExact();
597 if (Constant *ShAmt = ConstantFoldIntegerCast(C, A->getType(),
598 /*IsSigned*/ true, DL)) {
599 ShAmt = Constant::mergeUndefsWith(ShAmt, C);
600 Value *Shift =
601 OldShift->getOpcode() == Instruction::AShr
602 ? Builder.CreateAShr(A, ShAmt, OldShift->getName(), IsExact)
603 : Builder.CreateLShr(A, ShAmt, OldShift->getName(), IsExact);
604 return CastInst::CreateTruncOrBitCast(Shift, DestTy);
605 }
606 }
607 }
608 break;
609 }
610 default: break;
611 }
612
613 if (Instruction *NarrowOr = narrowFunnelShift(Trunc))
614 return NarrowOr;
615
616 return nullptr;
617 }
618
619 /// Try to narrow the width of a splat shuffle. This could be generalized to any
620 /// shuffle with a constant operand, but we limit the transform to avoid
621 /// creating a shuffle type that targets may not be able to lower effectively.
shrinkSplatShuffle(TruncInst & Trunc,InstCombiner::BuilderTy & Builder)622 static Instruction *shrinkSplatShuffle(TruncInst &Trunc,
623 InstCombiner::BuilderTy &Builder) {
624 auto *Shuf = dyn_cast<ShuffleVectorInst>(Trunc.getOperand(0));
625 if (Shuf && Shuf->hasOneUse() && match(Shuf->getOperand(1), m_Undef()) &&
626 all_equal(Shuf->getShuffleMask()) &&
627 Shuf->getType() == Shuf->getOperand(0)->getType()) {
628 // trunc (shuf X, Undef, SplatMask) --> shuf (trunc X), Poison, SplatMask
629 // trunc (shuf X, Poison, SplatMask) --> shuf (trunc X), Poison, SplatMask
630 Value *NarrowOp = Builder.CreateTrunc(Shuf->getOperand(0), Trunc.getType());
631 return new ShuffleVectorInst(NarrowOp, Shuf->getShuffleMask());
632 }
633
634 return nullptr;
635 }
636
637 /// Try to narrow the width of an insert element. This could be generalized for
638 /// any vector constant, but we limit the transform to insertion into undef to
639 /// avoid potential backend problems from unsupported insertion widths. This
640 /// could also be extended to handle the case of inserting a scalar constant
641 /// into a vector variable.
shrinkInsertElt(CastInst & Trunc,InstCombiner::BuilderTy & Builder)642 static Instruction *shrinkInsertElt(CastInst &Trunc,
643 InstCombiner::BuilderTy &Builder) {
644 Instruction::CastOps Opcode = Trunc.getOpcode();
645 assert((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
646 "Unexpected instruction for shrinking");
647
648 auto *InsElt = dyn_cast<InsertElementInst>(Trunc.getOperand(0));
649 if (!InsElt || !InsElt->hasOneUse())
650 return nullptr;
651
652 Type *DestTy = Trunc.getType();
653 Type *DestScalarTy = DestTy->getScalarType();
654 Value *VecOp = InsElt->getOperand(0);
655 Value *ScalarOp = InsElt->getOperand(1);
656 Value *Index = InsElt->getOperand(2);
657
658 if (match(VecOp, m_Undef())) {
659 // trunc (inselt undef, X, Index) --> inselt undef, (trunc X), Index
660 // fptrunc (inselt undef, X, Index) --> inselt undef, (fptrunc X), Index
661 UndefValue *NarrowUndef = UndefValue::get(DestTy);
662 Value *NarrowOp = Builder.CreateCast(Opcode, ScalarOp, DestScalarTy);
663 return InsertElementInst::Create(NarrowUndef, NarrowOp, Index);
664 }
665
666 return nullptr;
667 }
668
visitTrunc(TruncInst & Trunc)669 Instruction *InstCombinerImpl::visitTrunc(TruncInst &Trunc) {
670 if (Instruction *Result = commonCastTransforms(Trunc))
671 return Result;
672
673 Value *Src = Trunc.getOperand(0);
674 Type *DestTy = Trunc.getType(), *SrcTy = Src->getType();
675 unsigned DestWidth = DestTy->getScalarSizeInBits();
676 unsigned SrcWidth = SrcTy->getScalarSizeInBits();
677
678 // Attempt to truncate the entire input expression tree to the destination
679 // type. Only do this if the dest type is a simple type, don't convert the
680 // expression tree to something weird like i93 unless the source is also
681 // strange.
682 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&
683 canEvaluateTruncated(Src, DestTy, *this, &Trunc)) {
684
685 // If this cast is a truncate, evaluting in a different type always
686 // eliminates the cast, so it is always a win.
687 LLVM_DEBUG(
688 dbgs() << "ICE: EvaluateInDifferentType converting expression type"
689 " to avoid cast: "
690 << Trunc << '\n');
691 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
692 assert(Res->getType() == DestTy);
693 return replaceInstUsesWith(Trunc, Res);
694 }
695
696 // For integer types, check if we can shorten the entire input expression to
697 // DestWidth * 2, which won't allow removing the truncate, but reducing the
698 // width may enable further optimizations, e.g. allowing for larger
699 // vectorization factors.
700 if (auto *DestITy = dyn_cast<IntegerType>(DestTy)) {
701 if (DestWidth * 2 < SrcWidth) {
702 auto *NewDestTy = DestITy->getExtendedType();
703 if (shouldChangeType(SrcTy, NewDestTy) &&
704 canEvaluateTruncated(Src, NewDestTy, *this, &Trunc)) {
705 LLVM_DEBUG(
706 dbgs() << "ICE: EvaluateInDifferentType converting expression type"
707 " to reduce the width of operand of"
708 << Trunc << '\n');
709 Value *Res = EvaluateInDifferentType(Src, NewDestTy, false);
710 return new TruncInst(Res, DestTy);
711 }
712 }
713 }
714
715 // Test if the trunc is the user of a select which is part of a
716 // minimum or maximum operation. If so, don't do any more simplification.
717 // Even simplifying demanded bits can break the canonical form of a
718 // min/max.
719 Value *LHS, *RHS;
720 if (SelectInst *Sel = dyn_cast<SelectInst>(Src))
721 if (matchSelectPattern(Sel, LHS, RHS).Flavor != SPF_UNKNOWN)
722 return nullptr;
723
724 // See if we can simplify any instructions used by the input whose sole
725 // purpose is to compute bits we don't care about.
726 if (SimplifyDemandedInstructionBits(Trunc))
727 return &Trunc;
728
729 if (DestWidth == 1) {
730 Value *Zero = Constant::getNullValue(SrcTy);
731 if (DestTy->isIntegerTy()) {
732 // Canonicalize trunc x to i1 -> icmp ne (and x, 1), 0 (scalar only).
733 // TODO: We canonicalize to more instructions here because we are probably
734 // lacking equivalent analysis for trunc relative to icmp. There may also
735 // be codegen concerns. If those trunc limitations were removed, we could
736 // remove this transform.
737 Value *And = Builder.CreateAnd(Src, ConstantInt::get(SrcTy, 1));
738 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
739 }
740
741 // For vectors, we do not canonicalize all truncs to icmp, so optimize
742 // patterns that would be covered within visitICmpInst.
743 Value *X;
744 Constant *C;
745 if (match(Src, m_OneUse(m_LShr(m_Value(X), m_Constant(C))))) {
746 // trunc (lshr X, C) to i1 --> icmp ne (and X, C'), 0
747 Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));
748 Constant *MaskC = ConstantExpr::getShl(One, C);
749 Value *And = Builder.CreateAnd(X, MaskC);
750 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
751 }
752 if (match(Src, m_OneUse(m_c_Or(m_LShr(m_Value(X), m_ImmConstant(C)),
753 m_Deferred(X))))) {
754 // trunc (or (lshr X, C), X) to i1 --> icmp ne (and X, C'), 0
755 Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));
756 Constant *MaskC = ConstantExpr::getShl(One, C);
757 Value *And = Builder.CreateAnd(X, Builder.CreateOr(MaskC, One));
758 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);
759 }
760 }
761
762 Value *A, *B;
763 Constant *C;
764 if (match(Src, m_LShr(m_SExt(m_Value(A)), m_Constant(C)))) {
765 unsigned AWidth = A->getType()->getScalarSizeInBits();
766 unsigned MaxShiftAmt = SrcWidth - std::max(DestWidth, AWidth);
767 auto *OldSh = cast<Instruction>(Src);
768 bool IsExact = OldSh->isExact();
769
770 // If the shift is small enough, all zero bits created by the shift are
771 // removed by the trunc.
772 if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,
773 APInt(SrcWidth, MaxShiftAmt)))) {
774 auto GetNewShAmt = [&](unsigned Width) {
775 Constant *MaxAmt = ConstantInt::get(SrcTy, Width - 1, false);
776 Constant *Cmp =
777 ConstantFoldCompareInstOperands(ICmpInst::ICMP_ULT, C, MaxAmt, DL);
778 Constant *ShAmt = ConstantFoldSelectInstruction(Cmp, C, MaxAmt);
779 return ConstantFoldCastOperand(Instruction::Trunc, ShAmt, A->getType(),
780 DL);
781 };
782
783 // trunc (lshr (sext A), C) --> ashr A, C
784 if (A->getType() == DestTy) {
785 Constant *ShAmt = GetNewShAmt(DestWidth);
786 ShAmt = Constant::mergeUndefsWith(ShAmt, C);
787 return IsExact ? BinaryOperator::CreateExactAShr(A, ShAmt)
788 : BinaryOperator::CreateAShr(A, ShAmt);
789 }
790 // The types are mismatched, so create a cast after shifting:
791 // trunc (lshr (sext A), C) --> sext/trunc (ashr A, C)
792 if (Src->hasOneUse()) {
793 Constant *ShAmt = GetNewShAmt(AWidth);
794 Value *Shift = Builder.CreateAShr(A, ShAmt, "", IsExact);
795 return CastInst::CreateIntegerCast(Shift, DestTy, true);
796 }
797 }
798 // TODO: Mask high bits with 'and'.
799 }
800
801 if (Instruction *I = narrowBinOp(Trunc))
802 return I;
803
804 if (Instruction *I = shrinkSplatShuffle(Trunc, Builder))
805 return I;
806
807 if (Instruction *I = shrinkInsertElt(Trunc, Builder))
808 return I;
809
810 if (Src->hasOneUse() &&
811 (isa<VectorType>(SrcTy) || shouldChangeType(SrcTy, DestTy))) {
812 // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the
813 // dest type is native and cst < dest size.
814 if (match(Src, m_Shl(m_Value(A), m_Constant(C))) &&
815 !match(A, m_Shr(m_Value(), m_Constant()))) {
816 // Skip shifts of shift by constants. It undoes a combine in
817 // FoldShiftByConstant and is the extend in reg pattern.
818 APInt Threshold = APInt(C->getType()->getScalarSizeInBits(), DestWidth);
819 if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold))) {
820 Value *NewTrunc = Builder.CreateTrunc(A, DestTy, A->getName() + ".tr");
821 return BinaryOperator::Create(Instruction::Shl, NewTrunc,
822 ConstantExpr::getTrunc(C, DestTy));
823 }
824 }
825 }
826
827 if (Instruction *I = foldVecTruncToExtElt(Trunc, *this))
828 return I;
829
830 // Whenever an element is extracted from a vector, and then truncated,
831 // canonicalize by converting it to a bitcast followed by an
832 // extractelement.
833 //
834 // Example (little endian):
835 // trunc (extractelement <4 x i64> %X, 0) to i32
836 // --->
837 // extractelement <8 x i32> (bitcast <4 x i64> %X to <8 x i32>), i32 0
838 Value *VecOp;
839 ConstantInt *Cst;
840 if (match(Src, m_OneUse(m_ExtractElt(m_Value(VecOp), m_ConstantInt(Cst))))) {
841 auto *VecOpTy = cast<VectorType>(VecOp->getType());
842 auto VecElts = VecOpTy->getElementCount();
843
844 // A badly fit destination size would result in an invalid cast.
845 if (SrcWidth % DestWidth == 0) {
846 uint64_t TruncRatio = SrcWidth / DestWidth;
847 uint64_t BitCastNumElts = VecElts.getKnownMinValue() * TruncRatio;
848 uint64_t VecOpIdx = Cst->getZExtValue();
849 uint64_t NewIdx = DL.isBigEndian() ? (VecOpIdx + 1) * TruncRatio - 1
850 : VecOpIdx * TruncRatio;
851 assert(BitCastNumElts <= std::numeric_limits<uint32_t>::max() &&
852 "overflow 32-bits");
853
854 auto *BitCastTo =
855 VectorType::get(DestTy, BitCastNumElts, VecElts.isScalable());
856 Value *BitCast = Builder.CreateBitCast(VecOp, BitCastTo);
857 return ExtractElementInst::Create(BitCast, Builder.getInt32(NewIdx));
858 }
859 }
860
861 // trunc (ctlz_i32(zext(A), B) --> add(ctlz_i16(A, B), C)
862 if (match(Src, m_OneUse(m_Intrinsic<Intrinsic::ctlz>(m_ZExt(m_Value(A)),
863 m_Value(B))))) {
864 unsigned AWidth = A->getType()->getScalarSizeInBits();
865 if (AWidth == DestWidth && AWidth > Log2_32(SrcWidth)) {
866 Value *WidthDiff = ConstantInt::get(A->getType(), SrcWidth - AWidth);
867 Value *NarrowCtlz =
868 Builder.CreateIntrinsic(Intrinsic::ctlz, {Trunc.getType()}, {A, B});
869 return BinaryOperator::CreateAdd(NarrowCtlz, WidthDiff);
870 }
871 }
872
873 if (match(Src, m_VScale())) {
874 if (Trunc.getFunction() &&
875 Trunc.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
876 Attribute Attr =
877 Trunc.getFunction()->getFnAttribute(Attribute::VScaleRange);
878 if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
879 if (Log2_32(*MaxVScale) < DestWidth) {
880 Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
881 return replaceInstUsesWith(Trunc, VScale);
882 }
883 }
884 }
885 }
886
887 return nullptr;
888 }
889
transformZExtICmp(ICmpInst * Cmp,ZExtInst & Zext)890 Instruction *InstCombinerImpl::transformZExtICmp(ICmpInst *Cmp,
891 ZExtInst &Zext) {
892 // If we are just checking for a icmp eq of a single bit and zext'ing it
893 // to an integer, then shift the bit to the appropriate place and then
894 // cast to integer to avoid the comparison.
895
896 // FIXME: This set of transforms does not check for extra uses and/or creates
897 // an extra instruction (an optional final cast is not included
898 // in the transform comments). We may also want to favor icmp over
899 // shifts in cases of equal instructions because icmp has better
900 // analysis in general (invert the transform).
901
902 const APInt *Op1CV;
903 if (match(Cmp->getOperand(1), m_APInt(Op1CV))) {
904
905 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
906 if (Cmp->getPredicate() == ICmpInst::ICMP_SLT && Op1CV->isZero()) {
907 Value *In = Cmp->getOperand(0);
908 Value *Sh = ConstantInt::get(In->getType(),
909 In->getType()->getScalarSizeInBits() - 1);
910 In = Builder.CreateLShr(In, Sh, In->getName() + ".lobit");
911 if (In->getType() != Zext.getType())
912 In = Builder.CreateIntCast(In, Zext.getType(), false /*ZExt*/);
913
914 return replaceInstUsesWith(Zext, In);
915 }
916
917 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
918 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
919 // zext (X != 0) to i32 --> X iff X has only the low bit set.
920 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
921
922 if (Op1CV->isZero() && Cmp->isEquality()) {
923 // Exactly 1 possible 1? But not the high-bit because that is
924 // canonicalized to this form.
925 KnownBits Known = computeKnownBits(Cmp->getOperand(0), 0, &Zext);
926 APInt KnownZeroMask(~Known.Zero);
927 uint32_t ShAmt = KnownZeroMask.logBase2();
928 bool IsExpectShAmt = KnownZeroMask.isPowerOf2() &&
929 (Zext.getType()->getScalarSizeInBits() != ShAmt + 1);
930 if (IsExpectShAmt &&
931 (Cmp->getOperand(0)->getType() == Zext.getType() ||
932 Cmp->getPredicate() == ICmpInst::ICMP_NE || ShAmt == 0)) {
933 Value *In = Cmp->getOperand(0);
934 if (ShAmt) {
935 // Perform a logical shr by shiftamt.
936 // Insert the shift to put the result in the low bit.
937 In = Builder.CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),
938 In->getName() + ".lobit");
939 }
940
941 // Toggle the low bit for "X == 0".
942 if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)
943 In = Builder.CreateXor(In, ConstantInt::get(In->getType(), 1));
944
945 if (Zext.getType() == In->getType())
946 return replaceInstUsesWith(Zext, In);
947
948 Value *IntCast = Builder.CreateIntCast(In, Zext.getType(), false);
949 return replaceInstUsesWith(Zext, IntCast);
950 }
951 }
952 }
953
954 if (Cmp->isEquality() && Zext.getType() == Cmp->getOperand(0)->getType()) {
955 // Test if a bit is clear/set using a shifted-one mask:
956 // zext (icmp eq (and X, (1 << ShAmt)), 0) --> and (lshr (not X), ShAmt), 1
957 // zext (icmp ne (and X, (1 << ShAmt)), 0) --> and (lshr X, ShAmt), 1
958 Value *X, *ShAmt;
959 if (Cmp->hasOneUse() && match(Cmp->getOperand(1), m_ZeroInt()) &&
960 match(Cmp->getOperand(0),
961 m_OneUse(m_c_And(m_Shl(m_One(), m_Value(ShAmt)), m_Value(X))))) {
962 if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)
963 X = Builder.CreateNot(X);
964 Value *Lshr = Builder.CreateLShr(X, ShAmt);
965 Value *And1 = Builder.CreateAnd(Lshr, ConstantInt::get(X->getType(), 1));
966 return replaceInstUsesWith(Zext, And1);
967 }
968 }
969
970 return nullptr;
971 }
972
973 /// Determine if the specified value can be computed in the specified wider type
974 /// and produce the same low bits. If not, return false.
975 ///
976 /// If this function returns true, it can also return a non-zero number of bits
977 /// (in BitsToClear) which indicates that the value it computes is correct for
978 /// the zero extend, but that the additional BitsToClear bits need to be zero'd
979 /// out. For example, to promote something like:
980 ///
981 /// %B = trunc i64 %A to i32
982 /// %C = lshr i32 %B, 8
983 /// %E = zext i32 %C to i64
984 ///
985 /// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be
986 /// set to 8 to indicate that the promoted value needs to have bits 24-31
987 /// cleared in addition to bits 32-63. Since an 'and' will be generated to
988 /// clear the top bits anyway, doing this has no extra cost.
989 ///
990 /// This function works on both vectors and scalars.
canEvaluateZExtd(Value * V,Type * Ty,unsigned & BitsToClear,InstCombinerImpl & IC,Instruction * CxtI)991 static bool canEvaluateZExtd(Value *V, Type *Ty, unsigned &BitsToClear,
992 InstCombinerImpl &IC, Instruction *CxtI) {
993 BitsToClear = 0;
994 if (canAlwaysEvaluateInType(V, Ty))
995 return true;
996 if (canNotEvaluateInType(V, Ty))
997 return false;
998
999 auto *I = cast<Instruction>(V);
1000 unsigned Tmp;
1001 switch (I->getOpcode()) {
1002 case Instruction::ZExt: // zext(zext(x)) -> zext(x).
1003 case Instruction::SExt: // zext(sext(x)) -> sext(x).
1004 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)
1005 return true;
1006 case Instruction::And:
1007 case Instruction::Or:
1008 case Instruction::Xor:
1009 case Instruction::Add:
1010 case Instruction::Sub:
1011 case Instruction::Mul:
1012 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||
1013 !canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI))
1014 return false;
1015 // These can all be promoted if neither operand has 'bits to clear'.
1016 if (BitsToClear == 0 && Tmp == 0)
1017 return true;
1018
1019 // If the operation is an AND/OR/XOR and the bits to clear are zero in the
1020 // other side, BitsToClear is ok.
1021 if (Tmp == 0 && I->isBitwiseLogicOp()) {
1022 // We use MaskedValueIsZero here for generality, but the case we care
1023 // about the most is constant RHS.
1024 unsigned VSize = V->getType()->getScalarSizeInBits();
1025 if (IC.MaskedValueIsZero(I->getOperand(1),
1026 APInt::getHighBitsSet(VSize, BitsToClear),
1027 0, CxtI)) {
1028 // If this is an And instruction and all of the BitsToClear are
1029 // known to be zero we can reset BitsToClear.
1030 if (I->getOpcode() == Instruction::And)
1031 BitsToClear = 0;
1032 return true;
1033 }
1034 }
1035
1036 // Otherwise, we don't know how to analyze this BitsToClear case yet.
1037 return false;
1038
1039 case Instruction::Shl: {
1040 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the
1041 // upper bits we can reduce BitsToClear by the shift amount.
1042 const APInt *Amt;
1043 if (match(I->getOperand(1), m_APInt(Amt))) {
1044 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
1045 return false;
1046 uint64_t ShiftAmt = Amt->getZExtValue();
1047 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;
1048 return true;
1049 }
1050 return false;
1051 }
1052 case Instruction::LShr: {
1053 // We can promote lshr(x, cst) if we can promote x. This requires the
1054 // ultimate 'and' to clear out the high zero bits we're clearing out though.
1055 const APInt *Amt;
1056 if (match(I->getOperand(1), m_APInt(Amt))) {
1057 if (!canEvaluateZExtd(I->getOperand(0), Ty, BitsToClear, IC, CxtI))
1058 return false;
1059 BitsToClear += Amt->getZExtValue();
1060 if (BitsToClear > V->getType()->getScalarSizeInBits())
1061 BitsToClear = V->getType()->getScalarSizeInBits();
1062 return true;
1063 }
1064 // Cannot promote variable LSHR.
1065 return false;
1066 }
1067 case Instruction::Select:
1068 if (!canEvaluateZExtd(I->getOperand(1), Ty, Tmp, IC, CxtI) ||
1069 !canEvaluateZExtd(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||
1070 // TODO: If important, we could handle the case when the BitsToClear are
1071 // known zero in the disagreeing side.
1072 Tmp != BitsToClear)
1073 return false;
1074 return true;
1075
1076 case Instruction::PHI: {
1077 // We can change a phi if we can change all operands. Note that we never
1078 // get into trouble with cyclic PHIs here because we only consider
1079 // instructions with a single use.
1080 PHINode *PN = cast<PHINode>(I);
1081 if (!canEvaluateZExtd(PN->getIncomingValue(0), Ty, BitsToClear, IC, CxtI))
1082 return false;
1083 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
1084 if (!canEvaluateZExtd(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||
1085 // TODO: If important, we could handle the case when the BitsToClear
1086 // are known zero in the disagreeing input.
1087 Tmp != BitsToClear)
1088 return false;
1089 return true;
1090 }
1091 case Instruction::Call:
1092 // llvm.vscale() can always be executed in larger type, because the
1093 // value is automatically zero-extended.
1094 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1095 if (II->getIntrinsicID() == Intrinsic::vscale)
1096 return true;
1097 return false;
1098 default:
1099 // TODO: Can handle more cases here.
1100 return false;
1101 }
1102 }
1103
visitZExt(ZExtInst & Zext)1104 Instruction *InstCombinerImpl::visitZExt(ZExtInst &Zext) {
1105 // If this zero extend is only used by a truncate, let the truncate be
1106 // eliminated before we try to optimize this zext.
1107 if (Zext.hasOneUse() && isa<TruncInst>(Zext.user_back()) &&
1108 !isa<Constant>(Zext.getOperand(0)))
1109 return nullptr;
1110
1111 // If one of the common conversion will work, do it.
1112 if (Instruction *Result = commonCastTransforms(Zext))
1113 return Result;
1114
1115 Value *Src = Zext.getOperand(0);
1116 Type *SrcTy = Src->getType(), *DestTy = Zext.getType();
1117
1118 // Try to extend the entire expression tree to the wide destination type.
1119 unsigned BitsToClear;
1120 if (shouldChangeType(SrcTy, DestTy) &&
1121 canEvaluateZExtd(Src, DestTy, BitsToClear, *this, &Zext)) {
1122 assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&
1123 "Can't clear more bits than in SrcTy");
1124
1125 // Okay, we can transform this! Insert the new expression now.
1126 LLVM_DEBUG(
1127 dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1128 " to avoid zero extend: "
1129 << Zext << '\n');
1130 Value *Res = EvaluateInDifferentType(Src, DestTy, false);
1131 assert(Res->getType() == DestTy);
1132
1133 // Preserve debug values referring to Src if the zext is its last use.
1134 if (auto *SrcOp = dyn_cast<Instruction>(Src))
1135 if (SrcOp->hasOneUse())
1136 replaceAllDbgUsesWith(*SrcOp, *Res, Zext, DT);
1137
1138 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits() - BitsToClear;
1139 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
1140
1141 // If the high bits are already filled with zeros, just replace this
1142 // cast with the result.
1143 if (MaskedValueIsZero(Res,
1144 APInt::getHighBitsSet(DestBitSize,
1145 DestBitSize - SrcBitsKept),
1146 0, &Zext))
1147 return replaceInstUsesWith(Zext, Res);
1148
1149 // We need to emit an AND to clear the high bits.
1150 Constant *C = ConstantInt::get(Res->getType(),
1151 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));
1152 return BinaryOperator::CreateAnd(Res, C);
1153 }
1154
1155 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
1156 // types and if the sizes are just right we can convert this into a logical
1157 // 'and' which will be much cheaper than the pair of casts.
1158 if (auto *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
1159 // TODO: Subsume this into EvaluateInDifferentType.
1160
1161 // Get the sizes of the types involved. We know that the intermediate type
1162 // will be smaller than A or C, but don't know the relation between A and C.
1163 Value *A = CSrc->getOperand(0);
1164 unsigned SrcSize = A->getType()->getScalarSizeInBits();
1165 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
1166 unsigned DstSize = DestTy->getScalarSizeInBits();
1167 // If we're actually extending zero bits, then if
1168 // SrcSize < DstSize: zext(a & mask)
1169 // SrcSize == DstSize: a & mask
1170 // SrcSize > DstSize: trunc(a) & mask
1171 if (SrcSize < DstSize) {
1172 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
1173 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
1174 Value *And = Builder.CreateAnd(A, AndConst, CSrc->getName() + ".mask");
1175 return new ZExtInst(And, DestTy);
1176 }
1177
1178 if (SrcSize == DstSize) {
1179 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
1180 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
1181 AndValue));
1182 }
1183 if (SrcSize > DstSize) {
1184 Value *Trunc = Builder.CreateTrunc(A, DestTy);
1185 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
1186 return BinaryOperator::CreateAnd(Trunc,
1187 ConstantInt::get(Trunc->getType(),
1188 AndValue));
1189 }
1190 }
1191
1192 if (auto *Cmp = dyn_cast<ICmpInst>(Src))
1193 return transformZExtICmp(Cmp, Zext);
1194
1195 // zext(trunc(X) & C) -> (X & zext(C)).
1196 Constant *C;
1197 Value *X;
1198 if (match(Src, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&
1199 X->getType() == DestTy)
1200 return BinaryOperator::CreateAnd(X, Builder.CreateZExt(C, DestTy));
1201
1202 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).
1203 Value *And;
1204 if (match(Src, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&
1205 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&
1206 X->getType() == DestTy) {
1207 Value *ZC = Builder.CreateZExt(C, DestTy);
1208 return BinaryOperator::CreateXor(Builder.CreateAnd(X, ZC), ZC);
1209 }
1210
1211 // If we are truncating, masking, and then zexting back to the original type,
1212 // that's just a mask. This is not handled by canEvaluateZextd if the
1213 // intermediate values have extra uses. This could be generalized further for
1214 // a non-constant mask operand.
1215 // zext (and (trunc X), C) --> and X, (zext C)
1216 if (match(Src, m_And(m_Trunc(m_Value(X)), m_Constant(C))) &&
1217 X->getType() == DestTy) {
1218 Value *ZextC = Builder.CreateZExt(C, DestTy);
1219 return BinaryOperator::CreateAnd(X, ZextC);
1220 }
1221
1222 if (match(Src, m_VScale())) {
1223 if (Zext.getFunction() &&
1224 Zext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1225 Attribute Attr =
1226 Zext.getFunction()->getFnAttribute(Attribute::VScaleRange);
1227 if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
1228 unsigned TypeWidth = Src->getType()->getScalarSizeInBits();
1229 if (Log2_32(*MaxVScale) < TypeWidth) {
1230 Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
1231 return replaceInstUsesWith(Zext, VScale);
1232 }
1233 }
1234 }
1235 }
1236
1237 if (!Zext.hasNonNeg()) {
1238 // If this zero extend is only used by a shift, add nneg flag.
1239 if (Zext.hasOneUse() &&
1240 SrcTy->getScalarSizeInBits() >
1241 Log2_64_Ceil(DestTy->getScalarSizeInBits()) &&
1242 match(Zext.user_back(), m_Shift(m_Value(), m_Specific(&Zext)))) {
1243 Zext.setNonNeg();
1244 return &Zext;
1245 }
1246
1247 if (isKnownNonNegative(Src, SQ.getWithInstruction(&Zext))) {
1248 Zext.setNonNeg();
1249 return &Zext;
1250 }
1251 }
1252
1253 return nullptr;
1254 }
1255
1256 /// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.
transformSExtICmp(ICmpInst * Cmp,SExtInst & Sext)1257 Instruction *InstCombinerImpl::transformSExtICmp(ICmpInst *Cmp,
1258 SExtInst &Sext) {
1259 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1260 ICmpInst::Predicate Pred = Cmp->getPredicate();
1261
1262 // Don't bother if Op1 isn't of vector or integer type.
1263 if (!Op1->getType()->isIntOrIntVectorTy())
1264 return nullptr;
1265
1266 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_ZeroInt())) {
1267 // sext (x <s 0) --> ashr x, 31 (all ones if negative)
1268 Value *Sh = ConstantInt::get(Op0->getType(),
1269 Op0->getType()->getScalarSizeInBits() - 1);
1270 Value *In = Builder.CreateAShr(Op0, Sh, Op0->getName() + ".lobit");
1271 if (In->getType() != Sext.getType())
1272 In = Builder.CreateIntCast(In, Sext.getType(), true /*SExt*/);
1273
1274 return replaceInstUsesWith(Sext, In);
1275 }
1276
1277 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1278 // If we know that only one bit of the LHS of the icmp can be set and we
1279 // have an equality comparison with zero or a power of 2, we can transform
1280 // the icmp and sext into bitwise/integer operations.
1281 if (Cmp->hasOneUse() &&
1282 Cmp->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){
1283 KnownBits Known = computeKnownBits(Op0, 0, &Sext);
1284
1285 APInt KnownZeroMask(~Known.Zero);
1286 if (KnownZeroMask.isPowerOf2()) {
1287 Value *In = Cmp->getOperand(0);
1288
1289 // If the icmp tests for a known zero bit we can constant fold it.
1290 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {
1291 Value *V = Pred == ICmpInst::ICMP_NE ?
1292 ConstantInt::getAllOnesValue(Sext.getType()) :
1293 ConstantInt::getNullValue(Sext.getType());
1294 return replaceInstUsesWith(Sext, V);
1295 }
1296
1297 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {
1298 // sext ((x & 2^n) == 0) -> (x >> n) - 1
1299 // sext ((x & 2^n) != 2^n) -> (x >> n) - 1
1300 unsigned ShiftAmt = KnownZeroMask.countr_zero();
1301 // Perform a right shift to place the desired bit in the LSB.
1302 if (ShiftAmt)
1303 In = Builder.CreateLShr(In,
1304 ConstantInt::get(In->getType(), ShiftAmt));
1305
1306 // At this point "In" is either 1 or 0. Subtract 1 to turn
1307 // {1, 0} -> {0, -1}.
1308 In = Builder.CreateAdd(In,
1309 ConstantInt::getAllOnesValue(In->getType()),
1310 "sext");
1311 } else {
1312 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-1
1313 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-1
1314 unsigned ShiftAmt = KnownZeroMask.countl_zero();
1315 // Perform a left shift to place the desired bit in the MSB.
1316 if (ShiftAmt)
1317 In = Builder.CreateShl(In,
1318 ConstantInt::get(In->getType(), ShiftAmt));
1319
1320 // Distribute the bit over the whole bit width.
1321 In = Builder.CreateAShr(In, ConstantInt::get(In->getType(),
1322 KnownZeroMask.getBitWidth() - 1), "sext");
1323 }
1324
1325 if (Sext.getType() == In->getType())
1326 return replaceInstUsesWith(Sext, In);
1327 return CastInst::CreateIntegerCast(In, Sext.getType(), true/*SExt*/);
1328 }
1329 }
1330 }
1331
1332 return nullptr;
1333 }
1334
1335 /// Return true if we can take the specified value and return it as type Ty
1336 /// without inserting any new casts and without changing the value of the common
1337 /// low bits. This is used by code that tries to promote integer operations to
1338 /// a wider types will allow us to eliminate the extension.
1339 ///
1340 /// This function works on both vectors and scalars.
1341 ///
canEvaluateSExtd(Value * V,Type * Ty)1342 static bool canEvaluateSExtd(Value *V, Type *Ty) {
1343 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
1344 "Can't sign extend type to a smaller type");
1345 if (canAlwaysEvaluateInType(V, Ty))
1346 return true;
1347 if (canNotEvaluateInType(V, Ty))
1348 return false;
1349
1350 auto *I = cast<Instruction>(V);
1351 switch (I->getOpcode()) {
1352 case Instruction::SExt: // sext(sext(x)) -> sext(x)
1353 case Instruction::ZExt: // sext(zext(x)) -> zext(x)
1354 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)
1355 return true;
1356 case Instruction::And:
1357 case Instruction::Or:
1358 case Instruction::Xor:
1359 case Instruction::Add:
1360 case Instruction::Sub:
1361 case Instruction::Mul:
1362 // These operators can all arbitrarily be extended if their inputs can.
1363 return canEvaluateSExtd(I->getOperand(0), Ty) &&
1364 canEvaluateSExtd(I->getOperand(1), Ty);
1365
1366 //case Instruction::Shl: TODO
1367 //case Instruction::LShr: TODO
1368
1369 case Instruction::Select:
1370 return canEvaluateSExtd(I->getOperand(1), Ty) &&
1371 canEvaluateSExtd(I->getOperand(2), Ty);
1372
1373 case Instruction::PHI: {
1374 // We can change a phi if we can change all operands. Note that we never
1375 // get into trouble with cyclic PHIs here because we only consider
1376 // instructions with a single use.
1377 PHINode *PN = cast<PHINode>(I);
1378 for (Value *IncValue : PN->incoming_values())
1379 if (!canEvaluateSExtd(IncValue, Ty)) return false;
1380 return true;
1381 }
1382 default:
1383 // TODO: Can handle more cases here.
1384 break;
1385 }
1386
1387 return false;
1388 }
1389
visitSExt(SExtInst & Sext)1390 Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
1391 // If this sign extend is only used by a truncate, let the truncate be
1392 // eliminated before we try to optimize this sext.
1393 if (Sext.hasOneUse() && isa<TruncInst>(Sext.user_back()))
1394 return nullptr;
1395
1396 if (Instruction *I = commonCastTransforms(Sext))
1397 return I;
1398
1399 Value *Src = Sext.getOperand(0);
1400 Type *SrcTy = Src->getType(), *DestTy = Sext.getType();
1401 unsigned SrcBitSize = SrcTy->getScalarSizeInBits();
1402 unsigned DestBitSize = DestTy->getScalarSizeInBits();
1403
1404 // If the value being extended is zero or positive, use a zext instead.
1405 if (isKnownNonNegative(Src, SQ.getWithInstruction(&Sext))) {
1406 auto CI = CastInst::Create(Instruction::ZExt, Src, DestTy);
1407 CI->setNonNeg(true);
1408 return CI;
1409 }
1410
1411 // Try to extend the entire expression tree to the wide destination type.
1412 if (shouldChangeType(SrcTy, DestTy) && canEvaluateSExtd(Src, DestTy)) {
1413 // Okay, we can transform this! Insert the new expression now.
1414 LLVM_DEBUG(
1415 dbgs() << "ICE: EvaluateInDifferentType converting expression type"
1416 " to avoid sign extend: "
1417 << Sext << '\n');
1418 Value *Res = EvaluateInDifferentType(Src, DestTy, true);
1419 assert(Res->getType() == DestTy);
1420
1421 // If the high bits are already filled with sign bit, just replace this
1422 // cast with the result.
1423 if (ComputeNumSignBits(Res, 0, &Sext) > DestBitSize - SrcBitSize)
1424 return replaceInstUsesWith(Sext, Res);
1425
1426 // We need to emit a shl + ashr to do the sign extend.
1427 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize-SrcBitSize);
1428 return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),
1429 ShAmt);
1430 }
1431
1432 Value *X;
1433 if (match(Src, m_Trunc(m_Value(X)))) {
1434 // If the input has more sign bits than bits truncated, then convert
1435 // directly to final type.
1436 unsigned XBitSize = X->getType()->getScalarSizeInBits();
1437 if (ComputeNumSignBits(X, 0, &Sext) > XBitSize - SrcBitSize)
1438 return CastInst::CreateIntegerCast(X, DestTy, /* isSigned */ true);
1439
1440 // If input is a trunc from the destination type, then convert into shifts.
1441 if (Src->hasOneUse() && X->getType() == DestTy) {
1442 // sext (trunc X) --> ashr (shl X, C), C
1443 Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);
1444 return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);
1445 }
1446
1447 // If we are replacing shifted-in high zero bits with sign bits, convert
1448 // the logic shift to arithmetic shift and eliminate the cast to
1449 // intermediate type:
1450 // sext (trunc (lshr Y, C)) --> sext/trunc (ashr Y, C)
1451 Value *Y;
1452 if (Src->hasOneUse() &&
1453 match(X, m_LShr(m_Value(Y),
1454 m_SpecificIntAllowUndef(XBitSize - SrcBitSize)))) {
1455 Value *Ashr = Builder.CreateAShr(Y, XBitSize - SrcBitSize);
1456 return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);
1457 }
1458 }
1459
1460 if (auto *Cmp = dyn_cast<ICmpInst>(Src))
1461 return transformSExtICmp(Cmp, Sext);
1462
1463 // If the input is a shl/ashr pair of a same constant, then this is a sign
1464 // extension from a smaller value. If we could trust arbitrary bitwidth
1465 // integers, we could turn this into a truncate to the smaller bit and then
1466 // use a sext for the whole extension. Since we don't, look deeper and check
1467 // for a truncate. If the source and dest are the same type, eliminate the
1468 // trunc and extend and just do shifts. For example, turn:
1469 // %a = trunc i32 %i to i8
1470 // %b = shl i8 %a, C
1471 // %c = ashr i8 %b, C
1472 // %d = sext i8 %c to i32
1473 // into:
1474 // %a = shl i32 %i, 32-(8-C)
1475 // %d = ashr i32 %a, 32-(8-C)
1476 Value *A = nullptr;
1477 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.
1478 Constant *BA = nullptr, *CA = nullptr;
1479 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_Constant(BA)),
1480 m_ImmConstant(CA))) &&
1481 BA->isElementWiseEqual(CA) && A->getType() == DestTy) {
1482 Constant *WideCurrShAmt =
1483 ConstantFoldCastOperand(Instruction::SExt, CA, DestTy, DL);
1484 assert(WideCurrShAmt && "Constant folding of ImmConstant cannot fail");
1485 Constant *NumLowbitsLeft = ConstantExpr::getSub(
1486 ConstantInt::get(DestTy, SrcTy->getScalarSizeInBits()), WideCurrShAmt);
1487 Constant *NewShAmt = ConstantExpr::getSub(
1488 ConstantInt::get(DestTy, DestTy->getScalarSizeInBits()),
1489 NumLowbitsLeft);
1490 NewShAmt =
1491 Constant::mergeUndefsWith(Constant::mergeUndefsWith(NewShAmt, BA), CA);
1492 A = Builder.CreateShl(A, NewShAmt, Sext.getName());
1493 return BinaryOperator::CreateAShr(A, NewShAmt);
1494 }
1495
1496 // Splatting a bit of constant-index across a value:
1497 // sext (ashr (trunc iN X to iM), M-1) to iN --> ashr (shl X, N-M), N-1
1498 // If the dest type is different, use a cast (adjust use check).
1499 if (match(Src, m_OneUse(m_AShr(m_Trunc(m_Value(X)),
1500 m_SpecificInt(SrcBitSize - 1))))) {
1501 Type *XTy = X->getType();
1502 unsigned XBitSize = XTy->getScalarSizeInBits();
1503 Constant *ShlAmtC = ConstantInt::get(XTy, XBitSize - SrcBitSize);
1504 Constant *AshrAmtC = ConstantInt::get(XTy, XBitSize - 1);
1505 if (XTy == DestTy)
1506 return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShlAmtC),
1507 AshrAmtC);
1508 if (cast<BinaryOperator>(Src)->getOperand(0)->hasOneUse()) {
1509 Value *Ashr = Builder.CreateAShr(Builder.CreateShl(X, ShlAmtC), AshrAmtC);
1510 return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);
1511 }
1512 }
1513
1514 if (match(Src, m_VScale())) {
1515 if (Sext.getFunction() &&
1516 Sext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
1517 Attribute Attr =
1518 Sext.getFunction()->getFnAttribute(Attribute::VScaleRange);
1519 if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {
1520 if (Log2_32(*MaxVScale) < (SrcBitSize - 1)) {
1521 Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
1522 return replaceInstUsesWith(Sext, VScale);
1523 }
1524 }
1525 }
1526 }
1527
1528 return nullptr;
1529 }
1530
1531 /// Return a Constant* for the specified floating-point constant if it fits
1532 /// in the specified FP type without changing its value.
fitsInFPType(ConstantFP * CFP,const fltSemantics & Sem)1533 static bool fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
1534 bool losesInfo;
1535 APFloat F = CFP->getValueAPF();
1536 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
1537 return !losesInfo;
1538 }
1539
shrinkFPConstant(ConstantFP * CFP)1540 static Type *shrinkFPConstant(ConstantFP *CFP) {
1541 if (CFP->getType() == Type::getPPC_FP128Ty(CFP->getContext()))
1542 return nullptr; // No constant folding of this.
1543 // See if the value can be truncated to half and then reextended.
1544 if (fitsInFPType(CFP, APFloat::IEEEhalf()))
1545 return Type::getHalfTy(CFP->getContext());
1546 // See if the value can be truncated to float and then reextended.
1547 if (fitsInFPType(CFP, APFloat::IEEEsingle()))
1548 return Type::getFloatTy(CFP->getContext());
1549 if (CFP->getType()->isDoubleTy())
1550 return nullptr; // Won't shrink.
1551 if (fitsInFPType(CFP, APFloat::IEEEdouble()))
1552 return Type::getDoubleTy(CFP->getContext());
1553 // Don't try to shrink to various long double types.
1554 return nullptr;
1555 }
1556
1557 // Determine if this is a vector of ConstantFPs and if so, return the minimal
1558 // type we can safely truncate all elements to.
shrinkFPConstantVector(Value * V)1559 static Type *shrinkFPConstantVector(Value *V) {
1560 auto *CV = dyn_cast<Constant>(V);
1561 auto *CVVTy = dyn_cast<FixedVectorType>(V->getType());
1562 if (!CV || !CVVTy)
1563 return nullptr;
1564
1565 Type *MinType = nullptr;
1566
1567 unsigned NumElts = CVVTy->getNumElements();
1568
1569 // For fixed-width vectors we find the minimal type by looking
1570 // through the constant values of the vector.
1571 for (unsigned i = 0; i != NumElts; ++i) {
1572 if (isa<UndefValue>(CV->getAggregateElement(i)))
1573 continue;
1574
1575 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
1576 if (!CFP)
1577 return nullptr;
1578
1579 Type *T = shrinkFPConstant(CFP);
1580 if (!T)
1581 return nullptr;
1582
1583 // If we haven't found a type yet or this type has a larger mantissa than
1584 // our previous type, this is our new minimal type.
1585 if (!MinType || T->getFPMantissaWidth() > MinType->getFPMantissaWidth())
1586 MinType = T;
1587 }
1588
1589 // Make a vector type from the minimal type.
1590 return MinType ? FixedVectorType::get(MinType, NumElts) : nullptr;
1591 }
1592
1593 /// Find the minimum FP type we can safely truncate to.
getMinimumFPType(Value * V)1594 static Type *getMinimumFPType(Value *V) {
1595 if (auto *FPExt = dyn_cast<FPExtInst>(V))
1596 return FPExt->getOperand(0)->getType();
1597
1598 // If this value is a constant, return the constant in the smallest FP type
1599 // that can accurately represent it. This allows us to turn
1600 // (float)((double)X+2.0) into x+2.0f.
1601 if (auto *CFP = dyn_cast<ConstantFP>(V))
1602 if (Type *T = shrinkFPConstant(CFP))
1603 return T;
1604
1605 // We can only correctly find a minimum type for a scalable vector when it is
1606 // a splat. For splats of constant values the fpext is wrapped up as a
1607 // ConstantExpr.
1608 if (auto *FPCExt = dyn_cast<ConstantExpr>(V))
1609 if (FPCExt->getOpcode() == Instruction::FPExt)
1610 return FPCExt->getOperand(0)->getType();
1611
1612 // Try to shrink a vector of FP constants. This returns nullptr on scalable
1613 // vectors
1614 if (Type *T = shrinkFPConstantVector(V))
1615 return T;
1616
1617 return V->getType();
1618 }
1619
1620 /// Return true if the cast from integer to FP can be proven to be exact for all
1621 /// possible inputs (the conversion does not lose any precision).
isKnownExactCastIntToFP(CastInst & I,InstCombinerImpl & IC)1622 static bool isKnownExactCastIntToFP(CastInst &I, InstCombinerImpl &IC) {
1623 CastInst::CastOps Opcode = I.getOpcode();
1624 assert((Opcode == CastInst::SIToFP || Opcode == CastInst::UIToFP) &&
1625 "Unexpected cast");
1626 Value *Src = I.getOperand(0);
1627 Type *SrcTy = Src->getType();
1628 Type *FPTy = I.getType();
1629 bool IsSigned = Opcode == Instruction::SIToFP;
1630 int SrcSize = (int)SrcTy->getScalarSizeInBits() - IsSigned;
1631
1632 // Easy case - if the source integer type has less bits than the FP mantissa,
1633 // then the cast must be exact.
1634 int DestNumSigBits = FPTy->getFPMantissaWidth();
1635 if (SrcSize <= DestNumSigBits)
1636 return true;
1637
1638 // Cast from FP to integer and back to FP is independent of the intermediate
1639 // integer width because of poison on overflow.
1640 Value *F;
1641 if (match(Src, m_FPToSI(m_Value(F))) || match(Src, m_FPToUI(m_Value(F)))) {
1642 // If this is uitofp (fptosi F), the source needs an extra bit to avoid
1643 // potential rounding of negative FP input values.
1644 int SrcNumSigBits = F->getType()->getFPMantissaWidth();
1645 if (!IsSigned && match(Src, m_FPToSI(m_Value())))
1646 SrcNumSigBits++;
1647
1648 // [su]itofp (fpto[su]i F) --> exact if the source type has less or equal
1649 // significant bits than the destination (and make sure neither type is
1650 // weird -- ppc_fp128).
1651 if (SrcNumSigBits > 0 && DestNumSigBits > 0 &&
1652 SrcNumSigBits <= DestNumSigBits)
1653 return true;
1654 }
1655
1656 // TODO:
1657 // Try harder to find if the source integer type has less significant bits.
1658 // For example, compute number of sign bits.
1659 KnownBits SrcKnown = IC.computeKnownBits(Src, 0, &I);
1660 int SigBits = (int)SrcTy->getScalarSizeInBits() -
1661 SrcKnown.countMinLeadingZeros() -
1662 SrcKnown.countMinTrailingZeros();
1663 if (SigBits <= DestNumSigBits)
1664 return true;
1665
1666 return false;
1667 }
1668
visitFPTrunc(FPTruncInst & FPT)1669 Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {
1670 if (Instruction *I = commonCastTransforms(FPT))
1671 return I;
1672
1673 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to
1674 // simplify this expression to avoid one or more of the trunc/extend
1675 // operations if we can do so without changing the numerical results.
1676 //
1677 // The exact manner in which the widths of the operands interact to limit
1678 // what we can and cannot do safely varies from operation to operation, and
1679 // is explained below in the various case statements.
1680 Type *Ty = FPT.getType();
1681 auto *BO = dyn_cast<BinaryOperator>(FPT.getOperand(0));
1682 if (BO && BO->hasOneUse()) {
1683 Type *LHSMinType = getMinimumFPType(BO->getOperand(0));
1684 Type *RHSMinType = getMinimumFPType(BO->getOperand(1));
1685 unsigned OpWidth = BO->getType()->getFPMantissaWidth();
1686 unsigned LHSWidth = LHSMinType->getFPMantissaWidth();
1687 unsigned RHSWidth = RHSMinType->getFPMantissaWidth();
1688 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);
1689 unsigned DstWidth = Ty->getFPMantissaWidth();
1690 switch (BO->getOpcode()) {
1691 default: break;
1692 case Instruction::FAdd:
1693 case Instruction::FSub:
1694 // For addition and subtraction, the infinitely precise result can
1695 // essentially be arbitrarily wide; proving that double rounding
1696 // will not occur because the result of OpI is exact (as we will for
1697 // FMul, for example) is hopeless. However, we *can* nonetheless
1698 // frequently know that double rounding cannot occur (or that it is
1699 // innocuous) by taking advantage of the specific structure of
1700 // infinitely-precise results that admit double rounding.
1701 //
1702 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient
1703 // to represent both sources, we can guarantee that the double
1704 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,
1705 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."
1706 // for proof of this fact).
1707 //
1708 // Note: Figueroa does not consider the case where DstFormat !=
1709 // SrcFormat. It's possible (likely even!) that this analysis
1710 // could be tightened for those cases, but they are rare (the main
1711 // case of interest here is (float)((double)float + float)).
1712 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {
1713 Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1714 Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1715 Instruction *RI = BinaryOperator::Create(BO->getOpcode(), LHS, RHS);
1716 RI->copyFastMathFlags(BO);
1717 return RI;
1718 }
1719 break;
1720 case Instruction::FMul:
1721 // For multiplication, the infinitely precise result has at most
1722 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient
1723 // that such a value can be exactly represented, then no double
1724 // rounding can possibly occur; we can safely perform the operation
1725 // in the destination format if it can represent both sources.
1726 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {
1727 Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1728 Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1729 return BinaryOperator::CreateFMulFMF(LHS, RHS, BO);
1730 }
1731 break;
1732 case Instruction::FDiv:
1733 // For division, we use again use the bound from Figueroa's
1734 // dissertation. I am entirely certain that this bound can be
1735 // tightened in the unbalanced operand case by an analysis based on
1736 // the diophantine rational approximation bound, but the well-known
1737 // condition used here is a good conservative first pass.
1738 // TODO: Tighten bound via rigorous analysis of the unbalanced case.
1739 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {
1740 Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);
1741 Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);
1742 return BinaryOperator::CreateFDivFMF(LHS, RHS, BO);
1743 }
1744 break;
1745 case Instruction::FRem: {
1746 // Remainder is straightforward. Remainder is always exact, so the
1747 // type of OpI doesn't enter into things at all. We simply evaluate
1748 // in whichever source type is larger, then convert to the
1749 // destination type.
1750 if (SrcWidth == OpWidth)
1751 break;
1752 Value *LHS, *RHS;
1753 if (LHSWidth == SrcWidth) {
1754 LHS = Builder.CreateFPTrunc(BO->getOperand(0), LHSMinType);
1755 RHS = Builder.CreateFPTrunc(BO->getOperand(1), LHSMinType);
1756 } else {
1757 LHS = Builder.CreateFPTrunc(BO->getOperand(0), RHSMinType);
1758 RHS = Builder.CreateFPTrunc(BO->getOperand(1), RHSMinType);
1759 }
1760
1761 Value *ExactResult = Builder.CreateFRemFMF(LHS, RHS, BO);
1762 return CastInst::CreateFPCast(ExactResult, Ty);
1763 }
1764 }
1765 }
1766
1767 // (fptrunc (fneg x)) -> (fneg (fptrunc x))
1768 Value *X;
1769 Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));
1770 if (Op && Op->hasOneUse()) {
1771 // FIXME: The FMF should propagate from the fptrunc, not the source op.
1772 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1773 if (isa<FPMathOperator>(Op))
1774 Builder.setFastMathFlags(Op->getFastMathFlags());
1775
1776 if (match(Op, m_FNeg(m_Value(X)))) {
1777 Value *InnerTrunc = Builder.CreateFPTrunc(X, Ty);
1778
1779 return UnaryOperator::CreateFNegFMF(InnerTrunc, Op);
1780 }
1781
1782 // If we are truncating a select that has an extended operand, we can
1783 // narrow the other operand and do the select as a narrow op.
1784 Value *Cond, *X, *Y;
1785 if (match(Op, m_Select(m_Value(Cond), m_FPExt(m_Value(X)), m_Value(Y))) &&
1786 X->getType() == Ty) {
1787 // fptrunc (select Cond, (fpext X), Y --> select Cond, X, (fptrunc Y)
1788 Value *NarrowY = Builder.CreateFPTrunc(Y, Ty);
1789 Value *Sel = Builder.CreateSelect(Cond, X, NarrowY, "narrow.sel", Op);
1790 return replaceInstUsesWith(FPT, Sel);
1791 }
1792 if (match(Op, m_Select(m_Value(Cond), m_Value(Y), m_FPExt(m_Value(X)))) &&
1793 X->getType() == Ty) {
1794 // fptrunc (select Cond, Y, (fpext X) --> select Cond, (fptrunc Y), X
1795 Value *NarrowY = Builder.CreateFPTrunc(Y, Ty);
1796 Value *Sel = Builder.CreateSelect(Cond, NarrowY, X, "narrow.sel", Op);
1797 return replaceInstUsesWith(FPT, Sel);
1798 }
1799 }
1800
1801 if (auto *II = dyn_cast<IntrinsicInst>(FPT.getOperand(0))) {
1802 switch (II->getIntrinsicID()) {
1803 default: break;
1804 case Intrinsic::ceil:
1805 case Intrinsic::fabs:
1806 case Intrinsic::floor:
1807 case Intrinsic::nearbyint:
1808 case Intrinsic::rint:
1809 case Intrinsic::round:
1810 case Intrinsic::roundeven:
1811 case Intrinsic::trunc: {
1812 Value *Src = II->getArgOperand(0);
1813 if (!Src->hasOneUse())
1814 break;
1815
1816 // Except for fabs, this transformation requires the input of the unary FP
1817 // operation to be itself an fpext from the type to which we're
1818 // truncating.
1819 if (II->getIntrinsicID() != Intrinsic::fabs) {
1820 FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);
1821 if (!FPExtSrc || FPExtSrc->getSrcTy() != Ty)
1822 break;
1823 }
1824
1825 // Do unary FP operation on smaller type.
1826 // (fptrunc (fabs x)) -> (fabs (fptrunc x))
1827 Value *InnerTrunc = Builder.CreateFPTrunc(Src, Ty);
1828 Function *Overload = Intrinsic::getDeclaration(FPT.getModule(),
1829 II->getIntrinsicID(), Ty);
1830 SmallVector<OperandBundleDef, 1> OpBundles;
1831 II->getOperandBundlesAsDefs(OpBundles);
1832 CallInst *NewCI =
1833 CallInst::Create(Overload, {InnerTrunc}, OpBundles, II->getName());
1834 NewCI->copyFastMathFlags(II);
1835 return NewCI;
1836 }
1837 }
1838 }
1839
1840 if (Instruction *I = shrinkInsertElt(FPT, Builder))
1841 return I;
1842
1843 Value *Src = FPT.getOperand(0);
1844 if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {
1845 auto *FPCast = cast<CastInst>(Src);
1846 if (isKnownExactCastIntToFP(*FPCast, *this))
1847 return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
1848 }
1849
1850 return nullptr;
1851 }
1852
visitFPExt(CastInst & FPExt)1853 Instruction *InstCombinerImpl::visitFPExt(CastInst &FPExt) {
1854 // If the source operand is a cast from integer to FP and known exact, then
1855 // cast the integer operand directly to the destination type.
1856 Type *Ty = FPExt.getType();
1857 Value *Src = FPExt.getOperand(0);
1858 if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {
1859 auto *FPCast = cast<CastInst>(Src);
1860 if (isKnownExactCastIntToFP(*FPCast, *this))
1861 return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);
1862 }
1863
1864 return commonCastTransforms(FPExt);
1865 }
1866
1867 /// fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)
1868 /// This is safe if the intermediate type has enough bits in its mantissa to
1869 /// accurately represent all values of X. For example, this won't work with
1870 /// i64 -> float -> i64.
foldItoFPtoI(CastInst & FI)1871 Instruction *InstCombinerImpl::foldItoFPtoI(CastInst &FI) {
1872 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))
1873 return nullptr;
1874
1875 auto *OpI = cast<CastInst>(FI.getOperand(0));
1876 Value *X = OpI->getOperand(0);
1877 Type *XType = X->getType();
1878 Type *DestType = FI.getType();
1879 bool IsOutputSigned = isa<FPToSIInst>(FI);
1880
1881 // Since we can assume the conversion won't overflow, our decision as to
1882 // whether the input will fit in the float should depend on the minimum
1883 // of the input range and output range.
1884
1885 // This means this is also safe for a signed input and unsigned output, since
1886 // a negative input would lead to undefined behavior.
1887 if (!isKnownExactCastIntToFP(*OpI, *this)) {
1888 // The first cast may not round exactly based on the source integer width
1889 // and FP width, but the overflow UB rules can still allow this to fold.
1890 // If the destination type is narrow, that means the intermediate FP value
1891 // must be large enough to hold the source value exactly.
1892 // For example, (uint8_t)((float)(uint32_t 16777217) is undefined behavior.
1893 int OutputSize = (int)DestType->getScalarSizeInBits();
1894 if (OutputSize > OpI->getType()->getFPMantissaWidth())
1895 return nullptr;
1896 }
1897
1898 if (DestType->getScalarSizeInBits() > XType->getScalarSizeInBits()) {
1899 bool IsInputSigned = isa<SIToFPInst>(OpI);
1900 if (IsInputSigned && IsOutputSigned)
1901 return new SExtInst(X, DestType);
1902 return new ZExtInst(X, DestType);
1903 }
1904 if (DestType->getScalarSizeInBits() < XType->getScalarSizeInBits())
1905 return new TruncInst(X, DestType);
1906
1907 assert(XType == DestType && "Unexpected types for int to FP to int casts");
1908 return replaceInstUsesWith(FI, X);
1909 }
1910
visitFPToUI(FPToUIInst & FI)1911 Instruction *InstCombinerImpl::visitFPToUI(FPToUIInst &FI) {
1912 if (Instruction *I = foldItoFPtoI(FI))
1913 return I;
1914
1915 return commonCastTransforms(FI);
1916 }
1917
visitFPToSI(FPToSIInst & FI)1918 Instruction *InstCombinerImpl::visitFPToSI(FPToSIInst &FI) {
1919 if (Instruction *I = foldItoFPtoI(FI))
1920 return I;
1921
1922 return commonCastTransforms(FI);
1923 }
1924
visitUIToFP(CastInst & CI)1925 Instruction *InstCombinerImpl::visitUIToFP(CastInst &CI) {
1926 return commonCastTransforms(CI);
1927 }
1928
visitSIToFP(CastInst & CI)1929 Instruction *InstCombinerImpl::visitSIToFP(CastInst &CI) {
1930 return commonCastTransforms(CI);
1931 }
1932
visitIntToPtr(IntToPtrInst & CI)1933 Instruction *InstCombinerImpl::visitIntToPtr(IntToPtrInst &CI) {
1934 // If the source integer type is not the intptr_t type for this target, do a
1935 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the
1936 // cast to be exposed to other transforms.
1937 unsigned AS = CI.getAddressSpace();
1938 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=
1939 DL.getPointerSizeInBits(AS)) {
1940 Type *Ty = CI.getOperand(0)->getType()->getWithNewType(
1941 DL.getIntPtrType(CI.getContext(), AS));
1942 Value *P = Builder.CreateZExtOrTrunc(CI.getOperand(0), Ty);
1943 return new IntToPtrInst(P, CI.getType());
1944 }
1945
1946 if (Instruction *I = commonCastTransforms(CI))
1947 return I;
1948
1949 return nullptr;
1950 }
1951
visitPtrToInt(PtrToIntInst & CI)1952 Instruction *InstCombinerImpl::visitPtrToInt(PtrToIntInst &CI) {
1953 // If the destination integer type is not the intptr_t type for this target,
1954 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast
1955 // to be exposed to other transforms.
1956 Value *SrcOp = CI.getPointerOperand();
1957 Type *SrcTy = SrcOp->getType();
1958 Type *Ty = CI.getType();
1959 unsigned AS = CI.getPointerAddressSpace();
1960 unsigned TySize = Ty->getScalarSizeInBits();
1961 unsigned PtrSize = DL.getPointerSizeInBits(AS);
1962 if (TySize != PtrSize) {
1963 Type *IntPtrTy =
1964 SrcTy->getWithNewType(DL.getIntPtrType(CI.getContext(), AS));
1965 Value *P = Builder.CreatePtrToInt(SrcOp, IntPtrTy);
1966 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);
1967 }
1968
1969 // (ptrtoint (ptrmask P, M))
1970 // -> (and (ptrtoint P), M)
1971 // This is generally beneficial as `and` is better supported than `ptrmask`.
1972 Value *Ptr, *Mask;
1973 if (match(SrcOp, m_OneUse(m_Intrinsic<Intrinsic::ptrmask>(m_Value(Ptr),
1974 m_Value(Mask)))) &&
1975 Mask->getType() == Ty)
1976 return BinaryOperator::CreateAnd(Builder.CreatePtrToInt(Ptr, Ty), Mask);
1977
1978 if (auto *GEP = dyn_cast<GetElementPtrInst>(SrcOp)) {
1979 // Fold ptrtoint(gep null, x) to multiply + constant if the GEP has one use.
1980 // While this can increase the number of instructions it doesn't actually
1981 // increase the overall complexity since the arithmetic is just part of
1982 // the GEP otherwise.
1983 if (GEP->hasOneUse() &&
1984 isa<ConstantPointerNull>(GEP->getPointerOperand())) {
1985 return replaceInstUsesWith(CI,
1986 Builder.CreateIntCast(EmitGEPOffset(GEP), Ty,
1987 /*isSigned=*/false));
1988 }
1989 }
1990
1991 Value *Vec, *Scalar, *Index;
1992 if (match(SrcOp, m_OneUse(m_InsertElt(m_IntToPtr(m_Value(Vec)),
1993 m_Value(Scalar), m_Value(Index)))) &&
1994 Vec->getType() == Ty) {
1995 assert(Vec->getType()->getScalarSizeInBits() == PtrSize && "Wrong type");
1996 // Convert the scalar to int followed by insert to eliminate one cast:
1997 // p2i (ins (i2p Vec), Scalar, Index --> ins Vec, (p2i Scalar), Index
1998 Value *NewCast = Builder.CreatePtrToInt(Scalar, Ty->getScalarType());
1999 return InsertElementInst::Create(Vec, NewCast, Index);
2000 }
2001
2002 return commonCastTransforms(CI);
2003 }
2004
2005 /// This input value (which is known to have vector type) is being zero extended
2006 /// or truncated to the specified vector type. Since the zext/trunc is done
2007 /// using an integer type, we have a (bitcast(cast(bitcast))) pattern,
2008 /// endianness will impact which end of the vector that is extended or
2009 /// truncated.
2010 ///
2011 /// A vector is always stored with index 0 at the lowest address, which
2012 /// corresponds to the most significant bits for a big endian stored integer and
2013 /// the least significant bits for little endian. A trunc/zext of an integer
2014 /// impacts the big end of the integer. Thus, we need to add/remove elements at
2015 /// the front of the vector for big endian targets, and the back of the vector
2016 /// for little endian targets.
2017 ///
2018 /// Try to replace it with a shuffle (and vector/vector bitcast) if possible.
2019 ///
2020 /// The source and destination vector types may have different element types.
2021 static Instruction *
optimizeVectorResizeWithIntegerBitCasts(Value * InVal,VectorType * DestTy,InstCombinerImpl & IC)2022 optimizeVectorResizeWithIntegerBitCasts(Value *InVal, VectorType *DestTy,
2023 InstCombinerImpl &IC) {
2024 // We can only do this optimization if the output is a multiple of the input
2025 // element size, or the input is a multiple of the output element size.
2026 // Convert the input type to have the same element type as the output.
2027 VectorType *SrcTy = cast<VectorType>(InVal->getType());
2028
2029 if (SrcTy->getElementType() != DestTy->getElementType()) {
2030 // The input types don't need to be identical, but for now they must be the
2031 // same size. There is no specific reason we couldn't handle things like
2032 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten
2033 // there yet.
2034 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=
2035 DestTy->getElementType()->getPrimitiveSizeInBits())
2036 return nullptr;
2037
2038 SrcTy =
2039 FixedVectorType::get(DestTy->getElementType(),
2040 cast<FixedVectorType>(SrcTy)->getNumElements());
2041 InVal = IC.Builder.CreateBitCast(InVal, SrcTy);
2042 }
2043
2044 bool IsBigEndian = IC.getDataLayout().isBigEndian();
2045 unsigned SrcElts = cast<FixedVectorType>(SrcTy)->getNumElements();
2046 unsigned DestElts = cast<FixedVectorType>(DestTy)->getNumElements();
2047
2048 assert(SrcElts != DestElts && "Element counts should be different.");
2049
2050 // Now that the element types match, get the shuffle mask and RHS of the
2051 // shuffle to use, which depends on whether we're increasing or decreasing the
2052 // size of the input.
2053 auto ShuffleMaskStorage = llvm::to_vector<16>(llvm::seq<int>(0, SrcElts));
2054 ArrayRef<int> ShuffleMask;
2055 Value *V2;
2056
2057 if (SrcElts > DestElts) {
2058 // If we're shrinking the number of elements (rewriting an integer
2059 // truncate), just shuffle in the elements corresponding to the least
2060 // significant bits from the input and use poison as the second shuffle
2061 // input.
2062 V2 = PoisonValue::get(SrcTy);
2063 // Make sure the shuffle mask selects the "least significant bits" by
2064 // keeping elements from back of the src vector for big endian, and from the
2065 // front for little endian.
2066 ShuffleMask = ShuffleMaskStorage;
2067 if (IsBigEndian)
2068 ShuffleMask = ShuffleMask.take_back(DestElts);
2069 else
2070 ShuffleMask = ShuffleMask.take_front(DestElts);
2071 } else {
2072 // If we're increasing the number of elements (rewriting an integer zext),
2073 // shuffle in all of the elements from InVal. Fill the rest of the result
2074 // elements with zeros from a constant zero.
2075 V2 = Constant::getNullValue(SrcTy);
2076 // Use first elt from V2 when indicating zero in the shuffle mask.
2077 uint32_t NullElt = SrcElts;
2078 // Extend with null values in the "most significant bits" by adding elements
2079 // in front of the src vector for big endian, and at the back for little
2080 // endian.
2081 unsigned DeltaElts = DestElts - SrcElts;
2082 if (IsBigEndian)
2083 ShuffleMaskStorage.insert(ShuffleMaskStorage.begin(), DeltaElts, NullElt);
2084 else
2085 ShuffleMaskStorage.append(DeltaElts, NullElt);
2086 ShuffleMask = ShuffleMaskStorage;
2087 }
2088
2089 return new ShuffleVectorInst(InVal, V2, ShuffleMask);
2090 }
2091
isMultipleOfTypeSize(unsigned Value,Type * Ty)2092 static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {
2093 return Value % Ty->getPrimitiveSizeInBits() == 0;
2094 }
2095
getTypeSizeIndex(unsigned Value,Type * Ty)2096 static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {
2097 return Value / Ty->getPrimitiveSizeInBits();
2098 }
2099
2100 /// V is a value which is inserted into a vector of VecEltTy.
2101 /// Look through the value to see if we can decompose it into
2102 /// insertions into the vector. See the example in the comment for
2103 /// OptimizeIntegerToVectorInsertions for the pattern this handles.
2104 /// The type of V is always a non-zero multiple of VecEltTy's size.
2105 /// Shift is the number of bits between the lsb of V and the lsb of
2106 /// the vector.
2107 ///
2108 /// This returns false if the pattern can't be matched or true if it can,
2109 /// filling in Elements with the elements found here.
collectInsertionElements(Value * V,unsigned Shift,SmallVectorImpl<Value * > & Elements,Type * VecEltTy,bool isBigEndian)2110 static bool collectInsertionElements(Value *V, unsigned Shift,
2111 SmallVectorImpl<Value *> &Elements,
2112 Type *VecEltTy, bool isBigEndian) {
2113 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&
2114 "Shift should be a multiple of the element type size");
2115
2116 // Undef values never contribute useful bits to the result.
2117 if (isa<UndefValue>(V)) return true;
2118
2119 // If we got down to a value of the right type, we win, try inserting into the
2120 // right element.
2121 if (V->getType() == VecEltTy) {
2122 // Inserting null doesn't actually insert any elements.
2123 if (Constant *C = dyn_cast<Constant>(V))
2124 if (C->isNullValue())
2125 return true;
2126
2127 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);
2128 if (isBigEndian)
2129 ElementIndex = Elements.size() - ElementIndex - 1;
2130
2131 // Fail if multiple elements are inserted into this slot.
2132 if (Elements[ElementIndex])
2133 return false;
2134
2135 Elements[ElementIndex] = V;
2136 return true;
2137 }
2138
2139 if (Constant *C = dyn_cast<Constant>(V)) {
2140 // Figure out the # elements this provides, and bitcast it or slice it up
2141 // as required.
2142 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),
2143 VecEltTy);
2144 // If the constant is the size of a vector element, we just need to bitcast
2145 // it to the right type so it gets properly inserted.
2146 if (NumElts == 1)
2147 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),
2148 Shift, Elements, VecEltTy, isBigEndian);
2149
2150 // Okay, this is a constant that covers multiple elements. Slice it up into
2151 // pieces and insert each element-sized piece into the vector.
2152 if (!isa<IntegerType>(C->getType()))
2153 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),
2154 C->getType()->getPrimitiveSizeInBits()));
2155 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();
2156 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);
2157
2158 for (unsigned i = 0; i != NumElts; ++i) {
2159 unsigned ShiftI = i * ElementSize;
2160 Constant *Piece = ConstantFoldBinaryInstruction(
2161 Instruction::LShr, C, ConstantInt::get(C->getType(), ShiftI));
2162 if (!Piece)
2163 return false;
2164
2165 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);
2166 if (!collectInsertionElements(Piece, ShiftI + Shift, Elements, VecEltTy,
2167 isBigEndian))
2168 return false;
2169 }
2170 return true;
2171 }
2172
2173 if (!V->hasOneUse()) return false;
2174
2175 Instruction *I = dyn_cast<Instruction>(V);
2176 if (!I) return false;
2177 switch (I->getOpcode()) {
2178 default: return false; // Unhandled case.
2179 case Instruction::BitCast:
2180 if (I->getOperand(0)->getType()->isVectorTy())
2181 return false;
2182 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2183 isBigEndian);
2184 case Instruction::ZExt:
2185 if (!isMultipleOfTypeSize(
2186 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),
2187 VecEltTy))
2188 return false;
2189 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2190 isBigEndian);
2191 case Instruction::Or:
2192 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2193 isBigEndian) &&
2194 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,
2195 isBigEndian);
2196 case Instruction::Shl: {
2197 // Must be shifting by a constant that is a multiple of the element size.
2198 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));
2199 if (!CI) return false;
2200 Shift += CI->getZExtValue();
2201 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;
2202 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,
2203 isBigEndian);
2204 }
2205
2206 }
2207 }
2208
2209
2210 /// If the input is an 'or' instruction, we may be doing shifts and ors to
2211 /// assemble the elements of the vector manually.
2212 /// Try to rip the code out and replace it with insertelements. This is to
2213 /// optimize code like this:
2214 ///
2215 /// %tmp37 = bitcast float %inc to i32
2216 /// %tmp38 = zext i32 %tmp37 to i64
2217 /// %tmp31 = bitcast float %inc5 to i32
2218 /// %tmp32 = zext i32 %tmp31 to i64
2219 /// %tmp33 = shl i64 %tmp32, 32
2220 /// %ins35 = or i64 %tmp33, %tmp38
2221 /// %tmp43 = bitcast i64 %ins35 to <2 x float>
2222 ///
2223 /// Into two insertelements that do "buildvector{%inc, %inc5}".
optimizeIntegerToVectorInsertions(BitCastInst & CI,InstCombinerImpl & IC)2224 static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,
2225 InstCombinerImpl &IC) {
2226 auto *DestVecTy = cast<FixedVectorType>(CI.getType());
2227 Value *IntInput = CI.getOperand(0);
2228
2229 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());
2230 if (!collectInsertionElements(IntInput, 0, Elements,
2231 DestVecTy->getElementType(),
2232 IC.getDataLayout().isBigEndian()))
2233 return nullptr;
2234
2235 // If we succeeded, we know that all of the element are specified by Elements
2236 // or are zero if Elements has a null entry. Recast this as a set of
2237 // insertions.
2238 Value *Result = Constant::getNullValue(CI.getType());
2239 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {
2240 if (!Elements[i]) continue; // Unset element.
2241
2242 Result = IC.Builder.CreateInsertElement(Result, Elements[i],
2243 IC.Builder.getInt32(i));
2244 }
2245
2246 return Result;
2247 }
2248
2249 /// Canonicalize scalar bitcasts of extracted elements into a bitcast of the
2250 /// vector followed by extract element. The backend tends to handle bitcasts of
2251 /// vectors better than bitcasts of scalars because vector registers are
2252 /// usually not type-specific like scalar integer or scalar floating-point.
canonicalizeBitCastExtElt(BitCastInst & BitCast,InstCombinerImpl & IC)2253 static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,
2254 InstCombinerImpl &IC) {
2255 Value *VecOp, *Index;
2256 if (!match(BitCast.getOperand(0),
2257 m_OneUse(m_ExtractElt(m_Value(VecOp), m_Value(Index)))))
2258 return nullptr;
2259
2260 // The bitcast must be to a vectorizable type, otherwise we can't make a new
2261 // type to extract from.
2262 Type *DestType = BitCast.getType();
2263 VectorType *VecType = cast<VectorType>(VecOp->getType());
2264 if (VectorType::isValidElementType(DestType)) {
2265 auto *NewVecType = VectorType::get(DestType, VecType);
2266 auto *NewBC = IC.Builder.CreateBitCast(VecOp, NewVecType, "bc");
2267 return ExtractElementInst::Create(NewBC, Index);
2268 }
2269
2270 // Only solve DestType is vector to avoid inverse transform in visitBitCast.
2271 // bitcast (extractelement <1 x elt>, dest) -> bitcast(<1 x elt>, dest)
2272 auto *FixedVType = dyn_cast<FixedVectorType>(VecType);
2273 if (DestType->isVectorTy() && FixedVType && FixedVType->getNumElements() == 1)
2274 return CastInst::Create(Instruction::BitCast, VecOp, DestType);
2275
2276 return nullptr;
2277 }
2278
2279 /// Change the type of a bitwise logic operation if we can eliminate a bitcast.
foldBitCastBitwiseLogic(BitCastInst & BitCast,InstCombiner::BuilderTy & Builder)2280 static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,
2281 InstCombiner::BuilderTy &Builder) {
2282 Type *DestTy = BitCast.getType();
2283 BinaryOperator *BO;
2284
2285 if (!match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||
2286 !BO->isBitwiseLogicOp())
2287 return nullptr;
2288
2289 // FIXME: This transform is restricted to vector types to avoid backend
2290 // problems caused by creating potentially illegal operations. If a fix-up is
2291 // added to handle that situation, we can remove this check.
2292 if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())
2293 return nullptr;
2294
2295 if (DestTy->isFPOrFPVectorTy()) {
2296 Value *X, *Y;
2297 // bitcast(logic(bitcast(X), bitcast(Y))) -> bitcast'(logic(bitcast'(X), Y))
2298 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
2299 match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(Y))))) {
2300 if (X->getType()->isFPOrFPVectorTy() &&
2301 Y->getType()->isIntOrIntVectorTy()) {
2302 Value *CastedOp =
2303 Builder.CreateBitCast(BO->getOperand(0), Y->getType());
2304 Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, Y);
2305 return CastInst::CreateBitOrPointerCast(NewBO, DestTy);
2306 }
2307 if (X->getType()->isIntOrIntVectorTy() &&
2308 Y->getType()->isFPOrFPVectorTy()) {
2309 Value *CastedOp =
2310 Builder.CreateBitCast(BO->getOperand(1), X->getType());
2311 Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, X);
2312 return CastInst::CreateBitOrPointerCast(NewBO, DestTy);
2313 }
2314 }
2315 return nullptr;
2316 }
2317
2318 if (!DestTy->isIntOrIntVectorTy())
2319 return nullptr;
2320
2321 Value *X;
2322 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&
2323 X->getType() == DestTy && !isa<Constant>(X)) {
2324 // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))
2325 Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);
2326 return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);
2327 }
2328
2329 if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&
2330 X->getType() == DestTy && !isa<Constant>(X)) {
2331 // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)
2332 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
2333 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);
2334 }
2335
2336 // Canonicalize vector bitcasts to come before vector bitwise logic with a
2337 // constant. This eases recognition of special constants for later ops.
2338 // Example:
2339 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
2340 Constant *C;
2341 if (match(BO->getOperand(1), m_Constant(C))) {
2342 // bitcast (logic X, C) --> logic (bitcast X, C')
2343 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);
2344 Value *CastedC = Builder.CreateBitCast(C, DestTy);
2345 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, CastedC);
2346 }
2347
2348 return nullptr;
2349 }
2350
2351 /// Change the type of a select if we can eliminate a bitcast.
foldBitCastSelect(BitCastInst & BitCast,InstCombiner::BuilderTy & Builder)2352 static Instruction *foldBitCastSelect(BitCastInst &BitCast,
2353 InstCombiner::BuilderTy &Builder) {
2354 Value *Cond, *TVal, *FVal;
2355 if (!match(BitCast.getOperand(0),
2356 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
2357 return nullptr;
2358
2359 // A vector select must maintain the same number of elements in its operands.
2360 Type *CondTy = Cond->getType();
2361 Type *DestTy = BitCast.getType();
2362 if (auto *CondVTy = dyn_cast<VectorType>(CondTy))
2363 if (!DestTy->isVectorTy() ||
2364 CondVTy->getElementCount() !=
2365 cast<VectorType>(DestTy)->getElementCount())
2366 return nullptr;
2367
2368 // FIXME: This transform is restricted from changing the select between
2369 // scalars and vectors to avoid backend problems caused by creating
2370 // potentially illegal operations. If a fix-up is added to handle that
2371 // situation, we can remove this check.
2372 if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())
2373 return nullptr;
2374
2375 auto *Sel = cast<Instruction>(BitCast.getOperand(0));
2376 Value *X;
2377 if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
2378 !isa<Constant>(X)) {
2379 // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))
2380 Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);
2381 return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);
2382 }
2383
2384 if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&
2385 !isa<Constant>(X)) {
2386 // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)
2387 Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);
2388 return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);
2389 }
2390
2391 return nullptr;
2392 }
2393
2394 /// Check if all users of CI are StoreInsts.
hasStoreUsersOnly(CastInst & CI)2395 static bool hasStoreUsersOnly(CastInst &CI) {
2396 for (User *U : CI.users()) {
2397 if (!isa<StoreInst>(U))
2398 return false;
2399 }
2400 return true;
2401 }
2402
2403 /// This function handles following case
2404 ///
2405 /// A -> B cast
2406 /// PHI
2407 /// B -> A cast
2408 ///
2409 /// All the related PHI nodes can be replaced by new PHI nodes with type A.
2410 /// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
optimizeBitCastFromPhi(CastInst & CI,PHINode * PN)2411 Instruction *InstCombinerImpl::optimizeBitCastFromPhi(CastInst &CI,
2412 PHINode *PN) {
2413 // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.
2414 if (hasStoreUsersOnly(CI))
2415 return nullptr;
2416
2417 Value *Src = CI.getOperand(0);
2418 Type *SrcTy = Src->getType(); // Type B
2419 Type *DestTy = CI.getType(); // Type A
2420
2421 SmallVector<PHINode *, 4> PhiWorklist;
2422 SmallSetVector<PHINode *, 4> OldPhiNodes;
2423
2424 // Find all of the A->B casts and PHI nodes.
2425 // We need to inspect all related PHI nodes, but PHIs can be cyclic, so
2426 // OldPhiNodes is used to track all known PHI nodes, before adding a new
2427 // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
2428 PhiWorklist.push_back(PN);
2429 OldPhiNodes.insert(PN);
2430 while (!PhiWorklist.empty()) {
2431 auto *OldPN = PhiWorklist.pop_back_val();
2432 for (Value *IncValue : OldPN->incoming_values()) {
2433 if (isa<Constant>(IncValue))
2434 continue;
2435
2436 if (auto *LI = dyn_cast<LoadInst>(IncValue)) {
2437 // If there is a sequence of one or more load instructions, each loaded
2438 // value is used as address of later load instruction, bitcast is
2439 // necessary to change the value type, don't optimize it. For
2440 // simplicity we give up if the load address comes from another load.
2441 Value *Addr = LI->getOperand(0);
2442 if (Addr == &CI || isa<LoadInst>(Addr))
2443 return nullptr;
2444 // Don't tranform "load <256 x i32>, <256 x i32>*" to
2445 // "load x86_amx, x86_amx*", because x86_amx* is invalid.
2446 // TODO: Remove this check when bitcast between vector and x86_amx
2447 // is replaced with a specific intrinsic.
2448 if (DestTy->isX86_AMXTy())
2449 return nullptr;
2450 if (LI->hasOneUse() && LI->isSimple())
2451 continue;
2452 // If a LoadInst has more than one use, changing the type of loaded
2453 // value may create another bitcast.
2454 return nullptr;
2455 }
2456
2457 if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
2458 if (OldPhiNodes.insert(PNode))
2459 PhiWorklist.push_back(PNode);
2460 continue;
2461 }
2462
2463 auto *BCI = dyn_cast<BitCastInst>(IncValue);
2464 // We can't handle other instructions.
2465 if (!BCI)
2466 return nullptr;
2467
2468 // Verify it's a A->B cast.
2469 Type *TyA = BCI->getOperand(0)->getType();
2470 Type *TyB = BCI->getType();
2471 if (TyA != DestTy || TyB != SrcTy)
2472 return nullptr;
2473 }
2474 }
2475
2476 // Check that each user of each old PHI node is something that we can
2477 // rewrite, so that all of the old PHI nodes can be cleaned up afterwards.
2478 for (auto *OldPN : OldPhiNodes) {
2479 for (User *V : OldPN->users()) {
2480 if (auto *SI = dyn_cast<StoreInst>(V)) {
2481 if (!SI->isSimple() || SI->getOperand(0) != OldPN)
2482 return nullptr;
2483 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2484 // Verify it's a B->A cast.
2485 Type *TyB = BCI->getOperand(0)->getType();
2486 Type *TyA = BCI->getType();
2487 if (TyA != DestTy || TyB != SrcTy)
2488 return nullptr;
2489 } else if (auto *PHI = dyn_cast<PHINode>(V)) {
2490 // As long as the user is another old PHI node, then even if we don't
2491 // rewrite it, the PHI web we're considering won't have any users
2492 // outside itself, so it'll be dead.
2493 if (!OldPhiNodes.contains(PHI))
2494 return nullptr;
2495 } else {
2496 return nullptr;
2497 }
2498 }
2499 }
2500
2501 // For each old PHI node, create a corresponding new PHI node with a type A.
2502 SmallDenseMap<PHINode *, PHINode *> NewPNodes;
2503 for (auto *OldPN : OldPhiNodes) {
2504 Builder.SetInsertPoint(OldPN);
2505 PHINode *NewPN = Builder.CreatePHI(DestTy, OldPN->getNumOperands());
2506 NewPNodes[OldPN] = NewPN;
2507 }
2508
2509 // Fill in the operands of new PHI nodes.
2510 for (auto *OldPN : OldPhiNodes) {
2511 PHINode *NewPN = NewPNodes[OldPN];
2512 for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
2513 Value *V = OldPN->getOperand(j);
2514 Value *NewV = nullptr;
2515 if (auto *C = dyn_cast<Constant>(V)) {
2516 NewV = ConstantExpr::getBitCast(C, DestTy);
2517 } else if (auto *LI = dyn_cast<LoadInst>(V)) {
2518 // Explicitly perform load combine to make sure no opposing transform
2519 // can remove the bitcast in the meantime and trigger an infinite loop.
2520 Builder.SetInsertPoint(LI);
2521 NewV = combineLoadToNewType(*LI, DestTy);
2522 // Remove the old load and its use in the old phi, which itself becomes
2523 // dead once the whole transform finishes.
2524 replaceInstUsesWith(*LI, PoisonValue::get(LI->getType()));
2525 eraseInstFromFunction(*LI);
2526 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2527 NewV = BCI->getOperand(0);
2528 } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {
2529 NewV = NewPNodes[PrevPN];
2530 }
2531 assert(NewV);
2532 NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
2533 }
2534 }
2535
2536 // Traverse all accumulated PHI nodes and process its users,
2537 // which are Stores and BitcCasts. Without this processing
2538 // NewPHI nodes could be replicated and could lead to extra
2539 // moves generated after DeSSA.
2540 // If there is a store with type B, change it to type A.
2541
2542
2543 // Replace users of BitCast B->A with NewPHI. These will help
2544 // later to get rid off a closure formed by OldPHI nodes.
2545 Instruction *RetVal = nullptr;
2546 for (auto *OldPN : OldPhiNodes) {
2547 PHINode *NewPN = NewPNodes[OldPN];
2548 for (User *V : make_early_inc_range(OldPN->users())) {
2549 if (auto *SI = dyn_cast<StoreInst>(V)) {
2550 assert(SI->isSimple() && SI->getOperand(0) == OldPN);
2551 Builder.SetInsertPoint(SI);
2552 auto *NewBC =
2553 cast<BitCastInst>(Builder.CreateBitCast(NewPN, SrcTy));
2554 SI->setOperand(0, NewBC);
2555 Worklist.push(SI);
2556 assert(hasStoreUsersOnly(*NewBC));
2557 }
2558 else if (auto *BCI = dyn_cast<BitCastInst>(V)) {
2559 Type *TyB = BCI->getOperand(0)->getType();
2560 Type *TyA = BCI->getType();
2561 assert(TyA == DestTy && TyB == SrcTy);
2562 (void) TyA;
2563 (void) TyB;
2564 Instruction *I = replaceInstUsesWith(*BCI, NewPN);
2565 if (BCI == &CI)
2566 RetVal = I;
2567 } else if (auto *PHI = dyn_cast<PHINode>(V)) {
2568 assert(OldPhiNodes.contains(PHI));
2569 (void) PHI;
2570 } else {
2571 llvm_unreachable("all uses should be handled");
2572 }
2573 }
2574 }
2575
2576 return RetVal;
2577 }
2578
visitBitCast(BitCastInst & CI)2579 Instruction *InstCombinerImpl::visitBitCast(BitCastInst &CI) {
2580 // If the operands are integer typed then apply the integer transforms,
2581 // otherwise just apply the common ones.
2582 Value *Src = CI.getOperand(0);
2583 Type *SrcTy = Src->getType();
2584 Type *DestTy = CI.getType();
2585
2586 // Get rid of casts from one type to the same type. These are useless and can
2587 // be replaced by the operand.
2588 if (DestTy == Src->getType())
2589 return replaceInstUsesWith(CI, Src);
2590
2591 if (FixedVectorType *DestVTy = dyn_cast<FixedVectorType>(DestTy)) {
2592 // Beware: messing with this target-specific oddity may cause trouble.
2593 if (DestVTy->getNumElements() == 1 && SrcTy->isX86_MMXTy()) {
2594 Value *Elem = Builder.CreateBitCast(Src, DestVTy->getElementType());
2595 return InsertElementInst::Create(PoisonValue::get(DestTy), Elem,
2596 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2597 }
2598
2599 if (isa<IntegerType>(SrcTy)) {
2600 // If this is a cast from an integer to vector, check to see if the input
2601 // is a trunc or zext of a bitcast from vector. If so, we can replace all
2602 // the casts with a shuffle and (potentially) a bitcast.
2603 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {
2604 CastInst *SrcCast = cast<CastInst>(Src);
2605 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))
2606 if (isa<VectorType>(BCIn->getOperand(0)->getType()))
2607 if (Instruction *I = optimizeVectorResizeWithIntegerBitCasts(
2608 BCIn->getOperand(0), cast<VectorType>(DestTy), *this))
2609 return I;
2610 }
2611
2612 // If the input is an 'or' instruction, we may be doing shifts and ors to
2613 // assemble the elements of the vector manually. Try to rip the code out
2614 // and replace it with insertelements.
2615 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))
2616 return replaceInstUsesWith(CI, V);
2617 }
2618 }
2619
2620 if (FixedVectorType *SrcVTy = dyn_cast<FixedVectorType>(SrcTy)) {
2621 if (SrcVTy->getNumElements() == 1) {
2622 // If our destination is not a vector, then make this a straight
2623 // scalar-scalar cast.
2624 if (!DestTy->isVectorTy()) {
2625 Value *Elem =
2626 Builder.CreateExtractElement(Src,
2627 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));
2628 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
2629 }
2630
2631 // Otherwise, see if our source is an insert. If so, then use the scalar
2632 // component directly:
2633 // bitcast (inselt <1 x elt> V, X, 0) to <n x m> --> bitcast X to <n x m>
2634 if (auto *InsElt = dyn_cast<InsertElementInst>(Src))
2635 return new BitCastInst(InsElt->getOperand(1), DestTy);
2636 }
2637
2638 // Convert an artificial vector insert into more analyzable bitwise logic.
2639 unsigned BitWidth = DestTy->getScalarSizeInBits();
2640 Value *X, *Y;
2641 uint64_t IndexC;
2642 if (match(Src, m_OneUse(m_InsertElt(m_OneUse(m_BitCast(m_Value(X))),
2643 m_Value(Y), m_ConstantInt(IndexC)))) &&
2644 DestTy->isIntegerTy() && X->getType() == DestTy &&
2645 Y->getType()->isIntegerTy() && isDesirableIntType(BitWidth)) {
2646 // Adjust for big endian - the LSBs are at the high index.
2647 if (DL.isBigEndian())
2648 IndexC = SrcVTy->getNumElements() - 1 - IndexC;
2649
2650 // We only handle (endian-normalized) insert to index 0. Any other insert
2651 // would require a left-shift, so that is an extra instruction.
2652 if (IndexC == 0) {
2653 // bitcast (inselt (bitcast X), Y, 0) --> or (and X, MaskC), (zext Y)
2654 unsigned EltWidth = Y->getType()->getScalarSizeInBits();
2655 APInt MaskC = APInt::getHighBitsSet(BitWidth, BitWidth - EltWidth);
2656 Value *AndX = Builder.CreateAnd(X, MaskC);
2657 Value *ZextY = Builder.CreateZExt(Y, DestTy);
2658 return BinaryOperator::CreateOr(AndX, ZextY);
2659 }
2660 }
2661 }
2662
2663 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(Src)) {
2664 // Okay, we have (bitcast (shuffle ..)). Check to see if this is
2665 // a bitcast to a vector with the same # elts.
2666 Value *ShufOp0 = Shuf->getOperand(0);
2667 Value *ShufOp1 = Shuf->getOperand(1);
2668 auto ShufElts = cast<VectorType>(Shuf->getType())->getElementCount();
2669 auto SrcVecElts = cast<VectorType>(ShufOp0->getType())->getElementCount();
2670 if (Shuf->hasOneUse() && DestTy->isVectorTy() &&
2671 cast<VectorType>(DestTy)->getElementCount() == ShufElts &&
2672 ShufElts == SrcVecElts) {
2673 BitCastInst *Tmp;
2674 // If either of the operands is a cast from CI.getType(), then
2675 // evaluating the shuffle in the casted destination's type will allow
2676 // us to eliminate at least one cast.
2677 if (((Tmp = dyn_cast<BitCastInst>(ShufOp0)) &&
2678 Tmp->getOperand(0)->getType() == DestTy) ||
2679 ((Tmp = dyn_cast<BitCastInst>(ShufOp1)) &&
2680 Tmp->getOperand(0)->getType() == DestTy)) {
2681 Value *LHS = Builder.CreateBitCast(ShufOp0, DestTy);
2682 Value *RHS = Builder.CreateBitCast(ShufOp1, DestTy);
2683 // Return a new shuffle vector. Use the same element ID's, as we
2684 // know the vector types match #elts.
2685 return new ShuffleVectorInst(LHS, RHS, Shuf->getShuffleMask());
2686 }
2687 }
2688
2689 // A bitcasted-to-scalar and byte/bit reversing shuffle is better recognized
2690 // as a byte/bit swap:
2691 // bitcast <N x i8> (shuf X, undef, <N, N-1,...0>) -> bswap (bitcast X)
2692 // bitcast <N x i1> (shuf X, undef, <N, N-1,...0>) -> bitreverse (bitcast X)
2693 if (DestTy->isIntegerTy() && ShufElts.getKnownMinValue() % 2 == 0 &&
2694 Shuf->hasOneUse() && Shuf->isReverse()) {
2695 unsigned IntrinsicNum = 0;
2696 if (DL.isLegalInteger(DestTy->getScalarSizeInBits()) &&
2697 SrcTy->getScalarSizeInBits() == 8) {
2698 IntrinsicNum = Intrinsic::bswap;
2699 } else if (SrcTy->getScalarSizeInBits() == 1) {
2700 IntrinsicNum = Intrinsic::bitreverse;
2701 }
2702 if (IntrinsicNum != 0) {
2703 assert(ShufOp0->getType() == SrcTy && "Unexpected shuffle mask");
2704 assert(match(ShufOp1, m_Undef()) && "Unexpected shuffle op");
2705 Function *BswapOrBitreverse =
2706 Intrinsic::getDeclaration(CI.getModule(), IntrinsicNum, DestTy);
2707 Value *ScalarX = Builder.CreateBitCast(ShufOp0, DestTy);
2708 return CallInst::Create(BswapOrBitreverse, {ScalarX});
2709 }
2710 }
2711 }
2712
2713 // Handle the A->B->A cast, and there is an intervening PHI node.
2714 if (PHINode *PN = dyn_cast<PHINode>(Src))
2715 if (Instruction *I = optimizeBitCastFromPhi(CI, PN))
2716 return I;
2717
2718 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this))
2719 return I;
2720
2721 if (Instruction *I = foldBitCastBitwiseLogic(CI, Builder))
2722 return I;
2723
2724 if (Instruction *I = foldBitCastSelect(CI, Builder))
2725 return I;
2726
2727 return commonCastTransforms(CI);
2728 }
2729
visitAddrSpaceCast(AddrSpaceCastInst & CI)2730 Instruction *InstCombinerImpl::visitAddrSpaceCast(AddrSpaceCastInst &CI) {
2731 return commonCastTransforms(CI);
2732 }
2733