1	//===- InstCombineShifts.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 visitShl, visitLShr, and visitAShr functions.
10	//
11	//===----------------------------------------------------------------------===//
12
13	#include "InstCombineInternal.h"
14	#include "llvm/Analysis/InstructionSimplify.h"
15	#include "llvm/IR/IntrinsicInst.h"
16	#include "llvm/IR/PatternMatch.h"
17	#include "llvm/Transforms/InstCombine/InstCombiner.h"
18	using namespace llvm;
19	using namespace PatternMatch;
20
21	#define DEBUG_TYPE "instcombine"
22
23	bool canTryToConstantAddTwoShiftAmounts(Value *Sh0, Value *ShAmt0, Value *Sh1,
24	Value *ShAmt1) {
25	// We have two shift amounts from two different shifts. The types of those
26	// shift amounts may not match. If that's the case let's bailout now..
27	if (ShAmt0->getType() != ShAmt1->getType())
28	return false;
29
30	// As input, we have the following pattern:
31	// Sh0 (Sh1 X, Q), K
32	// We want to rewrite that as:
33	// Sh x, (Q+K) iff (Q+K) u< bitwidth(x)
34	// While we know that originally (Q+K) would not overflow
35	// (because 2 * (N-1) u<= iN -1), we have looked past extensions of
36	// shift amounts. so it may now overflow in smaller bitwidth.
37	// To ensure that does not happen, we need to ensure that the total maximal
38	// shift amount is still representable in that smaller bit width.
39	unsigned MaximalPossibleTotalShiftAmount =
40	(Sh0->getType()->getScalarSizeInBits() - 1) +
41	(Sh1->getType()->getScalarSizeInBits() - 1);
42	APInt MaximalRepresentableShiftAmount =
43	APInt::getAllOnes(numBits: ShAmt0->getType()->getScalarSizeInBits());
44	return MaximalRepresentableShiftAmount.uge(RHS: MaximalPossibleTotalShiftAmount);
45	}
46
47	// Given pattern:
48	// (x shiftopcode Q) shiftopcode K
49	// we should rewrite it as
50	// x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and
51	//
52	// This is valid for any shift, but they must be identical, and we must be
53	// careful in case we have (zext(Q)+zext(K)) and look past extensions,
54	// (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus.
55	//
56	// AnalyzeForSignBitExtraction indicates that we will only analyze whether this
57	// pattern has any 2 right-shifts that sum to 1 less than original bit width.
58	Value *InstCombinerImpl::reassociateShiftAmtsOfTwoSameDirectionShifts(
59	BinaryOperator *Sh0, const SimplifyQuery &SQ,
60	bool AnalyzeForSignBitExtraction) {
61	// Look for a shift of some instruction, ignore zext of shift amount if any.
62	Instruction *Sh0Op0;
63	Value *ShAmt0;
64	if (!match(V: Sh0,
65	P: m_Shift(L: m_Instruction(I&: Sh0Op0), R: m_ZExtOrSelf(Op: m_Value(V&: ShAmt0)))))
66	return nullptr;
67
68	// If there is a truncation between the two shifts, we must make note of it
69	// and look through it. The truncation imposes additional constraints on the
70	// transform.
71	Instruction *Sh1;
72	Value *Trunc = nullptr;
73	match(V: Sh0Op0,
74	P: m_CombineOr(L: m_CombineAnd(L: m_Trunc(Op: m_Instruction(I&: Sh1)), R: m_Value(V&: Trunc)),
75	R: m_Instruction(I&: Sh1)));
76
77	// Inner shift: (x shiftopcode ShAmt1)
78	// Like with other shift, ignore zext of shift amount if any.
79	Value *X, *ShAmt1;
80	if (!match(V: Sh1, P: m_Shift(L: m_Value(V&: X), R: m_ZExtOrSelf(Op: m_Value(V&: ShAmt1)))))
81	return nullptr;
82
83	// Verify that it would be safe to try to add those two shift amounts.
84	if (!canTryToConstantAddTwoShiftAmounts(Sh0, ShAmt0, Sh1, ShAmt1))
85	return nullptr;
86
87	// We are only looking for signbit extraction if we have two right shifts.
88	bool HadTwoRightShifts = match(V: Sh0, P: m_Shr(L: m_Value(), R: m_Value())) &&
89	match(V: Sh1, P: m_Shr(L: m_Value(), R: m_Value()));
90	// ... and if it's not two right-shifts, we know the answer already.
91	if (AnalyzeForSignBitExtraction && !HadTwoRightShifts)
92	return nullptr;
93
94	// The shift opcodes must be identical, unless we are just checking whether
95	// this pattern can be interpreted as a sign-bit-extraction.
96	Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode();
97	bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode();
98	if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction)
99	return nullptr;
100
101	// If we saw truncation, we'll need to produce extra instruction,
102	// and for that one of the operands of the shift must be one-use,
103	// unless of course we don't actually plan to produce any instructions here.
104	if (Trunc && !AnalyzeForSignBitExtraction &&
105	!match(V: Sh0, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value())))
106	return nullptr;
107
108	// Can we fold (ShAmt0+ShAmt1) ?
109	auto *NewShAmt = dyn_cast_or_null<Constant>(
110	Val: simplifyAddInst(LHS: ShAmt0, RHS: ShAmt1, /*isNSW=*/IsNSW: false, /*isNUW=*/IsNUW: false,
111	Q: SQ.getWithInstruction(I: Sh0)));
112	if (!NewShAmt)
113	return nullptr; // Did not simplify.
114	unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits();
115	unsigned XBitWidth = X->getType()->getScalarSizeInBits();
116	// Is the new shift amount smaller than the bit width of inner/new shift?
117	if (!match(V: NewShAmt, P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_ULT,
118	Threshold: APInt(NewShAmtBitWidth, XBitWidth))))
119	return nullptr; // FIXME: could perform constant-folding.
120
121	// If there was a truncation, and we have a right-shift, we can only fold if
122	// we are left with the original sign bit. Likewise, if we were just checking
123	// that this is a sighbit extraction, this is the place to check it.
124	// FIXME: zero shift amount is also legal here, but we can't *easily* check
125	// more than one predicate so it's not really worth it.
126	if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) {
127	// If it's not a sign bit extraction, then we're done.
128	if (!match(V: NewShAmt,
129	P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_EQ,
130	Threshold: APInt(NewShAmtBitWidth, XBitWidth - 1))))
131	return nullptr;
132	// If it is, and that was the question, return the base value.
133	if (AnalyzeForSignBitExtraction)
134	return X;
135	}
136
137	assert(IdenticalShOpcodes && "Should not get here with different shifts.");
138
139	if (NewShAmt->getType() != X->getType()) {
140	NewShAmt = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: NewShAmt,
141	DestTy: X->getType(), DL: SQ.DL);
142	if (!NewShAmt)
143	return nullptr;
144	}
145
146	// All good, we can do this fold.
147	BinaryOperator *NewShift = BinaryOperator::Create(Op: ShiftOpcode, S1: X, S2: NewShAmt);
148
149	// The flags can only be propagated if there wasn't a trunc.
150	if (!Trunc) {
151	// If the pattern did not involve trunc, and both of the original shifts
152	// had the same flag set, preserve the flag.
153	if (ShiftOpcode == Instruction::BinaryOps::Shl) {
154	NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() &&
155	Sh1->hasNoUnsignedWrap());
156	NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() &&
157	Sh1->hasNoSignedWrap());
158	} else {
159	NewShift->setIsExact(Sh0->isExact() && Sh1->isExact());
160	}
161	}
162
163	Instruction *Ret = NewShift;
164	if (Trunc) {
165	Builder.Insert(I: NewShift);
166	Ret = CastInst::Create(Instruction::Trunc, S: NewShift, Ty: Sh0->getType());
167	}
168
169	return Ret;
170	}
171
172	// If we have some pattern that leaves only some low bits set, and then performs
173	// left-shift of those bits, if none of the bits that are left after the final
174	// shift are modified by the mask, we can omit the mask.
175	//
176	// There are many variants to this pattern:
177	// a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt
178	// b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt
179	// c) (x & (-1 l>> MaskShAmt)) << ShiftShAmt
180	// d) (x & ((-1 << MaskShAmt) l>> MaskShAmt)) << ShiftShAmt
181	// e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt
182	// f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt
183	// All these patterns can be simplified to just:
184	// x << ShiftShAmt
185	// iff:
186	// a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x)
187	// c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt)
188	static Instruction *
189	dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift,
190	const SimplifyQuery &Q,
191	InstCombiner::BuilderTy &Builder) {
192	assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&
193	"The input must be 'shl'!");
194
195	Value *Masked, *ShiftShAmt;
196	match(V: OuterShift,
197	P: m_Shift(L: m_Value(V&: Masked), R: m_ZExtOrSelf(Op: m_Value(V&: ShiftShAmt))));
198
199	// *If* there is a truncation between an outer shift and a possibly-mask,
200	// then said truncation *must* be one-use, else we can't perform the fold.
201	Value *Trunc;
202	if (match(V: Masked, P: m_CombineAnd(L: m_Trunc(Op: m_Value(V&: Masked)), R: m_Value(V&: Trunc))) &&
203	!Trunc->hasOneUse())
204	return nullptr;
205
206	Type *NarrowestTy = OuterShift->getType();
207	Type *WidestTy = Masked->getType();
208	bool HadTrunc = WidestTy != NarrowestTy;
209
210	// The mask must be computed in a type twice as wide to ensure
211	// that no bits are lost if the sum-of-shifts is wider than the base type.
212	Type *ExtendedTy = WidestTy->getExtendedType();
213
214	Value *MaskShAmt;
215
216	// ((1 << MaskShAmt) - 1)
217	auto MaskA = m_Add(L: m_Shl(L: m_One(), R: m_Value(V&: MaskShAmt)), R: m_AllOnes());
218	// (~(-1 << maskNbits))
219	auto MaskB = m_Xor(L: m_Shl(L: m_AllOnes(), R: m_Value(V&: MaskShAmt)), R: m_AllOnes());
220	// (-1 l>> MaskShAmt)
221	auto MaskC = m_LShr(L: m_AllOnes(), R: m_Value(V&: MaskShAmt));
222	// ((-1 << MaskShAmt) l>> MaskShAmt)
223	auto MaskD =
224	m_LShr(L: m_Shl(L: m_AllOnes(), R: m_Value(V&: MaskShAmt)), R: m_Deferred(V: MaskShAmt));
225
226	Value *X;
227	Constant *NewMask;
228
229	if (match(V: Masked, P: m_c_And(L: m_CombineOr(L: MaskA, R: MaskB), R: m_Value(V&: X)))) {
230	// Peek through an optional zext of the shift amount.
231	match(V: MaskShAmt, P: m_ZExtOrSelf(Op: m_Value(V&: MaskShAmt)));
232
233	// Verify that it would be safe to try to add those two shift amounts.
234	if (!canTryToConstantAddTwoShiftAmounts(Sh0: OuterShift, ShAmt0: ShiftShAmt, Sh1: Masked,
235	ShAmt1: MaskShAmt))
236	return nullptr;
237
238	// Can we simplify (MaskShAmt+ShiftShAmt) ?
239	auto *SumOfShAmts = dyn_cast_or_null<Constant>(Val: simplifyAddInst(
240	LHS: MaskShAmt, RHS: ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
241	if (!SumOfShAmts)
242	return nullptr; // Did not simplify.
243	// In this pattern SumOfShAmts correlates with the number of low bits
244	// that shall remain in the root value (OuterShift).
245
246	// An extend of an undef value becomes zero because the high bits are never
247	// completely unknown. Replace the `undef` shift amounts with final
248	// shift bitwidth to ensure that the value remains undef when creating the
249	// subsequent shift op.
250	SumOfShAmts = Constant::replaceUndefsWith(
251	C: SumOfShAmts, Replacement: ConstantInt::get(Ty: SumOfShAmts->getType()->getScalarType(),
252	V: ExtendedTy->getScalarSizeInBits()));
253	auto *ExtendedSumOfShAmts = ConstantFoldCastOperand(
254	Opcode: Instruction::ZExt, C: SumOfShAmts, DestTy: ExtendedTy, DL: Q.DL);
255	if (!ExtendedSumOfShAmts)
256	return nullptr;
257
258	// And compute the mask as usual: ~(-1 << (SumOfShAmts))
259	auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(Ty: ExtendedTy);
260	auto *ExtendedInvertedMask =
261	ConstantExpr::getShl(C1: ExtendedAllOnes, C2: ExtendedSumOfShAmts);
262	NewMask = ConstantExpr::getNot(C: ExtendedInvertedMask);
263	} else if (match(V: Masked, P: m_c_And(L: m_CombineOr(L: MaskC, R: MaskD), R: m_Value(V&: X))) ||
264	match(V: Masked, P: m_Shr(L: m_Shl(L: m_Value(V&: X), R: m_Value(V&: MaskShAmt)),
265	R: m_Deferred(V: MaskShAmt)))) {
266	// Peek through an optional zext of the shift amount.
267	match(V: MaskShAmt, P: m_ZExtOrSelf(Op: m_Value(V&: MaskShAmt)));
268
269	// Verify that it would be safe to try to add those two shift amounts.
270	if (!canTryToConstantAddTwoShiftAmounts(Sh0: OuterShift, ShAmt0: ShiftShAmt, Sh1: Masked,
271	ShAmt1: MaskShAmt))
272	return nullptr;
273
274	// Can we simplify (ShiftShAmt-MaskShAmt) ?
275	auto *ShAmtsDiff = dyn_cast_or_null<Constant>(Val: simplifySubInst(
276	LHS: ShiftShAmt, RHS: MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
277	if (!ShAmtsDiff)
278	return nullptr; // Did not simplify.
279	// In this pattern ShAmtsDiff correlates with the number of high bits that
280	// shall be unset in the root value (OuterShift).
281
282	// An extend of an undef value becomes zero because the high bits are never
283	// completely unknown. Replace the `undef` shift amounts with negated
284	// bitwidth of innermost shift to ensure that the value remains undef when
285	// creating the subsequent shift op.
286	unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits();
287	ShAmtsDiff = Constant::replaceUndefsWith(
288	C: ShAmtsDiff, Replacement: ConstantInt::get(Ty: ShAmtsDiff->getType()->getScalarType(),
289	V: -WidestTyBitWidth));
290	auto *ExtendedNumHighBitsToClear = ConstantFoldCastOperand(
291	Opcode: Instruction::ZExt,
292	C: ConstantExpr::getSub(C1: ConstantInt::get(Ty: ShAmtsDiff->getType(),
293	V: WidestTyBitWidth,
294	/*isSigned=*/IsSigned: false),
295	C2: ShAmtsDiff),
296	DestTy: ExtendedTy, DL: Q.DL);
297	if (!ExtendedNumHighBitsToClear)
298	return nullptr;
299
300	// And compute the mask as usual: (-1 l>> (NumHighBitsToClear))
301	auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(Ty: ExtendedTy);
302	NewMask = ConstantFoldBinaryOpOperands(Opcode: Instruction::LShr, LHS: ExtendedAllOnes,
303	RHS: ExtendedNumHighBitsToClear, DL: Q.DL);
304	if (!NewMask)
305	return nullptr;
306	} else
307	return nullptr; // Don't know anything about this pattern.
308
309	NewMask = ConstantExpr::getTrunc(C: NewMask, Ty: NarrowestTy);
310
311	// Does this mask has any unset bits? If not then we can just not apply it.
312	bool NeedMask = !match(V: NewMask, P: m_AllOnes());
313
314	// If we need to apply a mask, there are several more restrictions we have.
315	if (NeedMask) {
316	// The old masking instruction must go away.
317	if (!Masked->hasOneUse())
318	return nullptr;
319	// The original "masking" instruction must not have been`ashr`.
320	if (match(V: Masked, P: m_AShr(L: m_Value(), R: m_Value())))
321	return nullptr;
322	}
323
324	// If we need to apply truncation, let's do it first, since we can.
325	// We have already ensured that the old truncation will go away.
326	if (HadTrunc)
327	X = Builder.CreateTrunc(V: X, DestTy: NarrowestTy);
328
329	// No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits.
330	// We didn't change the Type of this outermost shift, so we can just do it.
331	auto *NewShift = BinaryOperator::Create(Op: OuterShift->getOpcode(), S1: X,
332	S2: OuterShift->getOperand(i_nocapture: 1));
333	if (!NeedMask)
334	return NewShift;
335
336	Builder.Insert(I: NewShift);
337	return BinaryOperator::Create(Op: Instruction::And, S1: NewShift, S2: NewMask);
338	}
339
340	/// If we have a shift-by-constant of a bin op (bitwise logic op or add/sub w/
341	/// shl) that itself has a shift-by-constant operand with identical opcode, we
342	/// may be able to convert that into 2 independent shifts followed by the logic
343	/// op. This eliminates a use of an intermediate value (reduces dependency
344	/// chain).
345	static Instruction *foldShiftOfShiftedBinOp(BinaryOperator &I,
346	InstCombiner::BuilderTy &Builder) {
347	assert(I.isShift() && "Expected a shift as input");
348	auto *BinInst = dyn_cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 0));
349	if (!BinInst ||
350	(!BinInst->isBitwiseLogicOp() &&
351	BinInst->getOpcode() != Instruction::Add &&
352	BinInst->getOpcode() != Instruction::Sub) ||
353	!BinInst->hasOneUse())
354	return nullptr;
355
356	Constant *C0, *C1;
357	if (!match(V: I.getOperand(i_nocapture: 1), P: m_Constant(C&: C1)))
358	return nullptr;
359
360	Instruction::BinaryOps ShiftOpcode = I.getOpcode();
361	// Transform for add/sub only works with shl.
362	if ((BinInst->getOpcode() == Instruction::Add ||
363	BinInst->getOpcode() == Instruction::Sub) &&
364	ShiftOpcode != Instruction::Shl)
365	return nullptr;
366
367	Type *Ty = I.getType();
368
369	// Find a matching shift by constant. The fold is not valid if the sum
370	// of the shift values equals or exceeds bitwidth.
371	Value *X, *Y;
372	auto matchFirstShift = [&](Value *V, Value *W) {
373	unsigned Size = Ty->getScalarSizeInBits();
374	APInt Threshold(Size, Size);
375	return match(V, P: m_BinOp(Opcode: ShiftOpcode, L: m_Value(V&: X), R: m_Constant(C&: C0))) &&
376	(V->hasOneUse() || match(V: W, P: m_ImmConstant())) &&
377	match(V: ConstantExpr::getAdd(C1: C0, C2: C1),
378	P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_ULT, Threshold));
379	};
380
381	// Logic ops and Add are commutative, so check each operand for a match. Sub
382	// is not so we cannot reoder if we match operand(1) and need to keep the
383	// operands in their original positions.
384	bool FirstShiftIsOp1 = false;
385	if (matchFirstShift(BinInst->getOperand(i_nocapture: 0), BinInst->getOperand(i_nocapture: 1)))
386	Y = BinInst->getOperand(i_nocapture: 1);
387	else if (matchFirstShift(BinInst->getOperand(i_nocapture: 1), BinInst->getOperand(i_nocapture: 0))) {
388	Y = BinInst->getOperand(i_nocapture: 0);
389	FirstShiftIsOp1 = BinInst->getOpcode() == Instruction::Sub;
390	} else
391	return nullptr;
392
393	// shift (binop (shift X, C0), Y), C1 -> binop (shift X, C0+C1), (shift Y, C1)
394	Constant *ShiftSumC = ConstantExpr::getAdd(C1: C0, C2: C1);
395	Value *NewShift1 = Builder.CreateBinOp(Opc: ShiftOpcode, LHS: X, RHS: ShiftSumC);
396	Value *NewShift2 = Builder.CreateBinOp(Opc: ShiftOpcode, LHS: Y, RHS: C1);
397	Value *Op1 = FirstShiftIsOp1 ? NewShift2 : NewShift1;
398	Value *Op2 = FirstShiftIsOp1 ? NewShift1 : NewShift2;
399	return BinaryOperator::Create(Op: BinInst->getOpcode(), S1: Op1, S2: Op2);
400	}
401
402	Instruction *InstCombinerImpl::commonShiftTransforms(BinaryOperator &I) {
403	if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
404	return Phi;
405
406	Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
407	assert(Op0->getType() == Op1->getType());
408	Type *Ty = I.getType();
409
410	// If the shift amount is a one-use `sext`, we can demote it to `zext`.
411	Value *Y;
412	if (match(V: Op1, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: Y))))) {
413	Value *NewExt = Builder.CreateZExt(V: Y, DestTy: Ty, Name: Op1->getName());
414	return BinaryOperator::Create(Op: I.getOpcode(), S1: Op0, S2: NewExt);
415	}
416
417	// See if we can fold away this shift.
418	if (SimplifyDemandedInstructionBits(Inst&: I))
419	return &I;
420
421	// Try to fold constant and into select arguments.
422	if (isa<Constant>(Val: Op0))
423	if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1))
424	if (Instruction *R = FoldOpIntoSelect(Op&: I, SI))
425	return R;
426
427	if (Constant *CUI = dyn_cast<Constant>(Val: Op1))
428	if (Instruction *Res = FoldShiftByConstant(Op0, Op1: CUI, I))
429	return Res;
430
431	if (auto *NewShift = cast_or_null<Instruction>(
432	Val: reassociateShiftAmtsOfTwoSameDirectionShifts(Sh0: &I, SQ)))
433	return NewShift;
434
435	// Pre-shift a constant shifted by a variable amount with constant offset:
436	// C shift (A add nuw C1) --> (C shift C1) shift A
437	Value *A;
438	Constant *C, *C1;
439	if (match(V: Op0, P: m_Constant(C)) &&
440	match(V: Op1, P: m_NUWAddLike(L: m_Value(V&: A), R: m_Constant(C&: C1)))) {
441	Value *NewC = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: C, RHS: C1);
442	BinaryOperator *NewShiftOp = BinaryOperator::Create(Op: I.getOpcode(), S1: NewC, S2: A);
443	if (I.getOpcode() == Instruction::Shl) {
444	NewShiftOp->setHasNoSignedWrap(I.hasNoSignedWrap());
445	NewShiftOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
446	} else {
447	NewShiftOp->setIsExact(I.isExact());
448	}
449	return NewShiftOp;
450	}
451
452	unsigned BitWidth = Ty->getScalarSizeInBits();
453
454	const APInt *AC, *AddC;
455	// Try to pre-shift a constant shifted by a variable amount added with a
456	// negative number:
457	// C << (X - AddC) --> (C >> AddC) << X
458	// and
459	// C >> (X - AddC) --> (C << AddC) >> X
460	if (match(V: Op0, P: m_APInt(Res&: AC)) && match(V: Op1, P: m_Add(L: m_Value(V&: A), R: m_APInt(Res&: AddC))) &&
461	AddC->isNegative() && (-*AddC).ult(RHS: BitWidth)) {
462	assert(!AC->isZero() && "Expected simplify of shifted zero");
463	unsigned PosOffset = (-*AddC).getZExtValue();
464
465	auto isSuitableForPreShift = [PosOffset, &I, AC]() {
466	switch (I.getOpcode()) {
467	default:
468	return false;
469	case Instruction::Shl:
470	return (I.hasNoSignedWrap() || I.hasNoUnsignedWrap()) &&
471	AC->eq(RHS: AC->lshr(shiftAmt: PosOffset).shl(shiftAmt: PosOffset));
472	case Instruction::LShr:
473	return I.isExact() && AC->eq(RHS: AC->shl(shiftAmt: PosOffset).lshr(shiftAmt: PosOffset));
474	case Instruction::AShr:
475	return I.isExact() && AC->eq(RHS: AC->shl(shiftAmt: PosOffset).ashr(ShiftAmt: PosOffset));
476	}
477	};
478	if (isSuitableForPreShift()) {
479	Constant *NewC = ConstantInt::get(Ty, V: I.getOpcode() == Instruction::Shl
480	? AC->lshr(shiftAmt: PosOffset)
481	: AC->shl(shiftAmt: PosOffset));
482	BinaryOperator *NewShiftOp =
483	BinaryOperator::Create(Op: I.getOpcode(), S1: NewC, S2: A);
484	if (I.getOpcode() == Instruction::Shl) {
485	NewShiftOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
486	} else {
487	NewShiftOp->setIsExact();
488	}
489	return NewShiftOp;
490	}
491	}
492
493	// X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2.
494	// Because shifts by negative values (which could occur if A were negative)
495	// are undefined.
496	if (Op1->hasOneUse() && match(V: Op1, P: m_SRem(L: m_Value(V&: A), R: m_Constant(C))) &&
497	match(V: C, P: m_Power2())) {
498	// FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
499	// demand the sign bit (and many others) here??
500	Constant *Mask = ConstantExpr::getSub(C1: C, C2: ConstantInt::get(Ty, V: 1));
501	Value *Rem = Builder.CreateAnd(LHS: A, RHS: Mask, Name: Op1->getName());
502	return replaceOperand(I, OpNum: 1, V: Rem);
503	}
504
505	if (Instruction *Logic = foldShiftOfShiftedBinOp(I, Builder))
506	return Logic;
507
508	if (match(V: Op1, P: m_Or(L: m_Value(), R: m_SpecificInt(V: BitWidth - 1))))
509	return replaceOperand(I, OpNum: 1, V: ConstantInt::get(Ty, V: BitWidth - 1));
510
511	return nullptr;
512	}
513
514	/// Return true if we can simplify two logical (either left or right) shifts
515	/// that have constant shift amounts: OuterShift (InnerShift X, C1), C2.
516	static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl,
517	Instruction *InnerShift,
518	InstCombinerImpl &IC, Instruction *CxtI) {
519	assert(InnerShift->isLogicalShift() && "Unexpected instruction type");
520
521	// We need constant scalar or constant splat shifts.
522	const APInt *InnerShiftConst;
523	if (!match(V: InnerShift->getOperand(i: 1), P: m_APInt(Res&: InnerShiftConst)))
524	return false;
525
526	// Two logical shifts in the same direction:
527	// shl (shl X, C1), C2 --> shl X, C1 + C2
528	// lshr (lshr X, C1), C2 --> lshr X, C1 + C2
529	bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
530	if (IsInnerShl == IsOuterShl)
531	return true;
532
533	// Equal shift amounts in opposite directions become bitwise 'and':
534	// lshr (shl X, C), C --> and X, C'
535	// shl (lshr X, C), C --> and X, C'
536	if (*InnerShiftConst == OuterShAmt)
537	return true;
538
539	// If the 2nd shift is bigger than the 1st, we can fold:
540	// lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3
541	// shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3
542	// but it isn't profitable unless we know the and'd out bits are already zero.
543	// Also, check that the inner shift is valid (less than the type width) or
544	// we'll crash trying to produce the bit mask for the 'and'.
545	unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits();
546	if (InnerShiftConst->ugt(RHS: OuterShAmt) && InnerShiftConst->ult(RHS: TypeWidth)) {
547	unsigned InnerShAmt = InnerShiftConst->getZExtValue();
548	unsigned MaskShift =
549	IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt;
550	APInt Mask = APInt::getLowBitsSet(numBits: TypeWidth, loBitsSet: OuterShAmt) << MaskShift;
551	if (IC.MaskedValueIsZero(V: InnerShift->getOperand(i: 0), Mask, Depth: 0, CxtI))
552	return true;
553	}
554
555	return false;
556	}
557
558	/// See if we can compute the specified value, but shifted logically to the left
559	/// or right by some number of bits. This should return true if the expression
560	/// can be computed for the same cost as the current expression tree. This is
561	/// used to eliminate extraneous shifting from things like:
562	/// %C = shl i128 %A, 64
563	/// %D = shl i128 %B, 96
564	/// %E = or i128 %C, %D
565	/// %F = lshr i128 %E, 64
566	/// where the client will ask if E can be computed shifted right by 64-bits. If
567	/// this succeeds, getShiftedValue() will be called to produce the value.
568	static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift,
569	InstCombinerImpl &IC, Instruction *CxtI) {
570	// We can always evaluate immediate constants.
571	if (match(V, P: m_ImmConstant()))
572	return true;
573
574	Instruction *I = dyn_cast<Instruction>(Val: V);
575	if (!I) return false;
576
577	// We can't mutate something that has multiple uses: doing so would
578	// require duplicating the instruction in general, which isn't profitable.
579	if (!I->hasOneUse()) return false;
580
581	switch (I->getOpcode()) {
582	default: return false;
583	case Instruction::And:
584	case Instruction::Or:
585	case Instruction::Xor:
586	// Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
587	return canEvaluateShifted(V: I->getOperand(i: 0), NumBits, IsLeftShift, IC, CxtI: I) &&
588	canEvaluateShifted(V: I->getOperand(i: 1), NumBits, IsLeftShift, IC, CxtI: I);
589
590	case Instruction::Shl:
591	case Instruction::LShr:
592	return canEvaluateShiftedShift(OuterShAmt: NumBits, IsOuterShl: IsLeftShift, InnerShift: I, IC, CxtI);
593
594	case Instruction::Select: {
595	SelectInst *SI = cast<SelectInst>(Val: I);
596	Value *TrueVal = SI->getTrueValue();
597	Value *FalseVal = SI->getFalseValue();
598	return canEvaluateShifted(V: TrueVal, NumBits, IsLeftShift, IC, CxtI: SI) &&
599	canEvaluateShifted(V: FalseVal, NumBits, IsLeftShift, IC, CxtI: SI);
600	}
601	case Instruction::PHI: {
602	// We can change a phi if we can change all operands. Note that we never
603	// get into trouble with cyclic PHIs here because we only consider
604	// instructions with a single use.
605	PHINode *PN = cast<PHINode>(Val: I);
606	for (Value *IncValue : PN->incoming_values())
607	if (!canEvaluateShifted(V: IncValue, NumBits, IsLeftShift, IC, CxtI: PN))
608	return false;
609	return true;
610	}
611	case Instruction::Mul: {
612	const APInt *MulConst;
613	// We can fold (shr (mul X, -(1 << C)), C) -> (and (neg X), C`)
614	return !IsLeftShift && match(V: I->getOperand(i: 1), P: m_APInt(Res&: MulConst)) &&
615	MulConst->isNegatedPowerOf2() && MulConst->countr_zero() == NumBits;
616	}
617	}
618	}
619
620	/// Fold OuterShift (InnerShift X, C1), C2.
621	/// See canEvaluateShiftedShift() for the constraints on these instructions.
622	static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt,
623	bool IsOuterShl,
624	InstCombiner::BuilderTy &Builder) {
625	bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
626	Type *ShType = InnerShift->getType();
627	unsigned TypeWidth = ShType->getScalarSizeInBits();
628
629	// We only accept shifts-by-a-constant in canEvaluateShifted().
630	const APInt *C1;
631	match(V: InnerShift->getOperand(i_nocapture: 1), P: m_APInt(Res&: C1));
632	unsigned InnerShAmt = C1->getZExtValue();
633
634	// Change the shift amount and clear the appropriate IR flags.
635	auto NewInnerShift = [&](unsigned ShAmt) {
636	InnerShift->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: ShType, V: ShAmt));
637	if (IsInnerShl) {
638	InnerShift->setHasNoUnsignedWrap(false);
639	InnerShift->setHasNoSignedWrap(false);
640	} else {
641	InnerShift->setIsExact(false);
642	}
643	return InnerShift;
644	};
645
646	// Two logical shifts in the same direction:
647	// shl (shl X, C1), C2 --> shl X, C1 + C2
648	// lshr (lshr X, C1), C2 --> lshr X, C1 + C2
649	if (IsInnerShl == IsOuterShl) {
650	// If this is an oversized composite shift, then unsigned shifts get 0.
651	if (InnerShAmt + OuterShAmt >= TypeWidth)
652	return Constant::getNullValue(Ty: ShType);
653
654	return NewInnerShift(InnerShAmt + OuterShAmt);
655	}
656
657	// Equal shift amounts in opposite directions become bitwise 'and':
658	// lshr (shl X, C), C --> and X, C'
659	// shl (lshr X, C), C --> and X, C'
660	if (InnerShAmt == OuterShAmt) {
661	APInt Mask = IsInnerShl
662	? APInt::getLowBitsSet(numBits: TypeWidth, loBitsSet: TypeWidth - OuterShAmt)
663	: APInt::getHighBitsSet(numBits: TypeWidth, hiBitsSet: TypeWidth - OuterShAmt);
664	Value *And = Builder.CreateAnd(LHS: InnerShift->getOperand(i_nocapture: 0),
665	RHS: ConstantInt::get(Ty: ShType, V: Mask));
666	if (auto *AndI = dyn_cast<Instruction>(Val: And)) {
667	AndI->moveBefore(MovePos: InnerShift);
668	AndI->takeName(V: InnerShift);
669	}
670	return And;
671	}
672
673	assert(InnerShAmt > OuterShAmt &&
674	"Unexpected opposite direction logical shift pair");
675
676	// In general, we would need an 'and' for this transform, but
677	// canEvaluateShiftedShift() guarantees that the masked-off bits are not used.
678	// lshr (shl X, C1), C2 --> shl X, C1 - C2
679	// shl (lshr X, C1), C2 --> lshr X, C1 - C2
680	return NewInnerShift(InnerShAmt - OuterShAmt);
681	}
682
683	/// When canEvaluateShifted() returns true for an expression, this function
684	/// inserts the new computation that produces the shifted value.
685	static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
686	InstCombinerImpl &IC, const DataLayout &DL) {
687	// We can always evaluate constants shifted.
688	if (Constant *C = dyn_cast<Constant>(Val: V)) {
689	if (isLeftShift)
690	return IC.Builder.CreateShl(LHS: C, RHS: NumBits);
691	else
692	return IC.Builder.CreateLShr(LHS: C, RHS: NumBits);
693	}
694
695	Instruction *I = cast<Instruction>(Val: V);
696	IC.addToWorklist(I);
697
698	switch (I->getOpcode()) {
699	default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
700	case Instruction::And:
701	case Instruction::Or:
702	case Instruction::Xor:
703	// Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
704	I->setOperand(
705	i: 0, Val: getShiftedValue(V: I->getOperand(i: 0), NumBits, isLeftShift, IC, DL));
706	I->setOperand(
707	i: 1, Val: getShiftedValue(V: I->getOperand(i: 1), NumBits, isLeftShift, IC, DL));
708	return I;
709
710	case Instruction::Shl:
711	case Instruction::LShr:
712	return foldShiftedShift(InnerShift: cast<BinaryOperator>(Val: I), OuterShAmt: NumBits, IsOuterShl: isLeftShift,
713	Builder&: IC.Builder);
714
715	case Instruction::Select:
716	I->setOperand(
717	i: 1, Val: getShiftedValue(V: I->getOperand(i: 1), NumBits, isLeftShift, IC, DL));
718	I->setOperand(
719	i: 2, Val: getShiftedValue(V: I->getOperand(i: 2), NumBits, isLeftShift, IC, DL));
720	return I;
721	case Instruction::PHI: {
722	// We can change a phi if we can change all operands. Note that we never
723	// get into trouble with cyclic PHIs here because we only consider
724	// instructions with a single use.
725	PHINode *PN = cast<PHINode>(Val: I);
726	for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
727	PN->setIncomingValue(i, V: getShiftedValue(V: PN->getIncomingValue(i), NumBits,
728	isLeftShift, IC, DL));
729	return PN;
730	}
731	case Instruction::Mul: {
732	assert(!isLeftShift && "Unexpected shift direction!");
733	auto *Neg = BinaryOperator::CreateNeg(Op: I->getOperand(i: 0));
734	IC.InsertNewInstWith(New: Neg, Old: I->getIterator());
735	unsigned TypeWidth = I->getType()->getScalarSizeInBits();
736	APInt Mask = APInt::getLowBitsSet(numBits: TypeWidth, loBitsSet: TypeWidth - NumBits);
737	auto *And = BinaryOperator::CreateAnd(V1: Neg,
738	V2: ConstantInt::get(Ty: I->getType(), V: Mask));
739	And->takeName(V: I);
740	return IC.InsertNewInstWith(New: And, Old: I->getIterator());
741	}
742	}
743	}
744
745	// If this is a bitwise operator or add with a constant RHS we might be able
746	// to pull it through a shift.
747	static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift,
748	BinaryOperator *BO) {
749	switch (BO->getOpcode()) {
750	default:
751	return false; // Do not perform transform!
752	case Instruction::Add:
753	return Shift.getOpcode() == Instruction::Shl;
754	case Instruction::Or:
755	case Instruction::And:
756	return true;
757	case Instruction::Xor:
758	// Do not change a 'not' of logical shift because that would create a normal
759	// 'xor'. The 'not' is likely better for analysis, SCEV, and codegen.
760	return !(Shift.isLogicalShift() && match(V: BO, P: m_Not(V: m_Value())));
761	}
762	}
763
764	Instruction *InstCombinerImpl::FoldShiftByConstant(Value *Op0, Constant *C1,
765	BinaryOperator &I) {
766	// (C2 << X) << C1 --> (C2 << C1) << X
767	// (C2 >> X) >> C1 --> (C2 >> C1) >> X
768	Constant *C2;
769	Value *X;
770	if (match(V: Op0, P: m_BinOp(Opcode: I.getOpcode(), L: m_ImmConstant(C&: C2), R: m_Value(V&: X))))
771	return BinaryOperator::Create(
772	Op: I.getOpcode(), S1: Builder.CreateBinOp(Opc: I.getOpcode(), LHS: C2, RHS: C1), S2: X);
773
774	bool IsLeftShift = I.getOpcode() == Instruction::Shl;
775	Type *Ty = I.getType();
776	unsigned TypeBits = Ty->getScalarSizeInBits();
777
778	// (X / +DivC) >> (Width - 1) --> ext (X <= -DivC)
779	// (X / -DivC) >> (Width - 1) --> ext (X >= +DivC)
780	const APInt *DivC;
781	if (!IsLeftShift && match(V: C1, P: m_SpecificIntAllowPoison(V: TypeBits - 1)) &&
782	match(V: Op0, P: m_SDiv(L: m_Value(V&: X), R: m_APInt(Res&: DivC))) && !DivC->isZero() &&
783	!DivC->isMinSignedValue()) {
784	Constant *NegDivC = ConstantInt::get(Ty, V: -(*DivC));
785	ICmpInst::Predicate Pred =
786	DivC->isNegative() ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SLE;
787	Value *Cmp = Builder.CreateICmp(P: Pred, LHS: X, RHS: NegDivC);
788	auto ExtOpcode = (I.getOpcode() == Instruction::AShr) ? Instruction::SExt
789	: Instruction::ZExt;
790	return CastInst::Create(ExtOpcode, S: Cmp, Ty);
791	}
792
793	const APInt *Op1C;
794	if (!match(V: C1, P: m_APInt(Res&: Op1C)))
795	return nullptr;
796
797	assert(!Op1C->uge(TypeBits) &&
798	"Shift over the type width should have been removed already");
799
800	// See if we can propagate this shift into the input, this covers the trivial
801	// cast of lshr(shl(x,c1),c2) as well as other more complex cases.
802	if (I.getOpcode() != Instruction::AShr &&
803	canEvaluateShifted(V: Op0, NumBits: Op1C->getZExtValue(), IsLeftShift, IC&: *this, CxtI: &I)) {
804	LLVM_DEBUG(
805	dbgs() << "ICE: GetShiftedValue propagating shift through expression"
806	" to eliminate shift:\n IN: "
807	<< *Op0 << "\n SH: " << I << "\n");
808
809	return replaceInstUsesWith(
810	I, V: getShiftedValue(V: Op0, NumBits: Op1C->getZExtValue(), isLeftShift: IsLeftShift, IC&: *this, DL));
811	}
812
813	if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I))
814	return FoldedShift;
815
816	if (!Op0->hasOneUse())
817	return nullptr;
818
819	if (auto *Op0BO = dyn_cast<BinaryOperator>(Val: Op0)) {
820	// If the operand is a bitwise operator with a constant RHS, and the
821	// shift is the only use, we can pull it out of the shift.
822	const APInt *Op0C;
823	if (match(V: Op0BO->getOperand(i_nocapture: 1), P: m_APInt(Res&: Op0C))) {
824	if (canShiftBinOpWithConstantRHS(Shift&: I, BO: Op0BO)) {
825	Value *NewRHS =
826	Builder.CreateBinOp(Opc: I.getOpcode(), LHS: Op0BO->getOperand(i_nocapture: 1), RHS: C1);
827
828	Value *NewShift =
829	Builder.CreateBinOp(Opc: I.getOpcode(), LHS: Op0BO->getOperand(i_nocapture: 0), RHS: C1);
830	NewShift->takeName(V: Op0BO);
831
832	return BinaryOperator::Create(Op: Op0BO->getOpcode(), S1: NewShift, S2: NewRHS);
833	}
834	}
835	}
836
837	// If we have a select that conditionally executes some binary operator,
838	// see if we can pull it the select and operator through the shift.
839	//
840	// For example, turning:
841	// shl (select C, (add X, C1), X), C2
842	// Into:
843	// Y = shl X, C2
844	// select C, (add Y, C1 << C2), Y
845	Value *Cond;
846	BinaryOperator *TBO;
847	Value *FalseVal;
848	if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_OneUse(SubPattern: m_BinOp(I&: TBO)),
849	R: m_Value(V&: FalseVal)))) {
850	const APInt *C;
851	if (!isa<Constant>(Val: FalseVal) && TBO->getOperand(i_nocapture: 0) == FalseVal &&
852	match(V: TBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) &&
853	canShiftBinOpWithConstantRHS(Shift&: I, BO: TBO)) {
854	Value *NewRHS =
855	Builder.CreateBinOp(Opc: I.getOpcode(), LHS: TBO->getOperand(i_nocapture: 1), RHS: C1);
856
857	Value *NewShift = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: FalseVal, RHS: C1);
858	Value *NewOp = Builder.CreateBinOp(Opc: TBO->getOpcode(), LHS: NewShift, RHS: NewRHS);
859	return SelectInst::Create(C: Cond, S1: NewOp, S2: NewShift);
860	}
861	}
862
863	BinaryOperator *FBO;
864	Value *TrueVal;
865	if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TrueVal),
866	R: m_OneUse(SubPattern: m_BinOp(I&: FBO))))) {
867	const APInt *C;
868	if (!isa<Constant>(Val: TrueVal) && FBO->getOperand(i_nocapture: 0) == TrueVal &&
869	match(V: FBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) &&
870	canShiftBinOpWithConstantRHS(Shift&: I, BO: FBO)) {
871	Value *NewRHS =
872	Builder.CreateBinOp(Opc: I.getOpcode(), LHS: FBO->getOperand(i_nocapture: 1), RHS: C1);
873
874	Value *NewShift = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: TrueVal, RHS: C1);
875	Value *NewOp = Builder.CreateBinOp(Opc: FBO->getOpcode(), LHS: NewShift, RHS: NewRHS);
876	return SelectInst::Create(C: Cond, S1: NewShift, S2: NewOp);
877	}
878	}
879
880	return nullptr;
881	}
882
883	// Tries to perform
884	// (lshr (add (zext X), (zext Y)), K)
885	// -> (icmp ult (add X, Y), X)
886	// where
887	// - The add's operands are zexts from a K-bits integer to a bigger type.
888	// - The add is only used by the shr, or by iK (or narrower) truncates.
889	// - The lshr type has more than 2 bits (other types are boolean math).
890	// - K > 1
891	// note that
892	// - The resulting add cannot have nuw/nsw, else on overflow we get a
893	// poison value and the transform isn't legal anymore.
894	Instruction *InstCombinerImpl::foldLShrOverflowBit(BinaryOperator &I) {
895	assert(I.getOpcode() == Instruction::LShr);
896
897	Value *Add = I.getOperand(i_nocapture: 0);
898	Value *ShiftAmt = I.getOperand(i_nocapture: 1);
899	Type *Ty = I.getType();
900
901	if (Ty->getScalarSizeInBits() < 3)
902	return nullptr;
903
904	const APInt *ShAmtAPInt = nullptr;
905	Value *X = nullptr, *Y = nullptr;
906	if (!match(V: ShiftAmt, P: m_APInt(Res&: ShAmtAPInt)) ||
907	!match(V: Add,
908	P: m_Add(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))), R: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: Y))))))
909	return nullptr;
910
911	const unsigned ShAmt = ShAmtAPInt->getZExtValue();
912	if (ShAmt == 1)
913	return nullptr;
914
915	// X/Y are zexts from `ShAmt`-sized ints.
916	if (X->getType()->getScalarSizeInBits() != ShAmt ||
917	Y->getType()->getScalarSizeInBits() != ShAmt)
918	return nullptr;
919
920	// Make sure that `Add` is only used by `I` and `ShAmt`-truncates.
921	if (!Add->hasOneUse()) {
922	for (User *U : Add->users()) {
923	if (U == &I)
924	continue;
925
926	TruncInst *Trunc = dyn_cast<TruncInst>(Val: U);
927	if (!Trunc || Trunc->getType()->getScalarSizeInBits() > ShAmt)
928	return nullptr;
929	}
930	}
931
932	// Insert at Add so that the newly created `NarrowAdd` will dominate it's
933	// users (i.e. `Add`'s users).
934	Instruction *AddInst = cast<Instruction>(Val: Add);
935	Builder.SetInsertPoint(AddInst);
936
937	Value *NarrowAdd = Builder.CreateAdd(LHS: X, RHS: Y, Name: "add.narrowed");
938	Value *Overflow =
939	Builder.CreateICmpULT(LHS: NarrowAdd, RHS: X, Name: "add.narrowed.overflow");
940
941	// Replace the uses of the original add with a zext of the
942	// NarrowAdd's result. Note that all users at this stage are known to
943	// be ShAmt-sized truncs, or the lshr itself.
944	if (!Add->hasOneUse()) {
945	replaceInstUsesWith(I&: *AddInst, V: Builder.CreateZExt(V: NarrowAdd, DestTy: Ty));
946	eraseInstFromFunction(I&: *AddInst);
947	}
948
949	// Replace the LShr with a zext of the overflow check.
950	return new ZExtInst(Overflow, Ty);
951	}
952
953	// Try to set nuw/nsw flags on shl or exact flag on lshr/ashr using knownbits.
954	static bool setShiftFlags(BinaryOperator &I, const SimplifyQuery &Q) {
955	assert(I.isShift() && "Expected a shift as input");
956	// We already have all the flags.
957	if (I.getOpcode() == Instruction::Shl) {
958	if (I.hasNoUnsignedWrap() && I.hasNoSignedWrap())
959	return false;
960	} else {
961	if (I.isExact())
962	return false;
963
964	// shr (shl X, Y), Y
965	if (match(V: I.getOperand(i_nocapture: 0), P: m_Shl(L: m_Value(), R: m_Specific(V: I.getOperand(i_nocapture: 1))))) {
966	I.setIsExact();
967	return true;
968	}
969	}
970
971	// Compute what we know about shift count.
972	KnownBits KnownCnt = computeKnownBits(V: I.getOperand(i_nocapture: 1), /* Depth */ 0, Q);
973	unsigned BitWidth = KnownCnt.getBitWidth();
974	// Since shift produces a poison value if RHS is equal to or larger than the
975	// bit width, we can safely assume that RHS is less than the bit width.
976	uint64_t MaxCnt = KnownCnt.getMaxValue().getLimitedValue(Limit: BitWidth - 1);
977
978	KnownBits KnownAmt = computeKnownBits(V: I.getOperand(i_nocapture: 0), /* Depth */ 0, Q);
979	bool Changed = false;
980
981	if (I.getOpcode() == Instruction::Shl) {
982	// If we have as many leading zeros than maximum shift cnt we have nuw.
983	if (!I.hasNoUnsignedWrap() && MaxCnt <= KnownAmt.countMinLeadingZeros()) {
984	I.setHasNoUnsignedWrap();
985	Changed = true;
986	}
987	// If we have more sign bits than maximum shift cnt we have nsw.
988	if (!I.hasNoSignedWrap()) {
989	if (MaxCnt < KnownAmt.countMinSignBits() ||
990	MaxCnt < ComputeNumSignBits(Op: I.getOperand(i_nocapture: 0), DL: Q.DL, /*Depth*/ 0, AC: Q.AC,
991	CxtI: Q.CxtI, DT: Q.DT)) {
992	I.setHasNoSignedWrap();
993	Changed = true;
994	}
995	}
996	return Changed;
997	}
998
999	// If we have at least as many trailing zeros as maximum count then we have
1000	// exact.
1001	Changed = MaxCnt <= KnownAmt.countMinTrailingZeros();
1002	I.setIsExact(Changed);
1003
1004	return Changed;
1005	}
1006
1007	Instruction *InstCombinerImpl::visitShl(BinaryOperator &I) {
1008	const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
1009
1010	if (Value *V = simplifyShlInst(Op0: I.getOperand(i_nocapture: 0), Op1: I.getOperand(i_nocapture: 1),
1011	IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(), Q))
1012	return replaceInstUsesWith(I, V);
1013
1014	if (Instruction *X = foldVectorBinop(Inst&: I))
1015	return X;
1016
1017	if (Instruction *V = commonShiftTransforms(I))
1018	return V;
1019
1020	if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(OuterShift: &I, Q, Builder))
1021	return V;
1022
1023	Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
1024	Type *Ty = I.getType();
1025	unsigned BitWidth = Ty->getScalarSizeInBits();
1026
1027	const APInt *C;
1028	if (match(V: Op1, P: m_APInt(Res&: C))) {
1029	unsigned ShAmtC = C->getZExtValue();
1030
1031	// shl (zext X), C --> zext (shl X, C)
1032	// This is only valid if X would have zeros shifted out.
1033	Value *X;
1034	if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))))) {
1035	unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1036	if (ShAmtC < SrcWidth &&
1037	MaskedValueIsZero(V: X, Mask: APInt::getHighBitsSet(numBits: SrcWidth, hiBitsSet: ShAmtC), Depth: 0, CxtI: &I))
1038	return new ZExtInst(Builder.CreateShl(LHS: X, RHS: ShAmtC), Ty);
1039	}
1040
1041	// (X >> C) << C --> X & (-1 << C)
1042	if (match(V: Op0, P: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1)))) {
1043	APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC));
1044	return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty, V: Mask));
1045	}
1046
1047	const APInt *C1;
1048	if (match(V: Op0, P: m_Exact(SubPattern: m_Shr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) &&
1049	C1->ult(RHS: BitWidth)) {
1050	unsigned ShrAmt = C1->getZExtValue();
1051	if (ShrAmt < ShAmtC) {
1052	// If C1 < C: (X >>?,exact C1) << C --> X << (C - C1)
1053	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmtC - ShrAmt);
1054	auto *NewShl = BinaryOperator::CreateShl(V1: X, V2: ShiftDiff);
1055	NewShl->setHasNoUnsignedWrap(
1056	I.hasNoUnsignedWrap() ||
1057	(ShrAmt &&
1058	cast<Instruction>(Val: Op0)->getOpcode() == Instruction::LShr &&
1059	I.hasNoSignedWrap()));
1060	NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
1061	return NewShl;
1062	}
1063	if (ShrAmt > ShAmtC) {
1064	// If C1 > C: (X >>?exact C1) << C --> X >>?exact (C1 - C)
1065	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShrAmt - ShAmtC);
1066	auto *NewShr = BinaryOperator::Create(
1067	Op: cast<BinaryOperator>(Val: Op0)->getOpcode(), S1: X, S2: ShiftDiff);
1068	NewShr->setIsExact(true);
1069	return NewShr;
1070	}
1071	}
1072
1073	if (match(V: Op0, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) &&
1074	C1->ult(RHS: BitWidth)) {
1075	unsigned ShrAmt = C1->getZExtValue();
1076	if (ShrAmt < ShAmtC) {
1077	// If C1 < C: (X >>? C1) << C --> (X << (C - C1)) & (-1 << C)
1078	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmtC - ShrAmt);
1079	auto *NewShl = BinaryOperator::CreateShl(V1: X, V2: ShiftDiff);
1080	NewShl->setHasNoUnsignedWrap(
1081	I.hasNoUnsignedWrap() ||
1082	(ShrAmt &&
1083	cast<Instruction>(Val: Op0)->getOpcode() == Instruction::LShr &&
1084	I.hasNoSignedWrap()));
1085	NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
1086	Builder.Insert(I: NewShl);
1087	APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC));
1088	return BinaryOperator::CreateAnd(V1: NewShl, V2: ConstantInt::get(Ty, V: Mask));
1089	}
1090	if (ShrAmt > ShAmtC) {
1091	// If C1 > C: (X >>? C1) << C --> (X >>? (C1 - C)) & (-1 << C)
1092	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShrAmt - ShAmtC);
1093	auto *OldShr = cast<BinaryOperator>(Val: Op0);
1094	auto *NewShr =
1095	BinaryOperator::Create(Op: OldShr->getOpcode(), S1: X, S2: ShiftDiff);
1096	NewShr->setIsExact(OldShr->isExact());
1097	Builder.Insert(I: NewShr);
1098	APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC));
1099	return BinaryOperator::CreateAnd(V1: NewShr, V2: ConstantInt::get(Ty, V: Mask));
1100	}
1101	}
1102
1103	// Similar to above, but look through an intermediate trunc instruction.
1104	BinaryOperator *Shr;
1105	if (match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_OneUse(SubPattern: m_BinOp(I&: Shr))))) &&
1106	match(V: Shr, P: m_Shr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) {
1107	// The larger shift direction survives through the transform.
1108	unsigned ShrAmtC = C1->getZExtValue();
1109	unsigned ShDiff = ShrAmtC > ShAmtC ? ShrAmtC - ShAmtC : ShAmtC - ShrAmtC;
1110	Constant *ShiftDiffC = ConstantInt::get(Ty: X->getType(), V: ShDiff);
1111	auto ShiftOpc = ShrAmtC > ShAmtC ? Shr->getOpcode() : Instruction::Shl;
1112
1113	// If C1 > C:
1114	// (trunc (X >> C1)) << C --> (trunc (X >> (C1 - C))) && (-1 << C)
1115	// If C > C1:
1116	// (trunc (X >> C1)) << C --> (trunc (X << (C - C1))) && (-1 << C)
1117	Value *NewShift = Builder.CreateBinOp(Opc: ShiftOpc, LHS: X, RHS: ShiftDiffC, Name: "sh.diff");
1118	Value *Trunc = Builder.CreateTrunc(V: NewShift, DestTy: Ty, Name: "tr.sh.diff");
1119	APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC));
1120	return BinaryOperator::CreateAnd(V1: Trunc, V2: ConstantInt::get(Ty, V: Mask));
1121	}
1122
1123	if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && C1->ult(RHS: BitWidth)) {
1124	unsigned AmtSum = ShAmtC + C1->getZExtValue();
1125	// Oversized shifts are simplified to zero in InstSimplify.
1126	if (AmtSum < BitWidth)
1127	// (X << C1) << C2 --> X << (C1 + C2)
1128	return BinaryOperator::CreateShl(V1: X, V2: ConstantInt::get(Ty, V: AmtSum));
1129	}
1130
1131	// If we have an opposite shift by the same amount, we may be able to
1132	// reorder binops and shifts to eliminate math/logic.
1133	auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) {
1134	switch (BinOpcode) {
1135	default:
1136	return false;
1137	case Instruction::Add:
1138	case Instruction::And:
1139	case Instruction::Or:
1140	case Instruction::Xor:
1141	case Instruction::Sub:
1142	// NOTE: Sub is not commutable and the tranforms below may not be valid
1143	// when the shift-right is operand 1 (RHS) of the sub.
1144	return true;
1145	}
1146	};
1147	BinaryOperator *Op0BO;
1148	if (match(V: Op0, P: m_OneUse(SubPattern: m_BinOp(I&: Op0BO))) &&
1149	isSuitableBinOpcode(Op0BO->getOpcode())) {
1150	// Commute so shift-right is on LHS of the binop.
1151	// (Y bop (X >> C)) << C -> ((X >> C) bop Y) << C
1152	// (Y bop ((X >> C) & CC)) << C -> (((X >> C) & CC) bop Y) << C
1153	Value *Shr = Op0BO->getOperand(i_nocapture: 0);
1154	Value *Y = Op0BO->getOperand(i_nocapture: 1);
1155	Value *X;
1156	const APInt *CC;
1157	if (Op0BO->isCommutative() && Y->hasOneUse() &&
1158	(match(V: Y, P: m_Shr(L: m_Value(), R: m_Specific(V: Op1))) ||
1159	match(V: Y, P: m_And(L: m_OneUse(SubPattern: m_Shr(L: m_Value(), R: m_Specific(V: Op1))),
1160	R: m_APInt(Res&: CC)))))
1161	std::swap(a&: Shr, b&: Y);
1162
1163	// ((X >> C) bop Y) << C -> (X bop (Y << C)) & (~0 << C)
1164	if (match(V: Shr, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1))))) {
1165	// Y << C
1166	Value *YS = Builder.CreateShl(LHS: Y, RHS: Op1, Name: Op0BO->getName());
1167	// (X bop (Y << C))
1168	Value *B =
1169	Builder.CreateBinOp(Opc: Op0BO->getOpcode(), LHS: X, RHS: YS, Name: Shr->getName());
1170	unsigned Op1Val = C->getLimitedValue(Limit: BitWidth);
1171	APInt Bits = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - Op1Val);
1172	Constant *Mask = ConstantInt::get(Ty, V: Bits);
1173	return BinaryOperator::CreateAnd(V1: B, V2: Mask);
1174	}
1175
1176	// (((X >> C) & CC) bop Y) << C -> (X & (CC << C)) bop (Y << C)
1177	if (match(V: Shr,
1178	P: m_OneUse(SubPattern: m_And(L: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1))),
1179	R: m_APInt(Res&: CC))))) {
1180	// Y << C
1181	Value *YS = Builder.CreateShl(LHS: Y, RHS: Op1, Name: Op0BO->getName());
1182	// X & (CC << C)
1183	Value *M = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty, V: CC->shl(ShiftAmt: *C)),
1184	Name: X->getName() + ".mask");
1185	return BinaryOperator::Create(Op: Op0BO->getOpcode(), S1: M, S2: YS);
1186	}
1187	}
1188
1189	// (C1 - X) << C --> (C1 << C) - (X << C)
1190	if (match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_APInt(Res&: C1), R: m_Value(V&: X))))) {
1191	Constant *NewLHS = ConstantInt::get(Ty, V: C1->shl(ShiftAmt: *C));
1192	Value *NewShift = Builder.CreateShl(LHS: X, RHS: Op1);
1193	return BinaryOperator::CreateSub(V1: NewLHS, V2: NewShift);
1194	}
1195	}
1196
1197	if (setShiftFlags(I, Q))
1198	return &I;
1199
1200	// Transform (x >> y) << y to x & (-1 << y)
1201	// Valid for any type of right-shift.
1202	Value *X;
1203	if (match(V: Op0, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1))))) {
1204	Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1205	Value *Mask = Builder.CreateShl(LHS: AllOnes, RHS: Op1);
1206	return BinaryOperator::CreateAnd(V1: Mask, V2: X);
1207	}
1208
1209	// Transform (-1 >> y) << y to -1 << y
1210	if (match(V: Op0, P: m_LShr(L: m_AllOnes(), R: m_Specific(V: Op1)))) {
1211	Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1212	return BinaryOperator::CreateShl(V1: AllOnes, V2: Op1);
1213	}
1214
1215	Constant *C1;
1216	if (match(V: Op1, P: m_Constant(C&: C1))) {
1217	Constant *C2;
1218	Value *X;
1219	// (X * C2) << C1 --> X * (C2 << C1)
1220	if (match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_Constant(C&: C2))))
1221	return BinaryOperator::CreateMul(V1: X, V2: ConstantExpr::getShl(C1: C2, C2: C1));
1222
1223	// shl (zext i1 X), C1 --> select (X, 1 << C1, 0)
1224	if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
1225	auto *NewC = ConstantExpr::getShl(C1: ConstantInt::get(Ty, V: 1), C2: C1);
1226	return SelectInst::Create(C: X, S1: NewC, S2: ConstantInt::getNullValue(Ty));
1227	}
1228	}
1229
1230	if (match(V: Op0, P: m_One())) {
1231	// (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1
1232	if (match(V: Op1, P: m_Sub(L: m_SpecificInt(V: BitWidth - 1), R: m_Value(V&: X))))
1233	return BinaryOperator::CreateLShr(
1234	V1: ConstantInt::get(Ty, V: APInt::getSignMask(BitWidth)), V2: X);
1235
1236	// Canonicalize "extract lowest set bit" using cttz to and-with-negate:
1237	// 1 << (cttz X) --> -X & X
1238	if (match(Op1,
1239	m_OneUse(m_Intrinsic<Intrinsic::cttz>(m_Value(X), m_Value())))) {
1240	Value *NegX = Builder.CreateNeg(V: X, Name: "neg");
1241	return BinaryOperator::CreateAnd(V1: NegX, V2: X);
1242	}
1243	}
1244
1245	return nullptr;
1246	}
1247
1248	Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
1249	if (Value *V = simplifyLShrInst(Op0: I.getOperand(i_nocapture: 0), Op1: I.getOperand(i_nocapture: 1), IsExact: I.isExact(),
1250	Q: SQ.getWithInstruction(I: &I)))
1251	return replaceInstUsesWith(I, V);
1252
1253	if (Instruction *X = foldVectorBinop(Inst&: I))
1254	return X;
1255
1256	if (Instruction *R = commonShiftTransforms(I))
1257	return R;
1258
1259	Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
1260	Type *Ty = I.getType();
1261	Value *X;
1262	const APInt *C;
1263	unsigned BitWidth = Ty->getScalarSizeInBits();
1264
1265	// (iN (~X) u>> (N - 1)) --> zext (X > -1)
1266	if (match(V: Op0, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: X)))) &&
1267	match(V: Op1, P: m_SpecificIntAllowPoison(V: BitWidth - 1)))
1268	return new ZExtInst(Builder.CreateIsNotNeg(Arg: X, Name: "isnotneg"), Ty);
1269
1270	if (match(V: Op1, P: m_APInt(Res&: C))) {
1271	unsigned ShAmtC = C->getZExtValue();
1272	auto *II = dyn_cast<IntrinsicInst>(Val: Op0);
1273	if (II && isPowerOf2_32(Value: BitWidth) && Log2_32(Value: BitWidth) == ShAmtC &&
1274	(II->getIntrinsicID() == Intrinsic::ctlz ||
1275	II->getIntrinsicID() == Intrinsic::cttz ||
1276	II->getIntrinsicID() == Intrinsic::ctpop)) {
1277	// ctlz.i32(x)>>5 --> zext(x == 0)
1278	// cttz.i32(x)>>5 --> zext(x == 0)
1279	// ctpop.i32(x)>>5 --> zext(x == -1)
1280	bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop;
1281	Constant *RHS = ConstantInt::getSigned(Ty, V: IsPop ? -1 : 0);
1282	Value *Cmp = Builder.CreateICmpEQ(LHS: II->getArgOperand(i: 0), RHS);
1283	return new ZExtInst(Cmp, Ty);
1284	}
1285
1286	Value *X;
1287	const APInt *C1;
1288	if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && C1->ult(RHS: BitWidth)) {
1289	if (C1->ult(RHS: ShAmtC)) {
1290	unsigned ShlAmtC = C1->getZExtValue();
1291	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmtC - ShlAmtC);
1292	if (cast<BinaryOperator>(Val: Op0)->hasNoUnsignedWrap()) {
1293	// (X <<nuw C1) >>u C --> X >>u (C - C1)
1294	auto *NewLShr = BinaryOperator::CreateLShr(V1: X, V2: ShiftDiff);
1295	NewLShr->setIsExact(I.isExact());
1296	return NewLShr;
1297	}
1298	if (Op0->hasOneUse()) {
1299	// (X << C1) >>u C --> (X >>u (C - C1)) & (-1 >> C)
1300	Value *NewLShr = Builder.CreateLShr(LHS: X, RHS: ShiftDiff, Name: "", isExact: I.isExact());
1301	APInt Mask(APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC));
1302	return BinaryOperator::CreateAnd(V1: NewLShr, V2: ConstantInt::get(Ty, V: Mask));
1303	}
1304	} else if (C1->ugt(RHS: ShAmtC)) {
1305	unsigned ShlAmtC = C1->getZExtValue();
1306	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShlAmtC - ShAmtC);
1307	if (cast<BinaryOperator>(Val: Op0)->hasNoUnsignedWrap()) {
1308	// (X <<nuw C1) >>u C --> X <<nuw/nsw (C1 - C)
1309	auto *NewShl = BinaryOperator::CreateShl(V1: X, V2: ShiftDiff);
1310	NewShl->setHasNoUnsignedWrap(true);
1311	NewShl->setHasNoSignedWrap(ShAmtC > 0);
1312	return NewShl;
1313	}
1314	if (Op0->hasOneUse()) {
1315	// (X << C1) >>u C --> X << (C1 - C) & (-1 >> C)
1316	Value *NewShl = Builder.CreateShl(LHS: X, RHS: ShiftDiff);
1317	APInt Mask(APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC));
1318	return BinaryOperator::CreateAnd(V1: NewShl, V2: ConstantInt::get(Ty, V: Mask));
1319	}
1320	} else {
1321	assert(*C1 == ShAmtC);
1322	// (X << C) >>u C --> X & (-1 >>u C)
1323	APInt Mask(APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC));
1324	return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty, V: Mask));
1325	}
1326	}
1327
1328	// ((X << C) + Y) >>u C --> (X + (Y >>u C)) & (-1 >>u C)
1329	// TODO: Consolidate with the more general transform that starts from shl
1330	// (the shifts are in the opposite order).
1331	Value *Y;
1332	if (match(V: Op0,
1333	P: m_OneUse(SubPattern: m_c_Add(L: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X), R: m_Specific(V: Op1))),
1334	R: m_Value(V&: Y))))) {
1335	Value *NewLshr = Builder.CreateLShr(LHS: Y, RHS: Op1);
1336	Value *NewAdd = Builder.CreateAdd(LHS: NewLshr, RHS: X);
1337	unsigned Op1Val = C->getLimitedValue(Limit: BitWidth);
1338	APInt Bits = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - Op1Val);
1339	Constant *Mask = ConstantInt::get(Ty, V: Bits);
1340	return BinaryOperator::CreateAnd(V1: NewAdd, V2: Mask);
1341	}
1342
1343	if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) &&
1344	(!Ty->isIntegerTy() || shouldChangeType(From: Ty, To: X->getType()))) {
1345	assert(ShAmtC < X->getType()->getScalarSizeInBits() &&
1346	"Big shift not simplified to zero?");
1347	// lshr (zext iM X to iN), C --> zext (lshr X, C) to iN
1348	Value *NewLShr = Builder.CreateLShr(LHS: X, RHS: ShAmtC);
1349	return new ZExtInst(NewLShr, Ty);
1350	}
1351
1352	if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X)))) {
1353	unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits();
1354	// lshr (sext i1 X to iN), C --> select (X, -1 >> C, 0)
1355	if (SrcTyBitWidth == 1) {
1356	auto *NewC = ConstantInt::get(
1357	Ty, V: APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC));
1358	return SelectInst::Create(C: X, S1: NewC, S2: ConstantInt::getNullValue(Ty));
1359	}
1360
1361	if ((!Ty->isIntegerTy() || shouldChangeType(From: Ty, To: X->getType())) &&
1362	Op0->hasOneUse()) {
1363	// Are we moving the sign bit to the low bit and widening with high
1364	// zeros? lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN
1365	if (ShAmtC == BitWidth - 1) {
1366	Value *NewLShr = Builder.CreateLShr(LHS: X, RHS: SrcTyBitWidth - 1);
1367	return new ZExtInst(NewLShr, Ty);
1368	}
1369
1370	// lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN
1371	if (ShAmtC == BitWidth - SrcTyBitWidth) {
1372	// The new shift amount can't be more than the narrow source type.
1373	unsigned NewShAmt = std::min(a: ShAmtC, b: SrcTyBitWidth - 1);
1374	Value *AShr = Builder.CreateAShr(LHS: X, RHS: NewShAmt);
1375	return new ZExtInst(AShr, Ty);
1376	}
1377	}
1378	}
1379
1380	if (ShAmtC == BitWidth - 1) {
1381	// lshr i32 or(X,-X), 31 --> zext (X != 0)
1382	if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))))
1383	return new ZExtInst(Builder.CreateIsNotNull(Arg: X), Ty);
1384
1385	// lshr i32 (X -nsw Y), 31 --> zext (X < Y)
1386	if (match(V: Op0, P: m_OneUse(SubPattern: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y)))))
1387	return new ZExtInst(Builder.CreateICmpSLT(LHS: X, RHS: Y), Ty);
1388
1389	// Check if a number is negative and odd:
1390	// lshr i32 (srem X, 2), 31 --> and (X >> 31), X
1391	if (match(V: Op0, P: m_OneUse(SubPattern: m_SRem(L: m_Value(V&: X), R: m_SpecificInt(V: 2))))) {
1392	Value *Signbit = Builder.CreateLShr(LHS: X, RHS: ShAmtC);
1393	return BinaryOperator::CreateAnd(V1: Signbit, V2: X);
1394	}
1395	}
1396
1397	// (X >>u C1) >>u C --> X >>u (C1 + C)
1398	if (match(V: Op0, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) {
1399	// Oversized shifts are simplified to zero in InstSimplify.
1400	unsigned AmtSum = ShAmtC + C1->getZExtValue();
1401	if (AmtSum < BitWidth)
1402	return BinaryOperator::CreateLShr(V1: X, V2: ConstantInt::get(Ty, V: AmtSum));
1403	}
1404
1405	Instruction *TruncSrc;
1406	if (match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_Instruction(I&: TruncSrc)))) &&
1407	match(V: TruncSrc, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) {
1408	unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1409	unsigned AmtSum = ShAmtC + C1->getZExtValue();
1410
1411	// If the combined shift fits in the source width:
1412	// (trunc (X >>u C1)) >>u C --> and (trunc (X >>u (C1 + C)), MaskC
1413	//
1414	// If the first shift covers the number of bits truncated, then the
1415	// mask instruction is eliminated (and so the use check is relaxed).
1416	if (AmtSum < SrcWidth &&
1417	(TruncSrc->hasOneUse() || C1->uge(RHS: SrcWidth - BitWidth))) {
1418	Value *SumShift = Builder.CreateLShr(LHS: X, RHS: AmtSum, Name: "sum.shift");
1419	Value *Trunc = Builder.CreateTrunc(V: SumShift, DestTy: Ty, Name: I.getName());
1420
1421	// If the first shift does not cover the number of bits truncated, then
1422	// we require a mask to get rid of high bits in the result.
1423	APInt MaskC = APInt::getAllOnes(numBits: BitWidth).lshr(shiftAmt: ShAmtC);
1424	return BinaryOperator::CreateAnd(V1: Trunc, V2: ConstantInt::get(Ty, V: MaskC));
1425	}
1426	}
1427
1428	const APInt *MulC;
1429	if (match(V: Op0, P: m_NUWMul(L: m_Value(V&: X), R: m_APInt(Res&: MulC)))) {
1430	// Look for a "splat" mul pattern - it replicates bits across each half of
1431	// a value, so a right shift is just a mask of the low bits:
1432	// lshr i[2N] (mul nuw X, (2^N)+1), N --> and iN X, (2^N)-1
1433	// TODO: Generalize to allow more than just half-width shifts?
1434	if (BitWidth > 2 && ShAmtC * 2 == BitWidth && (*MulC - 1).isPowerOf2() &&
1435	MulC->logBase2() == ShAmtC)
1436	return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty, V: *MulC - 2));
1437
1438	// The one-use check is not strictly necessary, but codegen may not be
1439	// able to invert the transform and perf may suffer with an extra mul
1440	// instruction.
1441	if (Op0->hasOneUse()) {
1442	APInt NewMulC = MulC->lshr(shiftAmt: ShAmtC);
1443	// if c is divisible by (1 << ShAmtC):
1444	// lshr (mul nuw x, MulC), ShAmtC -> mul nuw nsw x, (MulC >> ShAmtC)
1445	if (MulC->eq(RHS: NewMulC.shl(shiftAmt: ShAmtC))) {
1446	auto *NewMul =
1447	BinaryOperator::CreateNUWMul(V1: X, V2: ConstantInt::get(Ty, V: NewMulC));
1448	assert(ShAmtC != 0 &&
1449	"lshr X, 0 should be handled by simplifyLShrInst.");
1450	NewMul->setHasNoSignedWrap(true);
1451	return NewMul;
1452	}
1453	}
1454	}
1455
1456	// Try to narrow bswap.
1457	// In the case where the shift amount equals the bitwidth difference, the
1458	// shift is eliminated.
1459	if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::bswap>(
1460	m_OneUse(m_ZExt(m_Value(X))))))) {
1461	unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1462	unsigned WidthDiff = BitWidth - SrcWidth;
1463	if (SrcWidth % 16 == 0) {
1464	Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::ID: bswap, V: X);
1465	if (ShAmtC >= WidthDiff) {
1466	// (bswap (zext X)) >> C --> zext (bswap X >> C')
1467	Value *NewShift = Builder.CreateLShr(LHS: NarrowSwap, RHS: ShAmtC - WidthDiff);
1468	return new ZExtInst(NewShift, Ty);
1469	} else {
1470	// (bswap (zext X)) >> C --> (zext (bswap X)) << C'
1471	Value *NewZExt = Builder.CreateZExt(V: NarrowSwap, DestTy: Ty);
1472	Constant *ShiftDiff = ConstantInt::get(Ty, V: WidthDiff - ShAmtC);
1473	return BinaryOperator::CreateShl(V1: NewZExt, V2: ShiftDiff);
1474	}
1475	}
1476	}
1477
1478	// Reduce add-carry of bools to logic:
1479	// ((zext BoolX) + (zext BoolY)) >> 1 --> zext (BoolX && BoolY)
1480	Value *BoolX, *BoolY;
1481	if (ShAmtC == 1 && match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
1482	match(V: X, P: m_ZExt(Op: m_Value(V&: BoolX))) && match(V: Y, P: m_ZExt(Op: m_Value(V&: BoolY))) &&
1483	BoolX->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
1484	BoolY->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
1485	(X->hasOneUse() || Y->hasOneUse() || Op0->hasOneUse())) {
1486	Value *And = Builder.CreateAnd(LHS: BoolX, RHS: BoolY);
1487	return new ZExtInst(And, Ty);
1488	}
1489	}
1490
1491	const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
1492	if (setShiftFlags(I, Q))
1493	return &I;
1494
1495	// Transform (x << y) >> y to x & (-1 >> y)
1496	if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X), R: m_Specific(V: Op1))))) {
1497	Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1498	Value *Mask = Builder.CreateLShr(LHS: AllOnes, RHS: Op1);
1499	return BinaryOperator::CreateAnd(V1: Mask, V2: X);
1500	}
1501
1502	// Transform (-1 << y) >> y to -1 >> y
1503	if (match(V: Op0, P: m_Shl(L: m_AllOnes(), R: m_Specific(V: Op1)))) {
1504	Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1505	return BinaryOperator::CreateLShr(V1: AllOnes, V2: Op1);
1506	}
1507
1508	if (Instruction *Overflow = foldLShrOverflowBit(I))
1509	return Overflow;
1510
1511	return nullptr;
1512	}
1513
1514	Instruction *
1515	InstCombinerImpl::foldVariableSignZeroExtensionOfVariableHighBitExtract(
1516	BinaryOperator &OldAShr) {
1517	assert(OldAShr.getOpcode() == Instruction::AShr &&
1518	"Must be called with arithmetic right-shift instruction only.");
1519
1520	// Check that constant C is a splat of the element-wise bitwidth of V.
1521	auto BitWidthSplat = [](Constant *C, Value *V) {
1522	return match(
1523	V: C, P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_EQ,
1524	Threshold: APInt(C->getType()->getScalarSizeInBits(),
1525	V->getType()->getScalarSizeInBits())));
1526	};
1527
1528	// It should look like variable-length sign-extension on the outside:
1529	// (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits)
1530	Value *NBits;
1531	Instruction *MaybeTrunc;
1532	Constant *C1, *C2;
1533	if (!match(V: &OldAShr,
1534	P: m_AShr(L: m_Shl(L: m_Instruction(I&: MaybeTrunc),
1535	R: m_ZExtOrSelf(Op: m_Sub(L: m_Constant(C&: C1),
1536	R: m_ZExtOrSelf(Op: m_Value(V&: NBits))))),
1537	R: m_ZExtOrSelf(Op: m_Sub(L: m_Constant(C&: C2),
1538	R: m_ZExtOrSelf(Op: m_Deferred(V: NBits)))))) ||
1539	!BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr))
1540	return nullptr;
1541
1542	// There may or may not be a truncation after outer two shifts.
1543	Instruction *HighBitExtract;
1544	match(V: MaybeTrunc, P: m_TruncOrSelf(Op: m_Instruction(I&: HighBitExtract)));
1545	bool HadTrunc = MaybeTrunc != HighBitExtract;
1546
1547	// And finally, the innermost part of the pattern must be a right-shift.
1548	Value *X, *NumLowBitsToSkip;
1549	if (!match(V: HighBitExtract, P: m_Shr(L: m_Value(V&: X), R: m_Value(V&: NumLowBitsToSkip))))
1550	return nullptr;
1551
1552	// Said right-shift must extract high NBits bits - C0 must be it's bitwidth.
1553	Constant *C0;
1554	if (!match(V: NumLowBitsToSkip,
1555	P: m_ZExtOrSelf(
1556	Op: m_Sub(L: m_Constant(C&: C0), R: m_ZExtOrSelf(Op: m_Specific(V: NBits))))) ||
1557	!BitWidthSplat(C0, HighBitExtract))
1558	return nullptr;
1559
1560	// Since the NBits is identical for all shifts, if the outermost and
1561	// innermost shifts are identical, then outermost shifts are redundant.
1562	// If we had truncation, do keep it though.
1563	if (HighBitExtract->getOpcode() == OldAShr.getOpcode())
1564	return replaceInstUsesWith(I&: OldAShr, V: MaybeTrunc);
1565
1566	// Else, if there was a truncation, then we need to ensure that one
1567	// instruction will go away.
1568	if (HadTrunc && !match(V: &OldAShr, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value())))
1569	return nullptr;
1570
1571	// Finally, bypass two innermost shifts, and perform the outermost shift on
1572	// the operands of the innermost shift.
1573	Instruction *NewAShr =
1574	BinaryOperator::Create(Op: OldAShr.getOpcode(), S1: X, S2: NumLowBitsToSkip);
1575	NewAShr->copyIRFlags(V: HighBitExtract); // We can preserve 'exact'-ness.
1576	if (!HadTrunc)
1577	return NewAShr;
1578
1579	Builder.Insert(I: NewAShr);
1580	return TruncInst::CreateTruncOrBitCast(S: NewAShr, Ty: OldAShr.getType());
1581	}
1582
1583	Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) {
1584	if (Value *V = simplifyAShrInst(Op0: I.getOperand(i_nocapture: 0), Op1: I.getOperand(i_nocapture: 1), IsExact: I.isExact(),
1585	Q: SQ.getWithInstruction(I: &I)))
1586	return replaceInstUsesWith(I, V);
1587
1588	if (Instruction *X = foldVectorBinop(Inst&: I))
1589	return X;
1590
1591	if (Instruction *R = commonShiftTransforms(I))
1592	return R;
1593
1594	Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
1595	Type *Ty = I.getType();
1596	unsigned BitWidth = Ty->getScalarSizeInBits();
1597	const APInt *ShAmtAPInt;
1598	if (match(V: Op1, P: m_APInt(Res&: ShAmtAPInt)) && ShAmtAPInt->ult(RHS: BitWidth)) {
1599	unsigned ShAmt = ShAmtAPInt->getZExtValue();
1600
1601	// If the shift amount equals the difference in width of the destination
1602	// and source scalar types:
1603	// ashr (shl (zext X), C), C --> sext X
1604	Value *X;
1605	if (match(V: Op0, P: m_Shl(L: m_ZExt(Op: m_Value(V&: X)), R: m_Specific(V: Op1))) &&
1606	ShAmt == BitWidth - X->getType()->getScalarSizeInBits())
1607	return new SExtInst(X, Ty);
1608
1609	// We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However,
1610	// we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
1611	const APInt *ShOp1;
1612	if (match(V: Op0, P: m_NSWShl(L: m_Value(V&: X), R: m_APInt(Res&: ShOp1))) &&
1613	ShOp1->ult(RHS: BitWidth)) {
1614	unsigned ShlAmt = ShOp1->getZExtValue();
1615	if (ShlAmt < ShAmt) {
1616	// (X <<nsw C1) >>s C2 --> X >>s (C2 - C1)
1617	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmt - ShlAmt);
1618	auto *NewAShr = BinaryOperator::CreateAShr(V1: X, V2: ShiftDiff);
1619	NewAShr->setIsExact(I.isExact());
1620	return NewAShr;
1621	}
1622	if (ShlAmt > ShAmt) {
1623	// (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2)
1624	Constant *ShiftDiff = ConstantInt::get(Ty, V: ShlAmt - ShAmt);
1625	auto *NewShl = BinaryOperator::Create(Op: Instruction::Shl, S1: X, S2: ShiftDiff);
1626	NewShl->setHasNoSignedWrap(true);
1627	return NewShl;
1628	}
1629	}
1630
1631	if (match(V: Op0, P: m_AShr(L: m_Value(V&: X), R: m_APInt(Res&: ShOp1))) &&
1632	ShOp1->ult(RHS: BitWidth)) {
1633	unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
1634	// Oversized arithmetic shifts replicate the sign bit.
1635	AmtSum = std::min(a: AmtSum, b: BitWidth - 1);
1636	// (X >>s C1) >>s C2 --> X >>s (C1 + C2)
1637	return BinaryOperator::CreateAShr(V1: X, V2: ConstantInt::get(Ty, V: AmtSum));
1638	}
1639
1640	if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X)))) &&
1641	(Ty->isVectorTy() || shouldChangeType(From: Ty, To: X->getType()))) {
1642	// ashr (sext X), C --> sext (ashr X, C')
1643	Type *SrcTy = X->getType();
1644	ShAmt = std::min(a: ShAmt, b: SrcTy->getScalarSizeInBits() - 1);
1645	Value *NewSh = Builder.CreateAShr(LHS: X, RHS: ConstantInt::get(Ty: SrcTy, V: ShAmt));
1646	return new SExtInst(NewSh, Ty);
1647	}
1648
1649	if (ShAmt == BitWidth - 1) {
1650	// ashr i32 or(X,-X), 31 --> sext (X != 0)
1651	if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))))
1652	return new SExtInst(Builder.CreateIsNotNull(Arg: X), Ty);
1653
1654	// ashr i32 (X -nsw Y), 31 --> sext (X < Y)
1655	Value *Y;
1656	if (match(V: Op0, P: m_OneUse(SubPattern: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y)))))
1657	return new SExtInst(Builder.CreateICmpSLT(LHS: X, RHS: Y), Ty);
1658	}
1659	}
1660
1661	const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
1662	if (setShiftFlags(I, Q))
1663	return &I;
1664
1665	// Prefer `-(x & 1)` over `(x << (bitwidth(x)-1)) a>> (bitwidth(x)-1)`
1666	// as the pattern to splat the lowest bit.
1667	// FIXME: iff X is already masked, we don't need the one-use check.
1668	Value *X;
1669	if (match(V: Op1, P: m_SpecificIntAllowPoison(V: BitWidth - 1)) &&
1670	match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X),
1671	R: m_SpecificIntAllowPoison(V: BitWidth - 1))))) {
1672	Constant *Mask = ConstantInt::get(Ty, V: 1);
1673	// Retain the knowledge about the ignored lanes.
1674	Mask = Constant::mergeUndefsWith(
1675	C: Constant::mergeUndefsWith(C: Mask, Other: cast<Constant>(Val: Op1)),
1676	Other: cast<Constant>(Val: cast<Instruction>(Val: Op0)->getOperand(i: 1)));
1677	X = Builder.CreateAnd(LHS: X, RHS: Mask);
1678	return BinaryOperator::CreateNeg(Op: X);
1679	}
1680
1681	if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(OldAShr&: I))
1682	return R;
1683
1684	// See if we can turn a signed shr into an unsigned shr.
1685	if (MaskedValueIsZero(V: Op0, Mask: APInt::getSignMask(BitWidth), Depth: 0, CxtI: &I)) {
1686	Instruction *Lshr = BinaryOperator::CreateLShr(V1: Op0, V2: Op1);
1687	Lshr->setIsExact(I.isExact());
1688	return Lshr;
1689	}
1690
1691	// ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1
1692	if (match(V: Op0, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: X))))) {
1693	// Note that we must drop 'exact'-ness of the shift!
1694	// Note that we can't keep undef's in -1 vector constant!
1695	auto *NewAShr = Builder.CreateAShr(LHS: X, RHS: Op1, Name: Op0->getName() + ".not");
1696	return BinaryOperator::CreateNot(Op: NewAShr);
1697	}
1698
1699	return nullptr;
1700	}
1701