xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstCombineIntrinsic.cpp (revision 47ef2a131091508e049ab10cad7f91a3c1342cd9)
1 //===-- X86InstCombineIntrinsic.cpp - X86 specific InstCombine pass -------===//
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 /// \file
9 /// This file implements a TargetTransformInfo analysis pass specific to the
10 /// X86 target machine. It uses the target's detailed information to provide
11 /// more precise answers to certain TTI queries, while letting the target
12 /// independent and default TTI implementations handle the rest.
13 ///
14 //===----------------------------------------------------------------------===//
15 
16 #include "X86TargetTransformInfo.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/IntrinsicsX86.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 llvm::PatternMatch;
25 
26 #define DEBUG_TYPE "x86tti"
27 
28 /// Return a constant boolean vector that has true elements in all positions
29 /// where the input constant data vector has an element with the sign bit set.
30 static Constant *getNegativeIsTrueBoolVec(Constant *V, const DataLayout &DL) {
31   VectorType *IntTy = VectorType::getInteger(cast<VectorType>(V->getType()));
32   V = ConstantExpr::getBitCast(V, IntTy);
33   V = ConstantFoldCompareInstOperands(CmpInst::ICMP_SGT,
34                                       Constant::getNullValue(IntTy), V, DL);
35   assert(V && "Vector must be foldable");
36   return V;
37 }
38 
39 /// Convert the x86 XMM integer vector mask to a vector of bools based on
40 /// each element's most significant bit (the sign bit).
41 static Value *getBoolVecFromMask(Value *Mask, const DataLayout &DL) {
42   // Fold Constant Mask.
43   if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask))
44     return getNegativeIsTrueBoolVec(ConstantMask, DL);
45 
46   // Mask was extended from a boolean vector.
47   Value *ExtMask;
48   if (match(Mask, m_SExt(m_Value(ExtMask))) &&
49       ExtMask->getType()->isIntOrIntVectorTy(1))
50     return ExtMask;
51 
52   return nullptr;
53 }
54 
55 // TODO: If the x86 backend knew how to convert a bool vector mask back to an
56 // XMM register mask efficiently, we could transform all x86 masked intrinsics
57 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
58 static Instruction *simplifyX86MaskedLoad(IntrinsicInst &II, InstCombiner &IC) {
59   Value *Ptr = II.getOperand(0);
60   Value *Mask = II.getOperand(1);
61   Constant *ZeroVec = Constant::getNullValue(II.getType());
62 
63   // Zero Mask - masked load instruction creates a zero vector.
64   if (isa<ConstantAggregateZero>(Mask))
65     return IC.replaceInstUsesWith(II, ZeroVec);
66 
67   // The mask is constant or extended from a bool vector. Convert this x86
68   // intrinsic to the LLVM intrinsic to allow target-independent optimizations.
69   if (Value *BoolMask = getBoolVecFromMask(Mask, IC.getDataLayout())) {
70     // First, cast the x86 intrinsic scalar pointer to a vector pointer to match
71     // the LLVM intrinsic definition for the pointer argument.
72     unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
73     PointerType *VecPtrTy = PointerType::get(II.getType(), AddrSpace);
74     Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
75 
76     // The pass-through vector for an x86 masked load is a zero vector.
77     CallInst *NewMaskedLoad = IC.Builder.CreateMaskedLoad(
78         II.getType(), PtrCast, Align(1), BoolMask, ZeroVec);
79     return IC.replaceInstUsesWith(II, NewMaskedLoad);
80   }
81 
82   return nullptr;
83 }
84 
85 // TODO: If the x86 backend knew how to convert a bool vector mask back to an
86 // XMM register mask efficiently, we could transform all x86 masked intrinsics
87 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
88 static bool simplifyX86MaskedStore(IntrinsicInst &II, InstCombiner &IC) {
89   Value *Ptr = II.getOperand(0);
90   Value *Mask = II.getOperand(1);
91   Value *Vec = II.getOperand(2);
92 
93   // Zero Mask - this masked store instruction does nothing.
94   if (isa<ConstantAggregateZero>(Mask)) {
95     IC.eraseInstFromFunction(II);
96     return true;
97   }
98 
99   // The SSE2 version is too weird (eg, unaligned but non-temporal) to do
100   // anything else at this level.
101   if (II.getIntrinsicID() == Intrinsic::x86_sse2_maskmov_dqu)
102     return false;
103 
104   // The mask is constant or extended from a bool vector. Convert this x86
105   // intrinsic to the LLVM intrinsic to allow target-independent optimizations.
106   if (Value *BoolMask = getBoolVecFromMask(Mask, IC.getDataLayout())) {
107     unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
108     PointerType *VecPtrTy = PointerType::get(Vec->getType(), AddrSpace);
109     Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
110 
111     IC.Builder.CreateMaskedStore(Vec, PtrCast, Align(1), BoolMask);
112 
113     // 'Replace uses' doesn't work for stores. Erase the original masked store.
114     IC.eraseInstFromFunction(II);
115     return true;
116   }
117 
118   return false;
119 }
120 
121 static Value *simplifyX86immShift(const IntrinsicInst &II,
122                                   InstCombiner::BuilderTy &Builder) {
123   bool LogicalShift = false;
124   bool ShiftLeft = false;
125   bool IsImm = false;
126 
127   switch (II.getIntrinsicID()) {
128   default:
129     llvm_unreachable("Unexpected intrinsic!");
130   case Intrinsic::x86_sse2_psrai_d:
131   case Intrinsic::x86_sse2_psrai_w:
132   case Intrinsic::x86_avx2_psrai_d:
133   case Intrinsic::x86_avx2_psrai_w:
134   case Intrinsic::x86_avx512_psrai_q_128:
135   case Intrinsic::x86_avx512_psrai_q_256:
136   case Intrinsic::x86_avx512_psrai_d_512:
137   case Intrinsic::x86_avx512_psrai_q_512:
138   case Intrinsic::x86_avx512_psrai_w_512:
139     IsImm = true;
140     [[fallthrough]];
141   case Intrinsic::x86_sse2_psra_d:
142   case Intrinsic::x86_sse2_psra_w:
143   case Intrinsic::x86_avx2_psra_d:
144   case Intrinsic::x86_avx2_psra_w:
145   case Intrinsic::x86_avx512_psra_q_128:
146   case Intrinsic::x86_avx512_psra_q_256:
147   case Intrinsic::x86_avx512_psra_d_512:
148   case Intrinsic::x86_avx512_psra_q_512:
149   case Intrinsic::x86_avx512_psra_w_512:
150     LogicalShift = false;
151     ShiftLeft = false;
152     break;
153   case Intrinsic::x86_sse2_psrli_d:
154   case Intrinsic::x86_sse2_psrli_q:
155   case Intrinsic::x86_sse2_psrli_w:
156   case Intrinsic::x86_avx2_psrli_d:
157   case Intrinsic::x86_avx2_psrli_q:
158   case Intrinsic::x86_avx2_psrli_w:
159   case Intrinsic::x86_avx512_psrli_d_512:
160   case Intrinsic::x86_avx512_psrli_q_512:
161   case Intrinsic::x86_avx512_psrli_w_512:
162     IsImm = true;
163     [[fallthrough]];
164   case Intrinsic::x86_sse2_psrl_d:
165   case Intrinsic::x86_sse2_psrl_q:
166   case Intrinsic::x86_sse2_psrl_w:
167   case Intrinsic::x86_avx2_psrl_d:
168   case Intrinsic::x86_avx2_psrl_q:
169   case Intrinsic::x86_avx2_psrl_w:
170   case Intrinsic::x86_avx512_psrl_d_512:
171   case Intrinsic::x86_avx512_psrl_q_512:
172   case Intrinsic::x86_avx512_psrl_w_512:
173     LogicalShift = true;
174     ShiftLeft = false;
175     break;
176   case Intrinsic::x86_sse2_pslli_d:
177   case Intrinsic::x86_sse2_pslli_q:
178   case Intrinsic::x86_sse2_pslli_w:
179   case Intrinsic::x86_avx2_pslli_d:
180   case Intrinsic::x86_avx2_pslli_q:
181   case Intrinsic::x86_avx2_pslli_w:
182   case Intrinsic::x86_avx512_pslli_d_512:
183   case Intrinsic::x86_avx512_pslli_q_512:
184   case Intrinsic::x86_avx512_pslli_w_512:
185     IsImm = true;
186     [[fallthrough]];
187   case Intrinsic::x86_sse2_psll_d:
188   case Intrinsic::x86_sse2_psll_q:
189   case Intrinsic::x86_sse2_psll_w:
190   case Intrinsic::x86_avx2_psll_d:
191   case Intrinsic::x86_avx2_psll_q:
192   case Intrinsic::x86_avx2_psll_w:
193   case Intrinsic::x86_avx512_psll_d_512:
194   case Intrinsic::x86_avx512_psll_q_512:
195   case Intrinsic::x86_avx512_psll_w_512:
196     LogicalShift = true;
197     ShiftLeft = true;
198     break;
199   }
200   assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
201 
202   Value *Vec = II.getArgOperand(0);
203   Value *Amt = II.getArgOperand(1);
204   auto *VT = cast<FixedVectorType>(Vec->getType());
205   Type *SVT = VT->getElementType();
206   Type *AmtVT = Amt->getType();
207   unsigned VWidth = VT->getNumElements();
208   unsigned BitWidth = SVT->getPrimitiveSizeInBits();
209 
210   // If the shift amount is guaranteed to be in-range we can replace it with a
211   // generic shift. If its guaranteed to be out of range, logical shifts combine
212   // to zero and arithmetic shifts are clamped to (BitWidth - 1).
213   if (IsImm) {
214     assert(AmtVT->isIntegerTy(32) && "Unexpected shift-by-immediate type");
215     KnownBits KnownAmtBits =
216         llvm::computeKnownBits(Amt, II.getDataLayout());
217     if (KnownAmtBits.getMaxValue().ult(BitWidth)) {
218       Amt = Builder.CreateZExtOrTrunc(Amt, SVT);
219       Amt = Builder.CreateVectorSplat(VWidth, Amt);
220       return (LogicalShift ? (ShiftLeft ? Builder.CreateShl(Vec, Amt)
221                                         : Builder.CreateLShr(Vec, Amt))
222                            : Builder.CreateAShr(Vec, Amt));
223     }
224     if (KnownAmtBits.getMinValue().uge(BitWidth)) {
225       if (LogicalShift)
226         return ConstantAggregateZero::get(VT);
227       Amt = ConstantInt::get(SVT, BitWidth - 1);
228       return Builder.CreateAShr(Vec, Builder.CreateVectorSplat(VWidth, Amt));
229     }
230   } else {
231     // Ensure the first element has an in-range value and the rest of the
232     // elements in the bottom 64 bits are zero.
233     assert(AmtVT->isVectorTy() && AmtVT->getPrimitiveSizeInBits() == 128 &&
234            cast<VectorType>(AmtVT)->getElementType() == SVT &&
235            "Unexpected shift-by-scalar type");
236     unsigned NumAmtElts = cast<FixedVectorType>(AmtVT)->getNumElements();
237     APInt DemandedLower = APInt::getOneBitSet(NumAmtElts, 0);
238     APInt DemandedUpper = APInt::getBitsSet(NumAmtElts, 1, NumAmtElts / 2);
239     KnownBits KnownLowerBits = llvm::computeKnownBits(
240         Amt, DemandedLower, II.getDataLayout());
241     KnownBits KnownUpperBits = llvm::computeKnownBits(
242         Amt, DemandedUpper, II.getDataLayout());
243     if (KnownLowerBits.getMaxValue().ult(BitWidth) &&
244         (DemandedUpper.isZero() || KnownUpperBits.isZero())) {
245       SmallVector<int, 16> ZeroSplat(VWidth, 0);
246       Amt = Builder.CreateShuffleVector(Amt, ZeroSplat);
247       return (LogicalShift ? (ShiftLeft ? Builder.CreateShl(Vec, Amt)
248                                         : Builder.CreateLShr(Vec, Amt))
249                            : Builder.CreateAShr(Vec, Amt));
250     }
251   }
252 
253   // Simplify if count is constant vector.
254   auto *CDV = dyn_cast<ConstantDataVector>(Amt);
255   if (!CDV)
256     return nullptr;
257 
258   // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector
259   // operand to compute the shift amount.
260   assert(AmtVT->isVectorTy() && AmtVT->getPrimitiveSizeInBits() == 128 &&
261          cast<VectorType>(AmtVT)->getElementType() == SVT &&
262          "Unexpected shift-by-scalar type");
263 
264   // Concatenate the sub-elements to create the 64-bit value.
265   APInt Count(64, 0);
266   for (unsigned i = 0, NumSubElts = 64 / BitWidth; i != NumSubElts; ++i) {
267     unsigned SubEltIdx = (NumSubElts - 1) - i;
268     auto *SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx));
269     Count <<= BitWidth;
270     Count |= SubElt->getValue().zextOrTrunc(64);
271   }
272 
273   // If shift-by-zero then just return the original value.
274   if (Count.isZero())
275     return Vec;
276 
277   // Handle cases when Shift >= BitWidth.
278   if (Count.uge(BitWidth)) {
279     // If LogicalShift - just return zero.
280     if (LogicalShift)
281       return ConstantAggregateZero::get(VT);
282 
283     // If ArithmeticShift - clamp Shift to (BitWidth - 1).
284     Count = APInt(64, BitWidth - 1);
285   }
286 
287   // Get a constant vector of the same type as the first operand.
288   auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
289   auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt);
290 
291   if (ShiftLeft)
292     return Builder.CreateShl(Vec, ShiftVec);
293 
294   if (LogicalShift)
295     return Builder.CreateLShr(Vec, ShiftVec);
296 
297   return Builder.CreateAShr(Vec, ShiftVec);
298 }
299 
300 // Attempt to simplify AVX2 per-element shift intrinsics to a generic IR shift.
301 // Unlike the generic IR shifts, the intrinsics have defined behaviour for out
302 // of range shift amounts (logical - set to zero, arithmetic - splat sign bit).
303 static Value *simplifyX86varShift(const IntrinsicInst &II,
304                                   InstCombiner::BuilderTy &Builder) {
305   bool LogicalShift = false;
306   bool ShiftLeft = false;
307 
308   switch (II.getIntrinsicID()) {
309   default:
310     llvm_unreachable("Unexpected intrinsic!");
311   case Intrinsic::x86_avx2_psrav_d:
312   case Intrinsic::x86_avx2_psrav_d_256:
313   case Intrinsic::x86_avx512_psrav_q_128:
314   case Intrinsic::x86_avx512_psrav_q_256:
315   case Intrinsic::x86_avx512_psrav_d_512:
316   case Intrinsic::x86_avx512_psrav_q_512:
317   case Intrinsic::x86_avx512_psrav_w_128:
318   case Intrinsic::x86_avx512_psrav_w_256:
319   case Intrinsic::x86_avx512_psrav_w_512:
320     LogicalShift = false;
321     ShiftLeft = false;
322     break;
323   case Intrinsic::x86_avx2_psrlv_d:
324   case Intrinsic::x86_avx2_psrlv_d_256:
325   case Intrinsic::x86_avx2_psrlv_q:
326   case Intrinsic::x86_avx2_psrlv_q_256:
327   case Intrinsic::x86_avx512_psrlv_d_512:
328   case Intrinsic::x86_avx512_psrlv_q_512:
329   case Intrinsic::x86_avx512_psrlv_w_128:
330   case Intrinsic::x86_avx512_psrlv_w_256:
331   case Intrinsic::x86_avx512_psrlv_w_512:
332     LogicalShift = true;
333     ShiftLeft = false;
334     break;
335   case Intrinsic::x86_avx2_psllv_d:
336   case Intrinsic::x86_avx2_psllv_d_256:
337   case Intrinsic::x86_avx2_psllv_q:
338   case Intrinsic::x86_avx2_psllv_q_256:
339   case Intrinsic::x86_avx512_psllv_d_512:
340   case Intrinsic::x86_avx512_psllv_q_512:
341   case Intrinsic::x86_avx512_psllv_w_128:
342   case Intrinsic::x86_avx512_psllv_w_256:
343   case Intrinsic::x86_avx512_psllv_w_512:
344     LogicalShift = true;
345     ShiftLeft = true;
346     break;
347   }
348   assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
349 
350   Value *Vec = II.getArgOperand(0);
351   Value *Amt = II.getArgOperand(1);
352   auto *VT = cast<FixedVectorType>(II.getType());
353   Type *SVT = VT->getElementType();
354   int NumElts = VT->getNumElements();
355   int BitWidth = SVT->getIntegerBitWidth();
356 
357   // If the shift amount is guaranteed to be in-range we can replace it with a
358   // generic shift.
359   KnownBits KnownAmt =
360       llvm::computeKnownBits(Amt, II.getDataLayout());
361   if (KnownAmt.getMaxValue().ult(BitWidth)) {
362     return (LogicalShift ? (ShiftLeft ? Builder.CreateShl(Vec, Amt)
363                                       : Builder.CreateLShr(Vec, Amt))
364                          : Builder.CreateAShr(Vec, Amt));
365   }
366 
367   // Simplify if all shift amounts are constant/undef.
368   auto *CShift = dyn_cast<Constant>(Amt);
369   if (!CShift)
370     return nullptr;
371 
372   // Collect each element's shift amount.
373   // We also collect special cases: UNDEF = -1, OUT-OF-RANGE = BitWidth.
374   bool AnyOutOfRange = false;
375   SmallVector<int, 8> ShiftAmts;
376   for (int I = 0; I < NumElts; ++I) {
377     auto *CElt = CShift->getAggregateElement(I);
378     if (isa_and_nonnull<UndefValue>(CElt)) {
379       ShiftAmts.push_back(-1);
380       continue;
381     }
382 
383     auto *COp = dyn_cast_or_null<ConstantInt>(CElt);
384     if (!COp)
385       return nullptr;
386 
387     // Handle out of range shifts.
388     // If LogicalShift - set to BitWidth (special case).
389     // If ArithmeticShift - set to (BitWidth - 1) (sign splat).
390     APInt ShiftVal = COp->getValue();
391     if (ShiftVal.uge(BitWidth)) {
392       AnyOutOfRange = LogicalShift;
393       ShiftAmts.push_back(LogicalShift ? BitWidth : BitWidth - 1);
394       continue;
395     }
396 
397     ShiftAmts.push_back((int)ShiftVal.getZExtValue());
398   }
399 
400   // If all elements out of range or UNDEF, return vector of zeros/undefs.
401   // ArithmeticShift should only hit this if they are all UNDEF.
402   auto OutOfRange = [&](int Idx) { return (Idx < 0) || (BitWidth <= Idx); };
403   if (llvm::all_of(ShiftAmts, OutOfRange)) {
404     SmallVector<Constant *, 8> ConstantVec;
405     for (int Idx : ShiftAmts) {
406       if (Idx < 0) {
407         ConstantVec.push_back(UndefValue::get(SVT));
408       } else {
409         assert(LogicalShift && "Logical shift expected");
410         ConstantVec.push_back(ConstantInt::getNullValue(SVT));
411       }
412     }
413     return ConstantVector::get(ConstantVec);
414   }
415 
416   // We can't handle only some out of range values with generic logical shifts.
417   if (AnyOutOfRange)
418     return nullptr;
419 
420   // Build the shift amount constant vector.
421   SmallVector<Constant *, 8> ShiftVecAmts;
422   for (int Idx : ShiftAmts) {
423     if (Idx < 0)
424       ShiftVecAmts.push_back(UndefValue::get(SVT));
425     else
426       ShiftVecAmts.push_back(ConstantInt::get(SVT, Idx));
427   }
428   auto ShiftVec = ConstantVector::get(ShiftVecAmts);
429 
430   if (ShiftLeft)
431     return Builder.CreateShl(Vec, ShiftVec);
432 
433   if (LogicalShift)
434     return Builder.CreateLShr(Vec, ShiftVec);
435 
436   return Builder.CreateAShr(Vec, ShiftVec);
437 }
438 
439 static Value *simplifyX86pack(IntrinsicInst &II,
440                               InstCombiner::BuilderTy &Builder, bool IsSigned) {
441   Value *Arg0 = II.getArgOperand(0);
442   Value *Arg1 = II.getArgOperand(1);
443   Type *ResTy = II.getType();
444 
445   // Fast all undef handling.
446   if (isa<UndefValue>(Arg0) && isa<UndefValue>(Arg1))
447     return UndefValue::get(ResTy);
448 
449   auto *ArgTy = cast<FixedVectorType>(Arg0->getType());
450   unsigned NumLanes = ResTy->getPrimitiveSizeInBits() / 128;
451   unsigned NumSrcElts = ArgTy->getNumElements();
452   assert(cast<FixedVectorType>(ResTy)->getNumElements() == (2 * NumSrcElts) &&
453          "Unexpected packing types");
454 
455   unsigned NumSrcEltsPerLane = NumSrcElts / NumLanes;
456   unsigned DstScalarSizeInBits = ResTy->getScalarSizeInBits();
457   unsigned SrcScalarSizeInBits = ArgTy->getScalarSizeInBits();
458   assert(SrcScalarSizeInBits == (2 * DstScalarSizeInBits) &&
459          "Unexpected packing types");
460 
461   // Constant folding.
462   if (!isa<Constant>(Arg0) || !isa<Constant>(Arg1))
463     return nullptr;
464 
465   // Clamp Values - signed/unsigned both use signed clamp values, but they
466   // differ on the min/max values.
467   APInt MinValue, MaxValue;
468   if (IsSigned) {
469     // PACKSS: Truncate signed value with signed saturation.
470     // Source values less than dst minint are saturated to minint.
471     // Source values greater than dst maxint are saturated to maxint.
472     MinValue =
473         APInt::getSignedMinValue(DstScalarSizeInBits).sext(SrcScalarSizeInBits);
474     MaxValue =
475         APInt::getSignedMaxValue(DstScalarSizeInBits).sext(SrcScalarSizeInBits);
476   } else {
477     // PACKUS: Truncate signed value with unsigned saturation.
478     // Source values less than zero are saturated to zero.
479     // Source values greater than dst maxuint are saturated to maxuint.
480     MinValue = APInt::getZero(SrcScalarSizeInBits);
481     MaxValue = APInt::getLowBitsSet(SrcScalarSizeInBits, DstScalarSizeInBits);
482   }
483 
484   auto *MinC = Constant::getIntegerValue(ArgTy, MinValue);
485   auto *MaxC = Constant::getIntegerValue(ArgTy, MaxValue);
486   Arg0 = Builder.CreateSelect(Builder.CreateICmpSLT(Arg0, MinC), MinC, Arg0);
487   Arg1 = Builder.CreateSelect(Builder.CreateICmpSLT(Arg1, MinC), MinC, Arg1);
488   Arg0 = Builder.CreateSelect(Builder.CreateICmpSGT(Arg0, MaxC), MaxC, Arg0);
489   Arg1 = Builder.CreateSelect(Builder.CreateICmpSGT(Arg1, MaxC), MaxC, Arg1);
490 
491   // Shuffle clamped args together at the lane level.
492   SmallVector<int, 32> PackMask;
493   for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
494     for (unsigned Elt = 0; Elt != NumSrcEltsPerLane; ++Elt)
495       PackMask.push_back(Elt + (Lane * NumSrcEltsPerLane));
496     for (unsigned Elt = 0; Elt != NumSrcEltsPerLane; ++Elt)
497       PackMask.push_back(Elt + (Lane * NumSrcEltsPerLane) + NumSrcElts);
498   }
499   auto *Shuffle = Builder.CreateShuffleVector(Arg0, Arg1, PackMask);
500 
501   // Truncate to dst size.
502   return Builder.CreateTrunc(Shuffle, ResTy);
503 }
504 
505 static Value *simplifyX86pmulh(IntrinsicInst &II,
506                                InstCombiner::BuilderTy &Builder, bool IsSigned,
507                                bool IsRounding) {
508   Value *Arg0 = II.getArgOperand(0);
509   Value *Arg1 = II.getArgOperand(1);
510   auto *ResTy = cast<FixedVectorType>(II.getType());
511   auto *ArgTy = cast<FixedVectorType>(Arg0->getType());
512   assert(ArgTy == ResTy && ResTy->getScalarSizeInBits() == 16 &&
513          "Unexpected PMULH types");
514   assert((!IsRounding || IsSigned) && "PMULHRS instruction must be signed");
515 
516   // Multiply by undef -> zero (NOT undef!) as other arg could still be zero.
517   if (isa<UndefValue>(Arg0) || isa<UndefValue>(Arg1))
518     return ConstantAggregateZero::get(ResTy);
519 
520   // Multiply by zero.
521   if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1))
522     return ConstantAggregateZero::get(ResTy);
523 
524   // Multiply by one.
525   if (!IsRounding) {
526     if (match(Arg0, m_One()))
527       return IsSigned ? Builder.CreateAShr(Arg1, 15)
528                       : ConstantAggregateZero::get(ResTy);
529     if (match(Arg1, m_One()))
530       return IsSigned ? Builder.CreateAShr(Arg0, 15)
531                       : ConstantAggregateZero::get(ResTy);
532   }
533 
534   // Constant folding.
535   if (!isa<Constant>(Arg0) || !isa<Constant>(Arg1))
536     return nullptr;
537 
538   // Extend to twice the width and multiply.
539   auto Cast =
540       IsSigned ? Instruction::CastOps::SExt : Instruction::CastOps::ZExt;
541   auto *ExtTy = FixedVectorType::getExtendedElementVectorType(ArgTy);
542   Value *LHS = Builder.CreateCast(Cast, Arg0, ExtTy);
543   Value *RHS = Builder.CreateCast(Cast, Arg1, ExtTy);
544   Value *Mul = Builder.CreateMul(LHS, RHS);
545 
546   if (IsRounding) {
547     // PMULHRSW: truncate to vXi18 of the most significant bits, add one and
548     // extract bits[16:1].
549     auto *RndEltTy = IntegerType::get(ExtTy->getContext(), 18);
550     auto *RndTy = FixedVectorType::get(RndEltTy, ExtTy);
551     Mul = Builder.CreateLShr(Mul, 14);
552     Mul = Builder.CreateTrunc(Mul, RndTy);
553     Mul = Builder.CreateAdd(Mul, ConstantInt::get(RndTy, 1));
554     Mul = Builder.CreateLShr(Mul, 1);
555   } else {
556     // PMULH/PMULHU: extract the vXi16 most significant bits.
557     Mul = Builder.CreateLShr(Mul, 16);
558   }
559 
560   return Builder.CreateTrunc(Mul, ResTy);
561 }
562 
563 static Value *simplifyX86pmadd(IntrinsicInst &II,
564                                InstCombiner::BuilderTy &Builder,
565                                bool IsPMADDWD) {
566   Value *Arg0 = II.getArgOperand(0);
567   Value *Arg1 = II.getArgOperand(1);
568   auto *ResTy = cast<FixedVectorType>(II.getType());
569   [[maybe_unused]] auto *ArgTy = cast<FixedVectorType>(Arg0->getType());
570 
571   unsigned NumDstElts = ResTy->getNumElements();
572   assert(ArgTy->getNumElements() == (2 * NumDstElts) &&
573          ResTy->getScalarSizeInBits() == (2 * ArgTy->getScalarSizeInBits()) &&
574          "Unexpected PMADD types");
575 
576   // Multiply by undef -> zero (NOT undef!) as other arg could still be zero.
577   if (isa<UndefValue>(Arg0) || isa<UndefValue>(Arg1))
578     return ConstantAggregateZero::get(ResTy);
579 
580   // Multiply by zero.
581   if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1))
582     return ConstantAggregateZero::get(ResTy);
583 
584   // Constant folding.
585   if (!isa<Constant>(Arg0) || !isa<Constant>(Arg1))
586     return nullptr;
587 
588   // Split Lo/Hi elements pairs, extend and add together.
589   // PMADDWD(X,Y) =
590   // add(mul(sext(lhs[0]),sext(rhs[0])),mul(sext(lhs[1]),sext(rhs[1])))
591   // PMADDUBSW(X,Y) =
592   // sadd_sat(mul(zext(lhs[0]),sext(rhs[0])),mul(zext(lhs[1]),sext(rhs[1])))
593   SmallVector<int> LoMask, HiMask;
594   for (unsigned I = 0; I != NumDstElts; ++I) {
595     LoMask.push_back(2 * I + 0);
596     HiMask.push_back(2 * I + 1);
597   }
598 
599   auto *LHSLo = Builder.CreateShuffleVector(Arg0, LoMask);
600   auto *LHSHi = Builder.CreateShuffleVector(Arg0, HiMask);
601   auto *RHSLo = Builder.CreateShuffleVector(Arg1, LoMask);
602   auto *RHSHi = Builder.CreateShuffleVector(Arg1, HiMask);
603 
604   auto LHSCast =
605       IsPMADDWD ? Instruction::CastOps::SExt : Instruction::CastOps::ZExt;
606   LHSLo = Builder.CreateCast(LHSCast, LHSLo, ResTy);
607   LHSHi = Builder.CreateCast(LHSCast, LHSHi, ResTy);
608   RHSLo = Builder.CreateCast(Instruction::CastOps::SExt, RHSLo, ResTy);
609   RHSHi = Builder.CreateCast(Instruction::CastOps::SExt, RHSHi, ResTy);
610   Value *Lo = Builder.CreateMul(LHSLo, RHSLo);
611   Value *Hi = Builder.CreateMul(LHSHi, RHSHi);
612   return IsPMADDWD
613              ? Builder.CreateAdd(Lo, Hi)
614              : Builder.CreateIntrinsic(ResTy, Intrinsic::sadd_sat, {Lo, Hi});
615 }
616 
617 static Value *simplifyX86movmsk(const IntrinsicInst &II,
618                                 InstCombiner::BuilderTy &Builder) {
619   Value *Arg = II.getArgOperand(0);
620   Type *ResTy = II.getType();
621 
622   // movmsk(undef) -> zero as we must ensure the upper bits are zero.
623   if (isa<UndefValue>(Arg))
624     return Constant::getNullValue(ResTy);
625 
626   auto *ArgTy = dyn_cast<FixedVectorType>(Arg->getType());
627   // We can't easily peek through x86_mmx types.
628   if (!ArgTy)
629     return nullptr;
630 
631   // Expand MOVMSK to compare/bitcast/zext:
632   // e.g. PMOVMSKB(v16i8 x):
633   // %cmp = icmp slt <16 x i8> %x, zeroinitializer
634   // %int = bitcast <16 x i1> %cmp to i16
635   // %res = zext i16 %int to i32
636   unsigned NumElts = ArgTy->getNumElements();
637   Type *IntegerTy = Builder.getIntNTy(NumElts);
638 
639   Value *Res = Builder.CreateBitCast(Arg, VectorType::getInteger(ArgTy));
640   Res = Builder.CreateIsNeg(Res);
641   Res = Builder.CreateBitCast(Res, IntegerTy);
642   Res = Builder.CreateZExtOrTrunc(Res, ResTy);
643   return Res;
644 }
645 
646 static Value *simplifyX86addcarry(const IntrinsicInst &II,
647                                   InstCombiner::BuilderTy &Builder) {
648   Value *CarryIn = II.getArgOperand(0);
649   Value *Op1 = II.getArgOperand(1);
650   Value *Op2 = II.getArgOperand(2);
651   Type *RetTy = II.getType();
652   Type *OpTy = Op1->getType();
653   assert(RetTy->getStructElementType(0)->isIntegerTy(8) &&
654          RetTy->getStructElementType(1) == OpTy && OpTy == Op2->getType() &&
655          "Unexpected types for x86 addcarry");
656 
657   // If carry-in is zero, this is just an unsigned add with overflow.
658   if (match(CarryIn, m_ZeroInt())) {
659     Value *UAdd = Builder.CreateIntrinsic(Intrinsic::uadd_with_overflow, OpTy,
660                                           {Op1, Op2});
661     // The types have to be adjusted to match the x86 call types.
662     Value *UAddResult = Builder.CreateExtractValue(UAdd, 0);
663     Value *UAddOV = Builder.CreateZExt(Builder.CreateExtractValue(UAdd, 1),
664                                        Builder.getInt8Ty());
665     Value *Res = PoisonValue::get(RetTy);
666     Res = Builder.CreateInsertValue(Res, UAddOV, 0);
667     return Builder.CreateInsertValue(Res, UAddResult, 1);
668   }
669 
670   return nullptr;
671 }
672 
673 static Value *simplifyTernarylogic(const IntrinsicInst &II,
674                                    InstCombiner::BuilderTy &Builder) {
675 
676   auto *ArgImm = dyn_cast<ConstantInt>(II.getArgOperand(3));
677   if (!ArgImm || ArgImm->getValue().uge(256))
678     return nullptr;
679 
680   Value *ArgA = II.getArgOperand(0);
681   Value *ArgB = II.getArgOperand(1);
682   Value *ArgC = II.getArgOperand(2);
683 
684   Type *Ty = II.getType();
685 
686   auto Or = [&](auto Lhs, auto Rhs) -> std::pair<Value *, uint8_t> {
687     return {Builder.CreateOr(Lhs.first, Rhs.first), Lhs.second | Rhs.second};
688   };
689   auto Xor = [&](auto Lhs, auto Rhs) -> std::pair<Value *, uint8_t> {
690     return {Builder.CreateXor(Lhs.first, Rhs.first), Lhs.second ^ Rhs.second};
691   };
692   auto And = [&](auto Lhs, auto Rhs) -> std::pair<Value *, uint8_t> {
693     return {Builder.CreateAnd(Lhs.first, Rhs.first), Lhs.second & Rhs.second};
694   };
695   auto Not = [&](auto V) -> std::pair<Value *, uint8_t> {
696     return {Builder.CreateNot(V.first), ~V.second};
697   };
698   auto Nor = [&](auto Lhs, auto Rhs) { return Not(Or(Lhs, Rhs)); };
699   auto Xnor = [&](auto Lhs, auto Rhs) { return Not(Xor(Lhs, Rhs)); };
700   auto Nand = [&](auto Lhs, auto Rhs) { return Not(And(Lhs, Rhs)); };
701 
702   bool AIsConst = match(ArgA, m_ImmConstant());
703   bool BIsConst = match(ArgB, m_ImmConstant());
704   bool CIsConst = match(ArgC, m_ImmConstant());
705 
706   bool ABIsConst = AIsConst && BIsConst;
707   bool ACIsConst = AIsConst && CIsConst;
708   bool BCIsConst = BIsConst && CIsConst;
709   bool ABCIsConst = AIsConst && BIsConst && CIsConst;
710 
711   // Use for verification. Its a big table. Its difficult to go from Imm ->
712   // logic ops, but easy to verify that a set of logic ops is correct. We track
713   // the logic ops through the second value in the pair. At the end it should
714   // equal Imm.
715   std::pair<Value *, uint8_t> A = {ArgA, 0xf0};
716   std::pair<Value *, uint8_t> B = {ArgB, 0xcc};
717   std::pair<Value *, uint8_t> C = {ArgC, 0xaa};
718   std::pair<Value *, uint8_t> Res = {nullptr, 0};
719 
720   // Currently we only handle cases that convert directly to another instruction
721   // or cases where all the ops are constant.  This is because we don't properly
722   // handle creating ternary ops in the backend, so splitting them here may
723   // cause regressions. As the backend improves, uncomment more cases.
724 
725   uint8_t Imm = ArgImm->getValue().getZExtValue();
726   switch (Imm) {
727   case 0x0:
728     Res = {Constant::getNullValue(Ty), 0};
729     break;
730   case 0x1:
731     if (ABCIsConst)
732       Res = Nor(Or(A, B), C);
733     break;
734   case 0x2:
735     if (ABCIsConst)
736       Res = And(Nor(A, B), C);
737     break;
738   case 0x3:
739     if (ABIsConst)
740       Res = Nor(A, B);
741     break;
742   case 0x4:
743     if (ABCIsConst)
744       Res = And(Nor(A, C), B);
745     break;
746   case 0x5:
747     if (ACIsConst)
748       Res = Nor(A, C);
749     break;
750   case 0x6:
751     if (ABCIsConst)
752       Res = Nor(A, Xnor(B, C));
753     break;
754   case 0x7:
755     if (ABCIsConst)
756       Res = Nor(A, And(B, C));
757     break;
758   case 0x8:
759     if (ABCIsConst)
760       Res = Nor(A, Nand(B, C));
761     break;
762   case 0x9:
763     if (ABCIsConst)
764       Res = Nor(A, Xor(B, C));
765     break;
766   case 0xa:
767     if (ACIsConst)
768       Res = Nor(A, Not(C));
769     break;
770   case 0xb:
771     if (ABCIsConst)
772       Res = Nor(A, Nor(C, Not(B)));
773     break;
774   case 0xc:
775     if (ABIsConst)
776       Res = Nor(A, Not(B));
777     break;
778   case 0xd:
779     if (ABCIsConst)
780       Res = Nor(A, Nor(B, Not(C)));
781     break;
782   case 0xe:
783     if (ABCIsConst)
784       Res = Nor(A, Nor(B, C));
785     break;
786   case 0xf:
787     Res = Not(A);
788     break;
789   case 0x10:
790     if (ABCIsConst)
791       Res = And(A, Nor(B, C));
792     break;
793   case 0x11:
794     if (BCIsConst)
795       Res = Nor(B, C);
796     break;
797   case 0x12:
798     if (ABCIsConst)
799       Res = Nor(Xnor(A, C), B);
800     break;
801   case 0x13:
802     if (ABCIsConst)
803       Res = Nor(And(A, C), B);
804     break;
805   case 0x14:
806     if (ABCIsConst)
807       Res = Nor(Xnor(A, B), C);
808     break;
809   case 0x15:
810     if (ABCIsConst)
811       Res = Nor(And(A, B), C);
812     break;
813   case 0x16:
814     if (ABCIsConst)
815       Res = Xor(Xor(A, B), And(Nand(A, B), C));
816     break;
817   case 0x17:
818     if (ABCIsConst)
819       Res = Xor(Or(A, B), Or(Xnor(A, B), C));
820     break;
821   case 0x18:
822     if (ABCIsConst)
823       Res = Nor(Xnor(A, B), Xnor(A, C));
824     break;
825   case 0x19:
826     if (ABCIsConst)
827       Res = And(Nand(A, B), Xnor(B, C));
828     break;
829   case 0x1a:
830     if (ABCIsConst)
831       Res = Xor(A, Or(And(A, B), C));
832     break;
833   case 0x1b:
834     if (ABCIsConst)
835       Res = Xor(A, Or(Xnor(A, B), C));
836     break;
837   case 0x1c:
838     if (ABCIsConst)
839       Res = Xor(A, Or(And(A, C), B));
840     break;
841   case 0x1d:
842     if (ABCIsConst)
843       Res = Xor(A, Or(Xnor(A, C), B));
844     break;
845   case 0x1e:
846     if (ABCIsConst)
847       Res = Xor(A, Or(B, C));
848     break;
849   case 0x1f:
850     if (ABCIsConst)
851       Res = Nand(A, Or(B, C));
852     break;
853   case 0x20:
854     if (ABCIsConst)
855       Res = Nor(Nand(A, C), B);
856     break;
857   case 0x21:
858     if (ABCIsConst)
859       Res = Nor(Xor(A, C), B);
860     break;
861   case 0x22:
862     if (BCIsConst)
863       Res = Nor(B, Not(C));
864     break;
865   case 0x23:
866     if (ABCIsConst)
867       Res = Nor(B, Nor(C, Not(A)));
868     break;
869   case 0x24:
870     if (ABCIsConst)
871       Res = Nor(Xnor(A, B), Xor(A, C));
872     break;
873   case 0x25:
874     if (ABCIsConst)
875       Res = Xor(A, Nand(Nand(A, B), C));
876     break;
877   case 0x26:
878     if (ABCIsConst)
879       Res = And(Nand(A, B), Xor(B, C));
880     break;
881   case 0x27:
882     if (ABCIsConst)
883       Res = Xor(Or(Xnor(A, B), C), B);
884     break;
885   case 0x28:
886     if (ABCIsConst)
887       Res = And(Xor(A, B), C);
888     break;
889   case 0x29:
890     if (ABCIsConst)
891       Res = Xor(Xor(A, B), Nor(And(A, B), C));
892     break;
893   case 0x2a:
894     if (ABCIsConst)
895       Res = And(Nand(A, B), C);
896     break;
897   case 0x2b:
898     if (ABCIsConst)
899       Res = Xor(Or(Xnor(A, B), Xor(A, C)), A);
900     break;
901   case 0x2c:
902     if (ABCIsConst)
903       Res = Nor(Xnor(A, B), Nor(B, C));
904     break;
905   case 0x2d:
906     if (ABCIsConst)
907       Res = Xor(A, Or(B, Not(C)));
908     break;
909   case 0x2e:
910     if (ABCIsConst)
911       Res = Xor(A, Or(Xor(A, C), B));
912     break;
913   case 0x2f:
914     if (ABCIsConst)
915       Res = Nand(A, Or(B, Not(C)));
916     break;
917   case 0x30:
918     if (ABIsConst)
919       Res = Nor(B, Not(A));
920     break;
921   case 0x31:
922     if (ABCIsConst)
923       Res = Nor(Nor(A, Not(C)), B);
924     break;
925   case 0x32:
926     if (ABCIsConst)
927       Res = Nor(Nor(A, C), B);
928     break;
929   case 0x33:
930     Res = Not(B);
931     break;
932   case 0x34:
933     if (ABCIsConst)
934       Res = And(Xor(A, B), Nand(B, C));
935     break;
936   case 0x35:
937     if (ABCIsConst)
938       Res = Xor(B, Or(A, Xnor(B, C)));
939     break;
940   case 0x36:
941     if (ABCIsConst)
942       Res = Xor(Or(A, C), B);
943     break;
944   case 0x37:
945     if (ABCIsConst)
946       Res = Nand(Or(A, C), B);
947     break;
948   case 0x38:
949     if (ABCIsConst)
950       Res = Nor(Xnor(A, B), Nor(A, C));
951     break;
952   case 0x39:
953     if (ABCIsConst)
954       Res = Xor(Or(A, Not(C)), B);
955     break;
956   case 0x3a:
957     if (ABCIsConst)
958       Res = Xor(B, Or(A, Xor(B, C)));
959     break;
960   case 0x3b:
961     if (ABCIsConst)
962       Res = Nand(Or(A, Not(C)), B);
963     break;
964   case 0x3c:
965     Res = Xor(A, B);
966     break;
967   case 0x3d:
968     if (ABCIsConst)
969       Res = Xor(A, Or(Nor(A, C), B));
970     break;
971   case 0x3e:
972     if (ABCIsConst)
973       Res = Xor(A, Or(Nor(A, Not(C)), B));
974     break;
975   case 0x3f:
976     if (ABIsConst)
977       Res = Nand(A, B);
978     break;
979   case 0x40:
980     if (ABCIsConst)
981       Res = Nor(Nand(A, B), C);
982     break;
983   case 0x41:
984     if (ABCIsConst)
985       Res = Nor(Xor(A, B), C);
986     break;
987   case 0x42:
988     if (ABCIsConst)
989       Res = Nor(Xor(A, B), Xnor(A, C));
990     break;
991   case 0x43:
992     if (ABCIsConst)
993       Res = Xor(A, Nand(Nand(A, C), B));
994     break;
995   case 0x44:
996     if (BCIsConst)
997       Res = Nor(C, Not(B));
998     break;
999   case 0x45:
1000     if (ABCIsConst)
1001       Res = Nor(Nor(B, Not(A)), C);
1002     break;
1003   case 0x46:
1004     if (ABCIsConst)
1005       Res = Xor(Or(And(A, C), B), C);
1006     break;
1007   case 0x47:
1008     if (ABCIsConst)
1009       Res = Xor(Or(Xnor(A, C), B), C);
1010     break;
1011   case 0x48:
1012     if (ABCIsConst)
1013       Res = And(Xor(A, C), B);
1014     break;
1015   case 0x49:
1016     if (ABCIsConst)
1017       Res = Xor(Or(Xnor(A, B), And(A, C)), C);
1018     break;
1019   case 0x4a:
1020     if (ABCIsConst)
1021       Res = Nor(Xnor(A, C), Nor(B, C));
1022     break;
1023   case 0x4b:
1024     if (ABCIsConst)
1025       Res = Xor(A, Or(C, Not(B)));
1026     break;
1027   case 0x4c:
1028     if (ABCIsConst)
1029       Res = And(Nand(A, C), B);
1030     break;
1031   case 0x4d:
1032     if (ABCIsConst)
1033       Res = Xor(Or(Xor(A, B), Xnor(A, C)), A);
1034     break;
1035   case 0x4e:
1036     if (ABCIsConst)
1037       Res = Xor(A, Or(Xor(A, B), C));
1038     break;
1039   case 0x4f:
1040     if (ABCIsConst)
1041       Res = Nand(A, Nand(B, Not(C)));
1042     break;
1043   case 0x50:
1044     if (ACIsConst)
1045       Res = Nor(C, Not(A));
1046     break;
1047   case 0x51:
1048     if (ABCIsConst)
1049       Res = Nor(Nor(A, Not(B)), C);
1050     break;
1051   case 0x52:
1052     if (ABCIsConst)
1053       Res = And(Xor(A, C), Nand(B, C));
1054     break;
1055   case 0x53:
1056     if (ABCIsConst)
1057       Res = Xor(Or(Xnor(B, C), A), C);
1058     break;
1059   case 0x54:
1060     if (ABCIsConst)
1061       Res = Nor(Nor(A, B), C);
1062     break;
1063   case 0x55:
1064     Res = Not(C);
1065     break;
1066   case 0x56:
1067     if (ABCIsConst)
1068       Res = Xor(Or(A, B), C);
1069     break;
1070   case 0x57:
1071     if (ABCIsConst)
1072       Res = Nand(Or(A, B), C);
1073     break;
1074   case 0x58:
1075     if (ABCIsConst)
1076       Res = Nor(Nor(A, B), Xnor(A, C));
1077     break;
1078   case 0x59:
1079     if (ABCIsConst)
1080       Res = Xor(Or(A, Not(B)), C);
1081     break;
1082   case 0x5a:
1083     Res = Xor(A, C);
1084     break;
1085   case 0x5b:
1086     if (ABCIsConst)
1087       Res = Xor(A, Or(Nor(A, B), C));
1088     break;
1089   case 0x5c:
1090     if (ABCIsConst)
1091       Res = Xor(Or(Xor(B, C), A), C);
1092     break;
1093   case 0x5d:
1094     if (ABCIsConst)
1095       Res = Nand(Or(A, Not(B)), C);
1096     break;
1097   case 0x5e:
1098     if (ABCIsConst)
1099       Res = Xor(A, Or(Nor(A, Not(B)), C));
1100     break;
1101   case 0x5f:
1102     if (ACIsConst)
1103       Res = Nand(A, C);
1104     break;
1105   case 0x60:
1106     if (ABCIsConst)
1107       Res = And(A, Xor(B, C));
1108     break;
1109   case 0x61:
1110     if (ABCIsConst)
1111       Res = Xor(Or(Xnor(A, B), And(B, C)), C);
1112     break;
1113   case 0x62:
1114     if (ABCIsConst)
1115       Res = Nor(Nor(A, C), Xnor(B, C));
1116     break;
1117   case 0x63:
1118     if (ABCIsConst)
1119       Res = Xor(B, Or(C, Not(A)));
1120     break;
1121   case 0x64:
1122     if (ABCIsConst)
1123       Res = Nor(Nor(A, B), Xnor(B, C));
1124     break;
1125   case 0x65:
1126     if (ABCIsConst)
1127       Res = Xor(Or(B, Not(A)), C);
1128     break;
1129   case 0x66:
1130     Res = Xor(B, C);
1131     break;
1132   case 0x67:
1133     if (ABCIsConst)
1134       Res = Or(Nor(A, B), Xor(B, C));
1135     break;
1136   case 0x68:
1137     if (ABCIsConst)
1138       Res = Xor(Xor(A, B), Nor(Nor(A, B), C));
1139     break;
1140   case 0x69:
1141     if (ABCIsConst)
1142       Res = Xor(Xnor(A, B), C);
1143     break;
1144   case 0x6a:
1145     if (ABCIsConst)
1146       Res = Xor(And(A, B), C);
1147     break;
1148   case 0x6b:
1149     if (ABCIsConst)
1150       Res = Or(Nor(A, B), Xor(Xnor(A, B), C));
1151     break;
1152   case 0x6c:
1153     if (ABCIsConst)
1154       Res = Xor(And(A, C), B);
1155     break;
1156   case 0x6d:
1157     if (ABCIsConst)
1158       Res = Xor(Or(Xnor(A, B), Nor(A, C)), C);
1159     break;
1160   case 0x6e:
1161     if (ABCIsConst)
1162       Res = Or(Nor(A, Not(B)), Xor(B, C));
1163     break;
1164   case 0x6f:
1165     if (ABCIsConst)
1166       Res = Nand(A, Xnor(B, C));
1167     break;
1168   case 0x70:
1169     if (ABCIsConst)
1170       Res = And(A, Nand(B, C));
1171     break;
1172   case 0x71:
1173     if (ABCIsConst)
1174       Res = Xor(Nor(Xor(A, B), Xor(A, C)), A);
1175     break;
1176   case 0x72:
1177     if (ABCIsConst)
1178       Res = Xor(Or(Xor(A, B), C), B);
1179     break;
1180   case 0x73:
1181     if (ABCIsConst)
1182       Res = Nand(Nand(A, Not(C)), B);
1183     break;
1184   case 0x74:
1185     if (ABCIsConst)
1186       Res = Xor(Or(Xor(A, C), B), C);
1187     break;
1188   case 0x75:
1189     if (ABCIsConst)
1190       Res = Nand(Nand(A, Not(B)), C);
1191     break;
1192   case 0x76:
1193     if (ABCIsConst)
1194       Res = Xor(B, Or(Nor(B, Not(A)), C));
1195     break;
1196   case 0x77:
1197     if (BCIsConst)
1198       Res = Nand(B, C);
1199     break;
1200   case 0x78:
1201     if (ABCIsConst)
1202       Res = Xor(A, And(B, C));
1203     break;
1204   case 0x79:
1205     if (ABCIsConst)
1206       Res = Xor(Or(Xnor(A, B), Nor(B, C)), C);
1207     break;
1208   case 0x7a:
1209     if (ABCIsConst)
1210       Res = Or(Xor(A, C), Nor(B, Not(A)));
1211     break;
1212   case 0x7b:
1213     if (ABCIsConst)
1214       Res = Nand(Xnor(A, C), B);
1215     break;
1216   case 0x7c:
1217     if (ABCIsConst)
1218       Res = Or(Xor(A, B), Nor(C, Not(A)));
1219     break;
1220   case 0x7d:
1221     if (ABCIsConst)
1222       Res = Nand(Xnor(A, B), C);
1223     break;
1224   case 0x7e:
1225     if (ABCIsConst)
1226       Res = Or(Xor(A, B), Xor(A, C));
1227     break;
1228   case 0x7f:
1229     if (ABCIsConst)
1230       Res = Nand(And(A, B), C);
1231     break;
1232   case 0x80:
1233     if (ABCIsConst)
1234       Res = And(And(A, B), C);
1235     break;
1236   case 0x81:
1237     if (ABCIsConst)
1238       Res = Nor(Xor(A, B), Xor(A, C));
1239     break;
1240   case 0x82:
1241     if (ABCIsConst)
1242       Res = And(Xnor(A, B), C);
1243     break;
1244   case 0x83:
1245     if (ABCIsConst)
1246       Res = Nor(Xor(A, B), Nor(C, Not(A)));
1247     break;
1248   case 0x84:
1249     if (ABCIsConst)
1250       Res = And(Xnor(A, C), B);
1251     break;
1252   case 0x85:
1253     if (ABCIsConst)
1254       Res = Nor(Xor(A, C), Nor(B, Not(A)));
1255     break;
1256   case 0x86:
1257     if (ABCIsConst)
1258       Res = Xor(Nor(Xnor(A, B), Nor(B, C)), C);
1259     break;
1260   case 0x87:
1261     if (ABCIsConst)
1262       Res = Xor(A, Nand(B, C));
1263     break;
1264   case 0x88:
1265     Res = And(B, C);
1266     break;
1267   case 0x89:
1268     if (ABCIsConst)
1269       Res = Xor(B, Nor(Nor(B, Not(A)), C));
1270     break;
1271   case 0x8a:
1272     if (ABCIsConst)
1273       Res = And(Nand(A, Not(B)), C);
1274     break;
1275   case 0x8b:
1276     if (ABCIsConst)
1277       Res = Xor(Nor(Xor(A, C), B), C);
1278     break;
1279   case 0x8c:
1280     if (ABCIsConst)
1281       Res = And(Nand(A, Not(C)), B);
1282     break;
1283   case 0x8d:
1284     if (ABCIsConst)
1285       Res = Xor(Nor(Xor(A, B), C), B);
1286     break;
1287   case 0x8e:
1288     if (ABCIsConst)
1289       Res = Xor(Or(Xor(A, B), Xor(A, C)), A);
1290     break;
1291   case 0x8f:
1292     if (ABCIsConst)
1293       Res = Nand(A, Nand(B, C));
1294     break;
1295   case 0x90:
1296     if (ABCIsConst)
1297       Res = And(A, Xnor(B, C));
1298     break;
1299   case 0x91:
1300     if (ABCIsConst)
1301       Res = Nor(Nor(A, Not(B)), Xor(B, C));
1302     break;
1303   case 0x92:
1304     if (ABCIsConst)
1305       Res = Xor(Nor(Xnor(A, B), Nor(A, C)), C);
1306     break;
1307   case 0x93:
1308     if (ABCIsConst)
1309       Res = Xor(Nand(A, C), B);
1310     break;
1311   case 0x94:
1312     if (ABCIsConst)
1313       Res = Nor(Nor(A, B), Xor(Xnor(A, B), C));
1314     break;
1315   case 0x95:
1316     if (ABCIsConst)
1317       Res = Xor(Nand(A, B), C);
1318     break;
1319   case 0x96:
1320     if (ABCIsConst)
1321       Res = Xor(Xor(A, B), C);
1322     break;
1323   case 0x97:
1324     if (ABCIsConst)
1325       Res = Xor(Xor(A, B), Or(Nor(A, B), C));
1326     break;
1327   case 0x98:
1328     if (ABCIsConst)
1329       Res = Nor(Nor(A, B), Xor(B, C));
1330     break;
1331   case 0x99:
1332     if (BCIsConst)
1333       Res = Xnor(B, C);
1334     break;
1335   case 0x9a:
1336     if (ABCIsConst)
1337       Res = Xor(Nor(B, Not(A)), C);
1338     break;
1339   case 0x9b:
1340     if (ABCIsConst)
1341       Res = Or(Nor(A, B), Xnor(B, C));
1342     break;
1343   case 0x9c:
1344     if (ABCIsConst)
1345       Res = Xor(B, Nor(C, Not(A)));
1346     break;
1347   case 0x9d:
1348     if (ABCIsConst)
1349       Res = Or(Nor(A, C), Xnor(B, C));
1350     break;
1351   case 0x9e:
1352     if (ABCIsConst)
1353       Res = Xor(And(Xor(A, B), Nand(B, C)), C);
1354     break;
1355   case 0x9f:
1356     if (ABCIsConst)
1357       Res = Nand(A, Xor(B, C));
1358     break;
1359   case 0xa0:
1360     Res = And(A, C);
1361     break;
1362   case 0xa1:
1363     if (ABCIsConst)
1364       Res = Xor(A, Nor(Nor(A, Not(B)), C));
1365     break;
1366   case 0xa2:
1367     if (ABCIsConst)
1368       Res = And(Or(A, Not(B)), C);
1369     break;
1370   case 0xa3:
1371     if (ABCIsConst)
1372       Res = Xor(Nor(Xor(B, C), A), C);
1373     break;
1374   case 0xa4:
1375     if (ABCIsConst)
1376       Res = Xor(A, Nor(Nor(A, B), C));
1377     break;
1378   case 0xa5:
1379     if (ACIsConst)
1380       Res = Xnor(A, C);
1381     break;
1382   case 0xa6:
1383     if (ABCIsConst)
1384       Res = Xor(Nor(A, Not(B)), C);
1385     break;
1386   case 0xa7:
1387     if (ABCIsConst)
1388       Res = Or(Nor(A, B), Xnor(A, C));
1389     break;
1390   case 0xa8:
1391     if (ABCIsConst)
1392       Res = And(Or(A, B), C);
1393     break;
1394   case 0xa9:
1395     if (ABCIsConst)
1396       Res = Xor(Nor(A, B), C);
1397     break;
1398   case 0xaa:
1399     Res = C;
1400     break;
1401   case 0xab:
1402     if (ABCIsConst)
1403       Res = Or(Nor(A, B), C);
1404     break;
1405   case 0xac:
1406     if (ABCIsConst)
1407       Res = Xor(Nor(Xnor(B, C), A), C);
1408     break;
1409   case 0xad:
1410     if (ABCIsConst)
1411       Res = Or(Xnor(A, C), And(B, C));
1412     break;
1413   case 0xae:
1414     if (ABCIsConst)
1415       Res = Or(Nor(A, Not(B)), C);
1416     break;
1417   case 0xaf:
1418     if (ACIsConst)
1419       Res = Or(C, Not(A));
1420     break;
1421   case 0xb0:
1422     if (ABCIsConst)
1423       Res = And(A, Nand(B, Not(C)));
1424     break;
1425   case 0xb1:
1426     if (ABCIsConst)
1427       Res = Xor(A, Nor(Xor(A, B), C));
1428     break;
1429   case 0xb2:
1430     if (ABCIsConst)
1431       Res = Xor(Nor(Xor(A, B), Xnor(A, C)), A);
1432     break;
1433   case 0xb3:
1434     if (ABCIsConst)
1435       Res = Nand(Nand(A, C), B);
1436     break;
1437   case 0xb4:
1438     if (ABCIsConst)
1439       Res = Xor(A, Nor(C, Not(B)));
1440     break;
1441   case 0xb5:
1442     if (ABCIsConst)
1443       Res = Or(Xnor(A, C), Nor(B, C));
1444     break;
1445   case 0xb6:
1446     if (ABCIsConst)
1447       Res = Xor(And(Xor(A, B), Nand(A, C)), C);
1448     break;
1449   case 0xb7:
1450     if (ABCIsConst)
1451       Res = Nand(Xor(A, C), B);
1452     break;
1453   case 0xb8:
1454     if (ABCIsConst)
1455       Res = Xor(Nor(Xnor(A, C), B), C);
1456     break;
1457   case 0xb9:
1458     if (ABCIsConst)
1459       Res = Xor(Nor(And(A, C), B), C);
1460     break;
1461   case 0xba:
1462     if (ABCIsConst)
1463       Res = Or(Nor(B, Not(A)), C);
1464     break;
1465   case 0xbb:
1466     if (BCIsConst)
1467       Res = Or(C, Not(B));
1468     break;
1469   case 0xbc:
1470     if (ABCIsConst)
1471       Res = Xor(A, And(Nand(A, C), B));
1472     break;
1473   case 0xbd:
1474     if (ABCIsConst)
1475       Res = Or(Xor(A, B), Xnor(A, C));
1476     break;
1477   case 0xbe:
1478     if (ABCIsConst)
1479       Res = Or(Xor(A, B), C);
1480     break;
1481   case 0xbf:
1482     if (ABCIsConst)
1483       Res = Or(Nand(A, B), C);
1484     break;
1485   case 0xc0:
1486     Res = And(A, B);
1487     break;
1488   case 0xc1:
1489     if (ABCIsConst)
1490       Res = Xor(A, Nor(Nor(A, Not(C)), B));
1491     break;
1492   case 0xc2:
1493     if (ABCIsConst)
1494       Res = Xor(A, Nor(Nor(A, C), B));
1495     break;
1496   case 0xc3:
1497     if (ABIsConst)
1498       Res = Xnor(A, B);
1499     break;
1500   case 0xc4:
1501     if (ABCIsConst)
1502       Res = And(Or(A, Not(C)), B);
1503     break;
1504   case 0xc5:
1505     if (ABCIsConst)
1506       Res = Xor(B, Nor(A, Xor(B, C)));
1507     break;
1508   case 0xc6:
1509     if (ABCIsConst)
1510       Res = Xor(Nor(A, Not(C)), B);
1511     break;
1512   case 0xc7:
1513     if (ABCIsConst)
1514       Res = Or(Xnor(A, B), Nor(A, C));
1515     break;
1516   case 0xc8:
1517     if (ABCIsConst)
1518       Res = And(Or(A, C), B);
1519     break;
1520   case 0xc9:
1521     if (ABCIsConst)
1522       Res = Xor(Nor(A, C), B);
1523     break;
1524   case 0xca:
1525     if (ABCIsConst)
1526       Res = Xor(B, Nor(A, Xnor(B, C)));
1527     break;
1528   case 0xcb:
1529     if (ABCIsConst)
1530       Res = Or(Xnor(A, B), And(B, C));
1531     break;
1532   case 0xcc:
1533     Res = B;
1534     break;
1535   case 0xcd:
1536     if (ABCIsConst)
1537       Res = Or(Nor(A, C), B);
1538     break;
1539   case 0xce:
1540     if (ABCIsConst)
1541       Res = Or(Nor(A, Not(C)), B);
1542     break;
1543   case 0xcf:
1544     if (ABIsConst)
1545       Res = Or(B, Not(A));
1546     break;
1547   case 0xd0:
1548     if (ABCIsConst)
1549       Res = And(A, Or(B, Not(C)));
1550     break;
1551   case 0xd1:
1552     if (ABCIsConst)
1553       Res = Xor(A, Nor(Xor(A, C), B));
1554     break;
1555   case 0xd2:
1556     if (ABCIsConst)
1557       Res = Xor(A, Nor(B, Not(C)));
1558     break;
1559   case 0xd3:
1560     if (ABCIsConst)
1561       Res = Or(Xnor(A, B), Nor(B, C));
1562     break;
1563   case 0xd4:
1564     if (ABCIsConst)
1565       Res = Xor(Nor(Xnor(A, B), Xor(A, C)), A);
1566     break;
1567   case 0xd5:
1568     if (ABCIsConst)
1569       Res = Nand(Nand(A, B), C);
1570     break;
1571   case 0xd6:
1572     if (ABCIsConst)
1573       Res = Xor(Xor(A, B), Or(And(A, B), C));
1574     break;
1575   case 0xd7:
1576     if (ABCIsConst)
1577       Res = Nand(Xor(A, B), C);
1578     break;
1579   case 0xd8:
1580     if (ABCIsConst)
1581       Res = Xor(Nor(Xnor(A, B), C), B);
1582     break;
1583   case 0xd9:
1584     if (ABCIsConst)
1585       Res = Or(And(A, B), Xnor(B, C));
1586     break;
1587   case 0xda:
1588     if (ABCIsConst)
1589       Res = Xor(A, And(Nand(A, B), C));
1590     break;
1591   case 0xdb:
1592     if (ABCIsConst)
1593       Res = Or(Xnor(A, B), Xor(A, C));
1594     break;
1595   case 0xdc:
1596     if (ABCIsConst)
1597       Res = Or(B, Nor(C, Not(A)));
1598     break;
1599   case 0xdd:
1600     if (BCIsConst)
1601       Res = Or(B, Not(C));
1602     break;
1603   case 0xde:
1604     if (ABCIsConst)
1605       Res = Or(Xor(A, C), B);
1606     break;
1607   case 0xdf:
1608     if (ABCIsConst)
1609       Res = Or(Nand(A, C), B);
1610     break;
1611   case 0xe0:
1612     if (ABCIsConst)
1613       Res = And(A, Or(B, C));
1614     break;
1615   case 0xe1:
1616     if (ABCIsConst)
1617       Res = Xor(A, Nor(B, C));
1618     break;
1619   case 0xe2:
1620     if (ABCIsConst)
1621       Res = Xor(A, Nor(Xnor(A, C), B));
1622     break;
1623   case 0xe3:
1624     if (ABCIsConst)
1625       Res = Xor(A, Nor(And(A, C), B));
1626     break;
1627   case 0xe4:
1628     if (ABCIsConst)
1629       Res = Xor(A, Nor(Xnor(A, B), C));
1630     break;
1631   case 0xe5:
1632     if (ABCIsConst)
1633       Res = Xor(A, Nor(And(A, B), C));
1634     break;
1635   case 0xe6:
1636     if (ABCIsConst)
1637       Res = Or(And(A, B), Xor(B, C));
1638     break;
1639   case 0xe7:
1640     if (ABCIsConst)
1641       Res = Or(Xnor(A, B), Xnor(A, C));
1642     break;
1643   case 0xe8:
1644     if (ABCIsConst)
1645       Res = Xor(Or(A, B), Nor(Xnor(A, B), C));
1646     break;
1647   case 0xe9:
1648     if (ABCIsConst)
1649       Res = Xor(Xor(A, B), Nand(Nand(A, B), C));
1650     break;
1651   case 0xea:
1652     if (ABCIsConst)
1653       Res = Or(And(A, B), C);
1654     break;
1655   case 0xeb:
1656     if (ABCIsConst)
1657       Res = Or(Xnor(A, B), C);
1658     break;
1659   case 0xec:
1660     if (ABCIsConst)
1661       Res = Or(And(A, C), B);
1662     break;
1663   case 0xed:
1664     if (ABCIsConst)
1665       Res = Or(Xnor(A, C), B);
1666     break;
1667   case 0xee:
1668     Res = Or(B, C);
1669     break;
1670   case 0xef:
1671     if (ABCIsConst)
1672       Res = Nand(A, Nor(B, C));
1673     break;
1674   case 0xf0:
1675     Res = A;
1676     break;
1677   case 0xf1:
1678     if (ABCIsConst)
1679       Res = Or(A, Nor(B, C));
1680     break;
1681   case 0xf2:
1682     if (ABCIsConst)
1683       Res = Or(A, Nor(B, Not(C)));
1684     break;
1685   case 0xf3:
1686     if (ABIsConst)
1687       Res = Or(A, Not(B));
1688     break;
1689   case 0xf4:
1690     if (ABCIsConst)
1691       Res = Or(A, Nor(C, Not(B)));
1692     break;
1693   case 0xf5:
1694     if (ACIsConst)
1695       Res = Or(A, Not(C));
1696     break;
1697   case 0xf6:
1698     if (ABCIsConst)
1699       Res = Or(A, Xor(B, C));
1700     break;
1701   case 0xf7:
1702     if (ABCIsConst)
1703       Res = Or(A, Nand(B, C));
1704     break;
1705   case 0xf8:
1706     if (ABCIsConst)
1707       Res = Or(A, And(B, C));
1708     break;
1709   case 0xf9:
1710     if (ABCIsConst)
1711       Res = Or(A, Xnor(B, C));
1712     break;
1713   case 0xfa:
1714     Res = Or(A, C);
1715     break;
1716   case 0xfb:
1717     if (ABCIsConst)
1718       Res = Nand(Nor(A, C), B);
1719     break;
1720   case 0xfc:
1721     Res = Or(A, B);
1722     break;
1723   case 0xfd:
1724     if (ABCIsConst)
1725       Res = Nand(Nor(A, B), C);
1726     break;
1727   case 0xfe:
1728     if (ABCIsConst)
1729       Res = Or(Or(A, B), C);
1730     break;
1731   case 0xff:
1732     Res = {Constant::getAllOnesValue(Ty), 0xff};
1733     break;
1734   }
1735 
1736   assert((Res.first == nullptr || Res.second == Imm) &&
1737          "Simplification of ternary logic does not verify!");
1738   return Res.first;
1739 }
1740 
1741 static Value *simplifyX86insertps(const IntrinsicInst &II,
1742                                   InstCombiner::BuilderTy &Builder) {
1743   auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2));
1744   if (!CInt)
1745     return nullptr;
1746 
1747   auto *VecTy = cast<FixedVectorType>(II.getType());
1748   assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
1749 
1750   // The immediate permute control byte looks like this:
1751   //    [3:0] - zero mask for each 32-bit lane
1752   //    [5:4] - select one 32-bit destination lane
1753   //    [7:6] - select one 32-bit source lane
1754 
1755   uint8_t Imm = CInt->getZExtValue();
1756   uint8_t ZMask = Imm & 0xf;
1757   uint8_t DestLane = (Imm >> 4) & 0x3;
1758   uint8_t SourceLane = (Imm >> 6) & 0x3;
1759 
1760   ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
1761 
1762   // If all zero mask bits are set, this was just a weird way to
1763   // generate a zero vector.
1764   if (ZMask == 0xf)
1765     return ZeroVector;
1766 
1767   // Initialize by passing all of the first source bits through.
1768   int ShuffleMask[4] = {0, 1, 2, 3};
1769 
1770   // We may replace the second operand with the zero vector.
1771   Value *V1 = II.getArgOperand(1);
1772 
1773   if (ZMask) {
1774     // If the zero mask is being used with a single input or the zero mask
1775     // overrides the destination lane, this is a shuffle with the zero vector.
1776     if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
1777         (ZMask & (1 << DestLane))) {
1778       V1 = ZeroVector;
1779       // We may still move 32-bits of the first source vector from one lane
1780       // to another.
1781       ShuffleMask[DestLane] = SourceLane;
1782       // The zero mask may override the previous insert operation.
1783       for (unsigned i = 0; i < 4; ++i)
1784         if ((ZMask >> i) & 0x1)
1785           ShuffleMask[i] = i + 4;
1786     } else {
1787       // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
1788       return nullptr;
1789     }
1790   } else {
1791     // Replace the selected destination lane with the selected source lane.
1792     ShuffleMask[DestLane] = SourceLane + 4;
1793   }
1794 
1795   return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
1796 }
1797 
1798 /// Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding
1799 /// or conversion to a shuffle vector.
1800 static Value *simplifyX86extrq(IntrinsicInst &II, Value *Op0,
1801                                ConstantInt *CILength, ConstantInt *CIIndex,
1802                                InstCombiner::BuilderTy &Builder) {
1803   auto LowConstantHighUndef = [&](uint64_t Val) {
1804     Type *IntTy64 = Type::getInt64Ty(II.getContext());
1805     Constant *Args[] = {ConstantInt::get(IntTy64, Val),
1806                         UndefValue::get(IntTy64)};
1807     return ConstantVector::get(Args);
1808   };
1809 
1810   // See if we're dealing with constant values.
1811   auto *C0 = dyn_cast<Constant>(Op0);
1812   auto *CI0 =
1813       C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
1814          : nullptr;
1815 
1816   // Attempt to constant fold.
1817   if (CILength && CIIndex) {
1818     // From AMD documentation: "The bit index and field length are each six
1819     // bits in length other bits of the field are ignored."
1820     APInt APIndex = CIIndex->getValue().zextOrTrunc(6);
1821     APInt APLength = CILength->getValue().zextOrTrunc(6);
1822 
1823     unsigned Index = APIndex.getZExtValue();
1824 
1825     // From AMD documentation: "a value of zero in the field length is
1826     // defined as length of 64".
1827     unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
1828 
1829     // From AMD documentation: "If the sum of the bit index + length field
1830     // is greater than 64, the results are undefined".
1831     unsigned End = Index + Length;
1832 
1833     // Note that both field index and field length are 8-bit quantities.
1834     // Since variables 'Index' and 'Length' are unsigned values
1835     // obtained from zero-extending field index and field length
1836     // respectively, their sum should never wrap around.
1837     if (End > 64)
1838       return UndefValue::get(II.getType());
1839 
1840     // If we are inserting whole bytes, we can convert this to a shuffle.
1841     // Lowering can recognize EXTRQI shuffle masks.
1842     if ((Length % 8) == 0 && (Index % 8) == 0) {
1843       // Convert bit indices to byte indices.
1844       Length /= 8;
1845       Index /= 8;
1846 
1847       Type *IntTy8 = Type::getInt8Ty(II.getContext());
1848       auto *ShufTy = FixedVectorType::get(IntTy8, 16);
1849 
1850       SmallVector<int, 16> ShuffleMask;
1851       for (int i = 0; i != (int)Length; ++i)
1852         ShuffleMask.push_back(i + Index);
1853       for (int i = Length; i != 8; ++i)
1854         ShuffleMask.push_back(i + 16);
1855       for (int i = 8; i != 16; ++i)
1856         ShuffleMask.push_back(-1);
1857 
1858       Value *SV = Builder.CreateShuffleVector(
1859           Builder.CreateBitCast(Op0, ShufTy),
1860           ConstantAggregateZero::get(ShufTy), ShuffleMask);
1861       return Builder.CreateBitCast(SV, II.getType());
1862     }
1863 
1864     // Constant Fold - shift Index'th bit to lowest position and mask off
1865     // Length bits.
1866     if (CI0) {
1867       APInt Elt = CI0->getValue();
1868       Elt.lshrInPlace(Index);
1869       Elt = Elt.zextOrTrunc(Length);
1870       return LowConstantHighUndef(Elt.getZExtValue());
1871     }
1872 
1873     // If we were an EXTRQ call, we'll save registers if we convert to EXTRQI.
1874     if (II.getIntrinsicID() == Intrinsic::x86_sse4a_extrq) {
1875       Value *Args[] = {Op0, CILength, CIIndex};
1876       Module *M = II.getModule();
1877       Function *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_extrqi);
1878       return Builder.CreateCall(F, Args);
1879     }
1880   }
1881 
1882   // Constant Fold - extraction from zero is always {zero, undef}.
1883   if (CI0 && CI0->isZero())
1884     return LowConstantHighUndef(0);
1885 
1886   return nullptr;
1887 }
1888 
1889 /// Attempt to simplify SSE4A INSERTQ/INSERTQI instructions using constant
1890 /// folding or conversion to a shuffle vector.
1891 static Value *simplifyX86insertq(IntrinsicInst &II, Value *Op0, Value *Op1,
1892                                  APInt APLength, APInt APIndex,
1893                                  InstCombiner::BuilderTy &Builder) {
1894   // From AMD documentation: "The bit index and field length are each six bits
1895   // in length other bits of the field are ignored."
1896   APIndex = APIndex.zextOrTrunc(6);
1897   APLength = APLength.zextOrTrunc(6);
1898 
1899   // Attempt to constant fold.
1900   unsigned Index = APIndex.getZExtValue();
1901 
1902   // From AMD documentation: "a value of zero in the field length is
1903   // defined as length of 64".
1904   unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
1905 
1906   // From AMD documentation: "If the sum of the bit index + length field
1907   // is greater than 64, the results are undefined".
1908   unsigned End = Index + Length;
1909 
1910   // Note that both field index and field length are 8-bit quantities.
1911   // Since variables 'Index' and 'Length' are unsigned values
1912   // obtained from zero-extending field index and field length
1913   // respectively, their sum should never wrap around.
1914   if (End > 64)
1915     return UndefValue::get(II.getType());
1916 
1917   // If we are inserting whole bytes, we can convert this to a shuffle.
1918   // Lowering can recognize INSERTQI shuffle masks.
1919   if ((Length % 8) == 0 && (Index % 8) == 0) {
1920     // Convert bit indices to byte indices.
1921     Length /= 8;
1922     Index /= 8;
1923 
1924     Type *IntTy8 = Type::getInt8Ty(II.getContext());
1925     auto *ShufTy = FixedVectorType::get(IntTy8, 16);
1926 
1927     SmallVector<int, 16> ShuffleMask;
1928     for (int i = 0; i != (int)Index; ++i)
1929       ShuffleMask.push_back(i);
1930     for (int i = 0; i != (int)Length; ++i)
1931       ShuffleMask.push_back(i + 16);
1932     for (int i = Index + Length; i != 8; ++i)
1933       ShuffleMask.push_back(i);
1934     for (int i = 8; i != 16; ++i)
1935       ShuffleMask.push_back(-1);
1936 
1937     Value *SV = Builder.CreateShuffleVector(Builder.CreateBitCast(Op0, ShufTy),
1938                                             Builder.CreateBitCast(Op1, ShufTy),
1939                                             ShuffleMask);
1940     return Builder.CreateBitCast(SV, II.getType());
1941   }
1942 
1943   // See if we're dealing with constant values.
1944   auto *C0 = dyn_cast<Constant>(Op0);
1945   auto *C1 = dyn_cast<Constant>(Op1);
1946   auto *CI00 =
1947       C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
1948          : nullptr;
1949   auto *CI10 =
1950       C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
1951          : nullptr;
1952 
1953   // Constant Fold - insert bottom Length bits starting at the Index'th bit.
1954   if (CI00 && CI10) {
1955     APInt V00 = CI00->getValue();
1956     APInt V10 = CI10->getValue();
1957     APInt Mask = APInt::getLowBitsSet(64, Length).shl(Index);
1958     V00 = V00 & ~Mask;
1959     V10 = V10.zextOrTrunc(Length).zextOrTrunc(64).shl(Index);
1960     APInt Val = V00 | V10;
1961     Type *IntTy64 = Type::getInt64Ty(II.getContext());
1962     Constant *Args[] = {ConstantInt::get(IntTy64, Val.getZExtValue()),
1963                         UndefValue::get(IntTy64)};
1964     return ConstantVector::get(Args);
1965   }
1966 
1967   // If we were an INSERTQ call, we'll save demanded elements if we convert to
1968   // INSERTQI.
1969   if (II.getIntrinsicID() == Intrinsic::x86_sse4a_insertq) {
1970     Type *IntTy8 = Type::getInt8Ty(II.getContext());
1971     Constant *CILength = ConstantInt::get(IntTy8, Length, false);
1972     Constant *CIIndex = ConstantInt::get(IntTy8, Index, false);
1973 
1974     Value *Args[] = {Op0, Op1, CILength, CIIndex};
1975     Module *M = II.getModule();
1976     Function *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
1977     return Builder.CreateCall(F, Args);
1978   }
1979 
1980   return nullptr;
1981 }
1982 
1983 /// Attempt to convert pshufb* to shufflevector if the mask is constant.
1984 static Value *simplifyX86pshufb(const IntrinsicInst &II,
1985                                 InstCombiner::BuilderTy &Builder) {
1986   auto *V = dyn_cast<Constant>(II.getArgOperand(1));
1987   if (!V)
1988     return nullptr;
1989 
1990   auto *VecTy = cast<FixedVectorType>(II.getType());
1991   unsigned NumElts = VecTy->getNumElements();
1992   assert((NumElts == 16 || NumElts == 32 || NumElts == 64) &&
1993          "Unexpected number of elements in shuffle mask!");
1994 
1995   // Construct a shuffle mask from constant integers or UNDEFs.
1996   int Indexes[64];
1997 
1998   // Each byte in the shuffle control mask forms an index to permute the
1999   // corresponding byte in the destination operand.
2000   for (unsigned I = 0; I < NumElts; ++I) {
2001     Constant *COp = V->getAggregateElement(I);
2002     if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
2003       return nullptr;
2004 
2005     if (isa<UndefValue>(COp)) {
2006       Indexes[I] = -1;
2007       continue;
2008     }
2009 
2010     int8_t Index = cast<ConstantInt>(COp)->getValue().getZExtValue();
2011 
2012     // If the most significant bit (bit[7]) of each byte of the shuffle
2013     // control mask is set, then zero is written in the result byte.
2014     // The zero vector is in the right-hand side of the resulting
2015     // shufflevector.
2016 
2017     // The value of each index for the high 128-bit lane is the least
2018     // significant 4 bits of the respective shuffle control byte.
2019     Index = ((Index < 0) ? NumElts : Index & 0x0F) + (I & 0xF0);
2020     Indexes[I] = Index;
2021   }
2022 
2023   auto V1 = II.getArgOperand(0);
2024   auto V2 = Constant::getNullValue(VecTy);
2025   return Builder.CreateShuffleVector(V1, V2, ArrayRef(Indexes, NumElts));
2026 }
2027 
2028 /// Attempt to convert vpermilvar* to shufflevector if the mask is constant.
2029 static Value *simplifyX86vpermilvar(const IntrinsicInst &II,
2030                                     InstCombiner::BuilderTy &Builder) {
2031   auto *V = dyn_cast<Constant>(II.getArgOperand(1));
2032   if (!V)
2033     return nullptr;
2034 
2035   auto *VecTy = cast<FixedVectorType>(II.getType());
2036   unsigned NumElts = VecTy->getNumElements();
2037   bool IsPD = VecTy->getScalarType()->isDoubleTy();
2038   unsigned NumLaneElts = IsPD ? 2 : 4;
2039   assert(NumElts == 16 || NumElts == 8 || NumElts == 4 || NumElts == 2);
2040 
2041   // Construct a shuffle mask from constant integers or UNDEFs.
2042   int Indexes[16];
2043 
2044   // The intrinsics only read one or two bits, clear the rest.
2045   for (unsigned I = 0; I < NumElts; ++I) {
2046     Constant *COp = V->getAggregateElement(I);
2047     if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
2048       return nullptr;
2049 
2050     if (isa<UndefValue>(COp)) {
2051       Indexes[I] = -1;
2052       continue;
2053     }
2054 
2055     APInt Index = cast<ConstantInt>(COp)->getValue();
2056     Index = Index.zextOrTrunc(32).getLoBits(2);
2057 
2058     // The PD variants uses bit 1 to select per-lane element index, so
2059     // shift down to convert to generic shuffle mask index.
2060     if (IsPD)
2061       Index.lshrInPlace(1);
2062 
2063     // The _256 variants are a bit trickier since the mask bits always index
2064     // into the corresponding 128 half. In order to convert to a generic
2065     // shuffle, we have to make that explicit.
2066     Index += APInt(32, (I / NumLaneElts) * NumLaneElts);
2067 
2068     Indexes[I] = Index.getZExtValue();
2069   }
2070 
2071   auto V1 = II.getArgOperand(0);
2072   return Builder.CreateShuffleVector(V1, ArrayRef(Indexes, NumElts));
2073 }
2074 
2075 /// Attempt to convert vpermd/vpermps to shufflevector if the mask is constant.
2076 static Value *simplifyX86vpermv(const IntrinsicInst &II,
2077                                 InstCombiner::BuilderTy &Builder) {
2078   auto *V = dyn_cast<Constant>(II.getArgOperand(1));
2079   if (!V)
2080     return nullptr;
2081 
2082   auto *VecTy = cast<FixedVectorType>(II.getType());
2083   unsigned Size = VecTy->getNumElements();
2084   assert((Size == 4 || Size == 8 || Size == 16 || Size == 32 || Size == 64) &&
2085          "Unexpected shuffle mask size");
2086 
2087   // Construct a shuffle mask from constant integers or UNDEFs.
2088   int Indexes[64];
2089 
2090   for (unsigned I = 0; I < Size; ++I) {
2091     Constant *COp = V->getAggregateElement(I);
2092     if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
2093       return nullptr;
2094 
2095     if (isa<UndefValue>(COp)) {
2096       Indexes[I] = -1;
2097       continue;
2098     }
2099 
2100     uint32_t Index = cast<ConstantInt>(COp)->getZExtValue();
2101     Index &= Size - 1;
2102     Indexes[I] = Index;
2103   }
2104 
2105   auto V1 = II.getArgOperand(0);
2106   return Builder.CreateShuffleVector(V1, ArrayRef(Indexes, Size));
2107 }
2108 
2109 /// Attempt to convert vpermi2/vpermt2 to shufflevector if the mask is constant.
2110 static Value *simplifyX86vpermv3(const IntrinsicInst &II,
2111                                  InstCombiner::BuilderTy &Builder) {
2112   auto *V = dyn_cast<Constant>(II.getArgOperand(1));
2113   if (!V)
2114     return nullptr;
2115 
2116   auto *VecTy = cast<FixedVectorType>(II.getType());
2117   unsigned Size = VecTy->getNumElements();
2118   assert((Size == 2 || Size == 4 || Size == 8 || Size == 16 || Size == 32 ||
2119           Size == 64) &&
2120          "Unexpected shuffle mask size");
2121 
2122   // Construct a shuffle mask from constant integers or UNDEFs.
2123   int Indexes[64];
2124 
2125   for (unsigned I = 0; I < Size; ++I) {
2126     Constant *COp = V->getAggregateElement(I);
2127     if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
2128       return nullptr;
2129 
2130     if (isa<UndefValue>(COp)) {
2131       Indexes[I] = -1;
2132       continue;
2133     }
2134 
2135     uint32_t Index = cast<ConstantInt>(COp)->getZExtValue();
2136     Index &= (2 * Size) - 1;
2137     Indexes[I] = Index;
2138   }
2139 
2140   auto V1 = II.getArgOperand(0);
2141   auto V2 = II.getArgOperand(2);
2142   return Builder.CreateShuffleVector(V1, V2, ArrayRef(Indexes, Size));
2143 }
2144 
2145 std::optional<Instruction *>
2146 X86TTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
2147   auto SimplifyDemandedVectorEltsLow = [&IC](Value *Op, unsigned Width,
2148                                              unsigned DemandedWidth) {
2149     APInt UndefElts(Width, 0);
2150     APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
2151     return IC.SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
2152   };
2153 
2154   Intrinsic::ID IID = II.getIntrinsicID();
2155   switch (IID) {
2156   case Intrinsic::x86_bmi_bextr_32:
2157   case Intrinsic::x86_bmi_bextr_64:
2158   case Intrinsic::x86_tbm_bextri_u32:
2159   case Intrinsic::x86_tbm_bextri_u64:
2160     // If the RHS is a constant we can try some simplifications.
2161     if (auto *C = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
2162       uint64_t Shift = C->getZExtValue();
2163       uint64_t Length = (Shift >> 8) & 0xff;
2164       Shift &= 0xff;
2165       unsigned BitWidth = II.getType()->getIntegerBitWidth();
2166       // If the length is 0 or the shift is out of range, replace with zero.
2167       if (Length == 0 || Shift >= BitWidth) {
2168         return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
2169       }
2170       // If the LHS is also a constant, we can completely constant fold this.
2171       if (auto *InC = dyn_cast<ConstantInt>(II.getArgOperand(0))) {
2172         uint64_t Result = InC->getZExtValue() >> Shift;
2173         if (Length > BitWidth)
2174           Length = BitWidth;
2175         Result &= maskTrailingOnes<uint64_t>(Length);
2176         return IC.replaceInstUsesWith(II,
2177                                       ConstantInt::get(II.getType(), Result));
2178       }
2179       // TODO should we turn this into 'and' if shift is 0? Or 'shl' if we
2180       // are only masking bits that a shift already cleared?
2181     }
2182     break;
2183 
2184   case Intrinsic::x86_bmi_bzhi_32:
2185   case Intrinsic::x86_bmi_bzhi_64:
2186     // If the RHS is a constant we can try some simplifications.
2187     if (auto *C = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
2188       uint64_t Index = C->getZExtValue() & 0xff;
2189       unsigned BitWidth = II.getType()->getIntegerBitWidth();
2190       if (Index >= BitWidth) {
2191         return IC.replaceInstUsesWith(II, II.getArgOperand(0));
2192       }
2193       if (Index == 0) {
2194         return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
2195       }
2196       // If the LHS is also a constant, we can completely constant fold this.
2197       if (auto *InC = dyn_cast<ConstantInt>(II.getArgOperand(0))) {
2198         uint64_t Result = InC->getZExtValue();
2199         Result &= maskTrailingOnes<uint64_t>(Index);
2200         return IC.replaceInstUsesWith(II,
2201                                       ConstantInt::get(II.getType(), Result));
2202       }
2203       // TODO should we convert this to an AND if the RHS is constant?
2204     }
2205     break;
2206   case Intrinsic::x86_bmi_pext_32:
2207   case Intrinsic::x86_bmi_pext_64:
2208     if (auto *MaskC = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
2209       if (MaskC->isNullValue()) {
2210         return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
2211       }
2212       if (MaskC->isAllOnesValue()) {
2213         return IC.replaceInstUsesWith(II, II.getArgOperand(0));
2214       }
2215 
2216       unsigned MaskIdx, MaskLen;
2217       if (MaskC->getValue().isShiftedMask(MaskIdx, MaskLen)) {
2218         // any single contingous sequence of 1s anywhere in the mask simply
2219         // describes a subset of the input bits shifted to the appropriate
2220         // position.  Replace with the straight forward IR.
2221         Value *Input = II.getArgOperand(0);
2222         Value *Masked = IC.Builder.CreateAnd(Input, II.getArgOperand(1));
2223         Value *ShiftAmt = ConstantInt::get(II.getType(), MaskIdx);
2224         Value *Shifted = IC.Builder.CreateLShr(Masked, ShiftAmt);
2225         return IC.replaceInstUsesWith(II, Shifted);
2226       }
2227 
2228       if (auto *SrcC = dyn_cast<ConstantInt>(II.getArgOperand(0))) {
2229         uint64_t Src = SrcC->getZExtValue();
2230         uint64_t Mask = MaskC->getZExtValue();
2231         uint64_t Result = 0;
2232         uint64_t BitToSet = 1;
2233 
2234         while (Mask) {
2235           // Isolate lowest set bit.
2236           uint64_t BitToTest = Mask & -Mask;
2237           if (BitToTest & Src)
2238             Result |= BitToSet;
2239 
2240           BitToSet <<= 1;
2241           // Clear lowest set bit.
2242           Mask &= Mask - 1;
2243         }
2244 
2245         return IC.replaceInstUsesWith(II,
2246                                       ConstantInt::get(II.getType(), Result));
2247       }
2248     }
2249     break;
2250   case Intrinsic::x86_bmi_pdep_32:
2251   case Intrinsic::x86_bmi_pdep_64:
2252     if (auto *MaskC = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
2253       if (MaskC->isNullValue()) {
2254         return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
2255       }
2256       if (MaskC->isAllOnesValue()) {
2257         return IC.replaceInstUsesWith(II, II.getArgOperand(0));
2258       }
2259 
2260       unsigned MaskIdx, MaskLen;
2261       if (MaskC->getValue().isShiftedMask(MaskIdx, MaskLen)) {
2262         // any single contingous sequence of 1s anywhere in the mask simply
2263         // describes a subset of the input bits shifted to the appropriate
2264         // position.  Replace with the straight forward IR.
2265         Value *Input = II.getArgOperand(0);
2266         Value *ShiftAmt = ConstantInt::get(II.getType(), MaskIdx);
2267         Value *Shifted = IC.Builder.CreateShl(Input, ShiftAmt);
2268         Value *Masked = IC.Builder.CreateAnd(Shifted, II.getArgOperand(1));
2269         return IC.replaceInstUsesWith(II, Masked);
2270       }
2271 
2272       if (auto *SrcC = dyn_cast<ConstantInt>(II.getArgOperand(0))) {
2273         uint64_t Src = SrcC->getZExtValue();
2274         uint64_t Mask = MaskC->getZExtValue();
2275         uint64_t Result = 0;
2276         uint64_t BitToTest = 1;
2277 
2278         while (Mask) {
2279           // Isolate lowest set bit.
2280           uint64_t BitToSet = Mask & -Mask;
2281           if (BitToTest & Src)
2282             Result |= BitToSet;
2283 
2284           BitToTest <<= 1;
2285           // Clear lowest set bit;
2286           Mask &= Mask - 1;
2287         }
2288 
2289         return IC.replaceInstUsesWith(II,
2290                                       ConstantInt::get(II.getType(), Result));
2291       }
2292     }
2293     break;
2294 
2295   case Intrinsic::x86_sse_cvtss2si:
2296   case Intrinsic::x86_sse_cvtss2si64:
2297   case Intrinsic::x86_sse_cvttss2si:
2298   case Intrinsic::x86_sse_cvttss2si64:
2299   case Intrinsic::x86_sse2_cvtsd2si:
2300   case Intrinsic::x86_sse2_cvtsd2si64:
2301   case Intrinsic::x86_sse2_cvttsd2si:
2302   case Intrinsic::x86_sse2_cvttsd2si64:
2303   case Intrinsic::x86_avx512_vcvtss2si32:
2304   case Intrinsic::x86_avx512_vcvtss2si64:
2305   case Intrinsic::x86_avx512_vcvtss2usi32:
2306   case Intrinsic::x86_avx512_vcvtss2usi64:
2307   case Intrinsic::x86_avx512_vcvtsd2si32:
2308   case Intrinsic::x86_avx512_vcvtsd2si64:
2309   case Intrinsic::x86_avx512_vcvtsd2usi32:
2310   case Intrinsic::x86_avx512_vcvtsd2usi64:
2311   case Intrinsic::x86_avx512_cvttss2si:
2312   case Intrinsic::x86_avx512_cvttss2si64:
2313   case Intrinsic::x86_avx512_cvttss2usi:
2314   case Intrinsic::x86_avx512_cvttss2usi64:
2315   case Intrinsic::x86_avx512_cvttsd2si:
2316   case Intrinsic::x86_avx512_cvttsd2si64:
2317   case Intrinsic::x86_avx512_cvttsd2usi:
2318   case Intrinsic::x86_avx512_cvttsd2usi64: {
2319     // These intrinsics only demand the 0th element of their input vectors. If
2320     // we can simplify the input based on that, do so now.
2321     Value *Arg = II.getArgOperand(0);
2322     unsigned VWidth = cast<FixedVectorType>(Arg->getType())->getNumElements();
2323     if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) {
2324       return IC.replaceOperand(II, 0, V);
2325     }
2326     break;
2327   }
2328 
2329   case Intrinsic::x86_mmx_pmovmskb:
2330   case Intrinsic::x86_sse_movmsk_ps:
2331   case Intrinsic::x86_sse2_movmsk_pd:
2332   case Intrinsic::x86_sse2_pmovmskb_128:
2333   case Intrinsic::x86_avx_movmsk_pd_256:
2334   case Intrinsic::x86_avx_movmsk_ps_256:
2335   case Intrinsic::x86_avx2_pmovmskb:
2336     if (Value *V = simplifyX86movmsk(II, IC.Builder)) {
2337       return IC.replaceInstUsesWith(II, V);
2338     }
2339     break;
2340 
2341   case Intrinsic::x86_sse_comieq_ss:
2342   case Intrinsic::x86_sse_comige_ss:
2343   case Intrinsic::x86_sse_comigt_ss:
2344   case Intrinsic::x86_sse_comile_ss:
2345   case Intrinsic::x86_sse_comilt_ss:
2346   case Intrinsic::x86_sse_comineq_ss:
2347   case Intrinsic::x86_sse_ucomieq_ss:
2348   case Intrinsic::x86_sse_ucomige_ss:
2349   case Intrinsic::x86_sse_ucomigt_ss:
2350   case Intrinsic::x86_sse_ucomile_ss:
2351   case Intrinsic::x86_sse_ucomilt_ss:
2352   case Intrinsic::x86_sse_ucomineq_ss:
2353   case Intrinsic::x86_sse2_comieq_sd:
2354   case Intrinsic::x86_sse2_comige_sd:
2355   case Intrinsic::x86_sse2_comigt_sd:
2356   case Intrinsic::x86_sse2_comile_sd:
2357   case Intrinsic::x86_sse2_comilt_sd:
2358   case Intrinsic::x86_sse2_comineq_sd:
2359   case Intrinsic::x86_sse2_ucomieq_sd:
2360   case Intrinsic::x86_sse2_ucomige_sd:
2361   case Intrinsic::x86_sse2_ucomigt_sd:
2362   case Intrinsic::x86_sse2_ucomile_sd:
2363   case Intrinsic::x86_sse2_ucomilt_sd:
2364   case Intrinsic::x86_sse2_ucomineq_sd:
2365   case Intrinsic::x86_avx512_vcomi_ss:
2366   case Intrinsic::x86_avx512_vcomi_sd:
2367   case Intrinsic::x86_avx512_mask_cmp_ss:
2368   case Intrinsic::x86_avx512_mask_cmp_sd: {
2369     // These intrinsics only demand the 0th element of their input vectors. If
2370     // we can simplify the input based on that, do so now.
2371     bool MadeChange = false;
2372     Value *Arg0 = II.getArgOperand(0);
2373     Value *Arg1 = II.getArgOperand(1);
2374     unsigned VWidth = cast<FixedVectorType>(Arg0->getType())->getNumElements();
2375     if (Value *V = SimplifyDemandedVectorEltsLow(Arg0, VWidth, 1)) {
2376       IC.replaceOperand(II, 0, V);
2377       MadeChange = true;
2378     }
2379     if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, 1)) {
2380       IC.replaceOperand(II, 1, V);
2381       MadeChange = true;
2382     }
2383     if (MadeChange) {
2384       return &II;
2385     }
2386     break;
2387   }
2388 
2389   case Intrinsic::x86_avx512_add_ps_512:
2390   case Intrinsic::x86_avx512_div_ps_512:
2391   case Intrinsic::x86_avx512_mul_ps_512:
2392   case Intrinsic::x86_avx512_sub_ps_512:
2393   case Intrinsic::x86_avx512_add_pd_512:
2394   case Intrinsic::x86_avx512_div_pd_512:
2395   case Intrinsic::x86_avx512_mul_pd_512:
2396   case Intrinsic::x86_avx512_sub_pd_512:
2397     // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular
2398     // IR operations.
2399     if (auto *R = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
2400       if (R->getValue() == 4) {
2401         Value *Arg0 = II.getArgOperand(0);
2402         Value *Arg1 = II.getArgOperand(1);
2403 
2404         Value *V;
2405         switch (IID) {
2406         default:
2407           llvm_unreachable("Case stmts out of sync!");
2408         case Intrinsic::x86_avx512_add_ps_512:
2409         case Intrinsic::x86_avx512_add_pd_512:
2410           V = IC.Builder.CreateFAdd(Arg0, Arg1);
2411           break;
2412         case Intrinsic::x86_avx512_sub_ps_512:
2413         case Intrinsic::x86_avx512_sub_pd_512:
2414           V = IC.Builder.CreateFSub(Arg0, Arg1);
2415           break;
2416         case Intrinsic::x86_avx512_mul_ps_512:
2417         case Intrinsic::x86_avx512_mul_pd_512:
2418           V = IC.Builder.CreateFMul(Arg0, Arg1);
2419           break;
2420         case Intrinsic::x86_avx512_div_ps_512:
2421         case Intrinsic::x86_avx512_div_pd_512:
2422           V = IC.Builder.CreateFDiv(Arg0, Arg1);
2423           break;
2424         }
2425 
2426         return IC.replaceInstUsesWith(II, V);
2427       }
2428     }
2429     break;
2430 
2431   case Intrinsic::x86_avx512_mask_add_ss_round:
2432   case Intrinsic::x86_avx512_mask_div_ss_round:
2433   case Intrinsic::x86_avx512_mask_mul_ss_round:
2434   case Intrinsic::x86_avx512_mask_sub_ss_round:
2435   case Intrinsic::x86_avx512_mask_add_sd_round:
2436   case Intrinsic::x86_avx512_mask_div_sd_round:
2437   case Intrinsic::x86_avx512_mask_mul_sd_round:
2438   case Intrinsic::x86_avx512_mask_sub_sd_round:
2439     // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular
2440     // IR operations.
2441     if (auto *R = dyn_cast<ConstantInt>(II.getArgOperand(4))) {
2442       if (R->getValue() == 4) {
2443         // Extract the element as scalars.
2444         Value *Arg0 = II.getArgOperand(0);
2445         Value *Arg1 = II.getArgOperand(1);
2446         Value *LHS = IC.Builder.CreateExtractElement(Arg0, (uint64_t)0);
2447         Value *RHS = IC.Builder.CreateExtractElement(Arg1, (uint64_t)0);
2448 
2449         Value *V;
2450         switch (IID) {
2451         default:
2452           llvm_unreachable("Case stmts out of sync!");
2453         case Intrinsic::x86_avx512_mask_add_ss_round:
2454         case Intrinsic::x86_avx512_mask_add_sd_round:
2455           V = IC.Builder.CreateFAdd(LHS, RHS);
2456           break;
2457         case Intrinsic::x86_avx512_mask_sub_ss_round:
2458         case Intrinsic::x86_avx512_mask_sub_sd_round:
2459           V = IC.Builder.CreateFSub(LHS, RHS);
2460           break;
2461         case Intrinsic::x86_avx512_mask_mul_ss_round:
2462         case Intrinsic::x86_avx512_mask_mul_sd_round:
2463           V = IC.Builder.CreateFMul(LHS, RHS);
2464           break;
2465         case Intrinsic::x86_avx512_mask_div_ss_round:
2466         case Intrinsic::x86_avx512_mask_div_sd_round:
2467           V = IC.Builder.CreateFDiv(LHS, RHS);
2468           break;
2469         }
2470 
2471         // Handle the masking aspect of the intrinsic.
2472         Value *Mask = II.getArgOperand(3);
2473         auto *C = dyn_cast<ConstantInt>(Mask);
2474         // We don't need a select if we know the mask bit is a 1.
2475         if (!C || !C->getValue()[0]) {
2476           // Cast the mask to an i1 vector and then extract the lowest element.
2477           auto *MaskTy = FixedVectorType::get(
2478               IC.Builder.getInt1Ty(),
2479               cast<IntegerType>(Mask->getType())->getBitWidth());
2480           Mask = IC.Builder.CreateBitCast(Mask, MaskTy);
2481           Mask = IC.Builder.CreateExtractElement(Mask, (uint64_t)0);
2482           // Extract the lowest element from the passthru operand.
2483           Value *Passthru =
2484               IC.Builder.CreateExtractElement(II.getArgOperand(2), (uint64_t)0);
2485           V = IC.Builder.CreateSelect(Mask, V, Passthru);
2486         }
2487 
2488         // Insert the result back into the original argument 0.
2489         V = IC.Builder.CreateInsertElement(Arg0, V, (uint64_t)0);
2490 
2491         return IC.replaceInstUsesWith(II, V);
2492       }
2493     }
2494     break;
2495 
2496   // Constant fold ashr( <A x Bi>, Ci ).
2497   // Constant fold lshr( <A x Bi>, Ci ).
2498   // Constant fold shl( <A x Bi>, Ci ).
2499   case Intrinsic::x86_sse2_psrai_d:
2500   case Intrinsic::x86_sse2_psrai_w:
2501   case Intrinsic::x86_avx2_psrai_d:
2502   case Intrinsic::x86_avx2_psrai_w:
2503   case Intrinsic::x86_avx512_psrai_q_128:
2504   case Intrinsic::x86_avx512_psrai_q_256:
2505   case Intrinsic::x86_avx512_psrai_d_512:
2506   case Intrinsic::x86_avx512_psrai_q_512:
2507   case Intrinsic::x86_avx512_psrai_w_512:
2508   case Intrinsic::x86_sse2_psrli_d:
2509   case Intrinsic::x86_sse2_psrli_q:
2510   case Intrinsic::x86_sse2_psrli_w:
2511   case Intrinsic::x86_avx2_psrli_d:
2512   case Intrinsic::x86_avx2_psrli_q:
2513   case Intrinsic::x86_avx2_psrli_w:
2514   case Intrinsic::x86_avx512_psrli_d_512:
2515   case Intrinsic::x86_avx512_psrli_q_512:
2516   case Intrinsic::x86_avx512_psrli_w_512:
2517   case Intrinsic::x86_sse2_pslli_d:
2518   case Intrinsic::x86_sse2_pslli_q:
2519   case Intrinsic::x86_sse2_pslli_w:
2520   case Intrinsic::x86_avx2_pslli_d:
2521   case Intrinsic::x86_avx2_pslli_q:
2522   case Intrinsic::x86_avx2_pslli_w:
2523   case Intrinsic::x86_avx512_pslli_d_512:
2524   case Intrinsic::x86_avx512_pslli_q_512:
2525   case Intrinsic::x86_avx512_pslli_w_512:
2526     if (Value *V = simplifyX86immShift(II, IC.Builder)) {
2527       return IC.replaceInstUsesWith(II, V);
2528     }
2529     break;
2530 
2531   case Intrinsic::x86_sse2_psra_d:
2532   case Intrinsic::x86_sse2_psra_w:
2533   case Intrinsic::x86_avx2_psra_d:
2534   case Intrinsic::x86_avx2_psra_w:
2535   case Intrinsic::x86_avx512_psra_q_128:
2536   case Intrinsic::x86_avx512_psra_q_256:
2537   case Intrinsic::x86_avx512_psra_d_512:
2538   case Intrinsic::x86_avx512_psra_q_512:
2539   case Intrinsic::x86_avx512_psra_w_512:
2540   case Intrinsic::x86_sse2_psrl_d:
2541   case Intrinsic::x86_sse2_psrl_q:
2542   case Intrinsic::x86_sse2_psrl_w:
2543   case Intrinsic::x86_avx2_psrl_d:
2544   case Intrinsic::x86_avx2_psrl_q:
2545   case Intrinsic::x86_avx2_psrl_w:
2546   case Intrinsic::x86_avx512_psrl_d_512:
2547   case Intrinsic::x86_avx512_psrl_q_512:
2548   case Intrinsic::x86_avx512_psrl_w_512:
2549   case Intrinsic::x86_sse2_psll_d:
2550   case Intrinsic::x86_sse2_psll_q:
2551   case Intrinsic::x86_sse2_psll_w:
2552   case Intrinsic::x86_avx2_psll_d:
2553   case Intrinsic::x86_avx2_psll_q:
2554   case Intrinsic::x86_avx2_psll_w:
2555   case Intrinsic::x86_avx512_psll_d_512:
2556   case Intrinsic::x86_avx512_psll_q_512:
2557   case Intrinsic::x86_avx512_psll_w_512: {
2558     if (Value *V = simplifyX86immShift(II, IC.Builder)) {
2559       return IC.replaceInstUsesWith(II, V);
2560     }
2561 
2562     // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector
2563     // operand to compute the shift amount.
2564     Value *Arg1 = II.getArgOperand(1);
2565     assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 &&
2566            "Unexpected packed shift size");
2567     unsigned VWidth = cast<FixedVectorType>(Arg1->getType())->getNumElements();
2568 
2569     if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) {
2570       return IC.replaceOperand(II, 1, V);
2571     }
2572     break;
2573   }
2574 
2575   case Intrinsic::x86_avx2_psllv_d:
2576   case Intrinsic::x86_avx2_psllv_d_256:
2577   case Intrinsic::x86_avx2_psllv_q:
2578   case Intrinsic::x86_avx2_psllv_q_256:
2579   case Intrinsic::x86_avx512_psllv_d_512:
2580   case Intrinsic::x86_avx512_psllv_q_512:
2581   case Intrinsic::x86_avx512_psllv_w_128:
2582   case Intrinsic::x86_avx512_psllv_w_256:
2583   case Intrinsic::x86_avx512_psllv_w_512:
2584   case Intrinsic::x86_avx2_psrav_d:
2585   case Intrinsic::x86_avx2_psrav_d_256:
2586   case Intrinsic::x86_avx512_psrav_q_128:
2587   case Intrinsic::x86_avx512_psrav_q_256:
2588   case Intrinsic::x86_avx512_psrav_d_512:
2589   case Intrinsic::x86_avx512_psrav_q_512:
2590   case Intrinsic::x86_avx512_psrav_w_128:
2591   case Intrinsic::x86_avx512_psrav_w_256:
2592   case Intrinsic::x86_avx512_psrav_w_512:
2593   case Intrinsic::x86_avx2_psrlv_d:
2594   case Intrinsic::x86_avx2_psrlv_d_256:
2595   case Intrinsic::x86_avx2_psrlv_q:
2596   case Intrinsic::x86_avx2_psrlv_q_256:
2597   case Intrinsic::x86_avx512_psrlv_d_512:
2598   case Intrinsic::x86_avx512_psrlv_q_512:
2599   case Intrinsic::x86_avx512_psrlv_w_128:
2600   case Intrinsic::x86_avx512_psrlv_w_256:
2601   case Intrinsic::x86_avx512_psrlv_w_512:
2602     if (Value *V = simplifyX86varShift(II, IC.Builder)) {
2603       return IC.replaceInstUsesWith(II, V);
2604     }
2605     break;
2606 
2607   case Intrinsic::x86_sse2_packssdw_128:
2608   case Intrinsic::x86_sse2_packsswb_128:
2609   case Intrinsic::x86_avx2_packssdw:
2610   case Intrinsic::x86_avx2_packsswb:
2611   case Intrinsic::x86_avx512_packssdw_512:
2612   case Intrinsic::x86_avx512_packsswb_512:
2613     if (Value *V = simplifyX86pack(II, IC.Builder, true)) {
2614       return IC.replaceInstUsesWith(II, V);
2615     }
2616     break;
2617 
2618   case Intrinsic::x86_sse2_packuswb_128:
2619   case Intrinsic::x86_sse41_packusdw:
2620   case Intrinsic::x86_avx2_packusdw:
2621   case Intrinsic::x86_avx2_packuswb:
2622   case Intrinsic::x86_avx512_packusdw_512:
2623   case Intrinsic::x86_avx512_packuswb_512:
2624     if (Value *V = simplifyX86pack(II, IC.Builder, false)) {
2625       return IC.replaceInstUsesWith(II, V);
2626     }
2627     break;
2628 
2629   case Intrinsic::x86_sse2_pmulh_w:
2630   case Intrinsic::x86_avx2_pmulh_w:
2631   case Intrinsic::x86_avx512_pmulh_w_512:
2632     if (Value *V = simplifyX86pmulh(II, IC.Builder, true, false)) {
2633       return IC.replaceInstUsesWith(II, V);
2634     }
2635     break;
2636 
2637   case Intrinsic::x86_sse2_pmulhu_w:
2638   case Intrinsic::x86_avx2_pmulhu_w:
2639   case Intrinsic::x86_avx512_pmulhu_w_512:
2640     if (Value *V = simplifyX86pmulh(II, IC.Builder, false, false)) {
2641       return IC.replaceInstUsesWith(II, V);
2642     }
2643     break;
2644 
2645   case Intrinsic::x86_ssse3_pmul_hr_sw_128:
2646   case Intrinsic::x86_avx2_pmul_hr_sw:
2647   case Intrinsic::x86_avx512_pmul_hr_sw_512:
2648     if (Value *V = simplifyX86pmulh(II, IC.Builder, true, true)) {
2649       return IC.replaceInstUsesWith(II, V);
2650     }
2651     break;
2652 
2653   case Intrinsic::x86_sse2_pmadd_wd:
2654   case Intrinsic::x86_avx2_pmadd_wd:
2655   case Intrinsic::x86_avx512_pmaddw_d_512:
2656     if (Value *V = simplifyX86pmadd(II, IC.Builder, true)) {
2657       return IC.replaceInstUsesWith(II, V);
2658     }
2659     break;
2660 
2661   case Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2662   case Intrinsic::x86_avx2_pmadd_ub_sw:
2663   case Intrinsic::x86_avx512_pmaddubs_w_512:
2664     if (Value *V = simplifyX86pmadd(II, IC.Builder, false)) {
2665       return IC.replaceInstUsesWith(II, V);
2666     }
2667     break;
2668 
2669   case Intrinsic::x86_pclmulqdq:
2670   case Intrinsic::x86_pclmulqdq_256:
2671   case Intrinsic::x86_pclmulqdq_512: {
2672     if (auto *C = dyn_cast<ConstantInt>(II.getArgOperand(2))) {
2673       unsigned Imm = C->getZExtValue();
2674 
2675       bool MadeChange = false;
2676       Value *Arg0 = II.getArgOperand(0);
2677       Value *Arg1 = II.getArgOperand(1);
2678       unsigned VWidth =
2679           cast<FixedVectorType>(Arg0->getType())->getNumElements();
2680 
2681       APInt UndefElts1(VWidth, 0);
2682       APInt DemandedElts1 =
2683           APInt::getSplat(VWidth, APInt(2, (Imm & 0x01) ? 2 : 1));
2684       if (Value *V =
2685               IC.SimplifyDemandedVectorElts(Arg0, DemandedElts1, UndefElts1)) {
2686         IC.replaceOperand(II, 0, V);
2687         MadeChange = true;
2688       }
2689 
2690       APInt UndefElts2(VWidth, 0);
2691       APInt DemandedElts2 =
2692           APInt::getSplat(VWidth, APInt(2, (Imm & 0x10) ? 2 : 1));
2693       if (Value *V =
2694               IC.SimplifyDemandedVectorElts(Arg1, DemandedElts2, UndefElts2)) {
2695         IC.replaceOperand(II, 1, V);
2696         MadeChange = true;
2697       }
2698 
2699       // If either input elements are undef, the result is zero.
2700       if (DemandedElts1.isSubsetOf(UndefElts1) ||
2701           DemandedElts2.isSubsetOf(UndefElts2)) {
2702         return IC.replaceInstUsesWith(II,
2703                                       ConstantAggregateZero::get(II.getType()));
2704       }
2705 
2706       if (MadeChange) {
2707         return &II;
2708       }
2709     }
2710     break;
2711   }
2712 
2713   case Intrinsic::x86_sse41_insertps:
2714     if (Value *V = simplifyX86insertps(II, IC.Builder)) {
2715       return IC.replaceInstUsesWith(II, V);
2716     }
2717     break;
2718 
2719   case Intrinsic::x86_sse4a_extrq: {
2720     Value *Op0 = II.getArgOperand(0);
2721     Value *Op1 = II.getArgOperand(1);
2722     unsigned VWidth0 = cast<FixedVectorType>(Op0->getType())->getNumElements();
2723     unsigned VWidth1 = cast<FixedVectorType>(Op1->getType())->getNumElements();
2724     assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
2725            Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 &&
2726            VWidth1 == 16 && "Unexpected operand sizes");
2727 
2728     // See if we're dealing with constant values.
2729     auto *C1 = dyn_cast<Constant>(Op1);
2730     auto *CILength =
2731         C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
2732            : nullptr;
2733     auto *CIIndex =
2734         C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1))
2735            : nullptr;
2736 
2737     // Attempt to simplify to a constant, shuffle vector or EXTRQI call.
2738     if (Value *V = simplifyX86extrq(II, Op0, CILength, CIIndex, IC.Builder)) {
2739       return IC.replaceInstUsesWith(II, V);
2740     }
2741 
2742     // EXTRQ only uses the lowest 64-bits of the first 128-bit vector
2743     // operands and the lowest 16-bits of the second.
2744     bool MadeChange = false;
2745     if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
2746       IC.replaceOperand(II, 0, V);
2747       MadeChange = true;
2748     }
2749     if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) {
2750       IC.replaceOperand(II, 1, V);
2751       MadeChange = true;
2752     }
2753     if (MadeChange) {
2754       return &II;
2755     }
2756     break;
2757   }
2758 
2759   case Intrinsic::x86_sse4a_extrqi: {
2760     // EXTRQI: Extract Length bits starting from Index. Zero pad the remaining
2761     // bits of the lower 64-bits. The upper 64-bits are undefined.
2762     Value *Op0 = II.getArgOperand(0);
2763     unsigned VWidth = cast<FixedVectorType>(Op0->getType())->getNumElements();
2764     assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 &&
2765            "Unexpected operand size");
2766 
2767     // See if we're dealing with constant values.
2768     auto *CILength = dyn_cast<ConstantInt>(II.getArgOperand(1));
2769     auto *CIIndex = dyn_cast<ConstantInt>(II.getArgOperand(2));
2770 
2771     // Attempt to simplify to a constant or shuffle vector.
2772     if (Value *V = simplifyX86extrq(II, Op0, CILength, CIIndex, IC.Builder)) {
2773       return IC.replaceInstUsesWith(II, V);
2774     }
2775 
2776     // EXTRQI only uses the lowest 64-bits of the first 128-bit vector
2777     // operand.
2778     if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) {
2779       return IC.replaceOperand(II, 0, V);
2780     }
2781     break;
2782   }
2783 
2784   case Intrinsic::x86_sse4a_insertq: {
2785     Value *Op0 = II.getArgOperand(0);
2786     Value *Op1 = II.getArgOperand(1);
2787     unsigned VWidth = cast<FixedVectorType>(Op0->getType())->getNumElements();
2788     assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
2789            Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 &&
2790            cast<FixedVectorType>(Op1->getType())->getNumElements() == 2 &&
2791            "Unexpected operand size");
2792 
2793     // See if we're dealing with constant values.
2794     auto *C1 = dyn_cast<Constant>(Op1);
2795     auto *CI11 =
2796         C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1))
2797            : nullptr;
2798 
2799     // Attempt to simplify to a constant, shuffle vector or INSERTQI call.
2800     if (CI11) {
2801       const APInt &V11 = CI11->getValue();
2802       APInt Len = V11.zextOrTrunc(6);
2803       APInt Idx = V11.lshr(8).zextOrTrunc(6);
2804       if (Value *V = simplifyX86insertq(II, Op0, Op1, Len, Idx, IC.Builder)) {
2805         return IC.replaceInstUsesWith(II, V);
2806       }
2807     }
2808 
2809     // INSERTQ only uses the lowest 64-bits of the first 128-bit vector
2810     // operand.
2811     if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) {
2812       return IC.replaceOperand(II, 0, V);
2813     }
2814     break;
2815   }
2816 
2817   case Intrinsic::x86_sse4a_insertqi: {
2818     // INSERTQI: Extract lowest Length bits from lower half of second source and
2819     // insert over first source starting at Index bit. The upper 64-bits are
2820     // undefined.
2821     Value *Op0 = II.getArgOperand(0);
2822     Value *Op1 = II.getArgOperand(1);
2823     unsigned VWidth0 = cast<FixedVectorType>(Op0->getType())->getNumElements();
2824     unsigned VWidth1 = cast<FixedVectorType>(Op1->getType())->getNumElements();
2825     assert(Op0->getType()->getPrimitiveSizeInBits() == 128 &&
2826            Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 &&
2827            VWidth1 == 2 && "Unexpected operand sizes");
2828 
2829     // See if we're dealing with constant values.
2830     auto *CILength = dyn_cast<ConstantInt>(II.getArgOperand(2));
2831     auto *CIIndex = dyn_cast<ConstantInt>(II.getArgOperand(3));
2832 
2833     // Attempt to simplify to a constant or shuffle vector.
2834     if (CILength && CIIndex) {
2835       APInt Len = CILength->getValue().zextOrTrunc(6);
2836       APInt Idx = CIIndex->getValue().zextOrTrunc(6);
2837       if (Value *V = simplifyX86insertq(II, Op0, Op1, Len, Idx, IC.Builder)) {
2838         return IC.replaceInstUsesWith(II, V);
2839       }
2840     }
2841 
2842     // INSERTQI only uses the lowest 64-bits of the first two 128-bit vector
2843     // operands.
2844     bool MadeChange = false;
2845     if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) {
2846       IC.replaceOperand(II, 0, V);
2847       MadeChange = true;
2848     }
2849     if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) {
2850       IC.replaceOperand(II, 1, V);
2851       MadeChange = true;
2852     }
2853     if (MadeChange) {
2854       return &II;
2855     }
2856     break;
2857   }
2858 
2859   case Intrinsic::x86_sse41_pblendvb:
2860   case Intrinsic::x86_sse41_blendvps:
2861   case Intrinsic::x86_sse41_blendvpd:
2862   case Intrinsic::x86_avx_blendv_ps_256:
2863   case Intrinsic::x86_avx_blendv_pd_256:
2864   case Intrinsic::x86_avx2_pblendvb: {
2865     // fold (blend A, A, Mask) -> A
2866     Value *Op0 = II.getArgOperand(0);
2867     Value *Op1 = II.getArgOperand(1);
2868     Value *Mask = II.getArgOperand(2);
2869     if (Op0 == Op1) {
2870       return IC.replaceInstUsesWith(II, Op0);
2871     }
2872 
2873     // Zero Mask - select 1st argument.
2874     if (isa<ConstantAggregateZero>(Mask)) {
2875       return IC.replaceInstUsesWith(II, Op0);
2876     }
2877 
2878     // Constant Mask - select 1st/2nd argument lane based on top bit of mask.
2879     if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask)) {
2880       Constant *NewSelector =
2881           getNegativeIsTrueBoolVec(ConstantMask, IC.getDataLayout());
2882       return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
2883     }
2884 
2885     Mask = InstCombiner::peekThroughBitcast(Mask);
2886 
2887     // Peek through a one-use shuffle - VectorCombine should have simplified
2888     // this for cases where we're splitting wider vectors to use blendv
2889     // intrinsics.
2890     Value *MaskSrc = nullptr;
2891     ArrayRef<int> ShuffleMask;
2892     if (match(Mask, m_OneUse(m_Shuffle(m_Value(MaskSrc), m_Undef(),
2893                                        m_Mask(ShuffleMask))))) {
2894       // Bail if the shuffle was irregular or contains undefs.
2895       int NumElts = cast<FixedVectorType>(MaskSrc->getType())->getNumElements();
2896       if (NumElts < (int)ShuffleMask.size() || !isPowerOf2_32(NumElts) ||
2897           any_of(ShuffleMask,
2898                  [NumElts](int M) { return M < 0 || M >= NumElts; }))
2899         break;
2900       Mask = InstCombiner::peekThroughBitcast(MaskSrc);
2901     }
2902 
2903     // Convert to a vector select if we can bypass casts and find a boolean
2904     // vector condition value.
2905     Value *BoolVec;
2906     if (match(Mask, m_SExt(m_Value(BoolVec))) &&
2907         BoolVec->getType()->isVectorTy() &&
2908         BoolVec->getType()->getScalarSizeInBits() == 1) {
2909       auto *MaskTy = cast<FixedVectorType>(Mask->getType());
2910       auto *OpTy = cast<FixedVectorType>(II.getType());
2911       unsigned NumMaskElts = MaskTy->getNumElements();
2912       unsigned NumOperandElts = OpTy->getNumElements();
2913 
2914       // If we peeked through a shuffle, reapply the shuffle to the bool vector.
2915       if (MaskSrc) {
2916         unsigned NumMaskSrcElts =
2917             cast<FixedVectorType>(MaskSrc->getType())->getNumElements();
2918         NumMaskElts = (ShuffleMask.size() * NumMaskElts) / NumMaskSrcElts;
2919         // Multiple mask bits maps to the same operand element - bail out.
2920         if (NumMaskElts > NumOperandElts)
2921           break;
2922         SmallVector<int> ScaledMask;
2923         if (!llvm::scaleShuffleMaskElts(NumMaskElts, ShuffleMask, ScaledMask))
2924           break;
2925         BoolVec = IC.Builder.CreateShuffleVector(BoolVec, ScaledMask);
2926         MaskTy = FixedVectorType::get(MaskTy->getElementType(), NumMaskElts);
2927       }
2928       assert(MaskTy->getPrimitiveSizeInBits() ==
2929                  OpTy->getPrimitiveSizeInBits() &&
2930              "Not expecting mask and operands with different sizes");
2931 
2932       if (NumMaskElts == NumOperandElts) {
2933         return SelectInst::Create(BoolVec, Op1, Op0);
2934       }
2935 
2936       // If the mask has less elements than the operands, each mask bit maps to
2937       // multiple elements of the operands. Bitcast back and forth.
2938       if (NumMaskElts < NumOperandElts) {
2939         Value *CastOp0 = IC.Builder.CreateBitCast(Op0, MaskTy);
2940         Value *CastOp1 = IC.Builder.CreateBitCast(Op1, MaskTy);
2941         Value *Sel = IC.Builder.CreateSelect(BoolVec, CastOp1, CastOp0);
2942         return new BitCastInst(Sel, II.getType());
2943       }
2944     }
2945 
2946     break;
2947   }
2948 
2949   case Intrinsic::x86_ssse3_pshuf_b_128:
2950   case Intrinsic::x86_avx2_pshuf_b:
2951   case Intrinsic::x86_avx512_pshuf_b_512:
2952     if (Value *V = simplifyX86pshufb(II, IC.Builder)) {
2953       return IC.replaceInstUsesWith(II, V);
2954     }
2955     break;
2956 
2957   case Intrinsic::x86_avx_vpermilvar_ps:
2958   case Intrinsic::x86_avx_vpermilvar_ps_256:
2959   case Intrinsic::x86_avx512_vpermilvar_ps_512:
2960   case Intrinsic::x86_avx_vpermilvar_pd:
2961   case Intrinsic::x86_avx_vpermilvar_pd_256:
2962   case Intrinsic::x86_avx512_vpermilvar_pd_512:
2963     if (Value *V = simplifyX86vpermilvar(II, IC.Builder)) {
2964       return IC.replaceInstUsesWith(II, V);
2965     }
2966     break;
2967 
2968   case Intrinsic::x86_avx2_permd:
2969   case Intrinsic::x86_avx2_permps:
2970   case Intrinsic::x86_avx512_permvar_df_256:
2971   case Intrinsic::x86_avx512_permvar_df_512:
2972   case Intrinsic::x86_avx512_permvar_di_256:
2973   case Intrinsic::x86_avx512_permvar_di_512:
2974   case Intrinsic::x86_avx512_permvar_hi_128:
2975   case Intrinsic::x86_avx512_permvar_hi_256:
2976   case Intrinsic::x86_avx512_permvar_hi_512:
2977   case Intrinsic::x86_avx512_permvar_qi_128:
2978   case Intrinsic::x86_avx512_permvar_qi_256:
2979   case Intrinsic::x86_avx512_permvar_qi_512:
2980   case Intrinsic::x86_avx512_permvar_sf_512:
2981   case Intrinsic::x86_avx512_permvar_si_512:
2982     if (Value *V = simplifyX86vpermv(II, IC.Builder)) {
2983       return IC.replaceInstUsesWith(II, V);
2984     }
2985     break;
2986 
2987   case Intrinsic::x86_avx512_vpermi2var_d_128:
2988   case Intrinsic::x86_avx512_vpermi2var_d_256:
2989   case Intrinsic::x86_avx512_vpermi2var_d_512:
2990   case Intrinsic::x86_avx512_vpermi2var_hi_128:
2991   case Intrinsic::x86_avx512_vpermi2var_hi_256:
2992   case Intrinsic::x86_avx512_vpermi2var_hi_512:
2993   case Intrinsic::x86_avx512_vpermi2var_pd_128:
2994   case Intrinsic::x86_avx512_vpermi2var_pd_256:
2995   case Intrinsic::x86_avx512_vpermi2var_pd_512:
2996   case Intrinsic::x86_avx512_vpermi2var_ps_128:
2997   case Intrinsic::x86_avx512_vpermi2var_ps_256:
2998   case Intrinsic::x86_avx512_vpermi2var_ps_512:
2999   case Intrinsic::x86_avx512_vpermi2var_q_128:
3000   case Intrinsic::x86_avx512_vpermi2var_q_256:
3001   case Intrinsic::x86_avx512_vpermi2var_q_512:
3002   case Intrinsic::x86_avx512_vpermi2var_qi_128:
3003   case Intrinsic::x86_avx512_vpermi2var_qi_256:
3004   case Intrinsic::x86_avx512_vpermi2var_qi_512:
3005     if (Value *V = simplifyX86vpermv3(II, IC.Builder)) {
3006       return IC.replaceInstUsesWith(II, V);
3007     }
3008     break;
3009 
3010   case Intrinsic::x86_avx_maskload_ps:
3011   case Intrinsic::x86_avx_maskload_pd:
3012   case Intrinsic::x86_avx_maskload_ps_256:
3013   case Intrinsic::x86_avx_maskload_pd_256:
3014   case Intrinsic::x86_avx2_maskload_d:
3015   case Intrinsic::x86_avx2_maskload_q:
3016   case Intrinsic::x86_avx2_maskload_d_256:
3017   case Intrinsic::x86_avx2_maskload_q_256:
3018     if (Instruction *I = simplifyX86MaskedLoad(II, IC)) {
3019       return I;
3020     }
3021     break;
3022 
3023   case Intrinsic::x86_sse2_maskmov_dqu:
3024   case Intrinsic::x86_avx_maskstore_ps:
3025   case Intrinsic::x86_avx_maskstore_pd:
3026   case Intrinsic::x86_avx_maskstore_ps_256:
3027   case Intrinsic::x86_avx_maskstore_pd_256:
3028   case Intrinsic::x86_avx2_maskstore_d:
3029   case Intrinsic::x86_avx2_maskstore_q:
3030   case Intrinsic::x86_avx2_maskstore_d_256:
3031   case Intrinsic::x86_avx2_maskstore_q_256:
3032     if (simplifyX86MaskedStore(II, IC)) {
3033       return nullptr;
3034     }
3035     break;
3036 
3037   case Intrinsic::x86_addcarry_32:
3038   case Intrinsic::x86_addcarry_64:
3039     if (Value *V = simplifyX86addcarry(II, IC.Builder)) {
3040       return IC.replaceInstUsesWith(II, V);
3041     }
3042     break;
3043 
3044   case Intrinsic::x86_avx512_pternlog_d_128:
3045   case Intrinsic::x86_avx512_pternlog_d_256:
3046   case Intrinsic::x86_avx512_pternlog_d_512:
3047   case Intrinsic::x86_avx512_pternlog_q_128:
3048   case Intrinsic::x86_avx512_pternlog_q_256:
3049   case Intrinsic::x86_avx512_pternlog_q_512:
3050     if (Value *V = simplifyTernarylogic(II, IC.Builder)) {
3051       return IC.replaceInstUsesWith(II, V);
3052     }
3053     break;
3054   default:
3055     break;
3056   }
3057   return std::nullopt;
3058 }
3059 
3060 std::optional<Value *> X86TTIImpl::simplifyDemandedUseBitsIntrinsic(
3061     InstCombiner &IC, IntrinsicInst &II, APInt DemandedMask, KnownBits &Known,
3062     bool &KnownBitsComputed) const {
3063   switch (II.getIntrinsicID()) {
3064   default:
3065     break;
3066   case Intrinsic::x86_mmx_pmovmskb:
3067   case Intrinsic::x86_sse_movmsk_ps:
3068   case Intrinsic::x86_sse2_movmsk_pd:
3069   case Intrinsic::x86_sse2_pmovmskb_128:
3070   case Intrinsic::x86_avx_movmsk_ps_256:
3071   case Intrinsic::x86_avx_movmsk_pd_256:
3072   case Intrinsic::x86_avx2_pmovmskb: {
3073     // MOVMSK copies the vector elements' sign bits to the low bits
3074     // and zeros the high bits.
3075     unsigned ArgWidth;
3076     if (II.getIntrinsicID() == Intrinsic::x86_mmx_pmovmskb) {
3077       ArgWidth = 8; // Arg is x86_mmx, but treated as <8 x i8>.
3078     } else {
3079       auto *ArgType = cast<FixedVectorType>(II.getArgOperand(0)->getType());
3080       ArgWidth = ArgType->getNumElements();
3081     }
3082 
3083     // If we don't need any of low bits then return zero,
3084     // we know that DemandedMask is non-zero already.
3085     APInt DemandedElts = DemandedMask.zextOrTrunc(ArgWidth);
3086     Type *VTy = II.getType();
3087     if (DemandedElts.isZero()) {
3088       return ConstantInt::getNullValue(VTy);
3089     }
3090 
3091     // We know that the upper bits are set to zero.
3092     Known.Zero.setBitsFrom(ArgWidth);
3093     KnownBitsComputed = true;
3094     break;
3095   }
3096   }
3097   return std::nullopt;
3098 }
3099 
3100 std::optional<Value *> X86TTIImpl::simplifyDemandedVectorEltsIntrinsic(
3101     InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts,
3102     APInt &UndefElts2, APInt &UndefElts3,
3103     std::function<void(Instruction *, unsigned, APInt, APInt &)>
3104         simplifyAndSetOp) const {
3105   unsigned VWidth = cast<FixedVectorType>(II.getType())->getNumElements();
3106   switch (II.getIntrinsicID()) {
3107   default:
3108     break;
3109   case Intrinsic::x86_xop_vfrcz_ss:
3110   case Intrinsic::x86_xop_vfrcz_sd:
3111     // The instructions for these intrinsics are speced to zero upper bits not
3112     // pass them through like other scalar intrinsics. So we shouldn't just
3113     // use Arg0 if DemandedElts[0] is clear like we do for other intrinsics.
3114     // Instead we should return a zero vector.
3115     if (!DemandedElts[0]) {
3116       IC.addToWorklist(&II);
3117       return ConstantAggregateZero::get(II.getType());
3118     }
3119 
3120     // Only the lower element is used.
3121     DemandedElts = 1;
3122     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3123 
3124     // Only the lower element is undefined. The high elements are zero.
3125     UndefElts = UndefElts[0];
3126     break;
3127 
3128   // Unary scalar-as-vector operations that work column-wise.
3129   case Intrinsic::x86_sse_rcp_ss:
3130   case Intrinsic::x86_sse_rsqrt_ss:
3131     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3132 
3133     // If lowest element of a scalar op isn't used then use Arg0.
3134     if (!DemandedElts[0]) {
3135       IC.addToWorklist(&II);
3136       return II.getArgOperand(0);
3137     }
3138     // TODO: If only low elt lower SQRT to FSQRT (with rounding/exceptions
3139     // checks).
3140     break;
3141 
3142   // Binary scalar-as-vector operations that work column-wise. The high
3143   // elements come from operand 0. The low element is a function of both
3144   // operands.
3145   case Intrinsic::x86_sse_min_ss:
3146   case Intrinsic::x86_sse_max_ss:
3147   case Intrinsic::x86_sse_cmp_ss:
3148   case Intrinsic::x86_sse2_min_sd:
3149   case Intrinsic::x86_sse2_max_sd:
3150   case Intrinsic::x86_sse2_cmp_sd: {
3151     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3152 
3153     // If lowest element of a scalar op isn't used then use Arg0.
3154     if (!DemandedElts[0]) {
3155       IC.addToWorklist(&II);
3156       return II.getArgOperand(0);
3157     }
3158 
3159     // Only lower element is used for operand 1.
3160     DemandedElts = 1;
3161     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts2);
3162 
3163     // Lower element is undefined if both lower elements are undefined.
3164     // Consider things like undef&0.  The result is known zero, not undef.
3165     if (!UndefElts2[0])
3166       UndefElts.clearBit(0);
3167 
3168     break;
3169   }
3170 
3171   // Binary scalar-as-vector operations that work column-wise. The high
3172   // elements come from operand 0 and the low element comes from operand 1.
3173   case Intrinsic::x86_sse41_round_ss:
3174   case Intrinsic::x86_sse41_round_sd: {
3175     // Don't use the low element of operand 0.
3176     APInt DemandedElts2 = DemandedElts;
3177     DemandedElts2.clearBit(0);
3178     simplifyAndSetOp(&II, 0, DemandedElts2, UndefElts);
3179 
3180     // If lowest element of a scalar op isn't used then use Arg0.
3181     if (!DemandedElts[0]) {
3182       IC.addToWorklist(&II);
3183       return II.getArgOperand(0);
3184     }
3185 
3186     // Only lower element is used for operand 1.
3187     DemandedElts = 1;
3188     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts2);
3189 
3190     // Take the high undef elements from operand 0 and take the lower element
3191     // from operand 1.
3192     UndefElts.clearBit(0);
3193     UndefElts |= UndefElts2[0];
3194     break;
3195   }
3196 
3197   // Three input scalar-as-vector operations that work column-wise. The high
3198   // elements come from operand 0 and the low element is a function of all
3199   // three inputs.
3200   case Intrinsic::x86_avx512_mask_add_ss_round:
3201   case Intrinsic::x86_avx512_mask_div_ss_round:
3202   case Intrinsic::x86_avx512_mask_mul_ss_round:
3203   case Intrinsic::x86_avx512_mask_sub_ss_round:
3204   case Intrinsic::x86_avx512_mask_max_ss_round:
3205   case Intrinsic::x86_avx512_mask_min_ss_round:
3206   case Intrinsic::x86_avx512_mask_add_sd_round:
3207   case Intrinsic::x86_avx512_mask_div_sd_round:
3208   case Intrinsic::x86_avx512_mask_mul_sd_round:
3209   case Intrinsic::x86_avx512_mask_sub_sd_round:
3210   case Intrinsic::x86_avx512_mask_max_sd_round:
3211   case Intrinsic::x86_avx512_mask_min_sd_round:
3212     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3213 
3214     // If lowest element of a scalar op isn't used then use Arg0.
3215     if (!DemandedElts[0]) {
3216       IC.addToWorklist(&II);
3217       return II.getArgOperand(0);
3218     }
3219 
3220     // Only lower element is used for operand 1 and 2.
3221     DemandedElts = 1;
3222     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts2);
3223     simplifyAndSetOp(&II, 2, DemandedElts, UndefElts3);
3224 
3225     // Lower element is undefined if all three lower elements are undefined.
3226     // Consider things like undef&0.  The result is known zero, not undef.
3227     if (!UndefElts2[0] || !UndefElts3[0])
3228       UndefElts.clearBit(0);
3229     break;
3230 
3231   // TODO: Add fmaddsub support?
3232   case Intrinsic::x86_sse3_addsub_pd:
3233   case Intrinsic::x86_sse3_addsub_ps:
3234   case Intrinsic::x86_avx_addsub_pd_256:
3235   case Intrinsic::x86_avx_addsub_ps_256: {
3236     // If none of the even or none of the odd lanes are required, turn this
3237     // into a generic FP math instruction.
3238     APInt SubMask = APInt::getSplat(VWidth, APInt(2, 0x1));
3239     APInt AddMask = APInt::getSplat(VWidth, APInt(2, 0x2));
3240     bool IsSubOnly = DemandedElts.isSubsetOf(SubMask);
3241     bool IsAddOnly = DemandedElts.isSubsetOf(AddMask);
3242     if (IsSubOnly || IsAddOnly) {
3243       assert((IsSubOnly ^ IsAddOnly) && "Can't be both add-only and sub-only");
3244       IRBuilderBase::InsertPointGuard Guard(IC.Builder);
3245       IC.Builder.SetInsertPoint(&II);
3246       Value *Arg0 = II.getArgOperand(0), *Arg1 = II.getArgOperand(1);
3247       return IC.Builder.CreateBinOp(
3248           IsSubOnly ? Instruction::FSub : Instruction::FAdd, Arg0, Arg1);
3249     }
3250 
3251     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3252     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts2);
3253     UndefElts &= UndefElts2;
3254     break;
3255   }
3256 
3257   // General per-element vector operations.
3258   case Intrinsic::x86_avx2_psllv_d:
3259   case Intrinsic::x86_avx2_psllv_d_256:
3260   case Intrinsic::x86_avx2_psllv_q:
3261   case Intrinsic::x86_avx2_psllv_q_256:
3262   case Intrinsic::x86_avx2_psrlv_d:
3263   case Intrinsic::x86_avx2_psrlv_d_256:
3264   case Intrinsic::x86_avx2_psrlv_q:
3265   case Intrinsic::x86_avx2_psrlv_q_256:
3266   case Intrinsic::x86_avx2_psrav_d:
3267   case Intrinsic::x86_avx2_psrav_d_256: {
3268     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3269     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts2);
3270     UndefElts &= UndefElts2;
3271     break;
3272   }
3273 
3274   case Intrinsic::x86_sse2_pmulh_w:
3275   case Intrinsic::x86_avx2_pmulh_w:
3276   case Intrinsic::x86_avx512_pmulh_w_512:
3277   case Intrinsic::x86_sse2_pmulhu_w:
3278   case Intrinsic::x86_avx2_pmulhu_w:
3279   case Intrinsic::x86_avx512_pmulhu_w_512:
3280   case Intrinsic::x86_ssse3_pmul_hr_sw_128:
3281   case Intrinsic::x86_avx2_pmul_hr_sw:
3282   case Intrinsic::x86_avx512_pmul_hr_sw_512: {
3283     simplifyAndSetOp(&II, 0, DemandedElts, UndefElts);
3284     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts2);
3285     // NOTE: mulh(undef,undef) != undef.
3286     break;
3287   }
3288 
3289   case Intrinsic::x86_sse2_packssdw_128:
3290   case Intrinsic::x86_sse2_packsswb_128:
3291   case Intrinsic::x86_sse2_packuswb_128:
3292   case Intrinsic::x86_sse41_packusdw:
3293   case Intrinsic::x86_avx2_packssdw:
3294   case Intrinsic::x86_avx2_packsswb:
3295   case Intrinsic::x86_avx2_packusdw:
3296   case Intrinsic::x86_avx2_packuswb:
3297   case Intrinsic::x86_avx512_packssdw_512:
3298   case Intrinsic::x86_avx512_packsswb_512:
3299   case Intrinsic::x86_avx512_packusdw_512:
3300   case Intrinsic::x86_avx512_packuswb_512: {
3301     auto *Ty0 = II.getArgOperand(0)->getType();
3302     unsigned InnerVWidth = cast<FixedVectorType>(Ty0)->getNumElements();
3303     assert(VWidth == (InnerVWidth * 2) && "Unexpected input size");
3304 
3305     unsigned NumLanes = Ty0->getPrimitiveSizeInBits() / 128;
3306     unsigned VWidthPerLane = VWidth / NumLanes;
3307     unsigned InnerVWidthPerLane = InnerVWidth / NumLanes;
3308 
3309     // Per lane, pack the elements of the first input and then the second.
3310     // e.g.
3311     // v8i16 PACK(v4i32 X, v4i32 Y) - (X[0..3],Y[0..3])
3312     // v32i8 PACK(v16i16 X, v16i16 Y) - (X[0..7],Y[0..7]),(X[8..15],Y[8..15])
3313     for (int OpNum = 0; OpNum != 2; ++OpNum) {
3314       APInt OpDemandedElts(InnerVWidth, 0);
3315       for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
3316         unsigned LaneIdx = Lane * VWidthPerLane;
3317         for (unsigned Elt = 0; Elt != InnerVWidthPerLane; ++Elt) {
3318           unsigned Idx = LaneIdx + Elt + InnerVWidthPerLane * OpNum;
3319           if (DemandedElts[Idx])
3320             OpDemandedElts.setBit((Lane * InnerVWidthPerLane) + Elt);
3321         }
3322       }
3323 
3324       // Demand elements from the operand.
3325       APInt OpUndefElts(InnerVWidth, 0);
3326       simplifyAndSetOp(&II, OpNum, OpDemandedElts, OpUndefElts);
3327 
3328       // Pack the operand's UNDEF elements, one lane at a time.
3329       OpUndefElts = OpUndefElts.zext(VWidth);
3330       for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
3331         APInt LaneElts = OpUndefElts.lshr(InnerVWidthPerLane * Lane);
3332         LaneElts = LaneElts.getLoBits(InnerVWidthPerLane);
3333         LaneElts <<= InnerVWidthPerLane * (2 * Lane + OpNum);
3334         UndefElts |= LaneElts;
3335       }
3336     }
3337     break;
3338   }
3339 
3340   case Intrinsic::x86_sse2_pmadd_wd:
3341   case Intrinsic::x86_avx2_pmadd_wd:
3342   case Intrinsic::x86_avx512_pmaddw_d_512:
3343   case Intrinsic::x86_ssse3_pmadd_ub_sw_128:
3344   case Intrinsic::x86_avx2_pmadd_ub_sw:
3345   case Intrinsic::x86_avx512_pmaddubs_w_512: {
3346     // PMADD - demand both src elements that map to each dst element.
3347     auto *ArgTy = II.getArgOperand(0)->getType();
3348     unsigned InnerVWidth = cast<FixedVectorType>(ArgTy)->getNumElements();
3349     assert((VWidth * 2) == InnerVWidth && "Unexpected input size");
3350     APInt OpDemandedElts = APIntOps::ScaleBitMask(DemandedElts, InnerVWidth);
3351     APInt Op0UndefElts(InnerVWidth, 0);
3352     APInt Op1UndefElts(InnerVWidth, 0);
3353     simplifyAndSetOp(&II, 0, OpDemandedElts, Op0UndefElts);
3354     simplifyAndSetOp(&II, 1, OpDemandedElts, Op1UndefElts);
3355     // NOTE: madd(undef,undef) != undef.
3356     break;
3357   }
3358 
3359   // PSHUFB
3360   case Intrinsic::x86_ssse3_pshuf_b_128:
3361   case Intrinsic::x86_avx2_pshuf_b:
3362   case Intrinsic::x86_avx512_pshuf_b_512:
3363   // PERMILVAR
3364   case Intrinsic::x86_avx_vpermilvar_ps:
3365   case Intrinsic::x86_avx_vpermilvar_ps_256:
3366   case Intrinsic::x86_avx512_vpermilvar_ps_512:
3367   case Intrinsic::x86_avx_vpermilvar_pd:
3368   case Intrinsic::x86_avx_vpermilvar_pd_256:
3369   case Intrinsic::x86_avx512_vpermilvar_pd_512:
3370   // PERMV
3371   case Intrinsic::x86_avx2_permd:
3372   case Intrinsic::x86_avx2_permps: {
3373     simplifyAndSetOp(&II, 1, DemandedElts, UndefElts);
3374     break;
3375   }
3376 
3377   // SSE4A instructions leave the upper 64-bits of the 128-bit result
3378   // in an undefined state.
3379   case Intrinsic::x86_sse4a_extrq:
3380   case Intrinsic::x86_sse4a_extrqi:
3381   case Intrinsic::x86_sse4a_insertq:
3382   case Intrinsic::x86_sse4a_insertqi:
3383     UndefElts.setHighBits(VWidth / 2);
3384     break;
3385   }
3386   return std::nullopt;
3387 }
3388