xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp (revision d56accc7c3dcc897489b6a07834763a03b9f3d68)
1 //===- ARMTargetTransformInfo.cpp - ARM specific TTI ----------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "ARMTargetTransformInfo.h"
10 #include "ARMSubtarget.h"
11 #include "MCTargetDesc/ARMAddressingModes.h"
12 #include "llvm/ADT/APInt.h"
13 #include "llvm/ADT/SmallVector.h"
14 #include "llvm/Analysis/LoopInfo.h"
15 #include "llvm/CodeGen/CostTable.h"
16 #include "llvm/CodeGen/ISDOpcodes.h"
17 #include "llvm/CodeGen/ValueTypes.h"
18 #include "llvm/IR/BasicBlock.h"
19 #include "llvm/IR/DataLayout.h"
20 #include "llvm/IR/DerivedTypes.h"
21 #include "llvm/IR/Instruction.h"
22 #include "llvm/IR/Instructions.h"
23 #include "llvm/IR/Intrinsics.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/IntrinsicsARM.h"
26 #include "llvm/IR/PatternMatch.h"
27 #include "llvm/IR/Type.h"
28 #include "llvm/MC/SubtargetFeature.h"
29 #include "llvm/Support/Casting.h"
30 #include "llvm/Support/KnownBits.h"
31 #include "llvm/Support/MachineValueType.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Transforms/InstCombine/InstCombiner.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/Transforms/Utils/LoopUtils.h"
36 #include <algorithm>
37 #include <cassert>
38 #include <cstdint>
39 #include <utility>
40 
41 using namespace llvm;
42 
43 #define DEBUG_TYPE "armtti"
44 
45 static cl::opt<bool> EnableMaskedLoadStores(
46   "enable-arm-maskedldst", cl::Hidden, cl::init(true),
47   cl::desc("Enable the generation of masked loads and stores"));
48 
49 static cl::opt<bool> DisableLowOverheadLoops(
50   "disable-arm-loloops", cl::Hidden, cl::init(false),
51   cl::desc("Disable the generation of low-overhead loops"));
52 
53 static cl::opt<bool>
54     AllowWLSLoops("allow-arm-wlsloops", cl::Hidden, cl::init(true),
55                   cl::desc("Enable the generation of WLS loops"));
56 
57 extern cl::opt<TailPredication::Mode> EnableTailPredication;
58 
59 extern cl::opt<bool> EnableMaskedGatherScatters;
60 
61 extern cl::opt<unsigned> MVEMaxSupportedInterleaveFactor;
62 
63 /// Convert a vector load intrinsic into a simple llvm load instruction.
64 /// This is beneficial when the underlying object being addressed comes
65 /// from a constant, since we get constant-folding for free.
66 static Value *simplifyNeonVld1(const IntrinsicInst &II, unsigned MemAlign,
67                                InstCombiner::BuilderTy &Builder) {
68   auto *IntrAlign = dyn_cast<ConstantInt>(II.getArgOperand(1));
69 
70   if (!IntrAlign)
71     return nullptr;
72 
73   unsigned Alignment = IntrAlign->getLimitedValue() < MemAlign
74                            ? MemAlign
75                            : IntrAlign->getLimitedValue();
76 
77   if (!isPowerOf2_32(Alignment))
78     return nullptr;
79 
80   auto *BCastInst = Builder.CreateBitCast(II.getArgOperand(0),
81                                           PointerType::get(II.getType(), 0));
82   return Builder.CreateAlignedLoad(II.getType(), BCastInst, Align(Alignment));
83 }
84 
85 bool ARMTTIImpl::areInlineCompatible(const Function *Caller,
86                                      const Function *Callee) const {
87   const TargetMachine &TM = getTLI()->getTargetMachine();
88   const FeatureBitset &CallerBits =
89       TM.getSubtargetImpl(*Caller)->getFeatureBits();
90   const FeatureBitset &CalleeBits =
91       TM.getSubtargetImpl(*Callee)->getFeatureBits();
92 
93   // To inline a callee, all features not in the allowed list must match exactly.
94   bool MatchExact = (CallerBits & ~InlineFeaturesAllowed) ==
95                     (CalleeBits & ~InlineFeaturesAllowed);
96   // For features in the allowed list, the callee's features must be a subset of
97   // the callers'.
98   bool MatchSubset = ((CallerBits & CalleeBits) & InlineFeaturesAllowed) ==
99                      (CalleeBits & InlineFeaturesAllowed);
100   return MatchExact && MatchSubset;
101 }
102 
103 TTI::AddressingModeKind
104 ARMTTIImpl::getPreferredAddressingMode(const Loop *L,
105                                        ScalarEvolution *SE) const {
106   if (ST->hasMVEIntegerOps())
107     return TTI::AMK_PostIndexed;
108 
109   if (L->getHeader()->getParent()->hasOptSize())
110     return TTI::AMK_None;
111 
112   if (ST->isMClass() && ST->isThumb2() &&
113       L->getNumBlocks() == 1)
114     return TTI::AMK_PreIndexed;
115 
116   return TTI::AMK_None;
117 }
118 
119 Optional<Instruction *>
120 ARMTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
121   using namespace PatternMatch;
122   Intrinsic::ID IID = II.getIntrinsicID();
123   switch (IID) {
124   default:
125     break;
126   case Intrinsic::arm_neon_vld1: {
127     Align MemAlign =
128         getKnownAlignment(II.getArgOperand(0), IC.getDataLayout(), &II,
129                           &IC.getAssumptionCache(), &IC.getDominatorTree());
130     if (Value *V = simplifyNeonVld1(II, MemAlign.value(), IC.Builder)) {
131       return IC.replaceInstUsesWith(II, V);
132     }
133     break;
134   }
135 
136   case Intrinsic::arm_neon_vld2:
137   case Intrinsic::arm_neon_vld3:
138   case Intrinsic::arm_neon_vld4:
139   case Intrinsic::arm_neon_vld2lane:
140   case Intrinsic::arm_neon_vld3lane:
141   case Intrinsic::arm_neon_vld4lane:
142   case Intrinsic::arm_neon_vst1:
143   case Intrinsic::arm_neon_vst2:
144   case Intrinsic::arm_neon_vst3:
145   case Intrinsic::arm_neon_vst4:
146   case Intrinsic::arm_neon_vst2lane:
147   case Intrinsic::arm_neon_vst3lane:
148   case Intrinsic::arm_neon_vst4lane: {
149     Align MemAlign =
150         getKnownAlignment(II.getArgOperand(0), IC.getDataLayout(), &II,
151                           &IC.getAssumptionCache(), &IC.getDominatorTree());
152     unsigned AlignArg = II.arg_size() - 1;
153     Value *AlignArgOp = II.getArgOperand(AlignArg);
154     MaybeAlign Align = cast<ConstantInt>(AlignArgOp)->getMaybeAlignValue();
155     if (Align && *Align < MemAlign) {
156       return IC.replaceOperand(
157           II, AlignArg,
158           ConstantInt::get(Type::getInt32Ty(II.getContext()), MemAlign.value(),
159                            false));
160     }
161     break;
162   }
163 
164   case Intrinsic::arm_mve_pred_i2v: {
165     Value *Arg = II.getArgOperand(0);
166     Value *ArgArg;
167     if (match(Arg, PatternMatch::m_Intrinsic<Intrinsic::arm_mve_pred_v2i>(
168                        PatternMatch::m_Value(ArgArg))) &&
169         II.getType() == ArgArg->getType()) {
170       return IC.replaceInstUsesWith(II, ArgArg);
171     }
172     Constant *XorMask;
173     if (match(Arg, m_Xor(PatternMatch::m_Intrinsic<Intrinsic::arm_mve_pred_v2i>(
174                              PatternMatch::m_Value(ArgArg)),
175                          PatternMatch::m_Constant(XorMask))) &&
176         II.getType() == ArgArg->getType()) {
177       if (auto *CI = dyn_cast<ConstantInt>(XorMask)) {
178         if (CI->getValue().trunc(16).isAllOnes()) {
179           auto TrueVector = IC.Builder.CreateVectorSplat(
180               cast<FixedVectorType>(II.getType())->getNumElements(),
181               IC.Builder.getTrue());
182           return BinaryOperator::Create(Instruction::Xor, ArgArg, TrueVector);
183         }
184       }
185     }
186     KnownBits ScalarKnown(32);
187     if (IC.SimplifyDemandedBits(&II, 0, APInt::getLowBitsSet(32, 16),
188                                 ScalarKnown, 0)) {
189       return &II;
190     }
191     break;
192   }
193   case Intrinsic::arm_mve_pred_v2i: {
194     Value *Arg = II.getArgOperand(0);
195     Value *ArgArg;
196     if (match(Arg, PatternMatch::m_Intrinsic<Intrinsic::arm_mve_pred_i2v>(
197                        PatternMatch::m_Value(ArgArg)))) {
198       return IC.replaceInstUsesWith(II, ArgArg);
199     }
200     if (!II.getMetadata(LLVMContext::MD_range)) {
201       Type *IntTy32 = Type::getInt32Ty(II.getContext());
202       Metadata *M[] = {
203           ConstantAsMetadata::get(ConstantInt::get(IntTy32, 0)),
204           ConstantAsMetadata::get(ConstantInt::get(IntTy32, 0x10000))};
205       II.setMetadata(LLVMContext::MD_range, MDNode::get(II.getContext(), M));
206       return &II;
207     }
208     break;
209   }
210   case Intrinsic::arm_mve_vadc:
211   case Intrinsic::arm_mve_vadc_predicated: {
212     unsigned CarryOp =
213         (II.getIntrinsicID() == Intrinsic::arm_mve_vadc_predicated) ? 3 : 2;
214     assert(II.getArgOperand(CarryOp)->getType()->getScalarSizeInBits() == 32 &&
215            "Bad type for intrinsic!");
216 
217     KnownBits CarryKnown(32);
218     if (IC.SimplifyDemandedBits(&II, CarryOp, APInt::getOneBitSet(32, 29),
219                                 CarryKnown)) {
220       return &II;
221     }
222     break;
223   }
224   case Intrinsic::arm_mve_vmldava: {
225     Instruction *I = cast<Instruction>(&II);
226     if (I->hasOneUse()) {
227       auto *User = cast<Instruction>(*I->user_begin());
228       Value *OpZ;
229       if (match(User, m_c_Add(m_Specific(I), m_Value(OpZ))) &&
230           match(I->getOperand(3), m_Zero())) {
231         Value *OpX = I->getOperand(4);
232         Value *OpY = I->getOperand(5);
233         Type *OpTy = OpX->getType();
234 
235         IC.Builder.SetInsertPoint(User);
236         Value *V =
237             IC.Builder.CreateIntrinsic(Intrinsic::arm_mve_vmldava, {OpTy},
238                                        {I->getOperand(0), I->getOperand(1),
239                                         I->getOperand(2), OpZ, OpX, OpY});
240 
241         IC.replaceInstUsesWith(*User, V);
242         return IC.eraseInstFromFunction(*User);
243       }
244     }
245     return None;
246   }
247   }
248   return None;
249 }
250 
251 Optional<Value *> ARMTTIImpl::simplifyDemandedVectorEltsIntrinsic(
252     InstCombiner &IC, IntrinsicInst &II, APInt OrigDemandedElts,
253     APInt &UndefElts, APInt &UndefElts2, APInt &UndefElts3,
254     std::function<void(Instruction *, unsigned, APInt, APInt &)>
255         SimplifyAndSetOp) const {
256 
257   // Compute the demanded bits for a narrowing MVE intrinsic. The TopOpc is the
258   // opcode specifying a Top/Bottom instruction, which can change between
259   // instructions.
260   auto SimplifyNarrowInstrTopBottom =[&](unsigned TopOpc) {
261     unsigned NumElts = cast<FixedVectorType>(II.getType())->getNumElements();
262     unsigned IsTop = cast<ConstantInt>(II.getOperand(TopOpc))->getZExtValue();
263 
264     // The only odd/even lanes of operand 0 will only be demanded depending
265     // on whether this is a top/bottom instruction.
266     APInt DemandedElts =
267         APInt::getSplat(NumElts, IsTop ? APInt::getLowBitsSet(2, 1)
268                                        : APInt::getHighBitsSet(2, 1));
269     SimplifyAndSetOp(&II, 0, OrigDemandedElts & DemandedElts, UndefElts);
270     // The other lanes will be defined from the inserted elements.
271     UndefElts &= APInt::getSplat(NumElts, !IsTop ? APInt::getLowBitsSet(2, 1)
272                                                  : APInt::getHighBitsSet(2, 1));
273     return None;
274   };
275 
276   switch (II.getIntrinsicID()) {
277   default:
278     break;
279   case Intrinsic::arm_mve_vcvt_narrow:
280     SimplifyNarrowInstrTopBottom(2);
281     break;
282   case Intrinsic::arm_mve_vqmovn:
283     SimplifyNarrowInstrTopBottom(4);
284     break;
285   case Intrinsic::arm_mve_vshrn:
286     SimplifyNarrowInstrTopBottom(7);
287     break;
288   }
289 
290   return None;
291 }
292 
293 InstructionCost ARMTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
294                                           TTI::TargetCostKind CostKind) {
295   assert(Ty->isIntegerTy());
296 
297  unsigned Bits = Ty->getPrimitiveSizeInBits();
298  if (Bits == 0 || Imm.getActiveBits() >= 64)
299    return 4;
300 
301   int64_t SImmVal = Imm.getSExtValue();
302   uint64_t ZImmVal = Imm.getZExtValue();
303   if (!ST->isThumb()) {
304     if ((SImmVal >= 0 && SImmVal < 65536) ||
305         (ARM_AM::getSOImmVal(ZImmVal) != -1) ||
306         (ARM_AM::getSOImmVal(~ZImmVal) != -1))
307       return 1;
308     return ST->hasV6T2Ops() ? 2 : 3;
309   }
310   if (ST->isThumb2()) {
311     if ((SImmVal >= 0 && SImmVal < 65536) ||
312         (ARM_AM::getT2SOImmVal(ZImmVal) != -1) ||
313         (ARM_AM::getT2SOImmVal(~ZImmVal) != -1))
314       return 1;
315     return ST->hasV6T2Ops() ? 2 : 3;
316   }
317   // Thumb1, any i8 imm cost 1.
318   if (Bits == 8 || (SImmVal >= 0 && SImmVal < 256))
319     return 1;
320   if ((~SImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal))
321     return 2;
322   // Load from constantpool.
323   return 3;
324 }
325 
326 // Constants smaller than 256 fit in the immediate field of
327 // Thumb1 instructions so we return a zero cost and 1 otherwise.
328 InstructionCost ARMTTIImpl::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx,
329                                                   const APInt &Imm, Type *Ty) {
330   if (Imm.isNonNegative() && Imm.getLimitedValue() < 256)
331     return 0;
332 
333   return 1;
334 }
335 
336 // Checks whether Inst is part of a min(max()) or max(min()) pattern
337 // that will match to an SSAT instruction. Returns the instruction being
338 // saturated, or null if no saturation pattern was found.
339 static Value *isSSATMinMaxPattern(Instruction *Inst, const APInt &Imm) {
340   Value *LHS, *RHS;
341   ConstantInt *C;
342   SelectPatternFlavor InstSPF = matchSelectPattern(Inst, LHS, RHS).Flavor;
343 
344   if (InstSPF == SPF_SMAX &&
345       PatternMatch::match(RHS, PatternMatch::m_ConstantInt(C)) &&
346       C->getValue() == Imm && Imm.isNegative() && Imm.isNegatedPowerOf2()) {
347 
348     auto isSSatMin = [&](Value *MinInst) {
349       if (isa<SelectInst>(MinInst)) {
350         Value *MinLHS, *MinRHS;
351         ConstantInt *MinC;
352         SelectPatternFlavor MinSPF =
353             matchSelectPattern(MinInst, MinLHS, MinRHS).Flavor;
354         if (MinSPF == SPF_SMIN &&
355             PatternMatch::match(MinRHS, PatternMatch::m_ConstantInt(MinC)) &&
356             MinC->getValue() == ((-Imm) - 1))
357           return true;
358       }
359       return false;
360     };
361 
362     if (isSSatMin(Inst->getOperand(1)))
363       return cast<Instruction>(Inst->getOperand(1))->getOperand(1);
364     if (Inst->hasNUses(2) &&
365         (isSSatMin(*Inst->user_begin()) || isSSatMin(*(++Inst->user_begin()))))
366       return Inst->getOperand(1);
367   }
368   return nullptr;
369 }
370 
371 // Look for a FP Saturation pattern, where the instruction can be simplified to
372 // a fptosi.sat. max(min(fptosi)). The constant in this case is always free.
373 static bool isFPSatMinMaxPattern(Instruction *Inst, const APInt &Imm) {
374   if (Imm.getBitWidth() != 64 ||
375       Imm != APInt::getHighBitsSet(64, 33)) // -2147483648
376     return false;
377   Value *FP = isSSATMinMaxPattern(Inst, Imm);
378   if (!FP && isa<ICmpInst>(Inst) && Inst->hasOneUse())
379     FP = isSSATMinMaxPattern(cast<Instruction>(*Inst->user_begin()), Imm);
380   if (!FP)
381     return false;
382   return isa<FPToSIInst>(FP);
383 }
384 
385 InstructionCost ARMTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
386                                               const APInt &Imm, Type *Ty,
387                                               TTI::TargetCostKind CostKind,
388                                               Instruction *Inst) {
389   // Division by a constant can be turned into multiplication, but only if we
390   // know it's constant. So it's not so much that the immediate is cheap (it's
391   // not), but that the alternative is worse.
392   // FIXME: this is probably unneeded with GlobalISel.
393   if ((Opcode == Instruction::SDiv || Opcode == Instruction::UDiv ||
394        Opcode == Instruction::SRem || Opcode == Instruction::URem) &&
395       Idx == 1)
396     return 0;
397 
398   // Leave any gep offsets for the CodeGenPrepare, which will do a better job at
399   // splitting any large offsets.
400   if (Opcode == Instruction::GetElementPtr && Idx != 0)
401     return 0;
402 
403   if (Opcode == Instruction::And) {
404     // UXTB/UXTH
405     if (Imm == 255 || Imm == 65535)
406       return 0;
407     // Conversion to BIC is free, and means we can use ~Imm instead.
408     return std::min(getIntImmCost(Imm, Ty, CostKind),
409                     getIntImmCost(~Imm, Ty, CostKind));
410   }
411 
412   if (Opcode == Instruction::Add)
413     // Conversion to SUB is free, and means we can use -Imm instead.
414     return std::min(getIntImmCost(Imm, Ty, CostKind),
415                     getIntImmCost(-Imm, Ty, CostKind));
416 
417   if (Opcode == Instruction::ICmp && Imm.isNegative() &&
418       Ty->getIntegerBitWidth() == 32) {
419     int64_t NegImm = -Imm.getSExtValue();
420     if (ST->isThumb2() && NegImm < 1<<12)
421       // icmp X, #-C -> cmn X, #C
422       return 0;
423     if (ST->isThumb() && NegImm < 1<<8)
424       // icmp X, #-C -> adds X, #C
425       return 0;
426   }
427 
428   // xor a, -1 can always be folded to MVN
429   if (Opcode == Instruction::Xor && Imm.isAllOnes())
430     return 0;
431 
432   // Ensures negative constant of min(max()) or max(min()) patterns that
433   // match to SSAT instructions don't get hoisted
434   if (Inst && ((ST->hasV6Ops() && !ST->isThumb()) || ST->isThumb2()) &&
435       Ty->getIntegerBitWidth() <= 32) {
436     if (isSSATMinMaxPattern(Inst, Imm) ||
437         (isa<ICmpInst>(Inst) && Inst->hasOneUse() &&
438          isSSATMinMaxPattern(cast<Instruction>(*Inst->user_begin()), Imm)))
439       return 0;
440   }
441 
442   if (Inst && ST->hasVFP2Base() && isFPSatMinMaxPattern(Inst, Imm))
443     return 0;
444 
445   // We can convert <= -1 to < 0, which is generally quite cheap.
446   if (Inst && Opcode == Instruction::ICmp && Idx == 1 && Imm.isAllOnesValue()) {
447     ICmpInst::Predicate Pred = cast<ICmpInst>(Inst)->getPredicate();
448     if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE)
449       return std::min(getIntImmCost(Imm, Ty, CostKind),
450                       getIntImmCost(Imm + 1, Ty, CostKind));
451   }
452 
453   return getIntImmCost(Imm, Ty, CostKind);
454 }
455 
456 InstructionCost ARMTTIImpl::getCFInstrCost(unsigned Opcode,
457                                            TTI::TargetCostKind CostKind,
458                                            const Instruction *I) {
459   if (CostKind == TTI::TCK_RecipThroughput &&
460       (ST->hasNEON() || ST->hasMVEIntegerOps())) {
461     // FIXME: The vectorizer is highly sensistive to the cost of these
462     // instructions, which suggests that it may be using the costs incorrectly.
463     // But, for now, just make them free to avoid performance regressions for
464     // vector targets.
465     return 0;
466   }
467   return BaseT::getCFInstrCost(Opcode, CostKind, I);
468 }
469 
470 InstructionCost ARMTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
471                                              Type *Src,
472                                              TTI::CastContextHint CCH,
473                                              TTI::TargetCostKind CostKind,
474                                              const Instruction *I) {
475   int ISD = TLI->InstructionOpcodeToISD(Opcode);
476   assert(ISD && "Invalid opcode");
477 
478   // TODO: Allow non-throughput costs that aren't binary.
479   auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost {
480     if (CostKind != TTI::TCK_RecipThroughput)
481       return Cost == 0 ? 0 : 1;
482     return Cost;
483   };
484   auto IsLegalFPType = [this](EVT VT) {
485     EVT EltVT = VT.getScalarType();
486     return (EltVT == MVT::f32 && ST->hasVFP2Base()) ||
487             (EltVT == MVT::f64 && ST->hasFP64()) ||
488             (EltVT == MVT::f16 && ST->hasFullFP16());
489   };
490 
491   EVT SrcTy = TLI->getValueType(DL, Src);
492   EVT DstTy = TLI->getValueType(DL, Dst);
493 
494   if (!SrcTy.isSimple() || !DstTy.isSimple())
495     return AdjustCost(
496         BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
497 
498   // Extending masked load/Truncating masked stores is expensive because we
499   // currently don't split them. This means that we'll likely end up
500   // loading/storing each element individually (hence the high cost).
501   if ((ST->hasMVEIntegerOps() &&
502        (Opcode == Instruction::Trunc || Opcode == Instruction::ZExt ||
503         Opcode == Instruction::SExt)) ||
504       (ST->hasMVEFloatOps() &&
505        (Opcode == Instruction::FPExt || Opcode == Instruction::FPTrunc) &&
506        IsLegalFPType(SrcTy) && IsLegalFPType(DstTy)))
507     if (CCH == TTI::CastContextHint::Masked && DstTy.getSizeInBits() > 128)
508       return 2 * DstTy.getVectorNumElements() *
509              ST->getMVEVectorCostFactor(CostKind);
510 
511   // The extend of other kinds of load is free
512   if (CCH == TTI::CastContextHint::Normal ||
513       CCH == TTI::CastContextHint::Masked) {
514     static const TypeConversionCostTblEntry LoadConversionTbl[] = {
515         {ISD::SIGN_EXTEND, MVT::i32, MVT::i16, 0},
516         {ISD::ZERO_EXTEND, MVT::i32, MVT::i16, 0},
517         {ISD::SIGN_EXTEND, MVT::i32, MVT::i8, 0},
518         {ISD::ZERO_EXTEND, MVT::i32, MVT::i8, 0},
519         {ISD::SIGN_EXTEND, MVT::i16, MVT::i8, 0},
520         {ISD::ZERO_EXTEND, MVT::i16, MVT::i8, 0},
521         {ISD::SIGN_EXTEND, MVT::i64, MVT::i32, 1},
522         {ISD::ZERO_EXTEND, MVT::i64, MVT::i32, 1},
523         {ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 1},
524         {ISD::ZERO_EXTEND, MVT::i64, MVT::i16, 1},
525         {ISD::SIGN_EXTEND, MVT::i64, MVT::i8, 1},
526         {ISD::ZERO_EXTEND, MVT::i64, MVT::i8, 1},
527     };
528     if (const auto *Entry = ConvertCostTableLookup(
529             LoadConversionTbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
530       return AdjustCost(Entry->Cost);
531 
532     static const TypeConversionCostTblEntry MVELoadConversionTbl[] = {
533         {ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 0},
534         {ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 0},
535         {ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 0},
536         {ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 0},
537         {ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 0},
538         {ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 0},
539         // The following extend from a legal type to an illegal type, so need to
540         // split the load. This introduced an extra load operation, but the
541         // extend is still "free".
542         {ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1},
543         {ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1},
544         {ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 3},
545         {ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 3},
546         {ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1},
547         {ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1},
548     };
549     if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
550       if (const auto *Entry =
551               ConvertCostTableLookup(MVELoadConversionTbl, ISD,
552                                      DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
553         return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
554     }
555 
556     static const TypeConversionCostTblEntry MVEFLoadConversionTbl[] = {
557         // FPExtends are similar but also require the VCVT instructions.
558         {ISD::FP_EXTEND, MVT::v4f32, MVT::v4f16, 1},
559         {ISD::FP_EXTEND, MVT::v8f32, MVT::v8f16, 3},
560     };
561     if (SrcTy.isVector() && ST->hasMVEFloatOps()) {
562       if (const auto *Entry =
563               ConvertCostTableLookup(MVEFLoadConversionTbl, ISD,
564                                      DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
565         return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
566     }
567 
568     // The truncate of a store is free. This is the mirror of extends above.
569     static const TypeConversionCostTblEntry MVEStoreConversionTbl[] = {
570         {ISD::TRUNCATE, MVT::v4i32, MVT::v4i16, 0},
571         {ISD::TRUNCATE, MVT::v4i32, MVT::v4i8, 0},
572         {ISD::TRUNCATE, MVT::v8i16, MVT::v8i8, 0},
573         {ISD::TRUNCATE, MVT::v8i32, MVT::v8i16, 1},
574         {ISD::TRUNCATE, MVT::v8i32, MVT::v8i8, 1},
575         {ISD::TRUNCATE, MVT::v16i32, MVT::v16i8, 3},
576         {ISD::TRUNCATE, MVT::v16i16, MVT::v16i8, 1},
577     };
578     if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
579       if (const auto *Entry =
580               ConvertCostTableLookup(MVEStoreConversionTbl, ISD,
581                                      SrcTy.getSimpleVT(), DstTy.getSimpleVT()))
582         return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
583     }
584 
585     static const TypeConversionCostTblEntry MVEFStoreConversionTbl[] = {
586         {ISD::FP_ROUND, MVT::v4f32, MVT::v4f16, 1},
587         {ISD::FP_ROUND, MVT::v8f32, MVT::v8f16, 3},
588     };
589     if (SrcTy.isVector() && ST->hasMVEFloatOps()) {
590       if (const auto *Entry =
591               ConvertCostTableLookup(MVEFStoreConversionTbl, ISD,
592                                      SrcTy.getSimpleVT(), DstTy.getSimpleVT()))
593         return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
594     }
595   }
596 
597   // NEON vector operations that can extend their inputs.
598   if ((ISD == ISD::SIGN_EXTEND || ISD == ISD::ZERO_EXTEND) &&
599       I && I->hasOneUse() && ST->hasNEON() && SrcTy.isVector()) {
600     static const TypeConversionCostTblEntry NEONDoubleWidthTbl[] = {
601       // vaddl
602       { ISD::ADD, MVT::v4i32, MVT::v4i16, 0 },
603       { ISD::ADD, MVT::v8i16, MVT::v8i8,  0 },
604       // vsubl
605       { ISD::SUB, MVT::v4i32, MVT::v4i16, 0 },
606       { ISD::SUB, MVT::v8i16, MVT::v8i8,  0 },
607       // vmull
608       { ISD::MUL, MVT::v4i32, MVT::v4i16, 0 },
609       { ISD::MUL, MVT::v8i16, MVT::v8i8,  0 },
610       // vshll
611       { ISD::SHL, MVT::v4i32, MVT::v4i16, 0 },
612       { ISD::SHL, MVT::v8i16, MVT::v8i8,  0 },
613     };
614 
615     auto *User = cast<Instruction>(*I->user_begin());
616     int UserISD = TLI->InstructionOpcodeToISD(User->getOpcode());
617     if (auto *Entry = ConvertCostTableLookup(NEONDoubleWidthTbl, UserISD,
618                                              DstTy.getSimpleVT(),
619                                              SrcTy.getSimpleVT())) {
620       return AdjustCost(Entry->Cost);
621     }
622   }
623 
624   // Single to/from double precision conversions.
625   if (Src->isVectorTy() && ST->hasNEON() &&
626       ((ISD == ISD::FP_ROUND && SrcTy.getScalarType() == MVT::f64 &&
627         DstTy.getScalarType() == MVT::f32) ||
628        (ISD == ISD::FP_EXTEND && SrcTy.getScalarType() == MVT::f32 &&
629         DstTy.getScalarType() == MVT::f64))) {
630     static const CostTblEntry NEONFltDblTbl[] = {
631         // Vector fptrunc/fpext conversions.
632         {ISD::FP_ROUND, MVT::v2f64, 2},
633         {ISD::FP_EXTEND, MVT::v2f32, 2},
634         {ISD::FP_EXTEND, MVT::v4f32, 4}};
635 
636     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
637     if (const auto *Entry = CostTableLookup(NEONFltDblTbl, ISD, LT.second))
638       return AdjustCost(LT.first * Entry->Cost);
639   }
640 
641   // Some arithmetic, load and store operations have specific instructions
642   // to cast up/down their types automatically at no extra cost.
643   // TODO: Get these tables to know at least what the related operations are.
644   static const TypeConversionCostTblEntry NEONVectorConversionTbl[] = {
645     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
646     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
647     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
648     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 1 },
649     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64, 0 },
650     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i32, 1 },
651 
652     // The number of vmovl instructions for the extension.
653     { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8,  1 },
654     { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8,  1 },
655     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8,  2 },
656     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8,  2 },
657     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i8,  3 },
658     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i8,  3 },
659     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
660     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
661     { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
662     { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
663     { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
664     { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
665     { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
666     { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 7 },
667     { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
668     { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 6 },
669     { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
670     { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6 },
671 
672     // Operations that we legalize using splitting.
673     { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i32, 6 },
674     { ISD::TRUNCATE,    MVT::v8i8, MVT::v8i32, 3 },
675 
676     // Vector float <-> i32 conversions.
677     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
678     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
679 
680     { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i8, 3 },
681     { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i8, 3 },
682     { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i16, 2 },
683     { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i16, 2 },
684     { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i32, 1 },
685     { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i32, 1 },
686     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1, 3 },
687     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1, 3 },
688     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8, 3 },
689     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8, 3 },
690     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
691     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
692     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 4 },
693     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 4 },
694     { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 2 },
695     { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 2 },
696     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 8 },
697     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 8 },
698     { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 4 },
699     { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 4 },
700 
701     { ISD::FP_TO_SINT,  MVT::v4i32, MVT::v4f32, 1 },
702     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f32, 1 },
703     { ISD::FP_TO_SINT,  MVT::v4i8, MVT::v4f32, 3 },
704     { ISD::FP_TO_UINT,  MVT::v4i8, MVT::v4f32, 3 },
705     { ISD::FP_TO_SINT,  MVT::v4i16, MVT::v4f32, 2 },
706     { ISD::FP_TO_UINT,  MVT::v4i16, MVT::v4f32, 2 },
707 
708     // Vector double <-> i32 conversions.
709     { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
710     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
711 
712     { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i8, 4 },
713     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i8, 4 },
714     { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i16, 3 },
715     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i16, 3 },
716     { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
717     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
718 
719     { ISD::FP_TO_SINT,  MVT::v2i32, MVT::v2f64, 2 },
720     { ISD::FP_TO_UINT,  MVT::v2i32, MVT::v2f64, 2 },
721     { ISD::FP_TO_SINT,  MVT::v8i16, MVT::v8f32, 4 },
722     { ISD::FP_TO_UINT,  MVT::v8i16, MVT::v8f32, 4 },
723     { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v16f32, 8 },
724     { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 8 }
725   };
726 
727   if (SrcTy.isVector() && ST->hasNEON()) {
728     if (const auto *Entry = ConvertCostTableLookup(NEONVectorConversionTbl, ISD,
729                                                    DstTy.getSimpleVT(),
730                                                    SrcTy.getSimpleVT()))
731       return AdjustCost(Entry->Cost);
732   }
733 
734   // Scalar float to integer conversions.
735   static const TypeConversionCostTblEntry NEONFloatConversionTbl[] = {
736     { ISD::FP_TO_SINT,  MVT::i1, MVT::f32, 2 },
737     { ISD::FP_TO_UINT,  MVT::i1, MVT::f32, 2 },
738     { ISD::FP_TO_SINT,  MVT::i1, MVT::f64, 2 },
739     { ISD::FP_TO_UINT,  MVT::i1, MVT::f64, 2 },
740     { ISD::FP_TO_SINT,  MVT::i8, MVT::f32, 2 },
741     { ISD::FP_TO_UINT,  MVT::i8, MVT::f32, 2 },
742     { ISD::FP_TO_SINT,  MVT::i8, MVT::f64, 2 },
743     { ISD::FP_TO_UINT,  MVT::i8, MVT::f64, 2 },
744     { ISD::FP_TO_SINT,  MVT::i16, MVT::f32, 2 },
745     { ISD::FP_TO_UINT,  MVT::i16, MVT::f32, 2 },
746     { ISD::FP_TO_SINT,  MVT::i16, MVT::f64, 2 },
747     { ISD::FP_TO_UINT,  MVT::i16, MVT::f64, 2 },
748     { ISD::FP_TO_SINT,  MVT::i32, MVT::f32, 2 },
749     { ISD::FP_TO_UINT,  MVT::i32, MVT::f32, 2 },
750     { ISD::FP_TO_SINT,  MVT::i32, MVT::f64, 2 },
751     { ISD::FP_TO_UINT,  MVT::i32, MVT::f64, 2 },
752     { ISD::FP_TO_SINT,  MVT::i64, MVT::f32, 10 },
753     { ISD::FP_TO_UINT,  MVT::i64, MVT::f32, 10 },
754     { ISD::FP_TO_SINT,  MVT::i64, MVT::f64, 10 },
755     { ISD::FP_TO_UINT,  MVT::i64, MVT::f64, 10 }
756   };
757   if (SrcTy.isFloatingPoint() && ST->hasNEON()) {
758     if (const auto *Entry = ConvertCostTableLookup(NEONFloatConversionTbl, ISD,
759                                                    DstTy.getSimpleVT(),
760                                                    SrcTy.getSimpleVT()))
761       return AdjustCost(Entry->Cost);
762   }
763 
764   // Scalar integer to float conversions.
765   static const TypeConversionCostTblEntry NEONIntegerConversionTbl[] = {
766     { ISD::SINT_TO_FP,  MVT::f32, MVT::i1, 2 },
767     { ISD::UINT_TO_FP,  MVT::f32, MVT::i1, 2 },
768     { ISD::SINT_TO_FP,  MVT::f64, MVT::i1, 2 },
769     { ISD::UINT_TO_FP,  MVT::f64, MVT::i1, 2 },
770     { ISD::SINT_TO_FP,  MVT::f32, MVT::i8, 2 },
771     { ISD::UINT_TO_FP,  MVT::f32, MVT::i8, 2 },
772     { ISD::SINT_TO_FP,  MVT::f64, MVT::i8, 2 },
773     { ISD::UINT_TO_FP,  MVT::f64, MVT::i8, 2 },
774     { ISD::SINT_TO_FP,  MVT::f32, MVT::i16, 2 },
775     { ISD::UINT_TO_FP,  MVT::f32, MVT::i16, 2 },
776     { ISD::SINT_TO_FP,  MVT::f64, MVT::i16, 2 },
777     { ISD::UINT_TO_FP,  MVT::f64, MVT::i16, 2 },
778     { ISD::SINT_TO_FP,  MVT::f32, MVT::i32, 2 },
779     { ISD::UINT_TO_FP,  MVT::f32, MVT::i32, 2 },
780     { ISD::SINT_TO_FP,  MVT::f64, MVT::i32, 2 },
781     { ISD::UINT_TO_FP,  MVT::f64, MVT::i32, 2 },
782     { ISD::SINT_TO_FP,  MVT::f32, MVT::i64, 10 },
783     { ISD::UINT_TO_FP,  MVT::f32, MVT::i64, 10 },
784     { ISD::SINT_TO_FP,  MVT::f64, MVT::i64, 10 },
785     { ISD::UINT_TO_FP,  MVT::f64, MVT::i64, 10 }
786   };
787 
788   if (SrcTy.isInteger() && ST->hasNEON()) {
789     if (const auto *Entry = ConvertCostTableLookup(NEONIntegerConversionTbl,
790                                                    ISD, DstTy.getSimpleVT(),
791                                                    SrcTy.getSimpleVT()))
792       return AdjustCost(Entry->Cost);
793   }
794 
795   // MVE extend costs, taken from codegen tests. i8->i16 or i16->i32 is one
796   // instruction, i8->i32 is two. i64 zexts are an VAND with a constant, sext
797   // are linearised so take more.
798   static const TypeConversionCostTblEntry MVEVectorConversionTbl[] = {
799     { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
800     { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i8, 1 },
801     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
802     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i8, 2 },
803     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i8, 10 },
804     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i8, 2 },
805     { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
806     { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i16, 1 },
807     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i16, 10 },
808     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i16, 2 },
809     { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i32, 8 },
810     { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i32, 2 },
811   };
812 
813   if (SrcTy.isVector() && ST->hasMVEIntegerOps()) {
814     if (const auto *Entry = ConvertCostTableLookup(MVEVectorConversionTbl,
815                                                    ISD, DstTy.getSimpleVT(),
816                                                    SrcTy.getSimpleVT()))
817       return Entry->Cost * ST->getMVEVectorCostFactor(CostKind);
818   }
819 
820   if (ISD == ISD::FP_ROUND || ISD == ISD::FP_EXTEND) {
821     // As general rule, fp converts that were not matched above are scalarized
822     // and cost 1 vcvt for each lane, so long as the instruction is available.
823     // If not it will become a series of function calls.
824     const InstructionCost CallCost =
825         getCallInstrCost(nullptr, Dst, {Src}, CostKind);
826     int Lanes = 1;
827     if (SrcTy.isFixedLengthVector())
828       Lanes = SrcTy.getVectorNumElements();
829 
830     if (IsLegalFPType(SrcTy) && IsLegalFPType(DstTy))
831       return Lanes;
832     else
833       return Lanes * CallCost;
834   }
835 
836   if (ISD == ISD::TRUNCATE && ST->hasMVEIntegerOps() &&
837       SrcTy.isFixedLengthVector()) {
838     // Treat a truncate with larger than legal source (128bits for MVE) as
839     // expensive, 2 instructions per lane.
840     if ((SrcTy.getScalarType() == MVT::i8 ||
841          SrcTy.getScalarType() == MVT::i16 ||
842          SrcTy.getScalarType() == MVT::i32) &&
843         SrcTy.getSizeInBits() > 128 &&
844         SrcTy.getSizeInBits() > DstTy.getSizeInBits())
845       return SrcTy.getVectorNumElements() * 2;
846   }
847 
848   // Scalar integer conversion costs.
849   static const TypeConversionCostTblEntry ARMIntegerConversionTbl[] = {
850     // i16 -> i64 requires two dependent operations.
851     { ISD::SIGN_EXTEND, MVT::i64, MVT::i16, 2 },
852 
853     // Truncates on i64 are assumed to be free.
854     { ISD::TRUNCATE,    MVT::i32, MVT::i64, 0 },
855     { ISD::TRUNCATE,    MVT::i16, MVT::i64, 0 },
856     { ISD::TRUNCATE,    MVT::i8,  MVT::i64, 0 },
857     { ISD::TRUNCATE,    MVT::i1,  MVT::i64, 0 }
858   };
859 
860   if (SrcTy.isInteger()) {
861     if (const auto *Entry = ConvertCostTableLookup(ARMIntegerConversionTbl, ISD,
862                                                    DstTy.getSimpleVT(),
863                                                    SrcTy.getSimpleVT()))
864       return AdjustCost(Entry->Cost);
865   }
866 
867   int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
868                      ? ST->getMVEVectorCostFactor(CostKind)
869                      : 1;
870   return AdjustCost(
871       BaseCost * BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
872 }
873 
874 InstructionCost ARMTTIImpl::getVectorInstrCost(unsigned Opcode, Type *ValTy,
875                                                unsigned Index) {
876   // Penalize inserting into an D-subregister. We end up with a three times
877   // lower estimated throughput on swift.
878   if (ST->hasSlowLoadDSubregister() && Opcode == Instruction::InsertElement &&
879       ValTy->isVectorTy() && ValTy->getScalarSizeInBits() <= 32)
880     return 3;
881 
882   if (ST->hasNEON() && (Opcode == Instruction::InsertElement ||
883                         Opcode == Instruction::ExtractElement)) {
884     // Cross-class copies are expensive on many microarchitectures,
885     // so assume they are expensive by default.
886     if (cast<VectorType>(ValTy)->getElementType()->isIntegerTy())
887       return 3;
888 
889     // Even if it's not a cross class copy, this likely leads to mixing
890     // of NEON and VFP code and should be therefore penalized.
891     if (ValTy->isVectorTy() &&
892         ValTy->getScalarSizeInBits() <= 32)
893       return std::max<InstructionCost>(
894           BaseT::getVectorInstrCost(Opcode, ValTy, Index), 2U);
895   }
896 
897   if (ST->hasMVEIntegerOps() && (Opcode == Instruction::InsertElement ||
898                                  Opcode == Instruction::ExtractElement)) {
899     // Integer cross-lane moves are more expensive than float, which can
900     // sometimes just be vmovs. Integer involve being passes to GPR registers,
901     // causing more of a delay.
902     std::pair<InstructionCost, MVT> LT =
903         getTLI()->getTypeLegalizationCost(DL, ValTy->getScalarType());
904     return LT.first * (ValTy->getScalarType()->isIntegerTy() ? 4 : 1);
905   }
906 
907   return BaseT::getVectorInstrCost(Opcode, ValTy, Index);
908 }
909 
910 InstructionCost ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
911                                                Type *CondTy,
912                                                CmpInst::Predicate VecPred,
913                                                TTI::TargetCostKind CostKind,
914                                                const Instruction *I) {
915   int ISD = TLI->InstructionOpcodeToISD(Opcode);
916 
917   // Thumb scalar code size cost for select.
918   if (CostKind == TTI::TCK_CodeSize && ISD == ISD::SELECT &&
919       ST->isThumb() && !ValTy->isVectorTy()) {
920     // Assume expensive structs.
921     if (TLI->getValueType(DL, ValTy, true) == MVT::Other)
922       return TTI::TCC_Expensive;
923 
924     // Select costs can vary because they:
925     // - may require one or more conditional mov (including an IT),
926     // - can't operate directly on immediates,
927     // - require live flags, which we can't copy around easily.
928     InstructionCost Cost = TLI->getTypeLegalizationCost(DL, ValTy).first;
929 
930     // Possible IT instruction for Thumb2, or more for Thumb1.
931     ++Cost;
932 
933     // i1 values may need rematerialising by using mov immediates and/or
934     // flag setting instructions.
935     if (ValTy->isIntegerTy(1))
936       ++Cost;
937 
938     return Cost;
939   }
940 
941   // If this is a vector min/max/abs, use the cost of that intrinsic directly
942   // instead. Hopefully when min/max intrinsics are more prevalent this code
943   // will not be needed.
944   const Instruction *Sel = I;
945   if ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && Sel &&
946       Sel->hasOneUse())
947     Sel = cast<Instruction>(Sel->user_back());
948   if (Sel && ValTy->isVectorTy() &&
949       (ValTy->isIntOrIntVectorTy() || ValTy->isFPOrFPVectorTy())) {
950     const Value *LHS, *RHS;
951     SelectPatternFlavor SPF = matchSelectPattern(Sel, LHS, RHS).Flavor;
952     unsigned IID = 0;
953     switch (SPF) {
954     case SPF_ABS:
955       IID = Intrinsic::abs;
956       break;
957     case SPF_SMIN:
958       IID = Intrinsic::smin;
959       break;
960     case SPF_SMAX:
961       IID = Intrinsic::smax;
962       break;
963     case SPF_UMIN:
964       IID = Intrinsic::umin;
965       break;
966     case SPF_UMAX:
967       IID = Intrinsic::umax;
968       break;
969     case SPF_FMINNUM:
970       IID = Intrinsic::minnum;
971       break;
972     case SPF_FMAXNUM:
973       IID = Intrinsic::maxnum;
974       break;
975     default:
976       break;
977     }
978     if (IID) {
979       // The ICmp is free, the select gets the cost of the min/max/etc
980       if (Sel != I)
981         return 0;
982       IntrinsicCostAttributes CostAttrs(IID, ValTy, {ValTy, ValTy});
983       return getIntrinsicInstrCost(CostAttrs, CostKind);
984     }
985   }
986 
987   // On NEON a vector select gets lowered to vbsl.
988   if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT && CondTy) {
989     // Lowering of some vector selects is currently far from perfect.
990     static const TypeConversionCostTblEntry NEONVectorSelectTbl[] = {
991       { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
992       { ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
993       { ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
994     };
995 
996     EVT SelCondTy = TLI->getValueType(DL, CondTy);
997     EVT SelValTy = TLI->getValueType(DL, ValTy);
998     if (SelCondTy.isSimple() && SelValTy.isSimple()) {
999       if (const auto *Entry = ConvertCostTableLookup(NEONVectorSelectTbl, ISD,
1000                                                      SelCondTy.getSimpleVT(),
1001                                                      SelValTy.getSimpleVT()))
1002         return Entry->Cost;
1003     }
1004 
1005     std::pair<InstructionCost, MVT> LT =
1006         TLI->getTypeLegalizationCost(DL, ValTy);
1007     return LT.first;
1008   }
1009 
1010   if (ST->hasMVEIntegerOps() && ValTy->isVectorTy() &&
1011       (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
1012       cast<FixedVectorType>(ValTy)->getNumElements() > 1) {
1013     FixedVectorType *VecValTy = cast<FixedVectorType>(ValTy);
1014     FixedVectorType *VecCondTy = dyn_cast_or_null<FixedVectorType>(CondTy);
1015     if (!VecCondTy)
1016       VecCondTy = cast<FixedVectorType>(CmpInst::makeCmpResultType(VecValTy));
1017 
1018     // If we don't have mve.fp any fp operations will need to be scalarized.
1019     if (Opcode == Instruction::FCmp && !ST->hasMVEFloatOps()) {
1020       // One scalaization insert, one scalarization extract and the cost of the
1021       // fcmps.
1022       return BaseT::getScalarizationOverhead(VecValTy, false, true) +
1023              BaseT::getScalarizationOverhead(VecCondTy, true, false) +
1024              VecValTy->getNumElements() *
1025                  getCmpSelInstrCost(Opcode, ValTy->getScalarType(),
1026                                     VecCondTy->getScalarType(), VecPred, CostKind,
1027                                     I);
1028     }
1029 
1030     std::pair<InstructionCost, MVT> LT =
1031         TLI->getTypeLegalizationCost(DL, ValTy);
1032     int BaseCost = ST->getMVEVectorCostFactor(CostKind);
1033     // There are two types - the input that specifies the type of the compare
1034     // and the output vXi1 type. Because we don't know how the output will be
1035     // split, we may need an expensive shuffle to get two in sync. This has the
1036     // effect of making larger than legal compares (v8i32 for example)
1037     // expensive.
1038     if (LT.second.getVectorNumElements() > 2) {
1039       if (LT.first > 1)
1040         return LT.first * BaseCost +
1041                BaseT::getScalarizationOverhead(VecCondTy, true, false);
1042       return BaseCost;
1043     }
1044   }
1045 
1046   // Default to cheap (throughput/size of 1 instruction) but adjust throughput
1047   // for "multiple beats" potentially needed by MVE instructions.
1048   int BaseCost = 1;
1049   if (ST->hasMVEIntegerOps() && ValTy->isVectorTy())
1050     BaseCost = ST->getMVEVectorCostFactor(CostKind);
1051 
1052   return BaseCost *
1053          BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
1054 }
1055 
1056 InstructionCost ARMTTIImpl::getAddressComputationCost(Type *Ty,
1057                                                       ScalarEvolution *SE,
1058                                                       const SCEV *Ptr) {
1059   // Address computations in vectorized code with non-consecutive addresses will
1060   // likely result in more instructions compared to scalar code where the
1061   // computation can more often be merged into the index mode. The resulting
1062   // extra micro-ops can significantly decrease throughput.
1063   unsigned NumVectorInstToHideOverhead = 10;
1064   int MaxMergeDistance = 64;
1065 
1066   if (ST->hasNEON()) {
1067     if (Ty->isVectorTy() && SE &&
1068         !BaseT::isConstantStridedAccessLessThan(SE, Ptr, MaxMergeDistance + 1))
1069       return NumVectorInstToHideOverhead;
1070 
1071     // In many cases the address computation is not merged into the instruction
1072     // addressing mode.
1073     return 1;
1074   }
1075   return BaseT::getAddressComputationCost(Ty, SE, Ptr);
1076 }
1077 
1078 bool ARMTTIImpl::isProfitableLSRChainElement(Instruction *I) {
1079   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1080     // If a VCTP is part of a chain, it's already profitable and shouldn't be
1081     // optimized, else LSR may block tail-predication.
1082     switch (II->getIntrinsicID()) {
1083     case Intrinsic::arm_mve_vctp8:
1084     case Intrinsic::arm_mve_vctp16:
1085     case Intrinsic::arm_mve_vctp32:
1086     case Intrinsic::arm_mve_vctp64:
1087       return true;
1088     default:
1089       break;
1090     }
1091   }
1092   return false;
1093 }
1094 
1095 bool ARMTTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
1096   if (!EnableMaskedLoadStores || !ST->hasMVEIntegerOps())
1097     return false;
1098 
1099   if (auto *VecTy = dyn_cast<FixedVectorType>(DataTy)) {
1100     // Don't support v2i1 yet.
1101     if (VecTy->getNumElements() == 2)
1102       return false;
1103 
1104     // We don't support extending fp types.
1105      unsigned VecWidth = DataTy->getPrimitiveSizeInBits();
1106     if (VecWidth != 128 && VecTy->getElementType()->isFloatingPointTy())
1107       return false;
1108   }
1109 
1110   unsigned EltWidth = DataTy->getScalarSizeInBits();
1111   return (EltWidth == 32 && Alignment >= 4) ||
1112          (EltWidth == 16 && Alignment >= 2) || (EltWidth == 8);
1113 }
1114 
1115 bool ARMTTIImpl::isLegalMaskedGather(Type *Ty, Align Alignment) {
1116   if (!EnableMaskedGatherScatters || !ST->hasMVEIntegerOps())
1117     return false;
1118 
1119   unsigned EltWidth = Ty->getScalarSizeInBits();
1120   return ((EltWidth == 32 && Alignment >= 4) ||
1121           (EltWidth == 16 && Alignment >= 2) || EltWidth == 8);
1122 }
1123 
1124 /// Given a memcpy/memset/memmove instruction, return the number of memory
1125 /// operations performed, via querying findOptimalMemOpLowering. Returns -1 if a
1126 /// call is used.
1127 int ARMTTIImpl::getNumMemOps(const IntrinsicInst *I) const {
1128   MemOp MOp;
1129   unsigned DstAddrSpace = ~0u;
1130   unsigned SrcAddrSpace = ~0u;
1131   const Function *F = I->getParent()->getParent();
1132 
1133   if (const auto *MC = dyn_cast<MemTransferInst>(I)) {
1134     ConstantInt *C = dyn_cast<ConstantInt>(MC->getLength());
1135     // If 'size' is not a constant, a library call will be generated.
1136     if (!C)
1137       return -1;
1138 
1139     const unsigned Size = C->getValue().getZExtValue();
1140     const Align DstAlign = *MC->getDestAlign();
1141     const Align SrcAlign = *MC->getSourceAlign();
1142 
1143     MOp = MemOp::Copy(Size, /*DstAlignCanChange*/ false, DstAlign, SrcAlign,
1144                       /*IsVolatile*/ false);
1145     DstAddrSpace = MC->getDestAddressSpace();
1146     SrcAddrSpace = MC->getSourceAddressSpace();
1147   }
1148   else if (const auto *MS = dyn_cast<MemSetInst>(I)) {
1149     ConstantInt *C = dyn_cast<ConstantInt>(MS->getLength());
1150     // If 'size' is not a constant, a library call will be generated.
1151     if (!C)
1152       return -1;
1153 
1154     const unsigned Size = C->getValue().getZExtValue();
1155     const Align DstAlign = *MS->getDestAlign();
1156 
1157     MOp = MemOp::Set(Size, /*DstAlignCanChange*/ false, DstAlign,
1158                      /*IsZeroMemset*/ false, /*IsVolatile*/ false);
1159     DstAddrSpace = MS->getDestAddressSpace();
1160   }
1161   else
1162     llvm_unreachable("Expected a memcpy/move or memset!");
1163 
1164   unsigned Limit, Factor = 2;
1165   switch(I->getIntrinsicID()) {
1166     case Intrinsic::memcpy:
1167       Limit = TLI->getMaxStoresPerMemcpy(F->hasMinSize());
1168       break;
1169     case Intrinsic::memmove:
1170       Limit = TLI->getMaxStoresPerMemmove(F->hasMinSize());
1171       break;
1172     case Intrinsic::memset:
1173       Limit = TLI->getMaxStoresPerMemset(F->hasMinSize());
1174       Factor = 1;
1175       break;
1176     default:
1177       llvm_unreachable("Expected a memcpy/move or memset!");
1178   }
1179 
1180   // MemOps will be poplulated with a list of data types that needs to be
1181   // loaded and stored. That's why we multiply the number of elements by 2 to
1182   // get the cost for this memcpy.
1183   std::vector<EVT> MemOps;
1184   if (getTLI()->findOptimalMemOpLowering(
1185           MemOps, Limit, MOp, DstAddrSpace,
1186           SrcAddrSpace, F->getAttributes()))
1187     return MemOps.size() * Factor;
1188 
1189   // If we can't find an optimal memop lowering, return the default cost
1190   return -1;
1191 }
1192 
1193 InstructionCost ARMTTIImpl::getMemcpyCost(const Instruction *I) {
1194   int NumOps = getNumMemOps(cast<IntrinsicInst>(I));
1195 
1196   // To model the cost of a library call, we assume 1 for the call, and
1197   // 3 for the argument setup.
1198   if (NumOps == -1)
1199     return 4;
1200   return NumOps;
1201 }
1202 
1203 InstructionCost ARMTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
1204                                            VectorType *Tp, ArrayRef<int> Mask,
1205                                            int Index, VectorType *SubTp) {
1206   Kind = improveShuffleKindFromMask(Kind, Mask);
1207   if (ST->hasNEON()) {
1208     if (Kind == TTI::SK_Broadcast) {
1209       static const CostTblEntry NEONDupTbl[] = {
1210           // VDUP handles these cases.
1211           {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
1212           {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
1213           {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
1214           {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
1215           {ISD::VECTOR_SHUFFLE, MVT::v4i16, 1},
1216           {ISD::VECTOR_SHUFFLE, MVT::v8i8, 1},
1217 
1218           {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
1219           {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
1220           {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
1221           {ISD::VECTOR_SHUFFLE, MVT::v16i8, 1}};
1222 
1223       std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
1224       if (const auto *Entry =
1225               CostTableLookup(NEONDupTbl, ISD::VECTOR_SHUFFLE, LT.second))
1226         return LT.first * Entry->Cost;
1227     }
1228     if (Kind == TTI::SK_Reverse) {
1229       static const CostTblEntry NEONShuffleTbl[] = {
1230           // Reverse shuffle cost one instruction if we are shuffling within a
1231           // double word (vrev) or two if we shuffle a quad word (vrev, vext).
1232           {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
1233           {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
1234           {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
1235           {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
1236           {ISD::VECTOR_SHUFFLE, MVT::v4i16, 1},
1237           {ISD::VECTOR_SHUFFLE, MVT::v8i8, 1},
1238 
1239           {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
1240           {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
1241           {ISD::VECTOR_SHUFFLE, MVT::v8i16, 2},
1242           {ISD::VECTOR_SHUFFLE, MVT::v16i8, 2}};
1243 
1244       std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
1245       if (const auto *Entry =
1246               CostTableLookup(NEONShuffleTbl, ISD::VECTOR_SHUFFLE, LT.second))
1247         return LT.first * Entry->Cost;
1248     }
1249     if (Kind == TTI::SK_Select) {
1250       static const CostTblEntry NEONSelShuffleTbl[] = {
1251           // Select shuffle cost table for ARM. Cost is the number of
1252           // instructions
1253           // required to create the shuffled vector.
1254 
1255           {ISD::VECTOR_SHUFFLE, MVT::v2f32, 1},
1256           {ISD::VECTOR_SHUFFLE, MVT::v2i64, 1},
1257           {ISD::VECTOR_SHUFFLE, MVT::v2f64, 1},
1258           {ISD::VECTOR_SHUFFLE, MVT::v2i32, 1},
1259 
1260           {ISD::VECTOR_SHUFFLE, MVT::v4i32, 2},
1261           {ISD::VECTOR_SHUFFLE, MVT::v4f32, 2},
1262           {ISD::VECTOR_SHUFFLE, MVT::v4i16, 2},
1263 
1264           {ISD::VECTOR_SHUFFLE, MVT::v8i16, 16},
1265 
1266           {ISD::VECTOR_SHUFFLE, MVT::v16i8, 32}};
1267 
1268       std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
1269       if (const auto *Entry = CostTableLookup(NEONSelShuffleTbl,
1270                                               ISD::VECTOR_SHUFFLE, LT.second))
1271         return LT.first * Entry->Cost;
1272     }
1273   }
1274   if (ST->hasMVEIntegerOps()) {
1275     if (Kind == TTI::SK_Broadcast) {
1276       static const CostTblEntry MVEDupTbl[] = {
1277           // VDUP handles these cases.
1278           {ISD::VECTOR_SHUFFLE, MVT::v4i32, 1},
1279           {ISD::VECTOR_SHUFFLE, MVT::v8i16, 1},
1280           {ISD::VECTOR_SHUFFLE, MVT::v16i8, 1},
1281           {ISD::VECTOR_SHUFFLE, MVT::v4f32, 1},
1282           {ISD::VECTOR_SHUFFLE, MVT::v8f16, 1}};
1283 
1284       std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
1285       if (const auto *Entry = CostTableLookup(MVEDupTbl, ISD::VECTOR_SHUFFLE,
1286                                               LT.second))
1287         return LT.first * Entry->Cost *
1288                ST->getMVEVectorCostFactor(TTI::TCK_RecipThroughput);
1289     }
1290 
1291     if (!Mask.empty()) {
1292       std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Tp);
1293       if (Mask.size() <= LT.second.getVectorNumElements() &&
1294           (isVREVMask(Mask, LT.second, 16) || isVREVMask(Mask, LT.second, 32) ||
1295            isVREVMask(Mask, LT.second, 64)))
1296         return ST->getMVEVectorCostFactor(TTI::TCK_RecipThroughput) * LT.first;
1297     }
1298   }
1299 
1300   int BaseCost = ST->hasMVEIntegerOps() && Tp->isVectorTy()
1301                      ? ST->getMVEVectorCostFactor(TTI::TCK_RecipThroughput)
1302                      : 1;
1303   return BaseCost * BaseT::getShuffleCost(Kind, Tp, Mask, Index, SubTp);
1304 }
1305 
1306 InstructionCost ARMTTIImpl::getArithmeticInstrCost(
1307     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
1308     TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
1309     TTI::OperandValueProperties Opd1PropInfo,
1310     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
1311     const Instruction *CxtI) {
1312   int ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
1313   if (ST->isThumb() && CostKind == TTI::TCK_CodeSize && Ty->isIntegerTy(1)) {
1314     // Make operations on i1 relatively expensive as this often involves
1315     // combining predicates. AND and XOR should be easier to handle with IT
1316     // blocks.
1317     switch (ISDOpcode) {
1318     default:
1319       break;
1320     case ISD::AND:
1321     case ISD::XOR:
1322       return 2;
1323     case ISD::OR:
1324       return 3;
1325     }
1326   }
1327 
1328   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
1329 
1330   if (ST->hasNEON()) {
1331     const unsigned FunctionCallDivCost = 20;
1332     const unsigned ReciprocalDivCost = 10;
1333     static const CostTblEntry CostTbl[] = {
1334       // Division.
1335       // These costs are somewhat random. Choose a cost of 20 to indicate that
1336       // vectorizing devision (added function call) is going to be very expensive.
1337       // Double registers types.
1338       { ISD::SDIV, MVT::v1i64, 1 * FunctionCallDivCost},
1339       { ISD::UDIV, MVT::v1i64, 1 * FunctionCallDivCost},
1340       { ISD::SREM, MVT::v1i64, 1 * FunctionCallDivCost},
1341       { ISD::UREM, MVT::v1i64, 1 * FunctionCallDivCost},
1342       { ISD::SDIV, MVT::v2i32, 2 * FunctionCallDivCost},
1343       { ISD::UDIV, MVT::v2i32, 2 * FunctionCallDivCost},
1344       { ISD::SREM, MVT::v2i32, 2 * FunctionCallDivCost},
1345       { ISD::UREM, MVT::v2i32, 2 * FunctionCallDivCost},
1346       { ISD::SDIV, MVT::v4i16,     ReciprocalDivCost},
1347       { ISD::UDIV, MVT::v4i16,     ReciprocalDivCost},
1348       { ISD::SREM, MVT::v4i16, 4 * FunctionCallDivCost},
1349       { ISD::UREM, MVT::v4i16, 4 * FunctionCallDivCost},
1350       { ISD::SDIV, MVT::v8i8,      ReciprocalDivCost},
1351       { ISD::UDIV, MVT::v8i8,      ReciprocalDivCost},
1352       { ISD::SREM, MVT::v8i8,  8 * FunctionCallDivCost},
1353       { ISD::UREM, MVT::v8i8,  8 * FunctionCallDivCost},
1354       // Quad register types.
1355       { ISD::SDIV, MVT::v2i64, 2 * FunctionCallDivCost},
1356       { ISD::UDIV, MVT::v2i64, 2 * FunctionCallDivCost},
1357       { ISD::SREM, MVT::v2i64, 2 * FunctionCallDivCost},
1358       { ISD::UREM, MVT::v2i64, 2 * FunctionCallDivCost},
1359       { ISD::SDIV, MVT::v4i32, 4 * FunctionCallDivCost},
1360       { ISD::UDIV, MVT::v4i32, 4 * FunctionCallDivCost},
1361       { ISD::SREM, MVT::v4i32, 4 * FunctionCallDivCost},
1362       { ISD::UREM, MVT::v4i32, 4 * FunctionCallDivCost},
1363       { ISD::SDIV, MVT::v8i16, 8 * FunctionCallDivCost},
1364       { ISD::UDIV, MVT::v8i16, 8 * FunctionCallDivCost},
1365       { ISD::SREM, MVT::v8i16, 8 * FunctionCallDivCost},
1366       { ISD::UREM, MVT::v8i16, 8 * FunctionCallDivCost},
1367       { ISD::SDIV, MVT::v16i8, 16 * FunctionCallDivCost},
1368       { ISD::UDIV, MVT::v16i8, 16 * FunctionCallDivCost},
1369       { ISD::SREM, MVT::v16i8, 16 * FunctionCallDivCost},
1370       { ISD::UREM, MVT::v16i8, 16 * FunctionCallDivCost},
1371       // Multiplication.
1372     };
1373 
1374     if (const auto *Entry = CostTableLookup(CostTbl, ISDOpcode, LT.second))
1375       return LT.first * Entry->Cost;
1376 
1377     InstructionCost Cost = BaseT::getArithmeticInstrCost(
1378         Opcode, Ty, CostKind, Op1Info, Op2Info, Opd1PropInfo, Opd2PropInfo);
1379 
1380     // This is somewhat of a hack. The problem that we are facing is that SROA
1381     // creates a sequence of shift, and, or instructions to construct values.
1382     // These sequences are recognized by the ISel and have zero-cost. Not so for
1383     // the vectorized code. Because we have support for v2i64 but not i64 those
1384     // sequences look particularly beneficial to vectorize.
1385     // To work around this we increase the cost of v2i64 operations to make them
1386     // seem less beneficial.
1387     if (LT.second == MVT::v2i64 &&
1388         Op2Info == TargetTransformInfo::OK_UniformConstantValue)
1389       Cost += 4;
1390 
1391     return Cost;
1392   }
1393 
1394   // If this operation is a shift on arm/thumb2, it might well be folded into
1395   // the following instruction, hence having a cost of 0.
1396   auto LooksLikeAFreeShift = [&]() {
1397     if (ST->isThumb1Only() || Ty->isVectorTy())
1398       return false;
1399 
1400     if (!CxtI || !CxtI->hasOneUse() || !CxtI->isShift())
1401       return false;
1402     if (Op2Info != TargetTransformInfo::OK_UniformConstantValue)
1403       return false;
1404 
1405     // Folded into a ADC/ADD/AND/BIC/CMP/EOR/MVN/ORR/ORN/RSB/SBC/SUB
1406     switch (cast<Instruction>(CxtI->user_back())->getOpcode()) {
1407     case Instruction::Add:
1408     case Instruction::Sub:
1409     case Instruction::And:
1410     case Instruction::Xor:
1411     case Instruction::Or:
1412     case Instruction::ICmp:
1413       return true;
1414     default:
1415       return false;
1416     }
1417   };
1418   if (LooksLikeAFreeShift())
1419     return 0;
1420 
1421   // Default to cheap (throughput/size of 1 instruction) but adjust throughput
1422   // for "multiple beats" potentially needed by MVE instructions.
1423   int BaseCost = 1;
1424   if (ST->hasMVEIntegerOps() && Ty->isVectorTy())
1425     BaseCost = ST->getMVEVectorCostFactor(CostKind);
1426 
1427   // The rest of this mostly follows what is done in BaseT::getArithmeticInstrCost,
1428   // without treating floats as more expensive that scalars or increasing the
1429   // costs for custom operations. The results is also multiplied by the
1430   // MVEVectorCostFactor where appropriate.
1431   if (TLI->isOperationLegalOrCustomOrPromote(ISDOpcode, LT.second))
1432     return LT.first * BaseCost;
1433 
1434   // Else this is expand, assume that we need to scalarize this op.
1435   if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
1436     unsigned Num = VTy->getNumElements();
1437     InstructionCost Cost =
1438         getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind);
1439     // Return the cost of multiple scalar invocation plus the cost of
1440     // inserting and extracting the values.
1441     SmallVector<Type *> Tys(Args.size(), Ty);
1442     return BaseT::getScalarizationOverhead(VTy, Args, Tys) + Num * Cost;
1443   }
1444 
1445   return BaseCost;
1446 }
1447 
1448 InstructionCost ARMTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
1449                                             MaybeAlign Alignment,
1450                                             unsigned AddressSpace,
1451                                             TTI::TargetCostKind CostKind,
1452                                             const Instruction *I) {
1453   // TODO: Handle other cost kinds.
1454   if (CostKind != TTI::TCK_RecipThroughput)
1455     return 1;
1456 
1457   // Type legalization can't handle structs
1458   if (TLI->getValueType(DL, Src, true) == MVT::Other)
1459     return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
1460                                   CostKind);
1461 
1462   if (ST->hasNEON() && Src->isVectorTy() &&
1463       (Alignment && *Alignment != Align(16)) &&
1464       cast<VectorType>(Src)->getElementType()->isDoubleTy()) {
1465     // Unaligned loads/stores are extremely inefficient.
1466     // We need 4 uops for vst.1/vld.1 vs 1uop for vldr/vstr.
1467     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
1468     return LT.first * 4;
1469   }
1470 
1471   // MVE can optimize a fpext(load(4xhalf)) using an extending integer load.
1472   // Same for stores.
1473   if (ST->hasMVEFloatOps() && isa<FixedVectorType>(Src) && I &&
1474       ((Opcode == Instruction::Load && I->hasOneUse() &&
1475         isa<FPExtInst>(*I->user_begin())) ||
1476        (Opcode == Instruction::Store && isa<FPTruncInst>(I->getOperand(0))))) {
1477     FixedVectorType *SrcVTy = cast<FixedVectorType>(Src);
1478     Type *DstTy =
1479         Opcode == Instruction::Load
1480             ? (*I->user_begin())->getType()
1481             : cast<Instruction>(I->getOperand(0))->getOperand(0)->getType();
1482     if (SrcVTy->getNumElements() == 4 && SrcVTy->getScalarType()->isHalfTy() &&
1483         DstTy->getScalarType()->isFloatTy())
1484       return ST->getMVEVectorCostFactor(CostKind);
1485   }
1486 
1487   int BaseCost = ST->hasMVEIntegerOps() && Src->isVectorTy()
1488                      ? ST->getMVEVectorCostFactor(CostKind)
1489                      : 1;
1490   return BaseCost * BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
1491                                            CostKind, I);
1492 }
1493 
1494 InstructionCost
1495 ARMTTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
1496                                   unsigned AddressSpace,
1497                                   TTI::TargetCostKind CostKind) {
1498   if (ST->hasMVEIntegerOps()) {
1499     if (Opcode == Instruction::Load && isLegalMaskedLoad(Src, Alignment))
1500       return ST->getMVEVectorCostFactor(CostKind);
1501     if (Opcode == Instruction::Store && isLegalMaskedStore(Src, Alignment))
1502       return ST->getMVEVectorCostFactor(CostKind);
1503   }
1504   if (!isa<FixedVectorType>(Src))
1505     return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
1506                                         CostKind);
1507   // Scalar cost, which is currently very high due to the efficiency of the
1508   // generated code.
1509   return cast<FixedVectorType>(Src)->getNumElements() * 8;
1510 }
1511 
1512 InstructionCost ARMTTIImpl::getInterleavedMemoryOpCost(
1513     unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
1514     Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
1515     bool UseMaskForCond, bool UseMaskForGaps) {
1516   assert(Factor >= 2 && "Invalid interleave factor");
1517   assert(isa<VectorType>(VecTy) && "Expect a vector type");
1518 
1519   // vldN/vstN doesn't support vector types of i64/f64 element.
1520   bool EltIs64Bits = DL.getTypeSizeInBits(VecTy->getScalarType()) == 64;
1521 
1522   if (Factor <= TLI->getMaxSupportedInterleaveFactor() && !EltIs64Bits &&
1523       !UseMaskForCond && !UseMaskForGaps) {
1524     unsigned NumElts = cast<FixedVectorType>(VecTy)->getNumElements();
1525     auto *SubVecTy =
1526         FixedVectorType::get(VecTy->getScalarType(), NumElts / Factor);
1527 
1528     // vldN/vstN only support legal vector types of size 64 or 128 in bits.
1529     // Accesses having vector types that are a multiple of 128 bits can be
1530     // matched to more than one vldN/vstN instruction.
1531     int BaseCost =
1532         ST->hasMVEIntegerOps() ? ST->getMVEVectorCostFactor(CostKind) : 1;
1533     if (NumElts % Factor == 0 &&
1534         TLI->isLegalInterleavedAccessType(Factor, SubVecTy, Alignment, DL))
1535       return Factor * BaseCost * TLI->getNumInterleavedAccesses(SubVecTy, DL);
1536 
1537     // Some smaller than legal interleaved patterns are cheap as we can make
1538     // use of the vmovn or vrev patterns to interleave a standard load. This is
1539     // true for v4i8, v8i8 and v4i16 at least (but not for v4f16 as it is
1540     // promoted differently). The cost of 2 here is then a load and vrev or
1541     // vmovn.
1542     if (ST->hasMVEIntegerOps() && Factor == 2 && NumElts / Factor > 2 &&
1543         VecTy->isIntOrIntVectorTy() &&
1544         DL.getTypeSizeInBits(SubVecTy).getFixedSize() <= 64)
1545       return 2 * BaseCost;
1546   }
1547 
1548   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
1549                                            Alignment, AddressSpace, CostKind,
1550                                            UseMaskForCond, UseMaskForGaps);
1551 }
1552 
1553 InstructionCost ARMTTIImpl::getGatherScatterOpCost(
1554     unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
1555     Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) {
1556   using namespace PatternMatch;
1557   if (!ST->hasMVEIntegerOps() || !EnableMaskedGatherScatters)
1558     return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
1559                                          Alignment, CostKind, I);
1560 
1561   assert(DataTy->isVectorTy() && "Can't do gather/scatters on scalar!");
1562   auto *VTy = cast<FixedVectorType>(DataTy);
1563 
1564   // TODO: Splitting, once we do that.
1565 
1566   unsigned NumElems = VTy->getNumElements();
1567   unsigned EltSize = VTy->getScalarSizeInBits();
1568   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, DataTy);
1569 
1570   // For now, it is assumed that for the MVE gather instructions the loads are
1571   // all effectively serialised. This means the cost is the scalar cost
1572   // multiplied by the number of elements being loaded. This is possibly very
1573   // conservative, but even so we still end up vectorising loops because the
1574   // cost per iteration for many loops is lower than for scalar loops.
1575   InstructionCost VectorCost =
1576       NumElems * LT.first * ST->getMVEVectorCostFactor(CostKind);
1577   // The scalarization cost should be a lot higher. We use the number of vector
1578   // elements plus the scalarization overhead.
1579   InstructionCost ScalarCost =
1580       NumElems * LT.first + BaseT::getScalarizationOverhead(VTy, true, false) +
1581       BaseT::getScalarizationOverhead(VTy, false, true);
1582 
1583   if (EltSize < 8 || Alignment < EltSize / 8)
1584     return ScalarCost;
1585 
1586   unsigned ExtSize = EltSize;
1587   // Check whether there's a single user that asks for an extended type
1588   if (I != nullptr) {
1589     // Dependent of the caller of this function, a gather instruction will
1590     // either have opcode Instruction::Load or be a call to the masked_gather
1591     // intrinsic
1592     if ((I->getOpcode() == Instruction::Load ||
1593          match(I, m_Intrinsic<Intrinsic::masked_gather>())) &&
1594         I->hasOneUse()) {
1595       const User *Us = *I->users().begin();
1596       if (isa<ZExtInst>(Us) || isa<SExtInst>(Us)) {
1597         // only allow valid type combinations
1598         unsigned TypeSize =
1599             cast<Instruction>(Us)->getType()->getScalarSizeInBits();
1600         if (((TypeSize == 32 && (EltSize == 8 || EltSize == 16)) ||
1601              (TypeSize == 16 && EltSize == 8)) &&
1602             TypeSize * NumElems == 128) {
1603           ExtSize = TypeSize;
1604         }
1605       }
1606     }
1607     // Check whether the input data needs to be truncated
1608     TruncInst *T;
1609     if ((I->getOpcode() == Instruction::Store ||
1610          match(I, m_Intrinsic<Intrinsic::masked_scatter>())) &&
1611         (T = dyn_cast<TruncInst>(I->getOperand(0)))) {
1612       // Only allow valid type combinations
1613       unsigned TypeSize = T->getOperand(0)->getType()->getScalarSizeInBits();
1614       if (((EltSize == 16 && TypeSize == 32) ||
1615            (EltSize == 8 && (TypeSize == 32 || TypeSize == 16))) &&
1616           TypeSize * NumElems == 128)
1617         ExtSize = TypeSize;
1618     }
1619   }
1620 
1621   if (ExtSize * NumElems != 128 || NumElems < 4)
1622     return ScalarCost;
1623 
1624   // Any (aligned) i32 gather will not need to be scalarised.
1625   if (ExtSize == 32)
1626     return VectorCost;
1627   // For smaller types, we need to ensure that the gep's inputs are correctly
1628   // extended from a small enough value. Other sizes (including i64) are
1629   // scalarized for now.
1630   if (ExtSize != 8 && ExtSize != 16)
1631     return ScalarCost;
1632 
1633   if (const auto *BC = dyn_cast<BitCastInst>(Ptr))
1634     Ptr = BC->getOperand(0);
1635   if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1636     if (GEP->getNumOperands() != 2)
1637       return ScalarCost;
1638     unsigned Scale = DL.getTypeAllocSize(GEP->getResultElementType());
1639     // Scale needs to be correct (which is only relevant for i16s).
1640     if (Scale != 1 && Scale * 8 != ExtSize)
1641       return ScalarCost;
1642     // And we need to zext (not sext) the indexes from a small enough type.
1643     if (const auto *ZExt = dyn_cast<ZExtInst>(GEP->getOperand(1))) {
1644       if (ZExt->getOperand(0)->getType()->getScalarSizeInBits() <= ExtSize)
1645         return VectorCost;
1646     }
1647     return ScalarCost;
1648   }
1649   return ScalarCost;
1650 }
1651 
1652 InstructionCost
1653 ARMTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
1654                                        Optional<FastMathFlags> FMF,
1655                                        TTI::TargetCostKind CostKind) {
1656   if (TTI::requiresOrderedReduction(FMF))
1657     return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
1658 
1659   EVT ValVT = TLI->getValueType(DL, ValTy);
1660   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1661   if (!ST->hasMVEIntegerOps() || !ValVT.isSimple() || ISD != ISD::ADD)
1662     return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
1663 
1664   std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
1665 
1666   static const CostTblEntry CostTblAdd[]{
1667       {ISD::ADD, MVT::v16i8, 1},
1668       {ISD::ADD, MVT::v8i16, 1},
1669       {ISD::ADD, MVT::v4i32, 1},
1670   };
1671   if (const auto *Entry = CostTableLookup(CostTblAdd, ISD, LT.second))
1672     return Entry->Cost * ST->getMVEVectorCostFactor(CostKind) * LT.first;
1673 
1674   return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
1675 }
1676 
1677 InstructionCost
1678 ARMTTIImpl::getExtendedAddReductionCost(bool IsMLA, bool IsUnsigned,
1679                                         Type *ResTy, VectorType *ValTy,
1680                                         TTI::TargetCostKind CostKind) {
1681   EVT ValVT = TLI->getValueType(DL, ValTy);
1682   EVT ResVT = TLI->getValueType(DL, ResTy);
1683 
1684   if (ST->hasMVEIntegerOps() && ValVT.isSimple() && ResVT.isSimple()) {
1685     std::pair<InstructionCost, MVT> LT =
1686         TLI->getTypeLegalizationCost(DL, ValTy);
1687 
1688     // The legal cases are:
1689     //   VADDV u/s 8/16/32
1690     //   VMLAV u/s 8/16/32
1691     //   VADDLV u/s 32
1692     //   VMLALV u/s 16/32
1693     // Codegen currently cannot always handle larger than legal vectors very
1694     // well, especially for predicated reductions where the mask needs to be
1695     // split, so restrict to 128bit or smaller input types.
1696     unsigned RevVTSize = ResVT.getSizeInBits();
1697     if (ValVT.getSizeInBits() <= 128 &&
1698         ((LT.second == MVT::v16i8 && RevVTSize <= 32) ||
1699          (LT.second == MVT::v8i16 && RevVTSize <= (IsMLA ? 64u : 32u)) ||
1700          (LT.second == MVT::v4i32 && RevVTSize <= 64)))
1701       return ST->getMVEVectorCostFactor(CostKind) * LT.first;
1702   }
1703 
1704   return BaseT::getExtendedAddReductionCost(IsMLA, IsUnsigned, ResTy, ValTy,
1705                                             CostKind);
1706 }
1707 
1708 InstructionCost
1709 ARMTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
1710                                   TTI::TargetCostKind CostKind) {
1711   switch (ICA.getID()) {
1712   case Intrinsic::get_active_lane_mask:
1713     // Currently we make a somewhat optimistic assumption that
1714     // active_lane_mask's are always free. In reality it may be freely folded
1715     // into a tail predicated loop, expanded into a VCPT or expanded into a lot
1716     // of add/icmp code. We may need to improve this in the future, but being
1717     // able to detect if it is free or not involves looking at a lot of other
1718     // code. We currently assume that the vectorizer inserted these, and knew
1719     // what it was doing in adding one.
1720     if (ST->hasMVEIntegerOps())
1721       return 0;
1722     break;
1723   case Intrinsic::sadd_sat:
1724   case Intrinsic::ssub_sat:
1725   case Intrinsic::uadd_sat:
1726   case Intrinsic::usub_sat: {
1727     if (!ST->hasMVEIntegerOps())
1728       break;
1729     Type *VT = ICA.getReturnType();
1730 
1731     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, VT);
1732     if (LT.second == MVT::v4i32 || LT.second == MVT::v8i16 ||
1733         LT.second == MVT::v16i8) {
1734       // This is a base cost of 1 for the vqadd, plus 3 extract shifts if we
1735       // need to extend the type, as it uses shr(qadd(shl, shl)).
1736       unsigned Instrs =
1737           LT.second.getScalarSizeInBits() == VT->getScalarSizeInBits() ? 1 : 4;
1738       return LT.first * ST->getMVEVectorCostFactor(CostKind) * Instrs;
1739     }
1740     break;
1741   }
1742   case Intrinsic::abs:
1743   case Intrinsic::smin:
1744   case Intrinsic::smax:
1745   case Intrinsic::umin:
1746   case Intrinsic::umax: {
1747     if (!ST->hasMVEIntegerOps())
1748       break;
1749     Type *VT = ICA.getReturnType();
1750 
1751     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, VT);
1752     if (LT.second == MVT::v4i32 || LT.second == MVT::v8i16 ||
1753         LT.second == MVT::v16i8)
1754       return LT.first * ST->getMVEVectorCostFactor(CostKind);
1755     break;
1756   }
1757   case Intrinsic::minnum:
1758   case Intrinsic::maxnum: {
1759     if (!ST->hasMVEFloatOps())
1760       break;
1761     Type *VT = ICA.getReturnType();
1762     std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, VT);
1763     if (LT.second == MVT::v4f32 || LT.second == MVT::v8f16)
1764       return LT.first * ST->getMVEVectorCostFactor(CostKind);
1765     break;
1766   }
1767   }
1768 
1769   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
1770 }
1771 
1772 bool ARMTTIImpl::isLoweredToCall(const Function *F) {
1773   if (!F->isIntrinsic())
1774     BaseT::isLoweredToCall(F);
1775 
1776   // Assume all Arm-specific intrinsics map to an instruction.
1777   if (F->getName().startswith("llvm.arm"))
1778     return false;
1779 
1780   switch (F->getIntrinsicID()) {
1781   default: break;
1782   case Intrinsic::powi:
1783   case Intrinsic::sin:
1784   case Intrinsic::cos:
1785   case Intrinsic::pow:
1786   case Intrinsic::log:
1787   case Intrinsic::log10:
1788   case Intrinsic::log2:
1789   case Intrinsic::exp:
1790   case Intrinsic::exp2:
1791     return true;
1792   case Intrinsic::sqrt:
1793   case Intrinsic::fabs:
1794   case Intrinsic::copysign:
1795   case Intrinsic::floor:
1796   case Intrinsic::ceil:
1797   case Intrinsic::trunc:
1798   case Intrinsic::rint:
1799   case Intrinsic::nearbyint:
1800   case Intrinsic::round:
1801   case Intrinsic::canonicalize:
1802   case Intrinsic::lround:
1803   case Intrinsic::llround:
1804   case Intrinsic::lrint:
1805   case Intrinsic::llrint:
1806     if (F->getReturnType()->isDoubleTy() && !ST->hasFP64())
1807       return true;
1808     if (F->getReturnType()->isHalfTy() && !ST->hasFullFP16())
1809       return true;
1810     // Some operations can be handled by vector instructions and assume
1811     // unsupported vectors will be expanded into supported scalar ones.
1812     // TODO Handle scalar operations properly.
1813     return !ST->hasFPARMv8Base() && !ST->hasVFP2Base();
1814   case Intrinsic::masked_store:
1815   case Intrinsic::masked_load:
1816   case Intrinsic::masked_gather:
1817   case Intrinsic::masked_scatter:
1818     return !ST->hasMVEIntegerOps();
1819   case Intrinsic::sadd_with_overflow:
1820   case Intrinsic::uadd_with_overflow:
1821   case Intrinsic::ssub_with_overflow:
1822   case Intrinsic::usub_with_overflow:
1823   case Intrinsic::sadd_sat:
1824   case Intrinsic::uadd_sat:
1825   case Intrinsic::ssub_sat:
1826   case Intrinsic::usub_sat:
1827     return false;
1828   }
1829 
1830   return BaseT::isLoweredToCall(F);
1831 }
1832 
1833 bool ARMTTIImpl::maybeLoweredToCall(Instruction &I) {
1834   unsigned ISD = TLI->InstructionOpcodeToISD(I.getOpcode());
1835   EVT VT = TLI->getValueType(DL, I.getType(), true);
1836   if (TLI->getOperationAction(ISD, VT) == TargetLowering::LibCall)
1837     return true;
1838 
1839   // Check if an intrinsic will be lowered to a call and assume that any
1840   // other CallInst will generate a bl.
1841   if (auto *Call = dyn_cast<CallInst>(&I)) {
1842     if (auto *II = dyn_cast<IntrinsicInst>(Call)) {
1843       switch(II->getIntrinsicID()) {
1844         case Intrinsic::memcpy:
1845         case Intrinsic::memset:
1846         case Intrinsic::memmove:
1847           return getNumMemOps(II) == -1;
1848         default:
1849           if (const Function *F = Call->getCalledFunction())
1850             return isLoweredToCall(F);
1851       }
1852     }
1853     return true;
1854   }
1855 
1856   // FPv5 provides conversions between integer, double-precision,
1857   // single-precision, and half-precision formats.
1858   switch (I.getOpcode()) {
1859   default:
1860     break;
1861   case Instruction::FPToSI:
1862   case Instruction::FPToUI:
1863   case Instruction::SIToFP:
1864   case Instruction::UIToFP:
1865   case Instruction::FPTrunc:
1866   case Instruction::FPExt:
1867     return !ST->hasFPARMv8Base();
1868   }
1869 
1870   // FIXME: Unfortunately the approach of checking the Operation Action does
1871   // not catch all cases of Legalization that use library calls. Our
1872   // Legalization step categorizes some transformations into library calls as
1873   // Custom, Expand or even Legal when doing type legalization. So for now
1874   // we have to special case for instance the SDIV of 64bit integers and the
1875   // use of floating point emulation.
1876   if (VT.isInteger() && VT.getSizeInBits() >= 64) {
1877     switch (ISD) {
1878     default:
1879       break;
1880     case ISD::SDIV:
1881     case ISD::UDIV:
1882     case ISD::SREM:
1883     case ISD::UREM:
1884     case ISD::SDIVREM:
1885     case ISD::UDIVREM:
1886       return true;
1887     }
1888   }
1889 
1890   // Assume all other non-float operations are supported.
1891   if (!VT.isFloatingPoint())
1892     return false;
1893 
1894   // We'll need a library call to handle most floats when using soft.
1895   if (TLI->useSoftFloat()) {
1896     switch (I.getOpcode()) {
1897     default:
1898       return true;
1899     case Instruction::Alloca:
1900     case Instruction::Load:
1901     case Instruction::Store:
1902     case Instruction::Select:
1903     case Instruction::PHI:
1904       return false;
1905     }
1906   }
1907 
1908   // We'll need a libcall to perform double precision operations on a single
1909   // precision only FPU.
1910   if (I.getType()->isDoubleTy() && !ST->hasFP64())
1911     return true;
1912 
1913   // Likewise for half precision arithmetic.
1914   if (I.getType()->isHalfTy() && !ST->hasFullFP16())
1915     return true;
1916 
1917   return false;
1918 }
1919 
1920 bool ARMTTIImpl::isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
1921                                           AssumptionCache &AC,
1922                                           TargetLibraryInfo *LibInfo,
1923                                           HardwareLoopInfo &HWLoopInfo) {
1924   // Low-overhead branches are only supported in the 'low-overhead branch'
1925   // extension of v8.1-m.
1926   if (!ST->hasLOB() || DisableLowOverheadLoops) {
1927     LLVM_DEBUG(dbgs() << "ARMHWLoops: Disabled\n");
1928     return false;
1929   }
1930 
1931   if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
1932     LLVM_DEBUG(dbgs() << "ARMHWLoops: No BETC\n");
1933     return false;
1934   }
1935 
1936   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1937   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1938     LLVM_DEBUG(dbgs() << "ARMHWLoops: Uncomputable BETC\n");
1939     return false;
1940   }
1941 
1942   const SCEV *TripCountSCEV =
1943     SE.getAddExpr(BackedgeTakenCount,
1944                   SE.getOne(BackedgeTakenCount->getType()));
1945 
1946   // We need to store the trip count in LR, a 32-bit register.
1947   if (SE.getUnsignedRangeMax(TripCountSCEV).getBitWidth() > 32) {
1948     LLVM_DEBUG(dbgs() << "ARMHWLoops: Trip count does not fit into 32bits\n");
1949     return false;
1950   }
1951 
1952   // Making a call will trash LR and clear LO_BRANCH_INFO, so there's little
1953   // point in generating a hardware loop if that's going to happen.
1954 
1955   auto IsHardwareLoopIntrinsic = [](Instruction &I) {
1956     if (auto *Call = dyn_cast<IntrinsicInst>(&I)) {
1957       switch (Call->getIntrinsicID()) {
1958       default:
1959         break;
1960       case Intrinsic::start_loop_iterations:
1961       case Intrinsic::test_start_loop_iterations:
1962       case Intrinsic::loop_decrement:
1963       case Intrinsic::loop_decrement_reg:
1964         return true;
1965       }
1966     }
1967     return false;
1968   };
1969 
1970   // Scan the instructions to see if there's any that we know will turn into a
1971   // call or if this loop is already a low-overhead loop or will become a tail
1972   // predicated loop.
1973   bool IsTailPredLoop = false;
1974   auto ScanLoop = [&](Loop *L) {
1975     for (auto *BB : L->getBlocks()) {
1976       for (auto &I : *BB) {
1977         if (maybeLoweredToCall(I) || IsHardwareLoopIntrinsic(I) ||
1978             isa<InlineAsm>(I)) {
1979           LLVM_DEBUG(dbgs() << "ARMHWLoops: Bad instruction: " << I << "\n");
1980           return false;
1981         }
1982         if (auto *II = dyn_cast<IntrinsicInst>(&I))
1983           IsTailPredLoop |=
1984               II->getIntrinsicID() == Intrinsic::get_active_lane_mask ||
1985               II->getIntrinsicID() == Intrinsic::arm_mve_vctp8 ||
1986               II->getIntrinsicID() == Intrinsic::arm_mve_vctp16 ||
1987               II->getIntrinsicID() == Intrinsic::arm_mve_vctp32 ||
1988               II->getIntrinsicID() == Intrinsic::arm_mve_vctp64;
1989       }
1990     }
1991     return true;
1992   };
1993 
1994   // Visit inner loops.
1995   for (auto Inner : *L)
1996     if (!ScanLoop(Inner))
1997       return false;
1998 
1999   if (!ScanLoop(L))
2000     return false;
2001 
2002   // TODO: Check whether the trip count calculation is expensive. If L is the
2003   // inner loop but we know it has a low trip count, calculating that trip
2004   // count (in the parent loop) may be detrimental.
2005 
2006   LLVMContext &C = L->getHeader()->getContext();
2007   HWLoopInfo.CounterInReg = true;
2008   HWLoopInfo.IsNestingLegal = false;
2009   HWLoopInfo.PerformEntryTest = AllowWLSLoops && !IsTailPredLoop;
2010   HWLoopInfo.CountType = Type::getInt32Ty(C);
2011   HWLoopInfo.LoopDecrement = ConstantInt::get(HWLoopInfo.CountType, 1);
2012   return true;
2013 }
2014 
2015 static bool canTailPredicateInstruction(Instruction &I, int &ICmpCount) {
2016   // We don't allow icmp's, and because we only look at single block loops,
2017   // we simply count the icmps, i.e. there should only be 1 for the backedge.
2018   if (isa<ICmpInst>(&I) && ++ICmpCount > 1)
2019     return false;
2020   // FIXME: This is a workaround for poor cost modelling. Min/Max intrinsics are
2021   // not currently canonical, but soon will be. Code without them uses icmp, and
2022   // so is not tail predicated as per the condition above. In order to get the
2023   // same performance we treat min and max the same as an icmp for tailpred
2024   // purposes for the moment (we often rely on non-tailpred and higher VF's to
2025   // pick more optimial instructions like VQDMULH. They need to be recognized
2026   // directly by the vectorizer).
2027   if (auto *II = dyn_cast<IntrinsicInst>(&I))
2028     if ((II->getIntrinsicID() == Intrinsic::smin ||
2029          II->getIntrinsicID() == Intrinsic::smax ||
2030          II->getIntrinsicID() == Intrinsic::umin ||
2031          II->getIntrinsicID() == Intrinsic::umax) &&
2032         ++ICmpCount > 1)
2033       return false;
2034 
2035   if (isa<FCmpInst>(&I))
2036     return false;
2037 
2038   // We could allow extending/narrowing FP loads/stores, but codegen is
2039   // too inefficient so reject this for now.
2040   if (isa<FPExtInst>(&I) || isa<FPTruncInst>(&I))
2041     return false;
2042 
2043   // Extends have to be extending-loads
2044   if (isa<SExtInst>(&I) || isa<ZExtInst>(&I) )
2045     if (!I.getOperand(0)->hasOneUse() || !isa<LoadInst>(I.getOperand(0)))
2046       return false;
2047 
2048   // Truncs have to be narrowing-stores
2049   if (isa<TruncInst>(&I) )
2050     if (!I.hasOneUse() || !isa<StoreInst>(*I.user_begin()))
2051       return false;
2052 
2053   return true;
2054 }
2055 
2056 // To set up a tail-predicated loop, we need to know the total number of
2057 // elements processed by that loop. Thus, we need to determine the element
2058 // size and:
2059 // 1) it should be uniform for all operations in the vector loop, so we
2060 //    e.g. don't want any widening/narrowing operations.
2061 // 2) it should be smaller than i64s because we don't have vector operations
2062 //    that work on i64s.
2063 // 3) we don't want elements to be reversed or shuffled, to make sure the
2064 //    tail-predication masks/predicates the right lanes.
2065 //
2066 static bool canTailPredicateLoop(Loop *L, LoopInfo *LI, ScalarEvolution &SE,
2067                                  const DataLayout &DL,
2068                                  const LoopAccessInfo *LAI) {
2069   LLVM_DEBUG(dbgs() << "Tail-predication: checking allowed instructions\n");
2070 
2071   // If there are live-out values, it is probably a reduction. We can predicate
2072   // most reduction operations freely under MVE using a combination of
2073   // prefer-predicated-reduction-select and inloop reductions. We limit this to
2074   // floating point and integer reductions, but don't check for operators
2075   // specifically here. If the value ends up not being a reduction (and so the
2076   // vectorizer cannot tailfold the loop), we should fall back to standard
2077   // vectorization automatically.
2078   SmallVector< Instruction *, 8 > LiveOuts;
2079   LiveOuts = llvm::findDefsUsedOutsideOfLoop(L);
2080   bool ReductionsDisabled =
2081       EnableTailPredication == TailPredication::EnabledNoReductions ||
2082       EnableTailPredication == TailPredication::ForceEnabledNoReductions;
2083 
2084   for (auto *I : LiveOuts) {
2085     if (!I->getType()->isIntegerTy() && !I->getType()->isFloatTy() &&
2086         !I->getType()->isHalfTy()) {
2087       LLVM_DEBUG(dbgs() << "Don't tail-predicate loop with non-integer/float "
2088                            "live-out value\n");
2089       return false;
2090     }
2091     if (ReductionsDisabled) {
2092       LLVM_DEBUG(dbgs() << "Reductions not enabled\n");
2093       return false;
2094     }
2095   }
2096 
2097   // Next, check that all instructions can be tail-predicated.
2098   PredicatedScalarEvolution PSE = LAI->getPSE();
2099   SmallVector<Instruction *, 16> LoadStores;
2100   int ICmpCount = 0;
2101 
2102   for (BasicBlock *BB : L->blocks()) {
2103     for (Instruction &I : BB->instructionsWithoutDebug()) {
2104       if (isa<PHINode>(&I))
2105         continue;
2106       if (!canTailPredicateInstruction(I, ICmpCount)) {
2107         LLVM_DEBUG(dbgs() << "Instruction not allowed: "; I.dump());
2108         return false;
2109       }
2110 
2111       Type *T  = I.getType();
2112       if (T->getScalarSizeInBits() > 32) {
2113         LLVM_DEBUG(dbgs() << "Unsupported Type: "; T->dump());
2114         return false;
2115       }
2116       if (isa<StoreInst>(I) || isa<LoadInst>(I)) {
2117         Value *Ptr = getLoadStorePointerOperand(&I);
2118         Type *AccessTy = getLoadStoreType(&I);
2119         int64_t NextStride = getPtrStride(PSE, AccessTy, Ptr, L);
2120         if (NextStride == 1) {
2121           // TODO: for now only allow consecutive strides of 1. We could support
2122           // other strides as long as it is uniform, but let's keep it simple
2123           // for now.
2124           continue;
2125         } else if (NextStride == -1 ||
2126                    (NextStride == 2 && MVEMaxSupportedInterleaveFactor >= 2) ||
2127                    (NextStride == 4 && MVEMaxSupportedInterleaveFactor >= 4)) {
2128           LLVM_DEBUG(dbgs()
2129                      << "Consecutive strides of 2 found, vld2/vstr2 can't "
2130                         "be tail-predicated\n.");
2131           return false;
2132           // TODO: don't tail predicate if there is a reversed load?
2133         } else if (EnableMaskedGatherScatters) {
2134           // Gather/scatters do allow loading from arbitrary strides, at
2135           // least if they are loop invariant.
2136           // TODO: Loop variant strides should in theory work, too, but
2137           // this requires further testing.
2138           const SCEV *PtrScev = PSE.getSE()->getSCEV(Ptr);
2139           if (auto AR = dyn_cast<SCEVAddRecExpr>(PtrScev)) {
2140             const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
2141             if (PSE.getSE()->isLoopInvariant(Step, L))
2142               continue;
2143           }
2144         }
2145         LLVM_DEBUG(dbgs() << "Bad stride found, can't "
2146                              "tail-predicate\n.");
2147         return false;
2148       }
2149     }
2150   }
2151 
2152   LLVM_DEBUG(dbgs() << "tail-predication: all instructions allowed!\n");
2153   return true;
2154 }
2155 
2156 bool ARMTTIImpl::preferPredicateOverEpilogue(Loop *L, LoopInfo *LI,
2157                                              ScalarEvolution &SE,
2158                                              AssumptionCache &AC,
2159                                              TargetLibraryInfo *TLI,
2160                                              DominatorTree *DT,
2161                                              const LoopAccessInfo *LAI) {
2162   if (!EnableTailPredication) {
2163     LLVM_DEBUG(dbgs() << "Tail-predication not enabled.\n");
2164     return false;
2165   }
2166 
2167   // Creating a predicated vector loop is the first step for generating a
2168   // tail-predicated hardware loop, for which we need the MVE masked
2169   // load/stores instructions:
2170   if (!ST->hasMVEIntegerOps())
2171     return false;
2172 
2173   // For now, restrict this to single block loops.
2174   if (L->getNumBlocks() > 1) {
2175     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: not a single block "
2176                          "loop.\n");
2177     return false;
2178   }
2179 
2180   assert(L->isInnermost() && "preferPredicateOverEpilogue: inner-loop expected");
2181 
2182   HardwareLoopInfo HWLoopInfo(L);
2183   if (!HWLoopInfo.canAnalyze(*LI)) {
2184     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
2185                          "analyzable.\n");
2186     return false;
2187   }
2188 
2189   // This checks if we have the low-overhead branch architecture
2190   // extension, and if we will create a hardware-loop:
2191   if (!isHardwareLoopProfitable(L, SE, AC, TLI, HWLoopInfo)) {
2192     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
2193                          "profitable.\n");
2194     return false;
2195   }
2196 
2197   if (!HWLoopInfo.isHardwareLoopCandidate(SE, *LI, *DT)) {
2198     LLVM_DEBUG(dbgs() << "preferPredicateOverEpilogue: hardware-loop is not "
2199                          "a candidate.\n");
2200     return false;
2201   }
2202 
2203   return canTailPredicateLoop(L, LI, SE, DL, LAI);
2204 }
2205 
2206 bool ARMTTIImpl::emitGetActiveLaneMask() const {
2207   if (!ST->hasMVEIntegerOps() || !EnableTailPredication)
2208     return false;
2209 
2210   // Intrinsic @llvm.get.active.lane.mask is supported.
2211   // It is used in the MVETailPredication pass, which requires the number of
2212   // elements processed by this vector loop to setup the tail-predicated
2213   // loop.
2214   return true;
2215 }
2216 void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
2217                                          TTI::UnrollingPreferences &UP,
2218                                          OptimizationRemarkEmitter *ORE) {
2219   // Enable Upper bound unrolling universally, not dependant upon the conditions
2220   // below.
2221   UP.UpperBound = true;
2222 
2223   // Only currently enable these preferences for M-Class cores.
2224   if (!ST->isMClass())
2225     return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP, ORE);
2226 
2227   // Disable loop unrolling for Oz and Os.
2228   UP.OptSizeThreshold = 0;
2229   UP.PartialOptSizeThreshold = 0;
2230   if (L->getHeader()->getParent()->hasOptSize())
2231     return;
2232 
2233   SmallVector<BasicBlock*, 4> ExitingBlocks;
2234   L->getExitingBlocks(ExitingBlocks);
2235   LLVM_DEBUG(dbgs() << "Loop has:\n"
2236                     << "Blocks: " << L->getNumBlocks() << "\n"
2237                     << "Exit blocks: " << ExitingBlocks.size() << "\n");
2238 
2239   // Only allow another exit other than the latch. This acts as an early exit
2240   // as it mirrors the profitability calculation of the runtime unroller.
2241   if (ExitingBlocks.size() > 2)
2242     return;
2243 
2244   // Limit the CFG of the loop body for targets with a branch predictor.
2245   // Allowing 4 blocks permits if-then-else diamonds in the body.
2246   if (ST->hasBranchPredictor() && L->getNumBlocks() > 4)
2247     return;
2248 
2249   // Don't unroll vectorized loops, including the remainder loop
2250   if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
2251     return;
2252 
2253   // Scan the loop: don't unroll loops with calls as this could prevent
2254   // inlining.
2255   InstructionCost Cost = 0;
2256   for (auto *BB : L->getBlocks()) {
2257     for (auto &I : *BB) {
2258       // Don't unroll vectorised loop. MVE does not benefit from it as much as
2259       // scalar code.
2260       if (I.getType()->isVectorTy())
2261         return;
2262 
2263       if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
2264         if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
2265           if (!isLoweredToCall(F))
2266             continue;
2267         }
2268         return;
2269       }
2270 
2271       SmallVector<const Value*, 4> Operands(I.operand_values());
2272       Cost +=
2273         getUserCost(&I, Operands, TargetTransformInfo::TCK_SizeAndLatency);
2274     }
2275   }
2276 
2277   // On v6m cores, there are very few registers available. We can easily end up
2278   // spilling and reloading more registers in an unrolled loop. Look at the
2279   // number of LCSSA phis as a rough measure of how many registers will need to
2280   // be live out of the loop, reducing the default unroll count if more than 1
2281   // value is needed.  In the long run, all of this should be being learnt by a
2282   // machine.
2283   unsigned UnrollCount = 4;
2284   if (ST->isThumb1Only()) {
2285     unsigned ExitingValues = 0;
2286     SmallVector<BasicBlock *, 4> ExitBlocks;
2287     L->getExitBlocks(ExitBlocks);
2288     for (auto *Exit : ExitBlocks) {
2289       // Count the number of LCSSA phis. Exclude values coming from GEP's as
2290       // only the last is expected to be needed for address operands.
2291       unsigned LiveOuts = count_if(Exit->phis(), [](auto &PH) {
2292         return PH.getNumOperands() != 1 ||
2293                !isa<GetElementPtrInst>(PH.getOperand(0));
2294       });
2295       ExitingValues = ExitingValues < LiveOuts ? LiveOuts : ExitingValues;
2296     }
2297     if (ExitingValues)
2298       UnrollCount /= ExitingValues;
2299     if (UnrollCount <= 1)
2300       return;
2301   }
2302 
2303   LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n");
2304   LLVM_DEBUG(dbgs() << "Default Runtime Unroll Count: " << UnrollCount << "\n");
2305 
2306   UP.Partial = true;
2307   UP.Runtime = true;
2308   UP.UnrollRemainder = true;
2309   UP.DefaultUnrollRuntimeCount = UnrollCount;
2310   UP.UnrollAndJam = true;
2311   UP.UnrollAndJamInnerLoopThreshold = 60;
2312 
2313   // Force unrolling small loops can be very useful because of the branch
2314   // taken cost of the backedge.
2315   if (Cost < 12)
2316     UP.Force = true;
2317 }
2318 
2319 void ARMTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
2320                                        TTI::PeelingPreferences &PP) {
2321   BaseT::getPeelingPreferences(L, SE, PP);
2322 }
2323 
2324 bool ARMTTIImpl::preferInLoopReduction(unsigned Opcode, Type *Ty,
2325                                        TTI::ReductionFlags Flags) const {
2326   if (!ST->hasMVEIntegerOps())
2327     return false;
2328 
2329   unsigned ScalarBits = Ty->getScalarSizeInBits();
2330   switch (Opcode) {
2331   case Instruction::Add:
2332     return ScalarBits <= 64;
2333   default:
2334     return false;
2335   }
2336 }
2337 
2338 bool ARMTTIImpl::preferPredicatedReductionSelect(
2339     unsigned Opcode, Type *Ty, TTI::ReductionFlags Flags) const {
2340   if (!ST->hasMVEIntegerOps())
2341     return false;
2342   return true;
2343 }
2344