xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
1 //===- AArch64LegalizerInfo.cpp ----------------------------------*- C++ -*-==//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 /// This file implements the targeting of the Machinelegalizer class for
10 /// AArch64.
11 /// \todo This should be generated by TableGen.
12 //===----------------------------------------------------------------------===//
13 
14 #include "AArch64LegalizerInfo.h"
15 #include "AArch64RegisterBankInfo.h"
16 #include "AArch64Subtarget.h"
17 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
18 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
19 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
20 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
21 #include "llvm/CodeGen/GlobalISel/Utils.h"
22 #include "llvm/CodeGen/MachineInstr.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetOpcodes.h"
25 #include "llvm/CodeGen/ValueTypes.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/IntrinsicsAArch64.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/Support/MathExtras.h"
31 #include <initializer_list>
32 
33 #define DEBUG_TYPE "aarch64-legalinfo"
34 
35 using namespace llvm;
36 using namespace LegalizeActions;
37 using namespace LegalizeMutations;
38 using namespace LegalityPredicates;
39 using namespace MIPatternMatch;
40 
41 AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
42     : ST(&ST) {
43   using namespace TargetOpcode;
44   const LLT p0 = LLT::pointer(0, 64);
45   const LLT s1 = LLT::scalar(1);
46   const LLT s8 = LLT::scalar(8);
47   const LLT s16 = LLT::scalar(16);
48   const LLT s32 = LLT::scalar(32);
49   const LLT s64 = LLT::scalar(64);
50   const LLT s128 = LLT::scalar(128);
51   const LLT v16s8 = LLT::fixed_vector(16, 8);
52   const LLT v8s8 = LLT::fixed_vector(8, 8);
53   const LLT v4s8 = LLT::fixed_vector(4, 8);
54   const LLT v8s16 = LLT::fixed_vector(8, 16);
55   const LLT v4s16 = LLT::fixed_vector(4, 16);
56   const LLT v2s16 = LLT::fixed_vector(2, 16);
57   const LLT v2s32 = LLT::fixed_vector(2, 32);
58   const LLT v4s32 = LLT::fixed_vector(4, 32);
59   const LLT v2s64 = LLT::fixed_vector(2, 64);
60   const LLT v2p0 = LLT::fixed_vector(2, p0);
61 
62   std::initializer_list<LLT> PackedVectorAllTypeList = {/* Begin 128bit types */
63                                                         v16s8, v8s16, v4s32,
64                                                         v2s64, v2p0,
65                                                         /* End 128bit types */
66                                                         /* Begin 64bit types */
67                                                         v8s8, v4s16, v2s32};
68 
69   const TargetMachine &TM = ST.getTargetLowering()->getTargetMachine();
70 
71   // FIXME: support subtargets which have neon/fp-armv8 disabled.
72   if (!ST.hasNEON() || !ST.hasFPARMv8()) {
73     getLegacyLegalizerInfo().computeTables();
74     return;
75   }
76 
77   // Some instructions only support s16 if the subtarget has full 16-bit FP
78   // support.
79   const bool HasFP16 = ST.hasFullFP16();
80   const LLT &MinFPScalar = HasFP16 ? s16 : s32;
81 
82   getActionDefinitionsBuilder({G_IMPLICIT_DEF, G_FREEZE})
83       .legalFor({p0, s1, s8, s16, s32, s64})
84       .legalFor(PackedVectorAllTypeList)
85       .widenScalarToNextPow2(0)
86       .clampScalar(0, s8, s64)
87       .fewerElementsIf(
88           [=](const LegalityQuery &Query) {
89             return Query.Types[0].isVector() &&
90                    (Query.Types[0].getElementType() != s64 ||
91                     Query.Types[0].getNumElements() != 2);
92           },
93           [=](const LegalityQuery &Query) {
94             LLT EltTy = Query.Types[0].getElementType();
95             if (EltTy == s64)
96               return std::make_pair(0, LLT::fixed_vector(2, 64));
97             return std::make_pair(0, EltTy);
98           });
99 
100   getActionDefinitionsBuilder(G_PHI)
101       .legalFor({p0, s16, s32, s64})
102       .legalFor(PackedVectorAllTypeList)
103       .widenScalarToNextPow2(0)
104       .clampScalar(0, s16, s64)
105       // Maximum: sN * k = 128
106       .clampMaxNumElements(0, s8, 16)
107       .clampMaxNumElements(0, s16, 8)
108       .clampMaxNumElements(0, s32, 4)
109       .clampMaxNumElements(0, s64, 2)
110       .clampMaxNumElements(0, p0, 2);
111 
112   getActionDefinitionsBuilder(G_BSWAP)
113       .legalFor({s32, s64, v4s32, v2s32, v2s64})
114       .widenScalarToNextPow2(0)
115       .clampScalar(0, s32, s64);
116 
117   getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR})
118       .legalFor({s32, s64, v2s32, v4s32, v4s16, v8s16, v16s8, v8s8})
119       .scalarizeIf(
120           [=](const LegalityQuery &Query) {
121             return Query.Opcode == G_MUL && Query.Types[0] == v2s64;
122           },
123           0)
124       .legalFor({v2s64})
125       .widenScalarToNextPow2(0)
126       .clampScalar(0, s32, s64)
127       .clampNumElements(0, v2s32, v4s32)
128       .clampNumElements(0, v2s64, v2s64)
129       .moreElementsToNextPow2(0);
130 
131   getActionDefinitionsBuilder({G_SHL, G_ASHR, G_LSHR})
132       .customIf([=](const LegalityQuery &Query) {
133         const auto &SrcTy = Query.Types[0];
134         const auto &AmtTy = Query.Types[1];
135         return !SrcTy.isVector() && SrcTy.getSizeInBits() == 32 &&
136                AmtTy.getSizeInBits() == 32;
137       })
138       .legalFor({
139           {s32, s32},
140           {s32, s64},
141           {s64, s64},
142           {v8s8, v8s8},
143           {v16s8, v16s8},
144           {v4s16, v4s16},
145           {v8s16, v8s16},
146           {v2s32, v2s32},
147           {v4s32, v4s32},
148           {v2s64, v2s64},
149       })
150       .widenScalarToNextPow2(0)
151       .clampScalar(1, s32, s64)
152       .clampScalar(0, s32, s64)
153       .clampNumElements(0, v2s32, v4s32)
154       .clampNumElements(0, v2s64, v2s64)
155       .moreElementsToNextPow2(0)
156       .minScalarSameAs(1, 0);
157 
158   getActionDefinitionsBuilder(G_PTR_ADD)
159       .legalFor({{p0, s64}, {v2p0, v2s64}})
160       .clampScalar(1, s64, s64);
161 
162   getActionDefinitionsBuilder(G_PTRMASK).legalFor({{p0, s64}});
163 
164   getActionDefinitionsBuilder({G_SDIV, G_UDIV})
165       .legalFor({s32, s64})
166       .libcallFor({s128})
167       .clampScalar(0, s32, s64)
168       .widenScalarToNextPow2(0)
169       .scalarize(0);
170 
171   getActionDefinitionsBuilder({G_SREM, G_UREM, G_SDIVREM, G_UDIVREM})
172       .lowerFor({s1, s8, s16, s32, s64, v2s64, v4s32, v2s32})
173       .widenScalarOrEltToNextPow2(0)
174       .clampScalarOrElt(0, s32, s64)
175       .clampNumElements(0, v2s32, v4s32)
176       .clampNumElements(0, v2s64, v2s64)
177       .moreElementsToNextPow2(0);
178 
179 
180   getActionDefinitionsBuilder({G_SMULO, G_UMULO})
181       .widenScalarToNextPow2(0, /*Min = */ 32)
182       .clampScalar(0, s32, s64)
183       .lowerIf(typeIs(1, s1));
184 
185   getActionDefinitionsBuilder({G_SMULH, G_UMULH})
186       .legalFor({s64, v8s16, v16s8, v4s32})
187       .lower();
188 
189   getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
190       .legalFor({v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
191       .clampNumElements(0, v8s8, v16s8)
192       .clampNumElements(0, v4s16, v8s16)
193       .clampNumElements(0, v2s32, v4s32)
194       // FIXME: This sholdn't be needed as v2s64 types are going to
195       // be expanded anyway, but G_ICMP doesn't support splitting vectors yet
196       .clampNumElements(0, v2s64, v2s64)
197       .lower();
198 
199   getActionDefinitionsBuilder(
200       {G_SADDE, G_SSUBE, G_UADDE, G_USUBE, G_SADDO, G_SSUBO, G_UADDO, G_USUBO})
201       .legalFor({{s32, s1}, {s64, s1}})
202       .clampScalar(0, s32, s64)
203       .widenScalarToNextPow2(0);
204 
205   getActionDefinitionsBuilder({G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FNEG})
206       .legalFor({MinFPScalar, s32, s64, v2s64, v4s32, v2s32})
207       .clampScalar(0, MinFPScalar, s64)
208       .clampNumElements(0, v2s32, v4s32)
209       .clampNumElements(0, v2s64, v2s64);
210 
211   getActionDefinitionsBuilder(G_FREM).libcallFor({s32, s64});
212 
213   getActionDefinitionsBuilder({G_FCEIL, G_FABS, G_FSQRT, G_FFLOOR, G_FRINT,
214                                G_FMA, G_INTRINSIC_TRUNC, G_INTRINSIC_ROUND,
215                                G_FNEARBYINT, G_INTRINSIC_LRINT})
216       // If we don't have full FP16 support, then scalarize the elements of
217       // vectors containing fp16 types.
218       .fewerElementsIf(
219           [=, &ST](const LegalityQuery &Query) {
220             const auto &Ty = Query.Types[0];
221             return Ty.isVector() && Ty.getElementType() == s16 &&
222                    !ST.hasFullFP16();
223           },
224           [=](const LegalityQuery &Query) { return std::make_pair(0, s16); })
225       // If we don't have full FP16 support, then widen s16 to s32 if we
226       // encounter it.
227       .widenScalarIf(
228           [=, &ST](const LegalityQuery &Query) {
229             return Query.Types[0] == s16 && !ST.hasFullFP16();
230           },
231           [=](const LegalityQuery &Query) { return std::make_pair(0, s32); })
232       .legalFor({s16, s32, s64, v2s32, v4s32, v2s64, v2s16, v4s16, v8s16});
233 
234   getActionDefinitionsBuilder(
235       {G_FCOS, G_FSIN, G_FLOG10, G_FLOG, G_FLOG2, G_FEXP, G_FEXP2, G_FPOW})
236       // We need a call for these, so we always need to scalarize.
237       .scalarize(0)
238       // Regardless of FP16 support, widen 16-bit elements to 32-bits.
239       .minScalar(0, s32)
240       .libcallFor({s32, s64, v2s32, v4s32, v2s64});
241 
242   getActionDefinitionsBuilder(G_INSERT)
243       .legalIf(all(typeInSet(0, {s32, s64, p0}),
244                    typeInSet(1, {s1, s8, s16, s32}), smallerThan(1, 0)))
245       .widenScalarToNextPow2(0)
246       .clampScalar(0, s32, s64)
247       .widenScalarToNextPow2(1)
248       .minScalar(1, s8)
249       .maxScalarIf(typeInSet(0, {s32}), 1, s16)
250       .maxScalarIf(typeInSet(0, {s64, p0}), 1, s32);
251 
252   getActionDefinitionsBuilder(G_EXTRACT)
253       .legalIf(all(typeInSet(0, {s16, s32, s64, p0}),
254                    typeInSet(1, {s32, s64, s128, p0}), smallerThan(0, 1)))
255       .widenScalarToNextPow2(1)
256       .clampScalar(1, s32, s128)
257       .widenScalarToNextPow2(0)
258       .minScalar(0, s16)
259       .maxScalarIf(typeInSet(1, {s32}), 0, s16)
260       .maxScalarIf(typeInSet(1, {s64, p0}), 0, s32)
261       .maxScalarIf(typeInSet(1, {s128}), 0, s64);
262 
263   getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
264       .lowerIf(atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Unordered))
265       .legalForTypesWithMemDesc({{s32, p0, s8, 8},
266                                  {s32, p0, s16, 8},
267                                  {s32, p0, s32, 8},
268                                  {s64, p0, s8, 2},
269                                  {s64, p0, s16, 2},
270                                  {s64, p0, s32, 4},
271                                  {s64, p0, s64, 8},
272                                  {p0, p0, s64, 8},
273                                  {v2s32, p0, s64, 8}})
274       .widenScalarToNextPow2(0)
275       .clampScalar(0, s32, s64)
276       // TODO: We could support sum-of-pow2's but the lowering code doesn't know
277       //       how to do that yet.
278       .unsupportedIfMemSizeNotPow2()
279       // Lower anything left over into G_*EXT and G_LOAD
280       .lower();
281 
282   auto IsPtrVecPred = [=](const LegalityQuery &Query) {
283     const LLT &ValTy = Query.Types[0];
284     if (!ValTy.isVector())
285       return false;
286     const LLT EltTy = ValTy.getElementType();
287     return EltTy.isPointer() && EltTy.getAddressSpace() == 0;
288   };
289 
290   getActionDefinitionsBuilder(G_LOAD)
291       .customIf([=](const LegalityQuery &Query) {
292         return Query.Types[0] == s128 &&
293                Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic;
294       })
295       .legalForTypesWithMemDesc({{s8, p0, s8, 8},
296                                  {s16, p0, s16, 8},
297                                  {s32, p0, s32, 8},
298                                  {s64, p0, s64, 8},
299                                  {p0, p0, s64, 8},
300                                  {s128, p0, s128, 8},
301                                  {v8s8, p0, s64, 8},
302                                  {v16s8, p0, s128, 8},
303                                  {v4s16, p0, s64, 8},
304                                  {v8s16, p0, s128, 8},
305                                  {v2s32, p0, s64, 8},
306                                  {v4s32, p0, s128, 8},
307                                  {v2s64, p0, s128, 8}})
308       // These extends are also legal
309       .legalForTypesWithMemDesc({{s32, p0, s8, 8}, {s32, p0, s16, 8}})
310       .widenScalarToNextPow2(0, /* MinSize = */8)
311       .lowerIfMemSizeNotPow2()
312       .clampScalar(0, s8, s64)
313       .narrowScalarIf([=](const LegalityQuery &Query) {
314         // Clamp extending load results to 32-bits.
315         return Query.Types[0].isScalar() &&
316           Query.Types[0] != Query.MMODescrs[0].MemoryTy &&
317           Query.Types[0].getSizeInBits() > 32;
318         },
319         changeTo(0, s32))
320       // Lower any any-extending loads left into G_ANYEXT and G_LOAD
321       .lowerIf([=](const LegalityQuery &Query) {
322         return Query.Types[0] != Query.MMODescrs[0].MemoryTy;
323       })
324       .clampMaxNumElements(0, s8, 16)
325       .clampMaxNumElements(0, s16, 8)
326       .clampMaxNumElements(0, s32, 4)
327       .clampMaxNumElements(0, s64, 2)
328       .clampMaxNumElements(0, p0, 2)
329       .customIf(IsPtrVecPred)
330       .scalarizeIf(typeIs(0, v2s16), 0);
331 
332   getActionDefinitionsBuilder(G_STORE)
333       .customIf([=](const LegalityQuery &Query) {
334         return Query.Types[0] == s128 &&
335                Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic;
336       })
337       .legalForTypesWithMemDesc({{s8, p0, s8, 8},
338                                  {s16, p0, s8, 8}, // truncstorei8 from s16
339                                  {s32, p0, s8, 8}, // truncstorei8 from s32
340                                  {s64, p0, s8, 8}, // truncstorei8 from s64
341                                  {s16, p0, s16, 8},
342                                  {s32, p0, s16, 8}, // truncstorei16 from s32
343                                  {s64, p0, s16, 8}, // truncstorei16 from s64
344                                  {s32, p0, s8, 8},
345                                  {s32, p0, s16, 8},
346                                  {s32, p0, s32, 8},
347                                  {s64, p0, s64, 8},
348                                  {s64, p0, s32, 8}, // truncstorei32 from s64
349                                  {p0, p0, s64, 8},
350                                  {s128, p0, s128, 8},
351                                  {v16s8, p0, s128, 8},
352                                  {v8s8, p0, s64, 8},
353                                  {v4s16, p0, s64, 8},
354                                  {v8s16, p0, s128, 8},
355                                  {v2s32, p0, s64, 8},
356                                  {v4s32, p0, s128, 8},
357                                  {v2s64, p0, s128, 8}})
358       .clampScalar(0, s8, s64)
359       .lowerIf([=](const LegalityQuery &Query) {
360         return Query.Types[0].isScalar() &&
361                Query.Types[0] != Query.MMODescrs[0].MemoryTy;
362       })
363       // Maximum: sN * k = 128
364       .clampMaxNumElements(0, s8, 16)
365       .clampMaxNumElements(0, s16, 8)
366       .clampMaxNumElements(0, s32, 4)
367       .clampMaxNumElements(0, s64, 2)
368       .clampMaxNumElements(0, p0, 2)
369       .lowerIfMemSizeNotPow2()
370       .customIf(IsPtrVecPred)
371       .scalarizeIf(typeIs(0, v2s16), 0);
372 
373   // Constants
374   getActionDefinitionsBuilder(G_CONSTANT)
375       .legalFor({p0, s8, s16, s32, s64})
376       .widenScalarToNextPow2(0)
377       .clampScalar(0, s8, s64);
378   getActionDefinitionsBuilder(G_FCONSTANT)
379       .legalIf([=](const LegalityQuery &Query) {
380         const auto &Ty = Query.Types[0];
381         if (HasFP16 && Ty == s16)
382           return true;
383         return Ty == s32 || Ty == s64 || Ty == s128;
384       })
385       .clampScalar(0, MinFPScalar, s128);
386 
387   getActionDefinitionsBuilder({G_ICMP, G_FCMP})
388       .legalFor({{s32, s32},
389                  {s32, s64},
390                  {s32, p0},
391                  {v4s32, v4s32},
392                  {v2s32, v2s32},
393                  {v2s64, v2s64},
394                  {v2s64, v2p0},
395                  {v4s16, v4s16},
396                  {v8s16, v8s16},
397                  {v8s8, v8s8},
398                  {v16s8, v16s8}})
399       .widenScalarOrEltToNextPow2(1)
400       .clampScalar(1, s32, s64)
401       .clampScalar(0, s32, s32)
402       .minScalarEltSameAsIf(
403           [=](const LegalityQuery &Query) {
404             const LLT &Ty = Query.Types[0];
405             const LLT &SrcTy = Query.Types[1];
406             return Ty.isVector() && !SrcTy.getElementType().isPointer() &&
407                    Ty.getElementType() != SrcTy.getElementType();
408           },
409           0, 1)
410       .minScalarOrEltIf(
411           [=](const LegalityQuery &Query) { return Query.Types[1] == v2s16; },
412           1, s32)
413       .minScalarOrEltIf(
414           [=](const LegalityQuery &Query) { return Query.Types[1] == v2p0; }, 0,
415           s64)
416       .clampNumElements(0, v2s32, v4s32);
417 
418   // Extensions
419   auto ExtLegalFunc = [=](const LegalityQuery &Query) {
420     unsigned DstSize = Query.Types[0].getSizeInBits();
421 
422     if (DstSize == 128 && !Query.Types[0].isVector())
423       return false; // Extending to a scalar s128 needs narrowing.
424 
425     // Make sure that we have something that will fit in a register, and
426     // make sure it's a power of 2.
427     if (DstSize < 8 || DstSize > 128 || !isPowerOf2_32(DstSize))
428       return false;
429 
430     const LLT &SrcTy = Query.Types[1];
431 
432     // Special case for s1.
433     if (SrcTy == s1)
434       return true;
435 
436     // Make sure we fit in a register otherwise. Don't bother checking that
437     // the source type is below 128 bits. We shouldn't be allowing anything
438     // through which is wider than the destination in the first place.
439     unsigned SrcSize = SrcTy.getSizeInBits();
440     if (SrcSize < 8 || !isPowerOf2_32(SrcSize))
441       return false;
442 
443     return true;
444   };
445   getActionDefinitionsBuilder({G_ZEXT, G_SEXT, G_ANYEXT})
446       .legalIf(ExtLegalFunc)
447       .clampScalar(0, s64, s64); // Just for s128, others are handled above.
448 
449   getActionDefinitionsBuilder(G_TRUNC)
450       .minScalarOrEltIf(
451           [=](const LegalityQuery &Query) { return Query.Types[0].isVector(); },
452           0, s8)
453       .customIf([=](const LegalityQuery &Query) {
454         LLT DstTy = Query.Types[0];
455         LLT SrcTy = Query.Types[1];
456         return DstTy == v8s8 && SrcTy.getSizeInBits() > 128;
457       })
458       .alwaysLegal();
459 
460   getActionDefinitionsBuilder(G_SEXT_INREG).legalFor({s32, s64}).lower();
461 
462   // FP conversions
463   getActionDefinitionsBuilder(G_FPTRUNC)
464       .legalFor(
465           {{s16, s32}, {s16, s64}, {s32, s64}, {v4s16, v4s32}, {v2s32, v2s64}})
466       .clampMaxNumElements(0, s32, 2);
467   getActionDefinitionsBuilder(G_FPEXT)
468       .legalFor(
469           {{s32, s16}, {s64, s16}, {s64, s32}, {v4s32, v4s16}, {v2s64, v2s32}})
470       .clampMaxNumElements(0, s64, 2);
471 
472   // Conversions
473   getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
474       .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})
475       .widenScalarToNextPow2(0)
476       .clampScalar(0, s32, s64)
477       .widenScalarToNextPow2(1)
478       .clampScalar(1, s32, s64);
479 
480   getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
481       .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})
482       .clampScalar(1, s32, s64)
483       .minScalarSameAs(1, 0)
484       .clampScalar(0, s32, s64)
485       .widenScalarToNextPow2(0);
486 
487   // Control-flow
488   getActionDefinitionsBuilder(G_BRCOND).legalFor({s1, s8, s16, s32});
489   getActionDefinitionsBuilder(G_BRINDIRECT).legalFor({p0});
490 
491   getActionDefinitionsBuilder(G_SELECT)
492       .legalFor({{s32, s1}, {s64, s1}, {p0, s1}})
493       .widenScalarToNextPow2(0)
494       .clampScalar(0, s32, s64)
495       .minScalarEltSameAsIf(all(isVector(0), isVector(1)), 1, 0)
496       .lowerIf(isVector(0));
497 
498   // Pointer-handling
499   getActionDefinitionsBuilder(G_FRAME_INDEX).legalFor({p0});
500 
501   if (TM.getCodeModel() == CodeModel::Small)
502     getActionDefinitionsBuilder(G_GLOBAL_VALUE).custom();
503   else
504     getActionDefinitionsBuilder(G_GLOBAL_VALUE).legalFor({p0});
505 
506   getActionDefinitionsBuilder(G_PTRTOINT)
507       .legalForCartesianProduct({s1, s8, s16, s32, s64}, {p0})
508       .legalFor({{v2s64, v2p0}})
509       .maxScalar(0, s64)
510       .widenScalarToNextPow2(0, /*Min*/ 8);
511 
512   getActionDefinitionsBuilder(G_INTTOPTR)
513       .unsupportedIf([&](const LegalityQuery &Query) {
514         return Query.Types[0].getSizeInBits() != Query.Types[1].getSizeInBits();
515       })
516       .legalFor({{p0, s64}, {v2p0, v2s64}});
517 
518   // Casts for 32 and 64-bit width type are just copies.
519   // Same for 128-bit width type, except they are on the FPR bank.
520   getActionDefinitionsBuilder(G_BITCAST)
521       // FIXME: This is wrong since G_BITCAST is not allowed to change the
522       // number of bits but it's what the previous code described and fixing
523       // it breaks tests.
524       .legalForCartesianProduct({s1, s8, s16, s32, s64, s128, v16s8, v8s8, v4s8,
525                                  v8s16, v4s16, v2s16, v4s32, v2s32, v2s64,
526                                  v2p0});
527 
528   getActionDefinitionsBuilder(G_VASTART).legalFor({p0});
529 
530   // va_list must be a pointer, but most sized types are pretty easy to handle
531   // as the destination.
532   getActionDefinitionsBuilder(G_VAARG)
533       .customForCartesianProduct({s8, s16, s32, s64, p0}, {p0})
534       .clampScalar(0, s8, s64)
535       .widenScalarToNextPow2(0, /*Min*/ 8);
536 
537   getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG_WITH_SUCCESS)
538       .lowerIf(
539           all(typeInSet(0, {s8, s16, s32, s64, s128}), typeIs(1, s1), typeIs(2, p0)));
540 
541   getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG)
542       .customIf([](const LegalityQuery &Query) {
543         return Query.Types[0].getSizeInBits() == 128;
544       })
545       .clampScalar(0, s32, s64)
546       .legalIf(all(typeInSet(0, {s32, s64}), typeIs(1, p0)));
547 
548   getActionDefinitionsBuilder(
549       {G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB, G_ATOMICRMW_AND,
550        G_ATOMICRMW_OR, G_ATOMICRMW_XOR, G_ATOMICRMW_MIN, G_ATOMICRMW_MAX,
551        G_ATOMICRMW_UMIN, G_ATOMICRMW_UMAX})
552       .clampScalar(0, s32, s64)
553       .legalIf(all(typeInSet(0, {s32, s64}), typeIs(1, p0)));
554 
555   getActionDefinitionsBuilder(G_BLOCK_ADDR).legalFor({p0});
556 
557   // Merge/Unmerge
558   for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
559     unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
560     unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
561     getActionDefinitionsBuilder(Op)
562         .widenScalarToNextPow2(LitTyIdx, 8)
563         .widenScalarToNextPow2(BigTyIdx, 32)
564         .clampScalar(LitTyIdx, s8, s64)
565         .clampScalar(BigTyIdx, s32, s128)
566         .legalIf([=](const LegalityQuery &Q) {
567           switch (Q.Types[BigTyIdx].getSizeInBits()) {
568           case 32:
569           case 64:
570           case 128:
571             break;
572           default:
573             return false;
574           }
575           switch (Q.Types[LitTyIdx].getSizeInBits()) {
576           case 8:
577           case 16:
578           case 32:
579           case 64:
580             return true;
581           default:
582             return false;
583           }
584         });
585   }
586 
587   getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
588       .unsupportedIf([=](const LegalityQuery &Query) {
589         const LLT &EltTy = Query.Types[1].getElementType();
590         return Query.Types[0] != EltTy;
591       })
592       .minScalar(2, s64)
593       .legalIf([=](const LegalityQuery &Query) {
594         const LLT &VecTy = Query.Types[1];
595         return VecTy == v2s16 || VecTy == v4s16 || VecTy == v8s16 ||
596                VecTy == v4s32 || VecTy == v2s64 || VecTy == v2s32 ||
597                VecTy == v8s8 || VecTy == v16s8 || VecTy == v2s32 ||
598                VecTy == v2p0;
599       })
600       .minScalarOrEltIf(
601           [=](const LegalityQuery &Query) {
602             // We want to promote to <M x s1> to <M x s64> if that wouldn't
603             // cause the total vec size to be > 128b.
604             return Query.Types[1].getNumElements() <= 2;
605           },
606           0, s64)
607       .minScalarOrEltIf(
608           [=](const LegalityQuery &Query) {
609             return Query.Types[1].getNumElements() <= 4;
610           },
611           0, s32)
612       .minScalarOrEltIf(
613           [=](const LegalityQuery &Query) {
614             return Query.Types[1].getNumElements() <= 8;
615           },
616           0, s16)
617       .minScalarOrEltIf(
618           [=](const LegalityQuery &Query) {
619             return Query.Types[1].getNumElements() <= 16;
620           },
621           0, s8)
622       .minScalarOrElt(0, s8) // Worst case, we need at least s8.
623       .clampMaxNumElements(1, s64, 2)
624       .clampMaxNumElements(1, s32, 4)
625       .clampMaxNumElements(1, s16, 8)
626       .clampMaxNumElements(1, p0, 2);
627 
628   getActionDefinitionsBuilder(G_INSERT_VECTOR_ELT)
629       .legalIf(typeInSet(0, {v8s16, v2s32, v4s32, v2s64}));
630 
631   getActionDefinitionsBuilder(G_BUILD_VECTOR)
632       .legalFor({{v8s8, s8},
633                  {v16s8, s8},
634                  {v2s16, s16},
635                  {v4s16, s16},
636                  {v8s16, s16},
637                  {v2s32, s32},
638                  {v4s32, s32},
639                  {v2p0, p0},
640                  {v2s64, s64}})
641       .clampNumElements(0, v4s32, v4s32)
642       .clampNumElements(0, v2s64, v2s64)
643       .minScalarOrElt(0, s8)
644       .minScalarSameAs(1, 0);
645 
646   getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC).lower();
647 
648   getActionDefinitionsBuilder(G_CTLZ)
649       .legalForCartesianProduct(
650           {s32, s64, v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
651       .scalarize(1);
652   getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF).lower();
653 
654   // TODO: Custom lowering for v2s32, v4s32, v2s64.
655   getActionDefinitionsBuilder(G_BITREVERSE)
656       .legalFor({s32, s64, v8s8, v16s8})
657       .widenScalarToNextPow2(0, /*Min = */ 32)
658       .clampScalar(0, s32, s64);
659 
660   getActionDefinitionsBuilder(G_CTTZ_ZERO_UNDEF).lower();
661 
662   // TODO: Handle vector types.
663   getActionDefinitionsBuilder(G_CTTZ)
664       .clampScalar(0, s32, s64)
665       .scalarSameSizeAs(1, 0)
666       .customFor({s32, s64});
667 
668   getActionDefinitionsBuilder(G_SHUFFLE_VECTOR)
669       .legalIf([=](const LegalityQuery &Query) {
670         const LLT &DstTy = Query.Types[0];
671         const LLT &SrcTy = Query.Types[1];
672         // For now just support the TBL2 variant which needs the source vectors
673         // to be the same size as the dest.
674         if (DstTy != SrcTy)
675           return false;
676         for (auto &Ty : {v2s32, v4s32, v2s64, v2p0, v16s8, v8s16}) {
677           if (DstTy == Ty)
678             return true;
679         }
680         return false;
681       })
682       // G_SHUFFLE_VECTOR can have scalar sources (from 1 x s vectors), we
683       // just want those lowered into G_BUILD_VECTOR
684       .lowerIf([=](const LegalityQuery &Query) {
685         return !Query.Types[1].isVector();
686       })
687       .moreElementsToNextPow2(0)
688       .clampNumElements(0, v4s32, v4s32)
689       .clampNumElements(0, v2s64, v2s64);
690 
691   getActionDefinitionsBuilder(G_CONCAT_VECTORS)
692       .legalFor({{v4s32, v2s32}, {v8s16, v4s16}, {v16s8, v8s8}});
693 
694   getActionDefinitionsBuilder(G_JUMP_TABLE).legalFor({{p0}, {s64}});
695 
696   getActionDefinitionsBuilder(G_BRJT).legalIf([=](const LegalityQuery &Query) {
697     return Query.Types[0] == p0 && Query.Types[1] == s64;
698   });
699 
700   getActionDefinitionsBuilder(G_DYN_STACKALLOC).lower();
701 
702   if (ST.hasMOPS()) {
703     // G_BZERO is not supported. Currently it is only emitted by
704     // PreLegalizerCombiner for G_MEMSET with zero constant.
705     getActionDefinitionsBuilder(G_BZERO).unsupported();
706 
707     getActionDefinitionsBuilder(G_MEMSET)
708         .legalForCartesianProduct({p0}, {s64}, {s64})
709         .customForCartesianProduct({p0}, {s8}, {s64})
710         .immIdx(0); // Inform verifier imm idx 0 is handled.
711 
712     getActionDefinitionsBuilder({G_MEMCPY, G_MEMMOVE})
713         .legalForCartesianProduct({p0}, {p0}, {s64})
714         .immIdx(0); // Inform verifier imm idx 0 is handled.
715 
716     // G_MEMCPY_INLINE does not have a tailcall immediate
717     getActionDefinitionsBuilder(G_MEMCPY_INLINE)
718         .legalForCartesianProduct({p0}, {p0}, {s64});
719 
720   } else {
721     getActionDefinitionsBuilder({G_BZERO, G_MEMCPY, G_MEMMOVE, G_MEMSET})
722         .libcall();
723   }
724 
725   // FIXME: Legal types are only legal with NEON.
726   getActionDefinitionsBuilder(G_ABS)
727       .lowerIf(isScalar(0))
728       .legalFor(PackedVectorAllTypeList);
729 
730   getActionDefinitionsBuilder(G_VECREDUCE_FADD)
731       // We only have FADDP to do reduction-like operations. Lower the rest.
732       .legalFor({{s32, v2s32}, {s64, v2s64}})
733       .clampMaxNumElements(1, s64, 2)
734       .clampMaxNumElements(1, s32, 2)
735       .lower();
736 
737   getActionDefinitionsBuilder(G_VECREDUCE_ADD)
738       .legalFor(
739           {{s8, v16s8}, {s16, v8s16}, {s32, v4s32}, {s32, v2s32}, {s64, v2s64}})
740       .clampMaxNumElements(1, s64, 2)
741       .clampMaxNumElements(1, s32, 4)
742       .lower();
743 
744   getActionDefinitionsBuilder(
745       {G_VECREDUCE_OR, G_VECREDUCE_AND, G_VECREDUCE_XOR})
746       // Try to break down into smaller vectors as long as they're at least 64
747       // bits. This lets us use vector operations for some parts of the
748       // reduction.
749       .fewerElementsIf(
750           [=](const LegalityQuery &Q) {
751             LLT SrcTy = Q.Types[1];
752             if (SrcTy.isScalar())
753               return false;
754             if (!isPowerOf2_32(SrcTy.getNumElements()))
755               return false;
756             // We can usually perform 64b vector operations.
757             return SrcTy.getSizeInBits() > 64;
758           },
759           [=](const LegalityQuery &Q) {
760             LLT SrcTy = Q.Types[1];
761             return std::make_pair(1, SrcTy.divide(2));
762           })
763       .scalarize(1)
764       .lower();
765 
766   getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT})
767       .lowerIf([=](const LegalityQuery &Q) { return Q.Types[0].isScalar(); });
768 
769   getActionDefinitionsBuilder({G_FSHL, G_FSHR}).lower();
770 
771   getActionDefinitionsBuilder(G_ROTR)
772       .legalFor({{s32, s64}, {s64, s64}})
773       .customIf([=](const LegalityQuery &Q) {
774         return Q.Types[0].isScalar() && Q.Types[1].getScalarSizeInBits() < 64;
775       })
776       .lower();
777   getActionDefinitionsBuilder(G_ROTL).lower();
778 
779   getActionDefinitionsBuilder({G_SBFX, G_UBFX})
780       .customFor({{s32, s32}, {s64, s64}});
781 
782   // TODO: Use generic lowering when custom lowering is not possible.
783   auto always = [=](const LegalityQuery &Q) { return true; };
784   getActionDefinitionsBuilder(G_CTPOP)
785       .legalFor({{v8s8, v8s8}, {v16s8, v16s8}})
786       .clampScalar(0, s32, s128)
787       .widenScalarToNextPow2(0)
788       .minScalarEltSameAsIf(always, 1, 0)
789       .maxScalarEltSameAsIf(always, 1, 0)
790       .customFor({{s32, s32},
791                   {s64, s64},
792                   {s128, s128},
793                   {v2s64, v2s64},
794                   {v2s32, v2s32},
795                   {v4s32, v4s32},
796                   {v4s16, v4s16},
797                   {v8s16, v8s16}});
798 
799   // TODO: Vector types.
800   getActionDefinitionsBuilder({G_SADDSAT, G_SSUBSAT}).lowerIf(isScalar(0));
801 
802   // TODO: Vector types.
803   getActionDefinitionsBuilder({G_FMAXNUM, G_FMINNUM})
804       .legalFor({MinFPScalar, s32, s64})
805       .libcallFor({s128})
806       .minScalar(0, MinFPScalar);
807 
808   // TODO: Vector types.
809   getActionDefinitionsBuilder({G_FMAXIMUM, G_FMINIMUM})
810       .legalFor({MinFPScalar, s32, s64})
811       .minScalar(0, MinFPScalar);
812 
813   // TODO: Libcall support for s128.
814   // TODO: s16 should be legal with full FP16 support.
815   getActionDefinitionsBuilder({G_LROUND, G_LLROUND})
816       .legalFor({{s64, s32}, {s64, s64}});
817 
818   getLegacyLegalizerInfo().computeTables();
819   verify(*ST.getInstrInfo());
820 }
821 
822 bool AArch64LegalizerInfo::legalizeCustom(LegalizerHelper &Helper,
823                                           MachineInstr &MI) const {
824   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
825   MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
826   GISelChangeObserver &Observer = Helper.Observer;
827   switch (MI.getOpcode()) {
828   default:
829     // No idea what to do.
830     return false;
831   case TargetOpcode::G_VAARG:
832     return legalizeVaArg(MI, MRI, MIRBuilder);
833   case TargetOpcode::G_LOAD:
834   case TargetOpcode::G_STORE:
835     return legalizeLoadStore(MI, MRI, MIRBuilder, Observer);
836   case TargetOpcode::G_SHL:
837   case TargetOpcode::G_ASHR:
838   case TargetOpcode::G_LSHR:
839     return legalizeShlAshrLshr(MI, MRI, MIRBuilder, Observer);
840   case TargetOpcode::G_GLOBAL_VALUE:
841     return legalizeSmallCMGlobalValue(MI, MRI, MIRBuilder, Observer);
842   case TargetOpcode::G_TRUNC:
843     return legalizeVectorTrunc(MI, Helper);
844   case TargetOpcode::G_SBFX:
845   case TargetOpcode::G_UBFX:
846     return legalizeBitfieldExtract(MI, MRI, Helper);
847   case TargetOpcode::G_ROTR:
848     return legalizeRotate(MI, MRI, Helper);
849   case TargetOpcode::G_CTPOP:
850     return legalizeCTPOP(MI, MRI, Helper);
851   case TargetOpcode::G_ATOMIC_CMPXCHG:
852     return legalizeAtomicCmpxchg128(MI, MRI, Helper);
853   case TargetOpcode::G_CTTZ:
854     return legalizeCTTZ(MI, Helper);
855   case TargetOpcode::G_BZERO:
856   case TargetOpcode::G_MEMCPY:
857   case TargetOpcode::G_MEMMOVE:
858   case TargetOpcode::G_MEMSET:
859     return legalizeMemOps(MI, Helper);
860   }
861 
862   llvm_unreachable("expected switch to return");
863 }
864 
865 bool AArch64LegalizerInfo::legalizeRotate(MachineInstr &MI,
866                                           MachineRegisterInfo &MRI,
867                                           LegalizerHelper &Helper) const {
868   // To allow for imported patterns to match, we ensure that the rotate amount
869   // is 64b with an extension.
870   Register AmtReg = MI.getOperand(2).getReg();
871   LLT AmtTy = MRI.getType(AmtReg);
872   (void)AmtTy;
873   assert(AmtTy.isScalar() && "Expected a scalar rotate");
874   assert(AmtTy.getSizeInBits() < 64 && "Expected this rotate to be legal");
875   auto NewAmt = Helper.MIRBuilder.buildSExt(LLT::scalar(64), AmtReg);
876   Helper.Observer.changingInstr(MI);
877   MI.getOperand(2).setReg(NewAmt.getReg(0));
878   Helper.Observer.changedInstr(MI);
879   return true;
880 }
881 
882 static void extractParts(Register Reg, MachineRegisterInfo &MRI,
883                          MachineIRBuilder &MIRBuilder, LLT Ty, int NumParts,
884                          SmallVectorImpl<Register> &VRegs) {
885   for (int I = 0; I < NumParts; ++I)
886     VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
887   MIRBuilder.buildUnmerge(VRegs, Reg);
888 }
889 
890 bool AArch64LegalizerInfo::legalizeVectorTrunc(
891     MachineInstr &MI, LegalizerHelper &Helper) const {
892   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
893   MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
894   // Similar to how operand splitting is done in SelectiondDAG, we can handle
895   // %res(v8s8) = G_TRUNC %in(v8s32) by generating:
896   //   %inlo(<4x s32>), %inhi(<4 x s32>) = G_UNMERGE %in(<8 x s32>)
897   //   %lo16(<4 x s16>) = G_TRUNC %inlo
898   //   %hi16(<4 x s16>) = G_TRUNC %inhi
899   //   %in16(<8 x s16>) = G_CONCAT_VECTORS %lo16, %hi16
900   //   %res(<8 x s8>) = G_TRUNC %in16
901 
902   Register DstReg = MI.getOperand(0).getReg();
903   Register SrcReg = MI.getOperand(1).getReg();
904   LLT DstTy = MRI.getType(DstReg);
905   LLT SrcTy = MRI.getType(SrcReg);
906   assert(isPowerOf2_32(DstTy.getSizeInBits()) &&
907          isPowerOf2_32(SrcTy.getSizeInBits()));
908 
909   // Split input type.
910   LLT SplitSrcTy =
911       SrcTy.changeElementCount(SrcTy.getElementCount().divideCoefficientBy(2));
912   // First, split the source into two smaller vectors.
913   SmallVector<Register, 2> SplitSrcs;
914   extractParts(SrcReg, MRI, MIRBuilder, SplitSrcTy, 2, SplitSrcs);
915 
916   // Truncate the splits into intermediate narrower elements.
917   LLT InterTy = SplitSrcTy.changeElementSize(DstTy.getScalarSizeInBits() * 2);
918   for (unsigned I = 0; I < SplitSrcs.size(); ++I)
919     SplitSrcs[I] = MIRBuilder.buildTrunc(InterTy, SplitSrcs[I]).getReg(0);
920 
921   auto Concat = MIRBuilder.buildConcatVectors(
922       DstTy.changeElementSize(DstTy.getScalarSizeInBits() * 2), SplitSrcs);
923 
924   Helper.Observer.changingInstr(MI);
925   MI.getOperand(1).setReg(Concat.getReg(0));
926   Helper.Observer.changedInstr(MI);
927   return true;
928 }
929 
930 bool AArch64LegalizerInfo::legalizeSmallCMGlobalValue(
931     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
932     GISelChangeObserver &Observer) const {
933   assert(MI.getOpcode() == TargetOpcode::G_GLOBAL_VALUE);
934   // We do this custom legalization to convert G_GLOBAL_VALUE into target ADRP +
935   // G_ADD_LOW instructions.
936   // By splitting this here, we can optimize accesses in the small code model by
937   // folding in the G_ADD_LOW into the load/store offset.
938   auto &GlobalOp = MI.getOperand(1);
939   const auto* GV = GlobalOp.getGlobal();
940   if (GV->isThreadLocal())
941     return true; // Don't want to modify TLS vars.
942 
943   auto &TM = ST->getTargetLowering()->getTargetMachine();
944   unsigned OpFlags = ST->ClassifyGlobalReference(GV, TM);
945 
946   if (OpFlags & AArch64II::MO_GOT)
947     return true;
948 
949   auto Offset = GlobalOp.getOffset();
950   Register DstReg = MI.getOperand(0).getReg();
951   auto ADRP = MIRBuilder.buildInstr(AArch64::ADRP, {LLT::pointer(0, 64)}, {})
952                   .addGlobalAddress(GV, Offset, OpFlags | AArch64II::MO_PAGE);
953   // Set the regclass on the dest reg too.
954   MRI.setRegClass(ADRP.getReg(0), &AArch64::GPR64RegClass);
955 
956   // MO_TAGGED on the page indicates a tagged address. Set the tag now. We do so
957   // by creating a MOVK that sets bits 48-63 of the register to (global address
958   // + 0x100000000 - PC) >> 48. The additional 0x100000000 offset here is to
959   // prevent an incorrect tag being generated during relocation when the the
960   // global appears before the code section. Without the offset, a global at
961   // `0x0f00'0000'0000'1000` (i.e. at `0x1000` with tag `0xf`) that's referenced
962   // by code at `0x2000` would result in `0x0f00'0000'0000'1000 - 0x2000 =
963   // 0x0eff'ffff'ffff'f000`, meaning the tag would be incorrectly set to `0xe`
964   // instead of `0xf`.
965   // This assumes that we're in the small code model so we can assume a binary
966   // size of <= 4GB, which makes the untagged PC relative offset positive. The
967   // binary must also be loaded into address range [0, 2^48). Both of these
968   // properties need to be ensured at runtime when using tagged addresses.
969   if (OpFlags & AArch64II::MO_TAGGED) {
970     assert(!Offset &&
971            "Should not have folded in an offset for a tagged global!");
972     ADRP = MIRBuilder.buildInstr(AArch64::MOVKXi, {LLT::pointer(0, 64)}, {ADRP})
973                .addGlobalAddress(GV, 0x100000000,
974                                  AArch64II::MO_PREL | AArch64II::MO_G3)
975                .addImm(48);
976     MRI.setRegClass(ADRP.getReg(0), &AArch64::GPR64RegClass);
977   }
978 
979   MIRBuilder.buildInstr(AArch64::G_ADD_LOW, {DstReg}, {ADRP})
980       .addGlobalAddress(GV, Offset,
981                         OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
982   MI.eraseFromParent();
983   return true;
984 }
985 
986 bool AArch64LegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
987                                              MachineInstr &MI) const {
988   switch (MI.getIntrinsicID()) {
989   case Intrinsic::vacopy: {
990     unsigned PtrSize = ST->isTargetILP32() ? 4 : 8;
991     unsigned VaListSize =
992       (ST->isTargetDarwin() || ST->isTargetWindows())
993           ? PtrSize
994           : ST->isTargetILP32() ? 20 : 32;
995 
996     MachineFunction &MF = *MI.getMF();
997     auto Val = MF.getRegInfo().createGenericVirtualRegister(
998         LLT::scalar(VaListSize * 8));
999     MachineIRBuilder MIB(MI);
1000     MIB.buildLoad(Val, MI.getOperand(2),
1001                   *MF.getMachineMemOperand(MachinePointerInfo(),
1002                                            MachineMemOperand::MOLoad,
1003                                            VaListSize, Align(PtrSize)));
1004     MIB.buildStore(Val, MI.getOperand(1),
1005                    *MF.getMachineMemOperand(MachinePointerInfo(),
1006                                             MachineMemOperand::MOStore,
1007                                             VaListSize, Align(PtrSize)));
1008     MI.eraseFromParent();
1009     return true;
1010   }
1011   case Intrinsic::get_dynamic_area_offset: {
1012     MachineIRBuilder &MIB = Helper.MIRBuilder;
1013     MIB.buildConstant(MI.getOperand(0).getReg(), 0);
1014     MI.eraseFromParent();
1015     return true;
1016   }
1017   case Intrinsic::aarch64_mops_memset_tag: {
1018     assert(MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS);
1019     // Zext the value to 64 bit
1020     MachineIRBuilder MIB(MI);
1021     auto &Value = MI.getOperand(3);
1022     Register ZExtValueReg = MIB.buildAnyExt(LLT::scalar(64), Value).getReg(0);
1023     Value.setReg(ZExtValueReg);
1024     return true;
1025   }
1026   }
1027 
1028   return true;
1029 }
1030 
1031 bool AArch64LegalizerInfo::legalizeShlAshrLshr(
1032     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
1033     GISelChangeObserver &Observer) const {
1034   assert(MI.getOpcode() == TargetOpcode::G_ASHR ||
1035          MI.getOpcode() == TargetOpcode::G_LSHR ||
1036          MI.getOpcode() == TargetOpcode::G_SHL);
1037   // If the shift amount is a G_CONSTANT, promote it to a 64 bit type so the
1038   // imported patterns can select it later. Either way, it will be legal.
1039   Register AmtReg = MI.getOperand(2).getReg();
1040   auto VRegAndVal = getIConstantVRegValWithLookThrough(AmtReg, MRI);
1041   if (!VRegAndVal)
1042     return true;
1043   // Check the shift amount is in range for an immediate form.
1044   int64_t Amount = VRegAndVal->Value.getSExtValue();
1045   if (Amount > 31)
1046     return true; // This will have to remain a register variant.
1047   auto ExtCst = MIRBuilder.buildConstant(LLT::scalar(64), Amount);
1048   Observer.changingInstr(MI);
1049   MI.getOperand(2).setReg(ExtCst.getReg(0));
1050   Observer.changedInstr(MI);
1051   return true;
1052 }
1053 
1054 static void matchLDPSTPAddrMode(Register Root, Register &Base, int &Offset,
1055                                 MachineRegisterInfo &MRI) {
1056   Base = Root;
1057   Offset = 0;
1058 
1059   Register NewBase;
1060   int64_t NewOffset;
1061   if (mi_match(Root, MRI, m_GPtrAdd(m_Reg(NewBase), m_ICst(NewOffset))) &&
1062       isShiftedInt<7, 3>(NewOffset)) {
1063     Base = NewBase;
1064     Offset = NewOffset;
1065   }
1066 }
1067 
1068 // FIXME: This should be removed and replaced with the generic bitcast legalize
1069 // action.
1070 bool AArch64LegalizerInfo::legalizeLoadStore(
1071     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
1072     GISelChangeObserver &Observer) const {
1073   assert(MI.getOpcode() == TargetOpcode::G_STORE ||
1074          MI.getOpcode() == TargetOpcode::G_LOAD);
1075   // Here we just try to handle vector loads/stores where our value type might
1076   // have pointer elements, which the SelectionDAG importer can't handle. To
1077   // allow the existing patterns for s64 to fire for p0, we just try to bitcast
1078   // the value to use s64 types.
1079 
1080   // Custom legalization requires the instruction, if not deleted, must be fully
1081   // legalized. In order to allow further legalization of the inst, we create
1082   // a new instruction and erase the existing one.
1083 
1084   Register ValReg = MI.getOperand(0).getReg();
1085   const LLT ValTy = MRI.getType(ValReg);
1086 
1087   if (ValTy == LLT::scalar(128)) {
1088     assert((*MI.memoperands_begin())->getSuccessOrdering() ==
1089                AtomicOrdering::Monotonic ||
1090            (*MI.memoperands_begin())->getSuccessOrdering() ==
1091                AtomicOrdering::Unordered);
1092     assert(ST->hasLSE2() && "ldp/stp not single copy atomic without +lse2");
1093     LLT s64 = LLT::scalar(64);
1094     MachineInstrBuilder NewI;
1095     if (MI.getOpcode() == TargetOpcode::G_LOAD) {
1096       NewI = MIRBuilder.buildInstr(AArch64::LDPXi, {s64, s64}, {});
1097       MIRBuilder.buildMerge(ValReg, {NewI->getOperand(0), NewI->getOperand(1)});
1098     } else {
1099       auto Split = MIRBuilder.buildUnmerge(s64, MI.getOperand(0));
1100       NewI = MIRBuilder.buildInstr(
1101           AArch64::STPXi, {}, {Split->getOperand(0), Split->getOperand(1)});
1102     }
1103     Register Base;
1104     int Offset;
1105     matchLDPSTPAddrMode(MI.getOperand(1).getReg(), Base, Offset, MRI);
1106     NewI.addUse(Base);
1107     NewI.addImm(Offset / 8);
1108 
1109     NewI.cloneMemRefs(MI);
1110     constrainSelectedInstRegOperands(*NewI, *ST->getInstrInfo(),
1111                                      *MRI.getTargetRegisterInfo(),
1112                                      *ST->getRegBankInfo());
1113     MI.eraseFromParent();
1114     return true;
1115   }
1116 
1117   if (!ValTy.isVector() || !ValTy.getElementType().isPointer() ||
1118       ValTy.getElementType().getAddressSpace() != 0) {
1119     LLVM_DEBUG(dbgs() << "Tried to do custom legalization on wrong load/store");
1120     return false;
1121   }
1122 
1123   unsigned PtrSize = ValTy.getElementType().getSizeInBits();
1124   const LLT NewTy = LLT::vector(ValTy.getElementCount(), PtrSize);
1125   auto &MMO = **MI.memoperands_begin();
1126   MMO.setType(NewTy);
1127 
1128   if (MI.getOpcode() == TargetOpcode::G_STORE) {
1129     auto Bitcast = MIRBuilder.buildBitcast(NewTy, ValReg);
1130     MIRBuilder.buildStore(Bitcast.getReg(0), MI.getOperand(1), MMO);
1131   } else {
1132     auto NewLoad = MIRBuilder.buildLoad(NewTy, MI.getOperand(1), MMO);
1133     MIRBuilder.buildBitcast(ValReg, NewLoad);
1134   }
1135   MI.eraseFromParent();
1136   return true;
1137 }
1138 
1139 bool AArch64LegalizerInfo::legalizeVaArg(MachineInstr &MI,
1140                                          MachineRegisterInfo &MRI,
1141                                          MachineIRBuilder &MIRBuilder) const {
1142   MachineFunction &MF = MIRBuilder.getMF();
1143   Align Alignment(MI.getOperand(2).getImm());
1144   Register Dst = MI.getOperand(0).getReg();
1145   Register ListPtr = MI.getOperand(1).getReg();
1146 
1147   LLT PtrTy = MRI.getType(ListPtr);
1148   LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits());
1149 
1150   const unsigned PtrSize = PtrTy.getSizeInBits() / 8;
1151   const Align PtrAlign = Align(PtrSize);
1152   auto List = MIRBuilder.buildLoad(
1153       PtrTy, ListPtr,
1154       *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
1155                                PtrTy, PtrAlign));
1156 
1157   MachineInstrBuilder DstPtr;
1158   if (Alignment > PtrAlign) {
1159     // Realign the list to the actual required alignment.
1160     auto AlignMinus1 =
1161         MIRBuilder.buildConstant(IntPtrTy, Alignment.value() - 1);
1162     auto ListTmp = MIRBuilder.buildPtrAdd(PtrTy, List, AlignMinus1.getReg(0));
1163     DstPtr = MIRBuilder.buildMaskLowPtrBits(PtrTy, ListTmp, Log2(Alignment));
1164   } else
1165     DstPtr = List;
1166 
1167   LLT ValTy = MRI.getType(Dst);
1168   uint64_t ValSize = ValTy.getSizeInBits() / 8;
1169   MIRBuilder.buildLoad(
1170       Dst, DstPtr,
1171       *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
1172                                ValTy, std::max(Alignment, PtrAlign)));
1173 
1174   auto Size = MIRBuilder.buildConstant(IntPtrTy, alignTo(ValSize, PtrAlign));
1175 
1176   auto NewList = MIRBuilder.buildPtrAdd(PtrTy, DstPtr, Size.getReg(0));
1177 
1178   MIRBuilder.buildStore(NewList, ListPtr,
1179                         *MF.getMachineMemOperand(MachinePointerInfo(),
1180                                                  MachineMemOperand::MOStore,
1181                                                  PtrTy, PtrAlign));
1182 
1183   MI.eraseFromParent();
1184   return true;
1185 }
1186 
1187 bool AArch64LegalizerInfo::legalizeBitfieldExtract(
1188     MachineInstr &MI, MachineRegisterInfo &MRI, LegalizerHelper &Helper) const {
1189   // Only legal if we can select immediate forms.
1190   // TODO: Lower this otherwise.
1191   return getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI) &&
1192          getIConstantVRegValWithLookThrough(MI.getOperand(3).getReg(), MRI);
1193 }
1194 
1195 bool AArch64LegalizerInfo::legalizeCTPOP(MachineInstr &MI,
1196                                          MachineRegisterInfo &MRI,
1197                                          LegalizerHelper &Helper) const {
1198   // While there is no integer popcount instruction, it can
1199   // be more efficiently lowered to the following sequence that uses
1200   // AdvSIMD registers/instructions as long as the copies to/from
1201   // the AdvSIMD registers are cheap.
1202   //  FMOV    D0, X0        // copy 64-bit int to vector, high bits zero'd
1203   //  CNT     V0.8B, V0.8B  // 8xbyte pop-counts
1204   //  ADDV    B0, V0.8B     // sum 8xbyte pop-counts
1205   //  UMOV    X0, V0.B[0]   // copy byte result back to integer reg
1206   //
1207   // For 128 bit vector popcounts, we lower to the following sequence:
1208   //  cnt.16b   v0, v0  // v8s16, v4s32, v2s64
1209   //  uaddlp.8h v0, v0  // v8s16, v4s32, v2s64
1210   //  uaddlp.4s v0, v0  //        v4s32, v2s64
1211   //  uaddlp.2d v0, v0  //               v2s64
1212   //
1213   // For 64 bit vector popcounts, we lower to the following sequence:
1214   //  cnt.8b    v0, v0  // v4s16, v2s32
1215   //  uaddlp.4h v0, v0  // v4s16, v2s32
1216   //  uaddlp.2s v0, v0  //        v2s32
1217 
1218   if (!ST->hasNEON() ||
1219       MI.getMF()->getFunction().hasFnAttribute(Attribute::NoImplicitFloat))
1220     return false;
1221   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
1222   Register Dst = MI.getOperand(0).getReg();
1223   Register Val = MI.getOperand(1).getReg();
1224   LLT Ty = MRI.getType(Val);
1225 
1226   assert(Ty == MRI.getType(Dst) &&
1227          "Expected src and dst to have the same type!");
1228   unsigned Size = Ty.getSizeInBits();
1229 
1230   // Pre-conditioning: widen Val up to the nearest vector type.
1231   // s32,s64,v4s16,v2s32 -> v8i8
1232   // v8s16,v4s32,v2s64 -> v16i8
1233   LLT VTy = Size == 128 ? LLT::fixed_vector(16, 8) : LLT::fixed_vector(8, 8);
1234   if (Ty.isScalar()) {
1235     assert((Size == 32 || Size == 64 || Size == 128) && "Expected only 32, 64, or 128 bit scalars!");
1236     if (Size == 32) {
1237       Val = MIRBuilder.buildZExt(LLT::scalar(64), Val).getReg(0);
1238     }
1239   }
1240   Val = MIRBuilder.buildBitcast(VTy, Val).getReg(0);
1241 
1242   // Count bits in each byte-sized lane.
1243   auto CTPOP = MIRBuilder.buildCTPOP(VTy, Val);
1244 
1245   // Sum across lanes.
1246   Register HSum = CTPOP.getReg(0);
1247   unsigned Opc;
1248   SmallVector<LLT> HAddTys;
1249   if (Ty.isScalar()) {
1250     Opc = Intrinsic::aarch64_neon_uaddlv;
1251     HAddTys.push_back(LLT::scalar(32));
1252   } else if (Ty == LLT::fixed_vector(8, 16)) {
1253     Opc = Intrinsic::aarch64_neon_uaddlp;
1254     HAddTys.push_back(LLT::fixed_vector(8, 16));
1255   } else if (Ty == LLT::fixed_vector(4, 32)) {
1256     Opc = Intrinsic::aarch64_neon_uaddlp;
1257     HAddTys.push_back(LLT::fixed_vector(8, 16));
1258     HAddTys.push_back(LLT::fixed_vector(4, 32));
1259   } else if (Ty == LLT::fixed_vector(2, 64)) {
1260     Opc = Intrinsic::aarch64_neon_uaddlp;
1261     HAddTys.push_back(LLT::fixed_vector(8, 16));
1262     HAddTys.push_back(LLT::fixed_vector(4, 32));
1263     HAddTys.push_back(LLT::fixed_vector(2, 64));
1264   } else if (Ty == LLT::fixed_vector(4, 16)) {
1265     Opc = Intrinsic::aarch64_neon_uaddlp;
1266     HAddTys.push_back(LLT::fixed_vector(4, 16));
1267   } else if (Ty == LLT::fixed_vector(2, 32)) {
1268     Opc = Intrinsic::aarch64_neon_uaddlp;
1269     HAddTys.push_back(LLT::fixed_vector(4, 16));
1270     HAddTys.push_back(LLT::fixed_vector(2, 32));
1271   } else
1272     llvm_unreachable("unexpected vector shape");
1273   MachineInstrBuilder UADD;
1274   for (LLT HTy : HAddTys) {
1275     UADD = MIRBuilder.buildIntrinsic(Opc, {HTy}, /*HasSideEffects =*/false)
1276                      .addUse(HSum);
1277     HSum = UADD.getReg(0);
1278   }
1279 
1280   // Post-conditioning.
1281   if (Ty.isScalar() && (Size == 64 || Size == 128))
1282     MIRBuilder.buildZExt(Dst, UADD);
1283   else
1284     UADD->getOperand(0).setReg(Dst);
1285   MI.eraseFromParent();
1286   return true;
1287 }
1288 
1289 bool AArch64LegalizerInfo::legalizeAtomicCmpxchg128(
1290     MachineInstr &MI, MachineRegisterInfo &MRI, LegalizerHelper &Helper) const {
1291   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
1292   LLT s64 = LLT::scalar(64);
1293   auto Addr = MI.getOperand(1).getReg();
1294   auto DesiredI = MIRBuilder.buildUnmerge({s64, s64}, MI.getOperand(2));
1295   auto NewI = MIRBuilder.buildUnmerge({s64, s64}, MI.getOperand(3));
1296   auto DstLo = MRI.createGenericVirtualRegister(s64);
1297   auto DstHi = MRI.createGenericVirtualRegister(s64);
1298 
1299   MachineInstrBuilder CAS;
1300   if (ST->hasLSE()) {
1301     // We have 128-bit CASP instructions taking XSeqPair registers, which are
1302     // s128. We need the merge/unmerge to bracket the expansion and pair up with
1303     // the rest of the MIR so we must reassemble the extracted registers into a
1304     // 128-bit known-regclass one with code like this:
1305     //
1306     //     %in1 = REG_SEQUENCE Lo, Hi    ; One for each input
1307     //     %out = CASP %in1, ...
1308     //     %OldLo = G_EXTRACT %out, 0
1309     //     %OldHi = G_EXTRACT %out, 64
1310     auto Ordering = (*MI.memoperands_begin())->getMergedOrdering();
1311     unsigned Opcode;
1312     switch (Ordering) {
1313     case AtomicOrdering::Acquire:
1314       Opcode = AArch64::CASPAX;
1315       break;
1316     case AtomicOrdering::Release:
1317       Opcode = AArch64::CASPLX;
1318       break;
1319     case AtomicOrdering::AcquireRelease:
1320     case AtomicOrdering::SequentiallyConsistent:
1321       Opcode = AArch64::CASPALX;
1322       break;
1323     default:
1324       Opcode = AArch64::CASPX;
1325       break;
1326     }
1327 
1328     LLT s128 = LLT::scalar(128);
1329     auto CASDst = MRI.createGenericVirtualRegister(s128);
1330     auto CASDesired = MRI.createGenericVirtualRegister(s128);
1331     auto CASNew = MRI.createGenericVirtualRegister(s128);
1332     MIRBuilder.buildInstr(TargetOpcode::REG_SEQUENCE, {CASDesired}, {})
1333         .addUse(DesiredI->getOperand(0).getReg())
1334         .addImm(AArch64::sube64)
1335         .addUse(DesiredI->getOperand(1).getReg())
1336         .addImm(AArch64::subo64);
1337     MIRBuilder.buildInstr(TargetOpcode::REG_SEQUENCE, {CASNew}, {})
1338         .addUse(NewI->getOperand(0).getReg())
1339         .addImm(AArch64::sube64)
1340         .addUse(NewI->getOperand(1).getReg())
1341         .addImm(AArch64::subo64);
1342 
1343     CAS = MIRBuilder.buildInstr(Opcode, {CASDst}, {CASDesired, CASNew, Addr});
1344 
1345     MIRBuilder.buildExtract({DstLo}, {CASDst}, 0);
1346     MIRBuilder.buildExtract({DstHi}, {CASDst}, 64);
1347   } else {
1348     // The -O0 CMP_SWAP_128 is friendlier to generate code for because LDXP/STXP
1349     // can take arbitrary registers so it just has the normal GPR64 operands the
1350     // rest of AArch64 is expecting.
1351     auto Ordering = (*MI.memoperands_begin())->getMergedOrdering();
1352     unsigned Opcode;
1353     switch (Ordering) {
1354     case AtomicOrdering::Acquire:
1355       Opcode = AArch64::CMP_SWAP_128_ACQUIRE;
1356       break;
1357     case AtomicOrdering::Release:
1358       Opcode = AArch64::CMP_SWAP_128_RELEASE;
1359       break;
1360     case AtomicOrdering::AcquireRelease:
1361     case AtomicOrdering::SequentiallyConsistent:
1362       Opcode = AArch64::CMP_SWAP_128;
1363       break;
1364     default:
1365       Opcode = AArch64::CMP_SWAP_128_MONOTONIC;
1366       break;
1367     }
1368 
1369     auto Scratch = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
1370     CAS = MIRBuilder.buildInstr(Opcode, {DstLo, DstHi, Scratch},
1371                                 {Addr, DesiredI->getOperand(0),
1372                                  DesiredI->getOperand(1), NewI->getOperand(0),
1373                                  NewI->getOperand(1)});
1374   }
1375 
1376   CAS.cloneMemRefs(MI);
1377   constrainSelectedInstRegOperands(*CAS, *ST->getInstrInfo(),
1378                                    *MRI.getTargetRegisterInfo(),
1379                                    *ST->getRegBankInfo());
1380 
1381   MIRBuilder.buildMerge(MI.getOperand(0), {DstLo, DstHi});
1382   MI.eraseFromParent();
1383   return true;
1384 }
1385 
1386 bool AArch64LegalizerInfo::legalizeCTTZ(MachineInstr &MI,
1387                                         LegalizerHelper &Helper) const {
1388   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
1389   MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
1390   LLT Ty = MRI.getType(MI.getOperand(1).getReg());
1391   auto BitReverse = MIRBuilder.buildBitReverse(Ty, MI.getOperand(1));
1392   MIRBuilder.buildCTLZ(MI.getOperand(0).getReg(), BitReverse);
1393   MI.eraseFromParent();
1394   return true;
1395 }
1396 
1397 bool AArch64LegalizerInfo::legalizeMemOps(MachineInstr &MI,
1398                                           LegalizerHelper &Helper) const {
1399   MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
1400 
1401   // Tagged version MOPSMemorySetTagged is legalised in legalizeIntrinsic
1402   if (MI.getOpcode() == TargetOpcode::G_MEMSET) {
1403     // Zext the value operand to 64 bit
1404     auto &Value = MI.getOperand(1);
1405     Register ZExtValueReg =
1406         MIRBuilder.buildAnyExt(LLT::scalar(64), Value).getReg(0);
1407     Value.setReg(ZExtValueReg);
1408     return true;
1409   }
1410 
1411   return false;
1412 }
1413