lib/Target/AArch64/AArch64LegalizerInfo.cpp

   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 "AArch64Subtarget.h"
  16 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
  17 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
  18 #include "llvm/CodeGen/GlobalISel/Utils.h"
  19 #include "llvm/CodeGen/MachineInstr.h"
  20 #include "llvm/CodeGen/MachineRegisterInfo.h"
  21 #include "llvm/CodeGen/TargetOpcodes.h"
  22 #include "llvm/CodeGen/ValueTypes.h"
  23 #include "llvm/IR/DerivedTypes.h"
  24 #include "llvm/IR/Type.h"
  25
  26 #define DEBUG_TYPE "aarch64-legalinfo"
  27
  28 using namespace llvm;
  29 using namespace LegalizeActions;
  30 using namespace LegalizeMutations;
  31 using namespace LegalityPredicates;
  32
  33 AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST) {
  34   using namespace TargetOpcode;
  35   const LLT p0 = LLT::pointer(0, 64);
  36   const LLT s1 = LLT::scalar(1);
  37   const LLT s8 = LLT::scalar(8);
  38   const LLT s16 = LLT::scalar(16);
  39   const LLT s32 = LLT::scalar(32);
  40   const LLT s64 = LLT::scalar(64);
  41   const LLT s128 = LLT::scalar(128);
  42   const LLT s256 = LLT::scalar(256);
  43   const LLT s512 = LLT::scalar(512);
  44   const LLT v16s8 = LLT::vector(16, 8);
  45   const LLT v8s8 = LLT::vector(8, 8);
  46   const LLT v4s8 = LLT::vector(4, 8);
  47   const LLT v8s16 = LLT::vector(8, 16);
  48   const LLT v4s16 = LLT::vector(4, 16);
  49   const LLT v2s16 = LLT::vector(2, 16);
  50   const LLT v2s32 = LLT::vector(2, 32);
  51   const LLT v4s32 = LLT::vector(4, 32);
  52   const LLT v2s64 = LLT::vector(2, 64);
  53   const LLT v2p0 = LLT::vector(2, p0);
  54
  55   // FIXME: support subtargets which have neon/fp-armv8 disabled.
  56   if (!ST.hasNEON() || !ST.hasFPARMv8()) {
  57     computeTables();
  58     return;
  59   }
  60
  61   getActionDefinitionsBuilder(G_IMPLICIT_DEF)
  62     .legalFor({p0, s1, s8, s16, s32, s64, v4s32, v2s64})
  63     .clampScalar(0, s1, s64)
  64     .widenScalarToNextPow2(0, 8)
  65     .fewerElementsIf(
  66       [=](const LegalityQuery &Query) {
  67         return Query.Types[0].isVector() &&
  68           (Query.Types[0].getElementType() != s64 ||
  69            Query.Types[0].getNumElements() != 2);
  70       },
  71       [=](const LegalityQuery &Query) {
  72         LLT EltTy = Query.Types[0].getElementType();
  73         if (EltTy == s64)
  74           return std::make_pair(0, LLT::vector(2, 64));
  75         return std::make_pair(0, EltTy);
  76       });
  77
  78   getActionDefinitionsBuilder(G_PHI)
  79       .legalFor({p0, s16, s32, s64, v2s32, v4s32, v2s64})
  80       .clampScalar(0, s16, s64)
  81       .widenScalarToNextPow2(0);
  82
  83   getActionDefinitionsBuilder(G_BSWAP)
  84       .legalFor({s32, s64, v4s32, v2s32, v2s64})
  85       .clampScalar(0, s16, s64)
  86       .widenScalarToNextPow2(0);
  87
  88   getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR})
  89       .legalFor({s32, s64, v2s32, v4s32, v2s64, v8s16, v16s8})
  90       .clampScalar(0, s32, s64)
  91       .widenScalarToNextPow2(0)
  92       .clampNumElements(0, v2s32, v4s32)
  93       .clampNumElements(0, v2s64, v2s64)
  94       .moreElementsToNextPow2(0);
  95
  96   getActionDefinitionsBuilder(G_SHL)
  97     .legalFor({{s32, s32}, {s64, s64},
  98                {v2s32, v2s32}, {v4s32, v4s32}, {v2s64, v2s64}})
  99     .clampScalar(1, s32, s64)
 100     .clampScalar(0, s32, s64)
 101     .widenScalarToNextPow2(0)
 102     .clampNumElements(0, v2s32, v4s32)
 103     .clampNumElements(0, v2s64, v2s64)
 104     .moreElementsToNextPow2(0)
 105     .minScalarSameAs(1, 0);
 106
 107   getActionDefinitionsBuilder(G_GEP)
 108       .legalFor({{p0, s64}})
 109       .clampScalar(1, s64, s64);
 110
 111   getActionDefinitionsBuilder(G_PTR_MASK).legalFor({p0});
 112
 113   getActionDefinitionsBuilder({G_SDIV, G_UDIV})
 114       .legalFor({s32, s64})
 115       .clampScalar(0, s32, s64)
 116       .widenScalarToNextPow2(0)
 117       .scalarize(0);
 118
 119   getActionDefinitionsBuilder({G_LSHR, G_ASHR})
 120       .customIf([=](const LegalityQuery &Query) {
 121         const auto &SrcTy = Query.Types[0];
 122         const auto &AmtTy = Query.Types[1];
 123         return !SrcTy.isVector() && SrcTy.getSizeInBits() == 32 &&
 124                AmtTy.getSizeInBits() == 32;
 125       })
 126       .legalFor(
 127           {{s32, s32}, {s32, s64}, {s64, s64}, {v2s32, v2s32}, {v4s32, v4s32}})
 128       .clampScalar(1, s32, s64)
 129       .clampScalar(0, s32, s64)
 130       .minScalarSameAs(1, 0);
 131
 132   getActionDefinitionsBuilder({G_SREM, G_UREM})
 133       .lowerFor({s1, s8, s16, s32, s64});
 134
 135   getActionDefinitionsBuilder({G_SMULO, G_UMULO})
 136       .lowerFor({{s64, s1}});
 137
 138   getActionDefinitionsBuilder({G_SMULH, G_UMULH}).legalFor({s32, s64});
 139
 140   getActionDefinitionsBuilder({G_UADDE, G_USUBE, G_SADDO, G_SSUBO, G_UADDO})
 141       .legalFor({{s32, s1}, {s64, s1}});
 142
 143   getActionDefinitionsBuilder({G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FNEG})
 144     .legalFor({s32, s64, v2s64, v4s32, v2s32});
 145
 146   getActionDefinitionsBuilder(G_FREM).libcallFor({s32, s64});
 147
 148   getActionDefinitionsBuilder({G_FCEIL, G_FABS, G_FSQRT, G_FFLOOR, G_FRINT,
 149                                G_FMA, G_INTRINSIC_TRUNC, G_INTRINSIC_ROUND,
 150                                G_FNEARBYINT})
 151       // If we don't have full FP16 support, then scalarize the elements of
 152       // vectors containing fp16 types.
 153       .fewerElementsIf(
 154           [=, &ST](const LegalityQuery &Query) {
 155             const auto &Ty = Query.Types[0];
 156             return Ty.isVector() && Ty.getElementType() == s16 &&
 157                    !ST.hasFullFP16();
 158           },
 159           [=](const LegalityQuery &Query) { return std::make_pair(0, s16); })
 160       // If we don't have full FP16 support, then widen s16 to s32 if we
 161       // encounter it.
 162       .widenScalarIf(
 163           [=, &ST](const LegalityQuery &Query) {
 164             return Query.Types[0] == s16 && !ST.hasFullFP16();
 165           },
 166           [=](const LegalityQuery &Query) { return std::make_pair(0, s32); })
 167       .legalFor({s16, s32, s64, v2s32, v4s32, v2s64, v2s16, v4s16, v8s16});
 168
 169   getActionDefinitionsBuilder(
 170       {G_FCOS, G_FSIN, G_FLOG10, G_FLOG, G_FLOG2, G_FEXP, G_FEXP2, G_FPOW})
 171       // We need a call for these, so we always need to scalarize.
 172       .scalarize(0)
 173       // Regardless of FP16 support, widen 16-bit elements to 32-bits.
 174       .minScalar(0, s32)
 175       .libcallFor({s32, s64, v2s32, v4s32, v2s64});
 176
 177   getActionDefinitionsBuilder(G_INSERT)
 178       .unsupportedIf([=](const LegalityQuery &Query) {
 179         return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits();
 180       })
 181       .legalIf([=](const LegalityQuery &Query) {
 182         const LLT &Ty0 = Query.Types[0];
 183         const LLT &Ty1 = Query.Types[1];
 184         if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0)
 185           return false;
 186         return isPowerOf2_32(Ty1.getSizeInBits()) &&
 187                (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8);
 188       })
 189       .clampScalar(0, s32, s64)
 190       .widenScalarToNextPow2(0)
 191       .maxScalarIf(typeInSet(0, {s32}), 1, s16)
 192       .maxScalarIf(typeInSet(0, {s64}), 1, s32)
 193       .widenScalarToNextPow2(1);
 194
 195   getActionDefinitionsBuilder(G_EXTRACT)
 196       .unsupportedIf([=](const LegalityQuery &Query) {
 197         return Query.Types[0].getSizeInBits() >= Query.Types[1].getSizeInBits();
 198       })
 199       .legalIf([=](const LegalityQuery &Query) {
 200         const LLT &Ty0 = Query.Types[0];
 201         const LLT &Ty1 = Query.Types[1];
 202         if (Ty1 != s32 && Ty1 != s64 && Ty1 != s128)
 203           return false;
 204         if (Ty1 == p0)
 205           return true;
 206         return isPowerOf2_32(Ty0.getSizeInBits()) &&
 207                (Ty0.getSizeInBits() == 1 || Ty0.getSizeInBits() >= 8);
 208       })
 209       .clampScalar(1, s32, s128)
 210       .widenScalarToNextPow2(1)
 211       .maxScalarIf(typeInSet(1, {s32}), 0, s16)
 212       .maxScalarIf(typeInSet(1, {s64}), 0, s32)
 213       .widenScalarToNextPow2(0);
 214
 215   getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
 216       .legalForTypesWithMemDesc({{s32, p0, 8, 8},
 217                                  {s32, p0, 16, 8},
 218                                  {s32, p0, 32, 8},
 219                                  {s64, p0, 8, 2},
 220                                  {s64, p0, 16, 2},
 221                                  {s64, p0, 32, 4},
 222                                  {s64, p0, 64, 8},
 223                                  {p0, p0, 64, 8},
 224                                  {v2s32, p0, 64, 8}})
 225       .clampScalar(0, s32, s64)
 226       .widenScalarToNextPow2(0)
 227       // TODO: We could support sum-of-pow2's but the lowering code doesn't know
 228       //       how to do that yet.
 229       .unsupportedIfMemSizeNotPow2()
 230       // Lower anything left over into G_*EXT and G_LOAD
 231       .lower();
 232
 233   auto IsPtrVecPred = [=](const LegalityQuery &Query) {
 234     const LLT &ValTy = Query.Types[0];
 235     if (!ValTy.isVector())
 236       return false;
 237     const LLT EltTy = ValTy.getElementType();
 238     return EltTy.isPointer() && EltTy.getAddressSpace() == 0;
 239   };
 240
 241   getActionDefinitionsBuilder(G_LOAD)
 242       .legalForTypesWithMemDesc({{s8, p0, 8, 8},
 243                                  {s16, p0, 16, 8},
 244                                  {s32, p0, 32, 8},
 245                                  {s64, p0, 64, 8},
 246                                  {p0, p0, 64, 8},
 247                                  {s128, p0, 128, 8},
 248                                  {v8s8, p0, 64, 8},
 249                                  {v16s8, p0, 128, 8},
 250                                  {v4s16, p0, 64, 8},
 251                                  {v8s16, p0, 128, 8},
 252                                  {v2s32, p0, 64, 8},
 253                                  {v4s32, p0, 128, 8},
 254                                  {v2s64, p0, 128, 8}})
 255       // These extends are also legal
 256       .legalForTypesWithMemDesc({{s32, p0, 8, 8},
 257                                  {s32, p0, 16, 8}})
 258       .clampScalar(0, s8, s64)
 259       .widenScalarToNextPow2(0)
 260       // TODO: We could support sum-of-pow2's but the lowering code doesn't know
 261       //       how to do that yet.
 262       .unsupportedIfMemSizeNotPow2()
 263       // Lower any any-extending loads left into G_ANYEXT and G_LOAD
 264       .lowerIf([=](const LegalityQuery &Query) {
 265         return Query.Types[0].getSizeInBits() != Query.MMODescrs[0].SizeInBits;
 266       })
 267       .clampMaxNumElements(0, s32, 2)
 268       .clampMaxNumElements(0, s64, 1)
 269       .customIf(IsPtrVecPred);
 270
 271   getActionDefinitionsBuilder(G_STORE)
 272       .legalForTypesWithMemDesc({{s8, p0, 8, 8},
 273                                  {s16, p0, 16, 8},
 274                                  {s32, p0, 32, 8},
 275                                  {s64, p0, 64, 8},
 276                                  {p0, p0, 64, 8},
 277                                  {s128, p0, 128, 8},
 278                                  {v16s8, p0, 128, 8},
 279                                  {v4s16, p0, 64, 8},
 280                                  {v8s16, p0, 128, 8},
 281                                  {v2s32, p0, 64, 8},
 282                                  {v4s32, p0, 128, 8},
 283                                  {v2s64, p0, 128, 8}})
 284       .clampScalar(0, s8, s64)
 285       .widenScalarToNextPow2(0)
 286       // TODO: We could support sum-of-pow2's but the lowering code doesn't know
 287       //       how to do that yet.
 288       .unsupportedIfMemSizeNotPow2()
 289       .lowerIf([=](const LegalityQuery &Query) {
 290         return Query.Types[0].isScalar() &&
 291                Query.Types[0].getSizeInBits() != Query.MMODescrs[0].SizeInBits;
 292       })
 293       .clampMaxNumElements(0, s32, 2)
 294       .clampMaxNumElements(0, s64, 1)
 295       .customIf(IsPtrVecPred);
 296
 297   // Constants
 298   getActionDefinitionsBuilder(G_CONSTANT)
 299     .legalFor({p0, s8, s16, s32, s64})
 300       .clampScalar(0, s8, s64)
 301       .widenScalarToNextPow2(0);
 302   getActionDefinitionsBuilder(G_FCONSTANT)
 303       .legalFor({s32, s64})
 304       .clampScalar(0, s32, s64);
 305
 306   getActionDefinitionsBuilder(G_ICMP)
 307       .legalFor({{s32, s32},
 308                  {s32, s64},
 309                  {s32, p0},
 310                  {v4s32, v4s32},
 311                  {v2s32, v2s32},
 312                  {v2s64, v2s64},
 313                  {v2s64, v2p0},
 314                  {v4s16, v4s16},
 315                  {v8s16, v8s16},
 316                  {v8s8, v8s8},
 317                  {v16s8, v16s8}})
 318       .clampScalar(1, s32, s64)
 319       .clampScalar(0, s32, s32)
 320       .minScalarEltSameAsIf(
 321           [=](const LegalityQuery &Query) {
 322             const LLT &Ty = Query.Types[0];
 323             const LLT &SrcTy = Query.Types[1];
 324             return Ty.isVector() && !SrcTy.getElementType().isPointer() &&
 325                    Ty.getElementType() != SrcTy.getElementType();
 326           },
 327           0, 1)
 328       .minScalarOrEltIf(
 329           [=](const LegalityQuery &Query) { return Query.Types[1] == v2s16; },
 330           1, s32)
 331       .minScalarOrEltIf(
 332           [=](const LegalityQuery &Query) { return Query.Types[1] == v2p0; }, 0,
 333           s64)
 334       .widenScalarOrEltToNextPow2(1);
 335
 336   getActionDefinitionsBuilder(G_FCMP)
 337       .legalFor({{s32, s32}, {s32, s64}})
 338       .clampScalar(0, s32, s32)
 339       .clampScalar(1, s32, s64)
 340       .widenScalarToNextPow2(1);
 341
 342   // Extensions
 343   auto ExtLegalFunc = [=](const LegalityQuery &Query) {
 344     unsigned DstSize = Query.Types[0].getSizeInBits();
 345
 346     if (DstSize == 128 && !Query.Types[0].isVector())
 347       return false; // Extending to a scalar s128 is not legal.
 348
 349     // Make sure that we have something that will fit in a register, and
 350     // make sure it's a power of 2.
 351     if (DstSize < 8 || DstSize > 128 || !isPowerOf2_32(DstSize))
 352       return false;
 353
 354     const LLT &SrcTy = Query.Types[1];
 355
 356     // Special case for s1.
 357     if (SrcTy == s1)
 358       return true;
 359
 360     // Make sure we fit in a register otherwise. Don't bother checking that
 361     // the source type is below 128 bits. We shouldn't be allowing anything
 362     // through which is wider than the destination in the first place.
 363     unsigned SrcSize = SrcTy.getSizeInBits();
 364     if (SrcSize < 8 || !isPowerOf2_32(SrcSize))
 365       return false;
 366
 367     return true;
 368   };
 369   getActionDefinitionsBuilder({G_ZEXT, G_ANYEXT}).legalIf(ExtLegalFunc);
 370   getActionDefinitionsBuilder(G_SEXT)
 371       .legalIf(ExtLegalFunc)
 372       .clampScalar(0, s64, s64); // Just for s128, others are handled above.
 373
 374   getActionDefinitionsBuilder(G_TRUNC).alwaysLegal();
 375
 376   // FP conversions
 377   getActionDefinitionsBuilder(G_FPTRUNC).legalFor(
 378       {{s16, s32}, {s16, s64}, {s32, s64}, {v4s16, v4s32}, {v2s32, v2s64}});
 379   getActionDefinitionsBuilder(G_FPEXT).legalFor(
 380       {{s32, s16}, {s64, s16}, {s64, s32}, {v4s32, v4s16}, {v2s64, v2s32}});
 381
 382   // Conversions
 383   getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
 384       .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})
 385       .clampScalar(0, s32, s64)
 386       .widenScalarToNextPow2(0)
 387       .clampScalar(1, s32, s64)
 388       .widenScalarToNextPow2(1);
 389
 390   getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
 391       .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})
 392       .clampScalar(1, s32, s64)
 393       .widenScalarToNextPow2(1)
 394       .clampScalar(0, s32, s64)
 395       .widenScalarToNextPow2(0);
 396
 397   // Control-flow
 398   getActionDefinitionsBuilder(G_BRCOND).legalFor({s1, s8, s16, s32});
 399   getActionDefinitionsBuilder(G_BRINDIRECT).legalFor({p0});
 400
 401   // Select
 402   // FIXME: We can probably do a bit better than just scalarizing vector
 403   // selects.
 404   getActionDefinitionsBuilder(G_SELECT)
 405       .legalFor({{s32, s1}, {s64, s1}, {p0, s1}})
 406       .clampScalar(0, s32, s64)
 407       .widenScalarToNextPow2(0)
 408       .scalarize(0);
 409
 410   // Pointer-handling
 411   getActionDefinitionsBuilder(G_FRAME_INDEX).legalFor({p0});
 412   getActionDefinitionsBuilder(G_GLOBAL_VALUE).legalFor({p0});
 413
 414   getActionDefinitionsBuilder(G_PTRTOINT)
 415       .legalForCartesianProduct({s1, s8, s16, s32, s64}, {p0})
 416       .maxScalar(0, s64)
 417       .widenScalarToNextPow2(0, /*Min*/ 8);
 418
 419   getActionDefinitionsBuilder(G_INTTOPTR)
 420       .unsupportedIf([&](const LegalityQuery &Query) {
 421         return Query.Types[0].getSizeInBits() != Query.Types[1].getSizeInBits();
 422       })
 423       .legalFor({{p0, s64}});
 424
 425   // Casts for 32 and 64-bit width type are just copies.
 426   // Same for 128-bit width type, except they are on the FPR bank.
 427   getActionDefinitionsBuilder(G_BITCAST)
 428       // FIXME: This is wrong since G_BITCAST is not allowed to change the
 429       // number of bits but it's what the previous code described and fixing
 430       // it breaks tests.
 431       .legalForCartesianProduct({s1, s8, s16, s32, s64, s128, v16s8, v8s8, v4s8,
 432                                  v8s16, v4s16, v2s16, v4s32, v2s32, v2s64,
 433                                  v2p0});
 434
 435   getActionDefinitionsBuilder(G_VASTART).legalFor({p0});
 436
 437   // va_list must be a pointer, but most sized types are pretty easy to handle
 438   // as the destination.
 439   getActionDefinitionsBuilder(G_VAARG)
 440       .customForCartesianProduct({s8, s16, s32, s64, p0}, {p0})
 441       .clampScalar(0, s8, s64)
 442       .widenScalarToNextPow2(0, /*Min*/ 8);
 443
 444   if (ST.hasLSE()) {
 445     getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG_WITH_SUCCESS)
 446         .lowerIf(all(
 447             typeInSet(0, {s8, s16, s32, s64}), typeIs(1, s1), typeIs(2, p0),
 448             atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Monotonic)));
 449
 450     getActionDefinitionsBuilder(
 451         {G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB, G_ATOMICRMW_AND,
 452          G_ATOMICRMW_OR, G_ATOMICRMW_XOR, G_ATOMICRMW_MIN, G_ATOMICRMW_MAX,
 453          G_ATOMICRMW_UMIN, G_ATOMICRMW_UMAX, G_ATOMIC_CMPXCHG})
 454         .legalIf(all(
 455             typeInSet(0, {s8, s16, s32, s64}), typeIs(1, p0),
 456             atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Monotonic)));
 457   }
 458
 459   getActionDefinitionsBuilder(G_BLOCK_ADDR).legalFor({p0});
 460
 461   // Merge/Unmerge
 462   for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
 463     unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
 464     unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;
 465
 466     auto notValidElt = [](const LegalityQuery &Query, unsigned TypeIdx) {
 467       const LLT &Ty = Query.Types[TypeIdx];
 468       if (Ty.isVector()) {
 469         const LLT &EltTy = Ty.getElementType();
 470         if (EltTy.getSizeInBits() < 8 || EltTy.getSizeInBits() > 64)
 471           return true;
 472         if (!isPowerOf2_32(EltTy.getSizeInBits()))
 473           return true;
 474       }
 475       return false;
 476     };
 477
 478     // FIXME: This rule is horrible, but specifies the same as what we had
 479     // before with the particularly strange definitions removed (e.g.
 480     // s8 = G_MERGE_VALUES s32, s32).
 481     // Part of the complexity comes from these ops being extremely flexible. For
 482     // example, you can build/decompose vectors with it, concatenate vectors,
 483     // etc. and in addition to this you can also bitcast with it at the same
 484     // time. We've been considering breaking it up into multiple ops to make it
 485     // more manageable throughout the backend.
 486     getActionDefinitionsBuilder(Op)
 487         // Break up vectors with weird elements into scalars
 488         .fewerElementsIf(
 489             [=](const LegalityQuery &Query) { return notValidElt(Query, 0); },
 490             scalarize(0))
 491         .fewerElementsIf(
 492             [=](const LegalityQuery &Query) { return notValidElt(Query, 1); },
 493             scalarize(1))
 494         // Clamp the big scalar to s8-s512 and make it either a power of 2, 192,
 495         // or 384.
 496         .clampScalar(BigTyIdx, s8, s512)
 497         .widenScalarIf(
 498             [=](const LegalityQuery &Query) {
 499               const LLT &Ty = Query.Types[BigTyIdx];
 500               return !isPowerOf2_32(Ty.getSizeInBits()) &&
 501                      Ty.getSizeInBits() % 64 != 0;
 502             },
 503             [=](const LegalityQuery &Query) {
 504               // Pick the next power of 2, or a multiple of 64 over 128.
 505               // Whichever is smaller.
 506               const LLT &Ty = Query.Types[BigTyIdx];
 507               unsigned NewSizeInBits = 1
 508                                        << Log2_32_Ceil(Ty.getSizeInBits() + 1);
 509               if (NewSizeInBits >= 256) {
 510                 unsigned RoundedTo = alignTo<64>(Ty.getSizeInBits() + 1);
 511                 if (RoundedTo < NewSizeInBits)
 512                   NewSizeInBits = RoundedTo;
 513               }
 514               return std::make_pair(BigTyIdx, LLT::scalar(NewSizeInBits));
 515             })
 516         // Clamp the little scalar to s8-s256 and make it a power of 2. It's not
 517         // worth considering the multiples of 64 since 2*192 and 2*384 are not
 518         // valid.
 519         .clampScalar(LitTyIdx, s8, s256)
 520         .widenScalarToNextPow2(LitTyIdx, /*Min*/ 8)
 521         // So at this point, we have s8, s16, s32, s64, s128, s192, s256, s384,
 522         // s512, <X x s8>, <X x s16>, <X x s32>, or <X x s64>.
 523         // At this point it's simple enough to accept the legal types.
 524         .legalIf([=](const LegalityQuery &Query) {
 525           const LLT &BigTy = Query.Types[BigTyIdx];
 526           const LLT &LitTy = Query.Types[LitTyIdx];
 527           if (BigTy.isVector() && BigTy.getSizeInBits() < 32)
 528             return false;
 529           if (LitTy.isVector() && LitTy.getSizeInBits() < 32)
 530             return false;
 531           return BigTy.getSizeInBits() % LitTy.getSizeInBits() == 0;
 532         })
 533         // Any vectors left are the wrong size. Scalarize them.
 534       .scalarize(0)
 535       .scalarize(1);
 536   }
 537
 538   getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
 539       .unsupportedIf([=](const LegalityQuery &Query) {
 540         const LLT &EltTy = Query.Types[1].getElementType();
 541         return Query.Types[0] != EltTy;
 542       })
 543       .minScalar(2, s64)
 544       .legalIf([=](const LegalityQuery &Query) {
 545         const LLT &VecTy = Query.Types[1];
 546         return VecTy == v2s16 || VecTy == v4s16 || VecTy == v8s16 ||
 547                VecTy == v4s32 || VecTy == v2s64 || VecTy == v2s32;
 548       });
 549
 550   getActionDefinitionsBuilder(G_INSERT_VECTOR_ELT)
 551       .legalIf([=](const LegalityQuery &Query) {
 552         const LLT &VecTy = Query.Types[0];
 553         // TODO: Support s8 and s16
 554         return VecTy == v2s32 || VecTy == v4s32 || VecTy == v2s64;
 555       });
 556
 557   getActionDefinitionsBuilder(G_BUILD_VECTOR)
 558       .legalFor({{v4s16, s16},
 559                  {v8s16, s16},
 560                  {v2s32, s32},
 561                  {v4s32, s32},
 562                  {v2p0, p0},
 563                  {v2s64, s64}})
 564       .clampNumElements(0, v4s32, v4s32)
 565       .clampNumElements(0, v2s64, v2s64)
 566
 567       // Deal with larger scalar types, which will be implicitly truncated.
 568       .legalIf([=](const LegalityQuery &Query) {
 569         return Query.Types[0].getScalarSizeInBits() <
 570                Query.Types[1].getSizeInBits();
 571       })
 572       .minScalarSameAs(1, 0);
 573
 574   getActionDefinitionsBuilder(G_CTLZ).legalForCartesianProduct(
 575       {s32, s64, v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
 576       .scalarize(1);
 577
 578   getActionDefinitionsBuilder(G_SHUFFLE_VECTOR)
 579       .legalIf([=](const LegalityQuery &Query) {
 580         const LLT &DstTy = Query.Types[0];
 581         const LLT &SrcTy = Query.Types[1];
 582         // For now just support the TBL2 variant which needs the source vectors
 583         // to be the same size as the dest.
 584         if (DstTy != SrcTy)
 585           return false;
 586         for (auto &Ty : {v2s32, v4s32, v2s64}) {
 587           if (DstTy == Ty)
 588             return true;
 589         }
 590         return false;
 591       })
 592       // G_SHUFFLE_VECTOR can have scalar sources (from 1 x s vectors), we
 593       // just want those lowered into G_BUILD_VECTOR
 594       .lowerIf([=](const LegalityQuery &Query) {
 595         return !Query.Types[1].isVector();
 596       })
 597       .clampNumElements(0, v4s32, v4s32)
 598       .clampNumElements(0, v2s64, v2s64);
 599
 600   getActionDefinitionsBuilder(G_CONCAT_VECTORS)
 601       .legalFor({{v4s32, v2s32}, {v8s16, v4s16}});
 602
 603   getActionDefinitionsBuilder(G_JUMP_TABLE)
 604     .legalFor({{p0}, {s64}});
 605
 606   getActionDefinitionsBuilder(G_BRJT).legalIf([=](const LegalityQuery &Query) {
 607     return Query.Types[0] == p0 && Query.Types[1] == s64;
 608   });
 609
 610   computeTables();
 611   verify(*ST.getInstrInfo());
 612 }
 613
 614 bool AArch64LegalizerInfo::legalizeCustom(MachineInstr &MI,
 615                                           MachineRegisterInfo &MRI,
 616                                           MachineIRBuilder &MIRBuilder,
 617                                           GISelChangeObserver &Observer) const {
 618   switch (MI.getOpcode()) {
 619   default:
 620     // No idea what to do.
 621     return false;
 622   case TargetOpcode::G_VAARG:
 623     return legalizeVaArg(MI, MRI, MIRBuilder);
 624   case TargetOpcode::G_LOAD:
 625   case TargetOpcode::G_STORE:
 626     return legalizeLoadStore(MI, MRI, MIRBuilder, Observer);
 627   case TargetOpcode::G_SHL:
 628   case TargetOpcode::G_ASHR:
 629   case TargetOpcode::G_LSHR:
 630     return legalizeShlAshrLshr(MI, MRI, MIRBuilder, Observer);
 631   }
 632
 633   llvm_unreachable("expected switch to return");
 634 }
 635
 636 bool AArch64LegalizerInfo::legalizeIntrinsic(
 637     MachineInstr &MI, MachineRegisterInfo &MRI,
 638     MachineIRBuilder &MIRBuilder) const {
 639   switch (MI.getIntrinsicID()) {
 640   case Intrinsic::memcpy:
 641   case Intrinsic::memset:
 642   case Intrinsic::memmove:
 643     if (createMemLibcall(MIRBuilder, MRI, MI) ==
 644         LegalizerHelper::UnableToLegalize)
 645       return false;
 646     MI.eraseFromParent();
 647     return true;
 648   default:
 649     break;
 650   }
 651   return true;
 652 }
 653
 654 bool AArch64LegalizerInfo::legalizeShlAshrLshr(
 655     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
 656     GISelChangeObserver &Observer) const {
 657   assert(MI.getOpcode() == TargetOpcode::G_ASHR ||
 658          MI.getOpcode() == TargetOpcode::G_LSHR ||
 659          MI.getOpcode() == TargetOpcode::G_SHL);
 660   // If the shift amount is a G_CONSTANT, promote it to a 64 bit type so the
 661   // imported patterns can select it later. Either way, it will be legal.
 662   Register AmtReg = MI.getOperand(2).getReg();
 663   auto *CstMI = MRI.getVRegDef(AmtReg);
 664   assert(CstMI && "expected to find a vreg def");
 665   if (CstMI->getOpcode() != TargetOpcode::G_CONSTANT)
 666     return true;
 667   // Check the shift amount is in range for an immediate form.
 668   unsigned Amount = CstMI->getOperand(1).getCImm()->getZExtValue();
 669   if (Amount > 31)
 670     return true; // This will have to remain a register variant.
 671   assert(MRI.getType(AmtReg).getSizeInBits() == 32);
 672   MIRBuilder.setInstr(MI);
 673   auto ExtCst = MIRBuilder.buildZExt(LLT::scalar(64), AmtReg);
 674   MI.getOperand(2).setReg(ExtCst.getReg(0));
 675   return true;
 676 }
 677
 678 bool AArch64LegalizerInfo::legalizeLoadStore(
 679     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
 680     GISelChangeObserver &Observer) const {
 681   assert(MI.getOpcode() == TargetOpcode::G_STORE ||
 682          MI.getOpcode() == TargetOpcode::G_LOAD);
 683   // Here we just try to handle vector loads/stores where our value type might
 684   // have pointer elements, which the SelectionDAG importer can't handle. To
 685   // allow the existing patterns for s64 to fire for p0, we just try to bitcast
 686   // the value to use s64 types.
 687
 688   // Custom legalization requires the instruction, if not deleted, must be fully
 689   // legalized. In order to allow further legalization of the inst, we create
 690   // a new instruction and erase the existing one.
 691
 692   unsigned ValReg = MI.getOperand(0).getReg();
 693   const LLT ValTy = MRI.getType(ValReg);
 694
 695   if (!ValTy.isVector() || !ValTy.getElementType().isPointer() ||
 696       ValTy.getElementType().getAddressSpace() != 0) {
 697     LLVM_DEBUG(dbgs() << "Tried to do custom legalization on wrong load/store");
 698     return false;
 699   }
 700
 701   MIRBuilder.setInstr(MI);
 702   unsigned PtrSize = ValTy.getElementType().getSizeInBits();
 703   const LLT NewTy = LLT::vector(ValTy.getNumElements(), PtrSize);
 704   auto &MMO = **MI.memoperands_begin();
 705   if (MI.getOpcode() == TargetOpcode::G_STORE) {
 706     auto Bitcast = MIRBuilder.buildBitcast({NewTy}, {ValReg});
 707     MIRBuilder.buildStore(Bitcast.getReg(0), MI.getOperand(1).getReg(), MMO);
 708   } else {
 709     unsigned NewReg = MRI.createGenericVirtualRegister(NewTy);
 710     auto NewLoad = MIRBuilder.buildLoad(NewReg, MI.getOperand(1).getReg(), MMO);
 711     MIRBuilder.buildBitcast({ValReg}, {NewLoad});
 712   }
 713   MI.eraseFromParent();
 714   return true;
 715 }
 716
 717 bool AArch64LegalizerInfo::legalizeVaArg(MachineInstr &MI,
 718                                          MachineRegisterInfo &MRI,
 719                                          MachineIRBuilder &MIRBuilder) const {
 720   MIRBuilder.setInstr(MI);
 721   MachineFunction &MF = MIRBuilder.getMF();
 722   unsigned Align = MI.getOperand(2).getImm();
 723   Register Dst = MI.getOperand(0).getReg();
 724   Register ListPtr = MI.getOperand(1).getReg();
 725
 726   LLT PtrTy = MRI.getType(ListPtr);
 727   LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits());
 728
 729   const unsigned PtrSize = PtrTy.getSizeInBits() / 8;
 730   Register List = MRI.createGenericVirtualRegister(PtrTy);
 731   MIRBuilder.buildLoad(
 732       List, ListPtr,
 733       *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
 734                                PtrSize, /* Align = */ PtrSize));
 735
 736   Register DstPtr;
 737   if (Align > PtrSize) {
 738     // Realign the list to the actual required alignment.
 739     auto AlignMinus1 = MIRBuilder.buildConstant(IntPtrTy, Align - 1);
 740
 741     auto ListTmp = MIRBuilder.buildGEP(PtrTy, List, AlignMinus1.getReg(0));
 742
 743     DstPtr = MRI.createGenericVirtualRegister(PtrTy);
 744     MIRBuilder.buildPtrMask(DstPtr, ListTmp, Log2_64(Align));
 745   } else
 746     DstPtr = List;
 747
 748   uint64_t ValSize = MRI.getType(Dst).getSizeInBits() / 8;
 749   MIRBuilder.buildLoad(
 750       Dst, DstPtr,
 751       *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
 752                                ValSize, std::max(Align, PtrSize)));
 753
 754   auto Size = MIRBuilder.buildConstant(IntPtrTy, alignTo(ValSize, PtrSize));
 755
 756   auto NewList = MIRBuilder.buildGEP(PtrTy, DstPtr, Size.getReg(0));
 757
 758   MIRBuilder.buildStore(
 759       NewList, ListPtr,
 760       *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOStore,
 761                                PtrSize, /* Align = */ PtrSize));
 762
 763   MI.eraseFromParent();
 764   return true;
 765 }