| 1 | //=- WebAssemblyISelLowering.cpp - WebAssembly DAG Lowering Implementation -==// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | /// |
| 9 | /// \file |
| 10 | /// This file implements the WebAssemblyTargetLowering class. |
| 11 | /// |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "WebAssemblyISelLowering.h" |
| 15 | #include "MCTargetDesc/WebAssemblyMCTargetDesc.h" |
| 16 | #include "Utils/WebAssemblyTypeUtilities.h" |
| 17 | #include "WebAssemblyMachineFunctionInfo.h" |
| 18 | #include "WebAssemblySubtarget.h" |
| 19 | #include "WebAssemblyTargetMachine.h" |
| 20 | #include "WebAssemblyUtilities.h" |
| 21 | #include "llvm/CodeGen/CallingConvLower.h" |
| 22 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 23 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 24 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
| 25 | #include "llvm/CodeGen/MachineModuleInfo.h" |
| 26 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 27 | #include "llvm/CodeGen/SDPatternMatch.h" |
| 28 | #include "llvm/CodeGen/SelectionDAG.h" |
| 29 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 30 | #include "llvm/IR/DiagnosticInfo.h" |
| 31 | #include "llvm/IR/DiagnosticPrinter.h" |
| 32 | #include "llvm/IR/Function.h" |
| 33 | #include "llvm/IR/IntrinsicInst.h" |
| 34 | #include "llvm/IR/Intrinsics.h" |
| 35 | #include "llvm/IR/IntrinsicsWebAssembly.h" |
| 36 | #include "llvm/Support/ErrorHandling.h" |
| 37 | #include "llvm/Support/KnownBits.h" |
| 38 | #include "llvm/Support/MathExtras.h" |
| 39 | #include "llvm/Target/TargetOptions.h" |
| 40 | using namespace llvm; |
| 41 | |
| 42 | #define DEBUG_TYPE "wasm-lower" |
| 43 | |
| 44 | WebAssemblyTargetLowering::WebAssemblyTargetLowering( |
| 45 | const TargetMachine &TM, const WebAssemblySubtarget &STI) |
| 46 | : TargetLowering(TM, STI), Subtarget(&STI) { |
| 47 | auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32; |
| 48 | |
| 49 | // Set the load count for memcmp expand optimization |
| 50 | MaxLoadsPerMemcmp = 8; |
| 51 | MaxLoadsPerMemcmpOptSize = 4; |
| 52 | |
| 53 | // Booleans always contain 0 or 1. |
| 54 | setBooleanContents(ZeroOrOneBooleanContent); |
| 55 | // Except in SIMD vectors |
| 56 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
| 57 | // We don't know the microarchitecture here, so just reduce register pressure. |
| 58 | setSchedulingPreference(Sched::RegPressure); |
| 59 | // Tell ISel that we have a stack pointer. |
| 60 | setStackPointerRegisterToSaveRestore( |
| 61 | Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32); |
| 62 | // Set up the register classes. |
| 63 | addRegisterClass(VT: MVT::i32, RC: &WebAssembly::I32RegClass); |
| 64 | addRegisterClass(VT: MVT::i64, RC: &WebAssembly::I64RegClass); |
| 65 | addRegisterClass(VT: MVT::f32, RC: &WebAssembly::F32RegClass); |
| 66 | addRegisterClass(VT: MVT::f64, RC: &WebAssembly::F64RegClass); |
| 67 | if (Subtarget->hasSIMD128()) { |
| 68 | addRegisterClass(VT: MVT::v16i8, RC: &WebAssembly::V128RegClass); |
| 69 | addRegisterClass(VT: MVT::v8i16, RC: &WebAssembly::V128RegClass); |
| 70 | addRegisterClass(VT: MVT::v4i32, RC: &WebAssembly::V128RegClass); |
| 71 | addRegisterClass(VT: MVT::v4f32, RC: &WebAssembly::V128RegClass); |
| 72 | addRegisterClass(VT: MVT::v2i64, RC: &WebAssembly::V128RegClass); |
| 73 | addRegisterClass(VT: MVT::v2f64, RC: &WebAssembly::V128RegClass); |
| 74 | } |
| 75 | if (Subtarget->hasFP16()) { |
| 76 | addRegisterClass(VT: MVT::v8f16, RC: &WebAssembly::V128RegClass); |
| 77 | } |
| 78 | if (Subtarget->hasReferenceTypes()) { |
| 79 | addRegisterClass(VT: MVT::externref, RC: &WebAssembly::EXTERNREFRegClass); |
| 80 | addRegisterClass(VT: MVT::funcref, RC: &WebAssembly::FUNCREFRegClass); |
| 81 | if (Subtarget->hasExceptionHandling()) { |
| 82 | addRegisterClass(VT: MVT::exnref, RC: &WebAssembly::EXNREFRegClass); |
| 83 | } |
| 84 | } |
| 85 | // Compute derived properties from the register classes. |
| 86 | computeRegisterProperties(TRI: Subtarget->getRegisterInfo()); |
| 87 | |
| 88 | // Transform loads and stores to pointers in address space 1 to loads and |
| 89 | // stores to WebAssembly global variables, outside linear memory. |
| 90 | for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) { |
| 91 | setOperationAction(Op: ISD::LOAD, VT: T, Action: Custom); |
| 92 | setOperationAction(Op: ISD::STORE, VT: T, Action: Custom); |
| 93 | } |
| 94 | if (Subtarget->hasSIMD128()) { |
| 95 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 96 | MVT::v2f64}) { |
| 97 | setOperationAction(Op: ISD::LOAD, VT: T, Action: Custom); |
| 98 | setOperationAction(Op: ISD::STORE, VT: T, Action: Custom); |
| 99 | } |
| 100 | } |
| 101 | if (Subtarget->hasFP16()) { |
| 102 | setOperationAction(Op: ISD::LOAD, VT: MVT::v8f16, Action: Custom); |
| 103 | setOperationAction(Op: ISD::STORE, VT: MVT::v8f16, Action: Custom); |
| 104 | } |
| 105 | if (Subtarget->hasReferenceTypes()) { |
| 106 | // We need custom load and store lowering for both externref, funcref and |
| 107 | // Other. The MVT::Other here represents tables of reference types. |
| 108 | for (auto T : {MVT::externref, MVT::funcref, MVT::Other}) { |
| 109 | setOperationAction(Op: ISD::LOAD, VT: T, Action: Custom); |
| 110 | setOperationAction(Op: ISD::STORE, VT: T, Action: Custom); |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | setOperationAction(Op: ISD::GlobalAddress, VT: MVTPtr, Action: Custom); |
| 115 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: MVTPtr, Action: Custom); |
| 116 | setOperationAction(Op: ISD::ExternalSymbol, VT: MVTPtr, Action: Custom); |
| 117 | setOperationAction(Op: ISD::JumpTable, VT: MVTPtr, Action: Custom); |
| 118 | setOperationAction(Op: ISD::BlockAddress, VT: MVTPtr, Action: Custom); |
| 119 | setOperationAction(Op: ISD::BRIND, VT: MVT::Other, Action: Custom); |
| 120 | setOperationAction(Op: ISD::CLEAR_CACHE, VT: MVT::Other, Action: Custom); |
| 121 | |
| 122 | // Take the default expansion for va_arg, va_copy, and va_end. There is no |
| 123 | // default action for va_start, so we do that custom. |
| 124 | setOperationAction(Op: ISD::VASTART, VT: MVT::Other, Action: Custom); |
| 125 | setOperationAction(Op: ISD::VAARG, VT: MVT::Other, Action: Expand); |
| 126 | setOperationAction(Op: ISD::VACOPY, VT: MVT::Other, Action: Expand); |
| 127 | setOperationAction(Op: ISD::VAEND, VT: MVT::Other, Action: Expand); |
| 128 | |
| 129 | for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64, MVT::v8f16}) { |
| 130 | if (!Subtarget->hasFP16() && T == MVT::v8f16) { |
| 131 | continue; |
| 132 | } |
| 133 | // Don't expand the floating-point types to constant pools. |
| 134 | setOperationAction(Op: ISD::ConstantFP, VT: T, Action: Legal); |
| 135 | // Expand floating-point comparisons. |
| 136 | for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE, |
| 137 | ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE}) |
| 138 | setCondCodeAction(CCs: CC, VT: T, Action: Expand); |
| 139 | // Expand floating-point library function operators. |
| 140 | for (auto Op : {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FMA}) |
| 141 | setOperationAction(Op, VT: T, Action: Expand); |
| 142 | // Expand vector FREM, but use a libcall rather than an expansion for scalar |
| 143 | if (MVT(T).isVector()) |
| 144 | setOperationAction(Op: ISD::FREM, VT: T, Action: Expand); |
| 145 | else |
| 146 | setOperationAction(Op: ISD::FREM, VT: T, Action: LibCall); |
| 147 | // Note supported floating-point library function operators that otherwise |
| 148 | // default to expand. |
| 149 | for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT, |
| 150 | ISD::FRINT, ISD::FROUNDEVEN}) |
| 151 | setOperationAction(Op, VT: T, Action: Legal); |
| 152 | // Support minimum and maximum, which otherwise default to expand. |
| 153 | setOperationAction(Op: ISD::FMINIMUM, VT: T, Action: Legal); |
| 154 | setOperationAction(Op: ISD::FMAXIMUM, VT: T, Action: Legal); |
| 155 | // When experimental v8f16 support is enabled these instructions don't need |
| 156 | // to be expanded. |
| 157 | if (T != MVT::v8f16) { |
| 158 | setOperationAction(Op: ISD::FP16_TO_FP, VT: T, Action: Expand); |
| 159 | setOperationAction(Op: ISD::FP_TO_FP16, VT: T, Action: Expand); |
| 160 | } |
| 161 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: T, MemVT: MVT::f16, Action: Expand); |
| 162 | setTruncStoreAction(ValVT: T, MemVT: MVT::f16, Action: Expand); |
| 163 | } |
| 164 | |
| 165 | // Expand unavailable integer operations. |
| 166 | for (auto Op : |
| 167 | {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU, |
| 168 | ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS, |
| 169 | ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) { |
| 170 | for (auto T : {MVT::i32, MVT::i64}) |
| 171 | setOperationAction(Op, VT: T, Action: Expand); |
| 172 | if (Subtarget->hasSIMD128()) |
| 173 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 174 | setOperationAction(Op, VT: T, Action: Expand); |
| 175 | } |
| 176 | |
| 177 | if (Subtarget->hasWideArithmetic()) { |
| 178 | setOperationAction(Op: ISD::ADD, VT: MVT::i128, Action: Custom); |
| 179 | setOperationAction(Op: ISD::SUB, VT: MVT::i128, Action: Custom); |
| 180 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i64, Action: Custom); |
| 181 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i64, Action: Custom); |
| 182 | setOperationAction(Op: ISD::UADDO, VT: MVT::i64, Action: Custom); |
| 183 | } |
| 184 | |
| 185 | if (Subtarget->hasNontrappingFPToInt()) |
| 186 | for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) |
| 187 | for (auto T : {MVT::i32, MVT::i64}) |
| 188 | setOperationAction(Op, VT: T, Action: Custom); |
| 189 | |
| 190 | if (Subtarget->hasRelaxedSIMD()) { |
| 191 | setOperationAction( |
| 192 | Ops: {ISD::FMINNUM, ISD::FMINIMUMNUM, ISD::FMAXNUM, ISD::FMAXIMUMNUM}, |
| 193 | VTs: {MVT::v4f32, MVT::v2f64}, Action: Custom); |
| 194 | } |
| 195 | // SIMD-specific configuration |
| 196 | if (Subtarget->hasSIMD128()) { |
| 197 | |
| 198 | setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); |
| 199 | |
| 200 | // Combine wide-vector muls, with extend inputs, to extmul_half. |
| 201 | setTargetDAGCombine(ISD::MUL); |
| 202 | |
| 203 | // Combine vector mask reductions into alltrue/anytrue |
| 204 | setTargetDAGCombine(ISD::SETCC); |
| 205 | |
| 206 | // Convert vector to integer bitcasts to bitmask |
| 207 | setTargetDAGCombine(ISD::BITCAST); |
| 208 | |
| 209 | // Hoist bitcasts out of shuffles |
| 210 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); |
| 211 | |
| 212 | // Combine extends of extract_subvectors into widening ops |
| 213 | setTargetDAGCombine({ISD::SIGN_EXTEND, ISD::ZERO_EXTEND}); |
| 214 | |
| 215 | // Combine int_to_fp or fp_extend of extract_vectors and vice versa into |
| 216 | // conversions ops |
| 217 | setTargetDAGCombine({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_EXTEND, |
| 218 | ISD::EXTRACT_SUBVECTOR}); |
| 219 | |
| 220 | // Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa |
| 221 | // into conversion ops |
| 222 | setTargetDAGCombine({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT, |
| 223 | ISD::FP_TO_SINT, ISD::FP_TO_UINT, ISD::FP_ROUND, |
| 224 | ISD::CONCAT_VECTORS}); |
| 225 | |
| 226 | setTargetDAGCombine(ISD::TRUNCATE); |
| 227 | |
| 228 | // Support saturating add/sub for i8x16 and i16x8 |
| 229 | for (auto Op : {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT}) |
| 230 | for (auto T : {MVT::v16i8, MVT::v8i16}) |
| 231 | setOperationAction(Op, VT: T, Action: Legal); |
| 232 | |
| 233 | // Support integer abs |
| 234 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 235 | setOperationAction(Op: ISD::ABS, VT: T, Action: Legal); |
| 236 | |
| 237 | // Custom lower BUILD_VECTORs to minimize number of replace_lanes |
| 238 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 239 | MVT::v2f64}) |
| 240 | setOperationAction(Op: ISD::BUILD_VECTOR, VT: T, Action: Custom); |
| 241 | |
| 242 | if (Subtarget->hasFP16()) |
| 243 | setOperationAction(Op: ISD::BUILD_VECTOR, VT: MVT::f16, Action: Custom); |
| 244 | |
| 245 | // We have custom shuffle lowering to expose the shuffle mask |
| 246 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 247 | MVT::v2f64}) |
| 248 | setOperationAction(Op: ISD::VECTOR_SHUFFLE, VT: T, Action: Custom); |
| 249 | |
| 250 | if (Subtarget->hasFP16()) |
| 251 | setOperationAction(Op: ISD::VECTOR_SHUFFLE, VT: MVT::v8f16, Action: Custom); |
| 252 | |
| 253 | // Support splatting |
| 254 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 255 | MVT::v2f64}) |
| 256 | setOperationAction(Op: ISD::SPLAT_VECTOR, VT: T, Action: Legal); |
| 257 | |
| 258 | setOperationAction(Ops: ISD::AVGCEILU, VTs: {MVT::v8i16, MVT::v16i8}, Action: Legal); |
| 259 | |
| 260 | // Custom lowering since wasm shifts must have a scalar shift amount |
| 261 | for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL}) |
| 262 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 263 | setOperationAction(Op, VT: T, Action: Custom); |
| 264 | |
| 265 | // Custom lower lane accesses to expand out variable indices |
| 266 | for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT}) |
| 267 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 268 | MVT::v2f64}) |
| 269 | setOperationAction(Op, VT: T, Action: Custom); |
| 270 | |
| 271 | // There is no i8x16.mul instruction |
| 272 | setOperationAction(Op: ISD::MUL, VT: MVT::v16i8, Action: Expand); |
| 273 | |
| 274 | // There is no vector conditional select instruction |
| 275 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 276 | MVT::v2f64}) |
| 277 | setOperationAction(Op: ISD::SELECT_CC, VT: T, Action: Expand); |
| 278 | |
| 279 | // Expand integer operations supported for scalars but not SIMD |
| 280 | for (auto Op : |
| 281 | {ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR}) |
| 282 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 283 | setOperationAction(Op, VT: T, Action: Expand); |
| 284 | |
| 285 | // But we do have integer min and max operations |
| 286 | for (auto Op : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}) |
| 287 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32}) |
| 288 | setOperationAction(Op, VT: T, Action: Legal); |
| 289 | |
| 290 | // And we have popcnt for i8x16. It can be used to expand ctlz/cttz. |
| 291 | setOperationAction(Op: ISD::CTPOP, VT: MVT::v16i8, Action: Legal); |
| 292 | setOperationAction(Op: ISD::CTLZ, VT: MVT::v16i8, Action: Expand); |
| 293 | setOperationAction(Op: ISD::CTTZ, VT: MVT::v16i8, Action: Expand); |
| 294 | |
| 295 | // Custom lower bit counting operations for other types to scalarize them. |
| 296 | for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP}) |
| 297 | for (auto T : {MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 298 | setOperationAction(Op, VT: T, Action: Custom); |
| 299 | |
| 300 | // Expand float operations supported for scalars but not SIMD |
| 301 | for (auto Op : {ISD::FCOPYSIGN, ISD::FLOG, ISD::FLOG2, ISD::FLOG10, |
| 302 | ISD::FEXP, ISD::FEXP2, ISD::FEXP10}) |
| 303 | for (auto T : {MVT::v4f32, MVT::v2f64}) |
| 304 | setOperationAction(Op, VT: T, Action: Expand); |
| 305 | |
| 306 | // Unsigned comparison operations are unavailable for i64x2 vectors. |
| 307 | for (auto CC : {ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE}) |
| 308 | setCondCodeAction(CCs: CC, VT: MVT::v2i64, Action: Custom); |
| 309 | |
| 310 | // 64x2 conversions are not in the spec |
| 311 | for (auto Op : |
| 312 | {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT}) |
| 313 | for (auto T : {MVT::v2i64, MVT::v2f64}) |
| 314 | setOperationAction(Op, VT: T, Action: Expand); |
| 315 | |
| 316 | // But saturating fp_to_int converstions are |
| 317 | for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) { |
| 318 | setOperationAction(Op, VT: MVT::v4i32, Action: Custom); |
| 319 | if (Subtarget->hasFP16()) { |
| 320 | setOperationAction(Op, VT: MVT::v8i16, Action: Custom); |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | // Support vector extending |
| 325 | for (auto T : MVT::integer_fixedlen_vector_valuetypes()) { |
| 326 | setOperationAction(Op: ISD::ANY_EXTEND_VECTOR_INREG, VT: T, Action: Custom); |
| 327 | setOperationAction(Op: ISD::SIGN_EXTEND_VECTOR_INREG, VT: T, Action: Custom); |
| 328 | setOperationAction(Op: ISD::ZERO_EXTEND_VECTOR_INREG, VT: T, Action: Custom); |
| 329 | } |
| 330 | |
| 331 | if (Subtarget->hasFP16()) { |
| 332 | setOperationAction(Op: ISD::FMA, VT: MVT::v8f16, Action: Legal); |
| 333 | } |
| 334 | |
| 335 | if (Subtarget->hasRelaxedSIMD()) { |
| 336 | setOperationAction(Op: ISD::FMULADD, VT: MVT::v4f32, Action: Legal); |
| 337 | setOperationAction(Op: ISD::FMULADD, VT: MVT::v2f64, Action: Legal); |
| 338 | } |
| 339 | |
| 340 | // Partial MLA reductions. |
| 341 | for (auto Op : {ISD::PARTIAL_REDUCE_SMLA, ISD::PARTIAL_REDUCE_UMLA}) { |
| 342 | setPartialReduceMLAAction(Opc: Op, AccVT: MVT::v4i32, InputVT: MVT::v16i8, Action: Legal); |
| 343 | setPartialReduceMLAAction(Opc: Op, AccVT: MVT::v4i32, InputVT: MVT::v8i16, Action: Legal); |
| 344 | } |
| 345 | } |
| 346 | |
| 347 | // As a special case, these operators use the type to mean the type to |
| 348 | // sign-extend from. |
| 349 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1, Action: Expand); |
| 350 | if (!Subtarget->hasSignExt()) { |
| 351 | // Sign extends are legal only when extending a vector extract |
| 352 | auto Action = Subtarget->hasSIMD128() ? Custom : Expand; |
| 353 | for (auto T : {MVT::i8, MVT::i16, MVT::i32}) |
| 354 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: T, Action); |
| 355 | } |
| 356 | for (auto T : MVT::integer_fixedlen_vector_valuetypes()) |
| 357 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: T, Action: Expand); |
| 358 | |
| 359 | // Dynamic stack allocation: use the default expansion. |
| 360 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
| 361 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand); |
| 362 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVTPtr, Action: Expand); |
| 363 | |
| 364 | setOperationAction(Op: ISD::FrameIndex, VT: MVT::i32, Action: Custom); |
| 365 | setOperationAction(Op: ISD::FrameIndex, VT: MVT::i64, Action: Custom); |
| 366 | setOperationAction(Op: ISD::CopyToReg, VT: MVT::Other, Action: Custom); |
| 367 | |
| 368 | // Expand these forms; we pattern-match the forms that we can handle in isel. |
| 369 | for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) |
| 370 | for (auto Op : {ISD::BR_CC, ISD::SELECT_CC}) |
| 371 | setOperationAction(Op, VT: T, Action: Expand); |
| 372 | |
| 373 | // We have custom switch handling. |
| 374 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Custom); |
| 375 | |
| 376 | // WebAssembly doesn't have: |
| 377 | // - Floating-point extending loads. |
| 378 | // - Floating-point truncating stores. |
| 379 | // - i1 extending loads. |
| 380 | // - truncating SIMD stores and most extending loads |
| 381 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
| 382 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
| 383 | for (auto T : MVT::integer_valuetypes()) |
| 384 | for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD}) |
| 385 | setLoadExtAction(ExtType: Ext, ValVT: T, MemVT: MVT::i1, Action: Promote); |
| 386 | if (Subtarget->hasSIMD128()) { |
| 387 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32, |
| 388 | MVT::v2f64}) { |
| 389 | for (auto MemT : MVT::fixedlen_vector_valuetypes()) { |
| 390 | if (MVT(T) != MemT) { |
| 391 | setTruncStoreAction(ValVT: T, MemVT: MemT, Action: Expand); |
| 392 | for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD}) |
| 393 | setLoadExtAction(ExtType: Ext, ValVT: T, MemVT: MemT, Action: Expand); |
| 394 | } |
| 395 | } |
| 396 | } |
| 397 | // But some vector extending loads are legal |
| 398 | for (auto Ext : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) { |
| 399 | setLoadExtAction(ExtType: Ext, ValVT: MVT::v8i16, MemVT: MVT::v8i8, Action: Legal); |
| 400 | setLoadExtAction(ExtType: Ext, ValVT: MVT::v4i32, MemVT: MVT::v4i16, Action: Legal); |
| 401 | setLoadExtAction(ExtType: Ext, ValVT: MVT::v2i64, MemVT: MVT::v2i32, Action: Legal); |
| 402 | } |
| 403 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::v2f64, MemVT: MVT::v2f32, Action: Legal); |
| 404 | } |
| 405 | |
| 406 | // Don't do anything clever with build_pairs |
| 407 | setOperationAction(Op: ISD::BUILD_PAIR, VT: MVT::i64, Action: Expand); |
| 408 | |
| 409 | // Trap lowers to wasm unreachable |
| 410 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Legal); |
| 411 | setOperationAction(Op: ISD::DEBUGTRAP, VT: MVT::Other, Action: Legal); |
| 412 | |
| 413 | // Exception handling intrinsics |
| 414 | setOperationAction(Op: ISD::INTRINSIC_WO_CHAIN, VT: MVT::Other, Action: Custom); |
| 415 | setOperationAction(Op: ISD::INTRINSIC_W_CHAIN, VT: MVT::Other, Action: Custom); |
| 416 | setOperationAction(Op: ISD::INTRINSIC_VOID, VT: MVT::Other, Action: Custom); |
| 417 | |
| 418 | setMaxAtomicSizeInBitsSupported(64); |
| 419 | |
| 420 | // Always convert switches to br_tables unless there is only one case, which |
| 421 | // is equivalent to a simple branch. This reduces code size for wasm, and we |
| 422 | // defer possible jump table optimizations to the VM. |
| 423 | setMinimumJumpTableEntries(2); |
| 424 | } |
| 425 | |
| 426 | MVT WebAssemblyTargetLowering::getPointerTy(const DataLayout &DL, |
| 427 | uint32_t AS) const { |
| 428 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF) |
| 429 | return MVT::externref; |
| 430 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF) |
| 431 | return MVT::funcref; |
| 432 | return TargetLowering::getPointerTy(DL, AS); |
| 433 | } |
| 434 | |
| 435 | MVT WebAssemblyTargetLowering::getPointerMemTy(const DataLayout &DL, |
| 436 | uint32_t AS) const { |
| 437 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF) |
| 438 | return MVT::externref; |
| 439 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF) |
| 440 | return MVT::funcref; |
| 441 | return TargetLowering::getPointerMemTy(DL, AS); |
| 442 | } |
| 443 | |
| 444 | TargetLowering::AtomicExpansionKind |
| 445 | WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR( |
| 446 | const AtomicRMWInst *AI) const { |
| 447 | // We have wasm instructions for these |
| 448 | switch (AI->getOperation()) { |
| 449 | case AtomicRMWInst::Add: |
| 450 | case AtomicRMWInst::Sub: |
| 451 | case AtomicRMWInst::And: |
| 452 | case AtomicRMWInst::Or: |
| 453 | case AtomicRMWInst::Xor: |
| 454 | case AtomicRMWInst::Xchg: |
| 455 | return AtomicExpansionKind::None; |
| 456 | default: |
| 457 | break; |
| 458 | } |
| 459 | return AtomicExpansionKind::CmpXChg; |
| 460 | } |
| 461 | |
| 462 | bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const { |
| 463 | // Implementation copied from X86TargetLowering. |
| 464 | unsigned Opc = VecOp.getOpcode(); |
| 465 | |
| 466 | // Assume target opcodes can't be scalarized. |
| 467 | // TODO - do we have any exceptions? |
| 468 | if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opcode: Opc)) |
| 469 | return false; |
| 470 | |
| 471 | // If the vector op is not supported, try to convert to scalar. |
| 472 | EVT VecVT = VecOp.getValueType(); |
| 473 | if (!isOperationLegalOrCustomOrPromote(Op: Opc, VT: VecVT)) |
| 474 | return true; |
| 475 | |
| 476 | // If the vector op is supported, but the scalar op is not, the transform may |
| 477 | // not be worthwhile. |
| 478 | EVT ScalarVT = VecVT.getScalarType(); |
| 479 | return isOperationLegalOrCustomOrPromote(Op: Opc, VT: ScalarVT); |
| 480 | } |
| 481 | |
| 482 | FastISel *WebAssemblyTargetLowering::createFastISel( |
| 483 | FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo, |
| 484 | const LibcallLoweringInfo *LibcallLowering) const { |
| 485 | return WebAssembly::createFastISel(funcInfo&: FuncInfo, libInfo: LibInfo, libcallLowering: LibcallLowering); |
| 486 | } |
| 487 | |
| 488 | MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/, |
| 489 | EVT VT) const { |
| 490 | unsigned BitWidth = NextPowerOf2(A: VT.getSizeInBits() - 1); |
| 491 | if (BitWidth > 1 && BitWidth < 8) |
| 492 | BitWidth = 8; |
| 493 | |
| 494 | if (BitWidth > 64) { |
| 495 | // The shift will be lowered to a libcall, and compiler-rt libcalls expect |
| 496 | // the count to be an i32. |
| 497 | BitWidth = 32; |
| 498 | assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) && |
| 499 | "32-bit shift counts ought to be enough for anyone" ); |
| 500 | } |
| 501 | |
| 502 | MVT Result = MVT::getIntegerVT(BitWidth); |
| 503 | assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE && |
| 504 | "Unable to represent scalar shift amount type" ); |
| 505 | return Result; |
| 506 | } |
| 507 | |
| 508 | // Lower an fp-to-int conversion operator from the LLVM opcode, which has an |
| 509 | // undefined result on invalid/overflow, to the WebAssembly opcode, which |
| 510 | // traps on invalid/overflow. |
| 511 | static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL, |
| 512 | MachineBasicBlock *BB, |
| 513 | const TargetInstrInfo &TII, |
| 514 | bool IsUnsigned, bool Int64, |
| 515 | bool Float64, unsigned LoweredOpcode) { |
| 516 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
| 517 | |
| 518 | Register OutReg = MI.getOperand(i: 0).getReg(); |
| 519 | Register InReg = MI.getOperand(i: 1).getReg(); |
| 520 | |
| 521 | unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32; |
| 522 | unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32; |
| 523 | unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32; |
| 524 | unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32; |
| 525 | unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32; |
| 526 | unsigned Eqz = WebAssembly::EQZ_I32; |
| 527 | unsigned And = WebAssembly::AND_I32; |
| 528 | int64_t Limit = Int64 ? INT64_MIN : INT32_MIN; |
| 529 | int64_t Substitute = IsUnsigned ? 0 : Limit; |
| 530 | double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit; |
| 531 | auto &Context = BB->getParent()->getFunction().getContext(); |
| 532 | Type *Ty = Float64 ? Type::getDoubleTy(C&: Context) : Type::getFloatTy(C&: Context); |
| 533 | |
| 534 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
| 535 | MachineFunction *F = BB->getParent(); |
| 536 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 537 | MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 538 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 539 | |
| 540 | MachineFunction::iterator It = ++BB->getIterator(); |
| 541 | F->insert(MBBI: It, MBB: FalseMBB); |
| 542 | F->insert(MBBI: It, MBB: TrueMBB); |
| 543 | F->insert(MBBI: It, MBB: DoneMBB); |
| 544 | |
| 545 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
| 546 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
| 547 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 548 | |
| 549 | BB->addSuccessor(Succ: TrueMBB); |
| 550 | BB->addSuccessor(Succ: FalseMBB); |
| 551 | TrueMBB->addSuccessor(Succ: DoneMBB); |
| 552 | FalseMBB->addSuccessor(Succ: DoneMBB); |
| 553 | |
| 554 | unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg; |
| 555 | Tmp0 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
| 556 | Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
| 557 | CmpReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 558 | EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 559 | FalseReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg)); |
| 560 | TrueReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg)); |
| 561 | |
| 562 | MI.eraseFromParent(); |
| 563 | // For signed numbers, we can do a single comparison to determine whether |
| 564 | // fabs(x) is within range. |
| 565 | if (IsUnsigned) { |
| 566 | Tmp0 = InReg; |
| 567 | } else { |
| 568 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Abs), DestReg: Tmp0).addReg(RegNo: InReg); |
| 569 | } |
| 570 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1) |
| 571 | .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: CmpVal))); |
| 572 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: LT), DestReg: CmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1); |
| 573 | |
| 574 | // For unsigned numbers, we have to do a separate comparison with zero. |
| 575 | if (IsUnsigned) { |
| 576 | Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
| 577 | Register SecondCmpReg = |
| 578 | MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 579 | Register AndReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 580 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1) |
| 581 | .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: 0.0))); |
| 582 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: GE), DestReg: SecondCmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1); |
| 583 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: And), DestReg: AndReg).addReg(RegNo: CmpReg).addReg(RegNo: SecondCmpReg); |
| 584 | CmpReg = AndReg; |
| 585 | } |
| 586 | |
| 587 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).addReg(RegNo: CmpReg); |
| 588 | |
| 589 | // Create the CFG diamond to select between doing the conversion or using |
| 590 | // the substitute value. |
| 591 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: TrueMBB).addReg(RegNo: EqzReg); |
| 592 | BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: LoweredOpcode), DestReg: FalseReg).addReg(RegNo: InReg); |
| 593 | BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB); |
| 594 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: IConst), DestReg: TrueReg).addImm(Val: Substitute); |
| 595 | BuildMI(BB&: *DoneMBB, I: DoneMBB->begin(), MIMD: DL, MCID: TII.get(Opcode: TargetOpcode::PHI), DestReg: OutReg) |
| 596 | .addReg(RegNo: FalseReg) |
| 597 | .addMBB(MBB: FalseMBB) |
| 598 | .addReg(RegNo: TrueReg) |
| 599 | .addMBB(MBB: TrueMBB); |
| 600 | |
| 601 | return DoneMBB; |
| 602 | } |
| 603 | |
| 604 | // Lower a `MEMCPY` instruction into a CFG triangle around a `MEMORY_COPY` |
| 605 | // instuction to handle the zero-length case. |
| 606 | static MachineBasicBlock *LowerMemcpy(MachineInstr &MI, DebugLoc DL, |
| 607 | MachineBasicBlock *BB, |
| 608 | const TargetInstrInfo &TII, bool Int64) { |
| 609 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
| 610 | |
| 611 | MachineOperand DstMem = MI.getOperand(i: 0); |
| 612 | MachineOperand SrcMem = MI.getOperand(i: 1); |
| 613 | MachineOperand Dst = MI.getOperand(i: 2); |
| 614 | MachineOperand Src = MI.getOperand(i: 3); |
| 615 | MachineOperand Len = MI.getOperand(i: 4); |
| 616 | |
| 617 | // If the length is a constant, we don't actually need the check. |
| 618 | if (MachineInstr *Def = MRI.getVRegDef(Reg: Len.getReg())) { |
| 619 | if (Def->getOpcode() == WebAssembly::CONST_I32 || |
| 620 | Def->getOpcode() == WebAssembly::CONST_I64) { |
| 621 | if (Def->getOperand(i: 1).getImm() == 0) { |
| 622 | // A zero-length memcpy is a no-op. |
| 623 | MI.eraseFromParent(); |
| 624 | return BB; |
| 625 | } |
| 626 | // A non-zero-length memcpy doesn't need a zero check. |
| 627 | unsigned MemoryCopy = |
| 628 | Int64 ? WebAssembly::MEMORY_COPY_A64 : WebAssembly::MEMORY_COPY_A32; |
| 629 | BuildMI(BB&: *BB, I&: MI, MIMD: DL, MCID: TII.get(Opcode: MemoryCopy)) |
| 630 | .add(MO: DstMem) |
| 631 | .add(MO: SrcMem) |
| 632 | .add(MO: Dst) |
| 633 | .add(MO: Src) |
| 634 | .add(MO: Len); |
| 635 | MI.eraseFromParent(); |
| 636 | return BB; |
| 637 | } |
| 638 | } |
| 639 | |
| 640 | // We're going to add an extra use to `Len` to test if it's zero; that |
| 641 | // use shouldn't be a kill, even if the original use is. |
| 642 | MachineOperand NoKillLen = Len; |
| 643 | NoKillLen.setIsKill(false); |
| 644 | |
| 645 | // Decide on which `MachineInstr` opcode we're going to use. |
| 646 | unsigned Eqz = Int64 ? WebAssembly::EQZ_I64 : WebAssembly::EQZ_I32; |
| 647 | unsigned MemoryCopy = |
| 648 | Int64 ? WebAssembly::MEMORY_COPY_A64 : WebAssembly::MEMORY_COPY_A32; |
| 649 | |
| 650 | // Create two new basic blocks; one for the new `memory.fill` that we can |
| 651 | // branch over, and one for the rest of the instructions after the original |
| 652 | // `memory.fill`. |
| 653 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
| 654 | MachineFunction *F = BB->getParent(); |
| 655 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 656 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 657 | |
| 658 | MachineFunction::iterator It = ++BB->getIterator(); |
| 659 | F->insert(MBBI: It, MBB: TrueMBB); |
| 660 | F->insert(MBBI: It, MBB: DoneMBB); |
| 661 | |
| 662 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
| 663 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
| 664 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 665 | |
| 666 | // Connect the CFG edges. |
| 667 | BB->addSuccessor(Succ: TrueMBB); |
| 668 | BB->addSuccessor(Succ: DoneMBB); |
| 669 | TrueMBB->addSuccessor(Succ: DoneMBB); |
| 670 | |
| 671 | // Create a virtual register for the `Eqz` result. |
| 672 | unsigned EqzReg; |
| 673 | EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 674 | |
| 675 | // Erase the original `memory.copy`. |
| 676 | MI.eraseFromParent(); |
| 677 | |
| 678 | // Test if `Len` is zero. |
| 679 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).add(MO: NoKillLen); |
| 680 | |
| 681 | // Insert a new `memory.copy`. |
| 682 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: MemoryCopy)) |
| 683 | .add(MO: DstMem) |
| 684 | .add(MO: SrcMem) |
| 685 | .add(MO: Dst) |
| 686 | .add(MO: Src) |
| 687 | .add(MO: Len); |
| 688 | |
| 689 | // Create the CFG triangle. |
| 690 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: DoneMBB).addReg(RegNo: EqzReg); |
| 691 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB); |
| 692 | |
| 693 | return DoneMBB; |
| 694 | } |
| 695 | |
| 696 | // Lower a `MEMSET` instruction into a CFG triangle around a `MEMORY_FILL` |
| 697 | // instuction to handle the zero-length case. |
| 698 | static MachineBasicBlock *LowerMemset(MachineInstr &MI, DebugLoc DL, |
| 699 | MachineBasicBlock *BB, |
| 700 | const TargetInstrInfo &TII, bool Int64) { |
| 701 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
| 702 | |
| 703 | MachineOperand Mem = MI.getOperand(i: 0); |
| 704 | MachineOperand Dst = MI.getOperand(i: 1); |
| 705 | MachineOperand Val = MI.getOperand(i: 2); |
| 706 | MachineOperand Len = MI.getOperand(i: 3); |
| 707 | |
| 708 | // If the length is a constant, we don't actually need the check. |
| 709 | if (MachineInstr *Def = MRI.getVRegDef(Reg: Len.getReg())) { |
| 710 | if (Def->getOpcode() == WebAssembly::CONST_I32 || |
| 711 | Def->getOpcode() == WebAssembly::CONST_I64) { |
| 712 | if (Def->getOperand(i: 1).getImm() == 0) { |
| 713 | // A zero-length memset is a no-op. |
| 714 | MI.eraseFromParent(); |
| 715 | return BB; |
| 716 | } |
| 717 | // A non-zero-length memset doesn't need a zero check. |
| 718 | unsigned MemoryFill = |
| 719 | Int64 ? WebAssembly::MEMORY_FILL_A64 : WebAssembly::MEMORY_FILL_A32; |
| 720 | BuildMI(BB&: *BB, I&: MI, MIMD: DL, MCID: TII.get(Opcode: MemoryFill)) |
| 721 | .add(MO: Mem) |
| 722 | .add(MO: Dst) |
| 723 | .add(MO: Val) |
| 724 | .add(MO: Len); |
| 725 | MI.eraseFromParent(); |
| 726 | return BB; |
| 727 | } |
| 728 | } |
| 729 | |
| 730 | // We're going to add an extra use to `Len` to test if it's zero; that |
| 731 | // use shouldn't be a kill, even if the original use is. |
| 732 | MachineOperand NoKillLen = Len; |
| 733 | NoKillLen.setIsKill(false); |
| 734 | |
| 735 | // Decide on which `MachineInstr` opcode we're going to use. |
| 736 | unsigned Eqz = Int64 ? WebAssembly::EQZ_I64 : WebAssembly::EQZ_I32; |
| 737 | unsigned MemoryFill = |
| 738 | Int64 ? WebAssembly::MEMORY_FILL_A64 : WebAssembly::MEMORY_FILL_A32; |
| 739 | |
| 740 | // Create two new basic blocks; one for the new `memory.fill` that we can |
| 741 | // branch over, and one for the rest of the instructions after the original |
| 742 | // `memory.fill`. |
| 743 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
| 744 | MachineFunction *F = BB->getParent(); |
| 745 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 746 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 747 | |
| 748 | MachineFunction::iterator It = ++BB->getIterator(); |
| 749 | F->insert(MBBI: It, MBB: TrueMBB); |
| 750 | F->insert(MBBI: It, MBB: DoneMBB); |
| 751 | |
| 752 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
| 753 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
| 754 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 755 | |
| 756 | // Connect the CFG edges. |
| 757 | BB->addSuccessor(Succ: TrueMBB); |
| 758 | BB->addSuccessor(Succ: DoneMBB); |
| 759 | TrueMBB->addSuccessor(Succ: DoneMBB); |
| 760 | |
| 761 | // Create a virtual register for the `Eqz` result. |
| 762 | unsigned EqzReg; |
| 763 | EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 764 | |
| 765 | // Erase the original `memory.fill`. |
| 766 | MI.eraseFromParent(); |
| 767 | |
| 768 | // Test if `Len` is zero. |
| 769 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).add(MO: NoKillLen); |
| 770 | |
| 771 | // Insert a new `memory.copy`. |
| 772 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: MemoryFill)).add(MO: Mem).add(MO: Dst).add(MO: Val).add(MO: Len); |
| 773 | |
| 774 | // Create the CFG triangle. |
| 775 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: DoneMBB).addReg(RegNo: EqzReg); |
| 776 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB); |
| 777 | |
| 778 | return DoneMBB; |
| 779 | } |
| 780 | |
| 781 | static MachineBasicBlock * |
| 782 | LowerCallResults(MachineInstr &CallResults, DebugLoc DL, MachineBasicBlock *BB, |
| 783 | const WebAssemblySubtarget *Subtarget, |
| 784 | const TargetInstrInfo &TII) { |
| 785 | MachineInstr &CallParams = *CallResults.getPrevNode(); |
| 786 | assert(CallParams.getOpcode() == WebAssembly::CALL_PARAMS); |
| 787 | assert(CallResults.getOpcode() == WebAssembly::CALL_RESULTS || |
| 788 | CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS); |
| 789 | |
| 790 | bool IsIndirect = |
| 791 | CallParams.getOperand(i: 0).isReg() || CallParams.getOperand(i: 0).isFI(); |
| 792 | bool IsRetCall = CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS; |
| 793 | |
| 794 | bool IsFuncrefCall = false; |
| 795 | if (IsIndirect && CallParams.getOperand(i: 0).isReg()) { |
| 796 | Register Reg = CallParams.getOperand(i: 0).getReg(); |
| 797 | const MachineFunction *MF = BB->getParent(); |
| 798 | const MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 799 | const TargetRegisterClass *TRC = MRI.getRegClass(Reg); |
| 800 | IsFuncrefCall = (TRC == &WebAssembly::FUNCREFRegClass); |
| 801 | assert(!IsFuncrefCall || Subtarget->hasReferenceTypes()); |
| 802 | } |
| 803 | |
| 804 | unsigned CallOp; |
| 805 | if (IsIndirect && IsRetCall) { |
| 806 | CallOp = WebAssembly::RET_CALL_INDIRECT; |
| 807 | } else if (IsIndirect) { |
| 808 | CallOp = WebAssembly::CALL_INDIRECT; |
| 809 | } else if (IsRetCall) { |
| 810 | CallOp = WebAssembly::RET_CALL; |
| 811 | } else { |
| 812 | CallOp = WebAssembly::CALL; |
| 813 | } |
| 814 | |
| 815 | MachineFunction &MF = *BB->getParent(); |
| 816 | const MCInstrDesc &MCID = TII.get(Opcode: CallOp); |
| 817 | MachineInstrBuilder MIB(MF, MF.CreateMachineInstr(MCID, DL)); |
| 818 | |
| 819 | // Move the function pointer to the end of the arguments for indirect calls |
| 820 | if (IsIndirect) { |
| 821 | auto FnPtr = CallParams.getOperand(i: 0); |
| 822 | CallParams.removeOperand(OpNo: 0); |
| 823 | |
| 824 | // For funcrefs, call_indirect is done through __funcref_call_table and the |
| 825 | // funcref is always installed in slot 0 of the table, therefore instead of |
| 826 | // having the function pointer added at the end of the params list, a zero |
| 827 | // (the index in |
| 828 | // __funcref_call_table is added). |
| 829 | if (IsFuncrefCall) { |
| 830 | Register RegZero = |
| 831 | MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 832 | MachineInstrBuilder MIBC0 = |
| 833 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::CONST_I32), DestReg: RegZero).addImm(Val: 0); |
| 834 | |
| 835 | BB->insert(I: CallResults.getIterator(), M: MIBC0); |
| 836 | MachineInstrBuilder(MF, CallParams).addReg(RegNo: RegZero); |
| 837 | } else |
| 838 | CallParams.addOperand(Op: FnPtr); |
| 839 | } |
| 840 | |
| 841 | for (auto Def : CallResults.defs()) |
| 842 | MIB.add(MO: Def); |
| 843 | |
| 844 | if (IsIndirect) { |
| 845 | // Placeholder for the type index. |
| 846 | // This gets replaced with the correct value in WebAssemblyMCInstLower.cpp |
| 847 | MIB.addImm(Val: 0); |
| 848 | // The table into which this call_indirect indexes. |
| 849 | MCSymbolWasm *Table = IsFuncrefCall |
| 850 | ? WebAssembly::getOrCreateFuncrefCallTableSymbol( |
| 851 | Ctx&: MF.getContext(), Subtarget) |
| 852 | : WebAssembly::getOrCreateFunctionTableSymbol( |
| 853 | Ctx&: MF.getContext(), Subtarget); |
| 854 | if (Subtarget->hasCallIndirectOverlong()) { |
| 855 | MIB.addSym(Sym: Table); |
| 856 | } else { |
| 857 | // For the MVP there is at most one table whose number is 0, but we can't |
| 858 | // write a table symbol or issue relocations. Instead we just ensure the |
| 859 | // table is live and write a zero. |
| 860 | Table->setNoStrip(); |
| 861 | MIB.addImm(Val: 0); |
| 862 | } |
| 863 | } |
| 864 | |
| 865 | for (auto Use : CallParams.uses()) |
| 866 | MIB.add(MO: Use); |
| 867 | |
| 868 | BB->insert(I: CallResults.getIterator(), M: MIB); |
| 869 | CallParams.eraseFromParent(); |
| 870 | CallResults.eraseFromParent(); |
| 871 | |
| 872 | // If this is a funcref call, to avoid hidden GC roots, we need to clear the |
| 873 | // table slot with ref.null upon call_indirect return. |
| 874 | // |
| 875 | // This generates the following code, which comes right after a call_indirect |
| 876 | // of a funcref: |
| 877 | // |
| 878 | // i32.const 0 |
| 879 | // ref.null func |
| 880 | // table.set __funcref_call_table |
| 881 | if (IsIndirect && IsFuncrefCall) { |
| 882 | MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol( |
| 883 | Ctx&: MF.getContext(), Subtarget); |
| 884 | Register RegZero = |
| 885 | MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 886 | MachineInstr *Const0 = |
| 887 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::CONST_I32), DestReg: RegZero).addImm(Val: 0); |
| 888 | BB->insertAfter(I: MIB.getInstr()->getIterator(), MI: Const0); |
| 889 | |
| 890 | Register RegFuncref = |
| 891 | MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::FUNCREFRegClass); |
| 892 | MachineInstr *RefNull = |
| 893 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::REF_NULL_FUNCREF), DestReg: RegFuncref); |
| 894 | BB->insertAfter(I: Const0->getIterator(), MI: RefNull); |
| 895 | |
| 896 | MachineInstr *TableSet = |
| 897 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::TABLE_SET_FUNCREF)) |
| 898 | .addSym(Sym: Table) |
| 899 | .addReg(RegNo: RegZero) |
| 900 | .addReg(RegNo: RegFuncref); |
| 901 | BB->insertAfter(I: RefNull->getIterator(), MI: TableSet); |
| 902 | } |
| 903 | |
| 904 | return BB; |
| 905 | } |
| 906 | |
| 907 | MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter( |
| 908 | MachineInstr &MI, MachineBasicBlock *BB) const { |
| 909 | const TargetInstrInfo &TII = *Subtarget->getInstrInfo(); |
| 910 | DebugLoc DL = MI.getDebugLoc(); |
| 911 | |
| 912 | switch (MI.getOpcode()) { |
| 913 | default: |
| 914 | llvm_unreachable("Unexpected instr type to insert" ); |
| 915 | case WebAssembly::FP_TO_SINT_I32_F32: |
| 916 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: false, Float64: false, |
| 917 | LoweredOpcode: WebAssembly::I32_TRUNC_S_F32); |
| 918 | case WebAssembly::FP_TO_UINT_I32_F32: |
| 919 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: false, Float64: false, |
| 920 | LoweredOpcode: WebAssembly::I32_TRUNC_U_F32); |
| 921 | case WebAssembly::FP_TO_SINT_I64_F32: |
| 922 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: true, Float64: false, |
| 923 | LoweredOpcode: WebAssembly::I64_TRUNC_S_F32); |
| 924 | case WebAssembly::FP_TO_UINT_I64_F32: |
| 925 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: true, Float64: false, |
| 926 | LoweredOpcode: WebAssembly::I64_TRUNC_U_F32); |
| 927 | case WebAssembly::FP_TO_SINT_I32_F64: |
| 928 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: false, Float64: true, |
| 929 | LoweredOpcode: WebAssembly::I32_TRUNC_S_F64); |
| 930 | case WebAssembly::FP_TO_UINT_I32_F64: |
| 931 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: false, Float64: true, |
| 932 | LoweredOpcode: WebAssembly::I32_TRUNC_U_F64); |
| 933 | case WebAssembly::FP_TO_SINT_I64_F64: |
| 934 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: true, Float64: true, |
| 935 | LoweredOpcode: WebAssembly::I64_TRUNC_S_F64); |
| 936 | case WebAssembly::FP_TO_UINT_I64_F64: |
| 937 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: true, Float64: true, |
| 938 | LoweredOpcode: WebAssembly::I64_TRUNC_U_F64); |
| 939 | case WebAssembly::MEMCPY_A32: |
| 940 | return LowerMemcpy(MI, DL, BB, TII, Int64: false); |
| 941 | case WebAssembly::MEMCPY_A64: |
| 942 | return LowerMemcpy(MI, DL, BB, TII, Int64: true); |
| 943 | case WebAssembly::MEMSET_A32: |
| 944 | return LowerMemset(MI, DL, BB, TII, Int64: false); |
| 945 | case WebAssembly::MEMSET_A64: |
| 946 | return LowerMemset(MI, DL, BB, TII, Int64: true); |
| 947 | case WebAssembly::CALL_RESULTS: |
| 948 | case WebAssembly::RET_CALL_RESULTS: |
| 949 | return LowerCallResults(CallResults&: MI, DL, BB, Subtarget, TII); |
| 950 | } |
| 951 | } |
| 952 | |
| 953 | std::pair<unsigned, const TargetRegisterClass *> |
| 954 | WebAssemblyTargetLowering::getRegForInlineAsmConstraint( |
| 955 | const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const { |
| 956 | // First, see if this is a constraint that directly corresponds to a |
| 957 | // WebAssembly register class. |
| 958 | if (Constraint.size() == 1) { |
| 959 | switch (Constraint[0]) { |
| 960 | case 'r': |
| 961 | assert(VT != MVT::iPTR && "Pointer MVT not expected here" ); |
| 962 | if (Subtarget->hasSIMD128() && VT.isVector()) { |
| 963 | if (VT.getSizeInBits() == 128) |
| 964 | return std::make_pair(x: 0U, y: &WebAssembly::V128RegClass); |
| 965 | } |
| 966 | if (VT.isInteger() && !VT.isVector()) { |
| 967 | if (VT.getSizeInBits() <= 32) |
| 968 | return std::make_pair(x: 0U, y: &WebAssembly::I32RegClass); |
| 969 | if (VT.getSizeInBits() <= 64) |
| 970 | return std::make_pair(x: 0U, y: &WebAssembly::I64RegClass); |
| 971 | } |
| 972 | if (VT.isFloatingPoint() && !VT.isVector()) { |
| 973 | switch (VT.getSizeInBits()) { |
| 974 | case 32: |
| 975 | return std::make_pair(x: 0U, y: &WebAssembly::F32RegClass); |
| 976 | case 64: |
| 977 | return std::make_pair(x: 0U, y: &WebAssembly::F64RegClass); |
| 978 | default: |
| 979 | break; |
| 980 | } |
| 981 | } |
| 982 | break; |
| 983 | default: |
| 984 | break; |
| 985 | } |
| 986 | } |
| 987 | |
| 988 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
| 989 | } |
| 990 | |
| 991 | bool WebAssemblyTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
| 992 | // Assume ctz is a relatively cheap operation. |
| 993 | return true; |
| 994 | } |
| 995 | |
| 996 | bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const { |
| 997 | // Assume clz is a relatively cheap operation. |
| 998 | return true; |
| 999 | } |
| 1000 | |
| 1001 | bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
| 1002 | const AddrMode &AM, |
| 1003 | Type *Ty, unsigned AS, |
| 1004 | Instruction *I) const { |
| 1005 | // WebAssembly offsets are added as unsigned without wrapping. The |
| 1006 | // isLegalAddressingMode gives us no way to determine if wrapping could be |
| 1007 | // happening, so we approximate this by accepting only non-negative offsets. |
| 1008 | if (AM.BaseOffs < 0) |
| 1009 | return false; |
| 1010 | |
| 1011 | // WebAssembly has no scale register operands. |
| 1012 | if (AM.Scale != 0) |
| 1013 | return false; |
| 1014 | |
| 1015 | // Everything else is legal. |
| 1016 | return true; |
| 1017 | } |
| 1018 | |
| 1019 | bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses( |
| 1020 | EVT /*VT*/, unsigned /*AddrSpace*/, Align /*Align*/, |
| 1021 | MachineMemOperand::Flags /*Flags*/, unsigned *Fast) const { |
| 1022 | // WebAssembly supports unaligned accesses, though it should be declared |
| 1023 | // with the p2align attribute on loads and stores which do so, and there |
| 1024 | // may be a performance impact. We tell LLVM they're "fast" because |
| 1025 | // for the kinds of things that LLVM uses this for (merging adjacent stores |
| 1026 | // of constants, etc.), WebAssembly implementations will either want the |
| 1027 | // unaligned access or they'll split anyway. |
| 1028 | if (Fast) |
| 1029 | *Fast = 1; |
| 1030 | return true; |
| 1031 | } |
| 1032 | |
| 1033 | bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT, |
| 1034 | AttributeList Attr) const { |
| 1035 | // The current thinking is that wasm engines will perform this optimization, |
| 1036 | // so we can save on code size. |
| 1037 | return true; |
| 1038 | } |
| 1039 | |
| 1040 | bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const { |
| 1041 | EVT ExtT = ExtVal.getValueType(); |
| 1042 | EVT MemT = cast<LoadSDNode>(Val: ExtVal->getOperand(Num: 0))->getValueType(ResNo: 0); |
| 1043 | return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) || |
| 1044 | (ExtT == MVT::v4i32 && MemT == MVT::v4i16) || |
| 1045 | (ExtT == MVT::v2i64 && MemT == MVT::v2i32); |
| 1046 | } |
| 1047 | |
| 1048 | bool WebAssemblyTargetLowering::isOffsetFoldingLegal( |
| 1049 | const GlobalAddressSDNode *GA) const { |
| 1050 | // Wasm doesn't support function addresses with offsets |
| 1051 | const GlobalValue *GV = GA->getGlobal(); |
| 1052 | return isa<Function>(Val: GV) ? false : TargetLowering::isOffsetFoldingLegal(GA); |
| 1053 | } |
| 1054 | |
| 1055 | EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL, |
| 1056 | LLVMContext &C, |
| 1057 | EVT VT) const { |
| 1058 | if (VT.isVector()) |
| 1059 | return VT.changeVectorElementTypeToInteger(); |
| 1060 | |
| 1061 | // So far, all branch instructions in Wasm take an I32 condition. |
| 1062 | // The default TargetLowering::getSetCCResultType returns the pointer size, |
| 1063 | // which would be useful to reduce instruction counts when testing |
| 1064 | // against 64-bit pointers/values if at some point Wasm supports that. |
| 1065 | return EVT::getIntegerVT(Context&: C, BitWidth: 32); |
| 1066 | } |
| 1067 | |
| 1068 | void WebAssemblyTargetLowering::getTgtMemIntrinsic( |
| 1069 | SmallVectorImpl<IntrinsicInfo> &Infos, const CallBase &I, |
| 1070 | MachineFunction &MF, unsigned Intrinsic) const { |
| 1071 | IntrinsicInfo Info; |
| 1072 | switch (Intrinsic) { |
| 1073 | case Intrinsic::wasm_memory_atomic_notify: |
| 1074 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1075 | Info.memVT = MVT::i32; |
| 1076 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1077 | Info.offset = 0; |
| 1078 | Info.align = Align(4); |
| 1079 | // atomic.notify instruction does not really load the memory specified with |
| 1080 | // this argument, but MachineMemOperand should either be load or store, so |
| 1081 | // we set this to a load. |
| 1082 | // FIXME Volatile isn't really correct, but currently all LLVM atomic |
| 1083 | // instructions are treated as volatiles in the backend, so we should be |
| 1084 | // consistent. The same applies for wasm_atomic_wait intrinsics too. |
| 1085 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
| 1086 | Infos.push_back(Elt: Info); |
| 1087 | return; |
| 1088 | case Intrinsic::wasm_memory_atomic_wait32: |
| 1089 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1090 | Info.memVT = MVT::i32; |
| 1091 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1092 | Info.offset = 0; |
| 1093 | Info.align = Align(4); |
| 1094 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
| 1095 | Infos.push_back(Elt: Info); |
| 1096 | return; |
| 1097 | case Intrinsic::wasm_memory_atomic_wait64: |
| 1098 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1099 | Info.memVT = MVT::i64; |
| 1100 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1101 | Info.offset = 0; |
| 1102 | Info.align = Align(8); |
| 1103 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
| 1104 | Infos.push_back(Elt: Info); |
| 1105 | return; |
| 1106 | case Intrinsic::wasm_loadf16_f32: |
| 1107 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1108 | Info.memVT = MVT::f16; |
| 1109 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1110 | Info.offset = 0; |
| 1111 | Info.align = Align(2); |
| 1112 | Info.flags = MachineMemOperand::MOLoad; |
| 1113 | Infos.push_back(Elt: Info); |
| 1114 | return; |
| 1115 | case Intrinsic::wasm_storef16_f32: |
| 1116 | Info.opc = ISD::INTRINSIC_VOID; |
| 1117 | Info.memVT = MVT::f16; |
| 1118 | Info.ptrVal = I.getArgOperand(i: 1); |
| 1119 | Info.offset = 0; |
| 1120 | Info.align = Align(2); |
| 1121 | Info.flags = MachineMemOperand::MOStore; |
| 1122 | Infos.push_back(Elt: Info); |
| 1123 | return; |
| 1124 | default: |
| 1125 | return; |
| 1126 | } |
| 1127 | } |
| 1128 | |
| 1129 | void WebAssemblyTargetLowering::computeKnownBitsForTargetNode( |
| 1130 | const SDValue Op, KnownBits &Known, const APInt &DemandedElts, |
| 1131 | const SelectionDAG &DAG, unsigned Depth) const { |
| 1132 | switch (Op.getOpcode()) { |
| 1133 | default: |
| 1134 | break; |
| 1135 | case ISD::INTRINSIC_WO_CHAIN: { |
| 1136 | unsigned IntNo = Op.getConstantOperandVal(i: 0); |
| 1137 | switch (IntNo) { |
| 1138 | default: |
| 1139 | break; |
| 1140 | case Intrinsic::wasm_bitmask: { |
| 1141 | unsigned BitWidth = Known.getBitWidth(); |
| 1142 | EVT VT = Op.getOperand(i: 1).getSimpleValueType(); |
| 1143 | unsigned PossibleBits = VT.getVectorNumElements(); |
| 1144 | APInt ZeroMask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - PossibleBits); |
| 1145 | Known.Zero |= ZeroMask; |
| 1146 | break; |
| 1147 | } |
| 1148 | } |
| 1149 | break; |
| 1150 | } |
| 1151 | case WebAssemblyISD::EXTEND_LOW_U: |
| 1152 | case WebAssemblyISD::EXTEND_HIGH_U: { |
| 1153 | // We know the high half, of each destination vector element, will be zero. |
| 1154 | SDValue SrcOp = Op.getOperand(i: 0); |
| 1155 | EVT VT = SrcOp.getSimpleValueType(); |
| 1156 | unsigned BitWidth = Known.getBitWidth(); |
| 1157 | if (VT == MVT::v8i8 || VT == MVT::v16i8) { |
| 1158 | assert(BitWidth >= 8 && "Unexpected width!" ); |
| 1159 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 8); |
| 1160 | Known.Zero |= Mask; |
| 1161 | } else if (VT == MVT::v4i16 || VT == MVT::v8i16) { |
| 1162 | assert(BitWidth >= 16 && "Unexpected width!" ); |
| 1163 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 16); |
| 1164 | Known.Zero |= Mask; |
| 1165 | } else if (VT == MVT::v2i32 || VT == MVT::v4i32) { |
| 1166 | assert(BitWidth >= 32 && "Unexpected width!" ); |
| 1167 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 32); |
| 1168 | Known.Zero |= Mask; |
| 1169 | } |
| 1170 | break; |
| 1171 | } |
| 1172 | // For 128-bit addition if the upper bits are all zero then it's known that |
| 1173 | // the upper bits of the result will have all bits guaranteed zero except the |
| 1174 | // first. |
| 1175 | case WebAssemblyISD::I64_ADD128: |
| 1176 | if (Op.getResNo() == 1) { |
| 1177 | SDValue LHS_HI = Op.getOperand(i: 1); |
| 1178 | SDValue RHS_HI = Op.getOperand(i: 3); |
| 1179 | if (isNullConstant(V: LHS_HI) && isNullConstant(V: RHS_HI)) |
| 1180 | Known.Zero.setBitsFrom(1); |
| 1181 | } |
| 1182 | break; |
| 1183 | } |
| 1184 | } |
| 1185 | |
| 1186 | TargetLoweringBase::LegalizeTypeAction |
| 1187 | WebAssemblyTargetLowering::getPreferredVectorAction(MVT VT) const { |
| 1188 | if (VT.isFixedLengthVector()) { |
| 1189 | MVT EltVT = VT.getVectorElementType(); |
| 1190 | // We have legal vector types with these lane types, so widening the |
| 1191 | // vector would let us use some of the lanes directly without having to |
| 1192 | // extend or truncate values. |
| 1193 | if (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 || |
| 1194 | EltVT == MVT::i64 || EltVT == MVT::f32 || EltVT == MVT::f64) |
| 1195 | return TypeWidenVector; |
| 1196 | } |
| 1197 | |
| 1198 | return TargetLoweringBase::getPreferredVectorAction(VT); |
| 1199 | } |
| 1200 | |
| 1201 | bool WebAssemblyTargetLowering::isFMAFasterThanFMulAndFAdd( |
| 1202 | const MachineFunction &MF, EVT VT) const { |
| 1203 | if (!Subtarget->hasFP16() || !VT.isVector()) |
| 1204 | return false; |
| 1205 | |
| 1206 | EVT ScalarVT = VT.getScalarType(); |
| 1207 | if (!ScalarVT.isSimple()) |
| 1208 | return false; |
| 1209 | |
| 1210 | return ScalarVT.getSimpleVT().SimpleTy == MVT::f16; |
| 1211 | } |
| 1212 | |
| 1213 | bool WebAssemblyTargetLowering::shouldSimplifyDemandedVectorElts( |
| 1214 | SDValue Op, const TargetLoweringOpt &TLO) const { |
| 1215 | // ISel process runs DAGCombiner after legalization; this step is called |
| 1216 | // SelectionDAG optimization phase. This post-legalization combining process |
| 1217 | // runs DAGCombiner on each node, and if there was a change to be made, |
| 1218 | // re-runs legalization again on it and its user nodes to make sure |
| 1219 | // everythiing is in a legalized state. |
| 1220 | // |
| 1221 | // The legalization calls lowering routines, and we do our custom lowering for |
| 1222 | // build_vectors (LowerBUILD_VECTOR), which converts undef vector elements |
| 1223 | // into zeros. But there is a set of routines in DAGCombiner that turns unused |
| 1224 | // (= not demanded) nodes into undef, among which SimplifyDemandedVectorElts |
| 1225 | // turns unused vector elements into undefs. But this routine does not work |
| 1226 | // with our custom LowerBUILD_VECTOR, which turns undefs into zeros. This |
| 1227 | // combination can result in a infinite loop, in which undefs are converted to |
| 1228 | // zeros in legalization and back to undefs in combining. |
| 1229 | // |
| 1230 | // So after DAG is legalized, we prevent SimplifyDemandedVectorElts from |
| 1231 | // running for build_vectors. |
| 1232 | if (Op.getOpcode() == ISD::BUILD_VECTOR && TLO.LegalOps && TLO.LegalTys) |
| 1233 | return false; |
| 1234 | return true; |
| 1235 | } |
| 1236 | |
| 1237 | //===----------------------------------------------------------------------===// |
| 1238 | // WebAssembly Lowering private implementation. |
| 1239 | //===----------------------------------------------------------------------===// |
| 1240 | |
| 1241 | //===----------------------------------------------------------------------===// |
| 1242 | // Lowering Code |
| 1243 | //===----------------------------------------------------------------------===// |
| 1244 | |
| 1245 | static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) { |
| 1246 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1247 | DAG.getContext()->diagnose( |
| 1248 | DI: DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc())); |
| 1249 | } |
| 1250 | |
| 1251 | // Test whether the given calling convention is supported. |
| 1252 | static bool callingConvSupported(CallingConv::ID CallConv) { |
| 1253 | // We currently support the language-independent target-independent |
| 1254 | // conventions. We don't yet have a way to annotate calls with properties like |
| 1255 | // "cold", and we don't have any call-clobbered registers, so these are mostly |
| 1256 | // all handled the same. |
| 1257 | return CallConv == CallingConv::C || CallConv == CallingConv::Fast || |
| 1258 | CallConv == CallingConv::Cold || |
| 1259 | CallConv == CallingConv::PreserveMost || |
| 1260 | CallConv == CallingConv::PreserveAll || |
| 1261 | CallConv == CallingConv::CXX_FAST_TLS || |
| 1262 | CallConv == CallingConv::WASM_EmscriptenInvoke || |
| 1263 | CallConv == CallingConv::Swift; |
| 1264 | } |
| 1265 | |
| 1266 | SDValue |
| 1267 | WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI, |
| 1268 | SmallVectorImpl<SDValue> &InVals) const { |
| 1269 | SelectionDAG &DAG = CLI.DAG; |
| 1270 | SDLoc DL = CLI.DL; |
| 1271 | SDValue Chain = CLI.Chain; |
| 1272 | SDValue Callee = CLI.Callee; |
| 1273 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1274 | auto Layout = MF.getDataLayout(); |
| 1275 | |
| 1276 | CallingConv::ID CallConv = CLI.CallConv; |
| 1277 | if (!callingConvSupported(CallConv)) |
| 1278 | fail(DL, DAG, |
| 1279 | Msg: "WebAssembly doesn't support language-specific or target-specific " |
| 1280 | "calling conventions yet" ); |
| 1281 | if (CLI.IsPatchPoint) |
| 1282 | fail(DL, DAG, Msg: "WebAssembly doesn't support patch point yet" ); |
| 1283 | |
| 1284 | if (CLI.IsTailCall) { |
| 1285 | auto NoTail = [&](const char *Msg) { |
| 1286 | if (CLI.CB && CLI.CB->isMustTailCall()) |
| 1287 | fail(DL, DAG, Msg); |
| 1288 | CLI.IsTailCall = false; |
| 1289 | }; |
| 1290 | |
| 1291 | if (!Subtarget->hasTailCall()) |
| 1292 | NoTail("WebAssembly 'tail-call' feature not enabled" ); |
| 1293 | |
| 1294 | // Varargs calls cannot be tail calls because the buffer is on the stack |
| 1295 | if (CLI.IsVarArg) |
| 1296 | NoTail("WebAssembly does not support varargs tail calls" ); |
| 1297 | |
| 1298 | // Do not tail call unless caller and callee return types match |
| 1299 | const Function &F = MF.getFunction(); |
| 1300 | const TargetMachine &TM = getTargetMachine(); |
| 1301 | Type *RetTy = F.getReturnType(); |
| 1302 | SmallVector<MVT, 4> CallerRetTys; |
| 1303 | SmallVector<MVT, 4> CalleeRetTys; |
| 1304 | computeLegalValueVTs(F, TM, Ty: RetTy, ValueVTs&: CallerRetTys); |
| 1305 | computeLegalValueVTs(F, TM, Ty: CLI.RetTy, ValueVTs&: CalleeRetTys); |
| 1306 | bool TypesMatch = CallerRetTys.size() == CalleeRetTys.size() && |
| 1307 | std::equal(first1: CallerRetTys.begin(), last1: CallerRetTys.end(), |
| 1308 | first2: CalleeRetTys.begin()); |
| 1309 | if (!TypesMatch) |
| 1310 | NoTail("WebAssembly tail call requires caller and callee return types to " |
| 1311 | "match" ); |
| 1312 | |
| 1313 | // If pointers to local stack values are passed, we cannot tail call |
| 1314 | if (CLI.CB) { |
| 1315 | for (auto &Arg : CLI.CB->args()) { |
| 1316 | Value *Val = Arg.get(); |
| 1317 | // Trace the value back through pointer operations |
| 1318 | while (true) { |
| 1319 | Value *Src = Val->stripPointerCastsAndAliases(); |
| 1320 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: Src)) |
| 1321 | Src = GEP->getPointerOperand(); |
| 1322 | if (Val == Src) |
| 1323 | break; |
| 1324 | Val = Src; |
| 1325 | } |
| 1326 | if (isa<AllocaInst>(Val)) { |
| 1327 | NoTail( |
| 1328 | "WebAssembly does not support tail calling with stack arguments" ); |
| 1329 | break; |
| 1330 | } |
| 1331 | } |
| 1332 | } |
| 1333 | } |
| 1334 | |
| 1335 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
| 1336 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
| 1337 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
| 1338 | |
| 1339 | // The generic code may have added an sret argument. If we're lowering an |
| 1340 | // invoke function, the ABI requires that the function pointer be the first |
| 1341 | // argument, so we may have to swap the arguments. |
| 1342 | if (CallConv == CallingConv::WASM_EmscriptenInvoke && Outs.size() >= 2 && |
| 1343 | Outs[0].Flags.isSRet()) { |
| 1344 | std::swap(a&: Outs[0], b&: Outs[1]); |
| 1345 | std::swap(a&: OutVals[0], b&: OutVals[1]); |
| 1346 | } |
| 1347 | |
| 1348 | bool HasSwiftSelfArg = false; |
| 1349 | bool HasSwiftErrorArg = false; |
| 1350 | unsigned NumFixedArgs = 0; |
| 1351 | for (unsigned I = 0; I < Outs.size(); ++I) { |
| 1352 | const ISD::OutputArg &Out = Outs[I]; |
| 1353 | SDValue &OutVal = OutVals[I]; |
| 1354 | HasSwiftSelfArg |= Out.Flags.isSwiftSelf(); |
| 1355 | HasSwiftErrorArg |= Out.Flags.isSwiftError(); |
| 1356 | if (Out.Flags.isNest()) |
| 1357 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments" ); |
| 1358 | if (Out.Flags.isInAlloca()) |
| 1359 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments" ); |
| 1360 | if (Out.Flags.isInConsecutiveRegs()) |
| 1361 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments" ); |
| 1362 | if (Out.Flags.isInConsecutiveRegsLast()) |
| 1363 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments" ); |
| 1364 | if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) { |
| 1365 | auto &MFI = MF.getFrameInfo(); |
| 1366 | int FI = MFI.CreateStackObject(Size: Out.Flags.getByValSize(), |
| 1367 | Alignment: Out.Flags.getNonZeroByValAlign(), |
| 1368 | /*isSS=*/isSpillSlot: false); |
| 1369 | SDValue SizeNode = |
| 1370 | DAG.getConstant(Val: Out.Flags.getByValSize(), DL, VT: MVT::i32); |
| 1371 | SDValue FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout)); |
| 1372 | Chain = DAG.getMemcpy(Chain, dl: DL, Dst: FINode, Src: OutVal, Size: SizeNode, |
| 1373 | Alignment: Out.Flags.getNonZeroByValAlign(), |
| 1374 | /*isVolatile*/ isVol: false, /*AlwaysInline=*/false, |
| 1375 | /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(), |
| 1376 | SrcPtrInfo: MachinePointerInfo()); |
| 1377 | OutVal = FINode; |
| 1378 | } |
| 1379 | // Count the number of fixed args *after* legalization. |
| 1380 | NumFixedArgs += !Out.Flags.isVarArg(); |
| 1381 | } |
| 1382 | |
| 1383 | bool IsVarArg = CLI.IsVarArg; |
| 1384 | auto PtrVT = getPointerTy(DL: Layout); |
| 1385 | |
| 1386 | // For swiftcc, emit additional swiftself and swifterror arguments |
| 1387 | // if there aren't. These additional arguments are also added for callee |
| 1388 | // signature They are necessary to match callee and caller signature for |
| 1389 | // indirect call. |
| 1390 | if (CallConv == CallingConv::Swift) { |
| 1391 | Type *PtrTy = PointerType::getUnqual(C&: *DAG.getContext()); |
| 1392 | if (!HasSwiftSelfArg) { |
| 1393 | NumFixedArgs++; |
| 1394 | ISD::ArgFlagsTy Flags; |
| 1395 | Flags.setSwiftSelf(); |
| 1396 | ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0); |
| 1397 | CLI.Outs.push_back(Elt: Arg); |
| 1398 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
| 1399 | CLI.OutVals.push_back(Elt: ArgVal); |
| 1400 | } |
| 1401 | if (!HasSwiftErrorArg) { |
| 1402 | NumFixedArgs++; |
| 1403 | ISD::ArgFlagsTy Flags; |
| 1404 | Flags.setSwiftError(); |
| 1405 | ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0); |
| 1406 | CLI.Outs.push_back(Elt: Arg); |
| 1407 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
| 1408 | CLI.OutVals.push_back(Elt: ArgVal); |
| 1409 | } |
| 1410 | } |
| 1411 | |
| 1412 | // Analyze operands of the call, assigning locations to each operand. |
| 1413 | SmallVector<CCValAssign, 16> ArgLocs; |
| 1414 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
| 1415 | |
| 1416 | if (IsVarArg) { |
| 1417 | // Outgoing non-fixed arguments are placed in a buffer. First |
| 1418 | // compute their offsets and the total amount of buffer space needed. |
| 1419 | for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) { |
| 1420 | const ISD::OutputArg &Out = Outs[I]; |
| 1421 | SDValue &Arg = OutVals[I]; |
| 1422 | EVT VT = Arg.getValueType(); |
| 1423 | assert(VT != MVT::iPTR && "Legalized args should be concrete" ); |
| 1424 | Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext()); |
| 1425 | Align Alignment = |
| 1426 | std::max(a: Out.Flags.getNonZeroOrigAlign(), b: Layout.getABITypeAlign(Ty)); |
| 1427 | unsigned Offset = |
| 1428 | CCInfo.AllocateStack(Size: Layout.getTypeAllocSize(Ty), Alignment); |
| 1429 | CCInfo.addLoc(V: CCValAssign::getMem(ValNo: ArgLocs.size(), ValVT: VT.getSimpleVT(), |
| 1430 | Offset, LocVT: VT.getSimpleVT(), |
| 1431 | HTP: CCValAssign::Full)); |
| 1432 | } |
| 1433 | } |
| 1434 | |
| 1435 | unsigned NumBytes = CCInfo.getAlignedCallFrameSize(); |
| 1436 | |
| 1437 | SDValue FINode; |
| 1438 | if (IsVarArg && NumBytes) { |
| 1439 | // For non-fixed arguments, next emit stores to store the argument values |
| 1440 | // to the stack buffer at the offsets computed above. |
| 1441 | MaybeAlign StackAlign = Layout.getStackAlignment(); |
| 1442 | assert(StackAlign && "data layout string is missing stack alignment" ); |
| 1443 | int FI = MF.getFrameInfo().CreateStackObject(Size: NumBytes, Alignment: *StackAlign, |
| 1444 | /*isSS=*/isSpillSlot: false); |
| 1445 | unsigned ValNo = 0; |
| 1446 | SmallVector<SDValue, 8> Chains; |
| 1447 | for (SDValue Arg : drop_begin(RangeOrContainer&: OutVals, N: NumFixedArgs)) { |
| 1448 | assert(ArgLocs[ValNo].getValNo() == ValNo && |
| 1449 | "ArgLocs should remain in order and only hold varargs args" ); |
| 1450 | unsigned Offset = ArgLocs[ValNo++].getLocMemOffset(); |
| 1451 | FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout)); |
| 1452 | SDValue Add = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: FINode, |
| 1453 | N2: DAG.getConstant(Val: Offset, DL, VT: PtrVT)); |
| 1454 | Chains.push_back( |
| 1455 | Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: Add, |
| 1456 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI, Offset))); |
| 1457 | } |
| 1458 | if (!Chains.empty()) |
| 1459 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: Chains); |
| 1460 | } else if (IsVarArg) { |
| 1461 | FINode = DAG.getIntPtrConstant(Val: 0, DL); |
| 1462 | } |
| 1463 | |
| 1464 | if (Callee->getOpcode() == ISD::GlobalAddress) { |
| 1465 | // If the callee is a GlobalAddress node (quite common, every direct call |
| 1466 | // is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress |
| 1467 | // doesn't at MO_GOT which is not needed for direct calls. |
| 1468 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Callee); |
| 1469 | Callee = DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, |
| 1470 | VT: getPointerTy(DL: DAG.getDataLayout()), |
| 1471 | offset: GA->getOffset()); |
| 1472 | Callee = DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, |
| 1473 | VT: getPointerTy(DL: DAG.getDataLayout()), Operand: Callee); |
| 1474 | } |
| 1475 | |
| 1476 | // Compute the operands for the CALLn node. |
| 1477 | SmallVector<SDValue, 16> Ops; |
| 1478 | Ops.push_back(Elt: Chain); |
| 1479 | Ops.push_back(Elt: Callee); |
| 1480 | |
| 1481 | // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs |
| 1482 | // isn't reliable. |
| 1483 | Ops.append(in_start: OutVals.begin(), |
| 1484 | in_end: IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end()); |
| 1485 | // Add a pointer to the vararg buffer. |
| 1486 | if (IsVarArg) |
| 1487 | Ops.push_back(Elt: FINode); |
| 1488 | |
| 1489 | SmallVector<EVT, 8> InTys; |
| 1490 | for (const auto &In : Ins) { |
| 1491 | assert(!In.Flags.isByVal() && "byval is not valid for return values" ); |
| 1492 | assert(!In.Flags.isNest() && "nest is not valid for return values" ); |
| 1493 | if (In.Flags.isInAlloca()) |
| 1494 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca return values" ); |
| 1495 | if (In.Flags.isInConsecutiveRegs()) |
| 1496 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs return values" ); |
| 1497 | if (In.Flags.isInConsecutiveRegsLast()) |
| 1498 | fail(DL, DAG, |
| 1499 | Msg: "WebAssembly hasn't implemented cons regs last return values" ); |
| 1500 | // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in |
| 1501 | // registers. |
| 1502 | InTys.push_back(Elt: In.VT); |
| 1503 | } |
| 1504 | |
| 1505 | // Lastly, if this is a call to a funcref we need to add an instruction |
| 1506 | // table.set to the chain and transform the call. |
| 1507 | if (CLI.CB && WebAssembly::isWebAssemblyFuncrefType( |
| 1508 | Ty: CLI.CB->getCalledOperand()->getType())) { |
| 1509 | // In the absence of function references proposal where a funcref call is |
| 1510 | // lowered to call_ref, using reference types we generate a table.set to set |
| 1511 | // the funcref to a special table used solely for this purpose, followed by |
| 1512 | // a call_indirect. Here we just generate the table set, and return the |
| 1513 | // SDValue of the table.set so that LowerCall can finalize the lowering by |
| 1514 | // generating the call_indirect. |
| 1515 | SDValue Chain = Ops[0]; |
| 1516 | |
| 1517 | MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol( |
| 1518 | Ctx&: MF.getContext(), Subtarget); |
| 1519 | SDValue Sym = DAG.getMCSymbol(Sym: Table, VT: PtrVT); |
| 1520 | SDValue TableSlot = DAG.getConstant(Val: 0, DL, VT: MVT::i32); |
| 1521 | SDValue TableSetOps[] = {Chain, Sym, TableSlot, Callee}; |
| 1522 | SDValue TableSet = DAG.getMemIntrinsicNode( |
| 1523 | Opcode: WebAssemblyISD::TABLE_SET, dl: DL, VTList: DAG.getVTList(VT: MVT::Other), Ops: TableSetOps, |
| 1524 | MemVT: MVT::funcref, |
| 1525 | // Machine Mem Operand args |
| 1526 | PtrInfo: MachinePointerInfo( |
| 1527 | WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF), |
| 1528 | Alignment: CLI.CB->getCalledOperand()->getPointerAlignment(DL: DAG.getDataLayout()), |
| 1529 | Flags: MachineMemOperand::MOStore); |
| 1530 | |
| 1531 | Ops[0] = TableSet; // The new chain is the TableSet itself |
| 1532 | } |
| 1533 | |
| 1534 | if (CLI.IsTailCall) { |
| 1535 | // ret_calls do not return values to the current frame |
| 1536 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
| 1537 | return DAG.getNode(Opcode: WebAssemblyISD::RET_CALL, DL, VTList: NodeTys, Ops); |
| 1538 | } |
| 1539 | |
| 1540 | InTys.push_back(Elt: MVT::Other); |
| 1541 | SDVTList InTyList = DAG.getVTList(VTs: InTys); |
| 1542 | SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::CALL, DL, VTList: InTyList, Ops); |
| 1543 | |
| 1544 | for (size_t I = 0; I < Ins.size(); ++I) |
| 1545 | InVals.push_back(Elt: Res.getValue(R: I)); |
| 1546 | |
| 1547 | // Return the chain |
| 1548 | return Res.getValue(R: Ins.size()); |
| 1549 | } |
| 1550 | |
| 1551 | bool WebAssemblyTargetLowering::CanLowerReturn( |
| 1552 | CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/, |
| 1553 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext & /*Context*/, |
| 1554 | const Type *RetTy) const { |
| 1555 | // WebAssembly can only handle returning tuples with multivalue enabled |
| 1556 | return WebAssembly::canLowerReturn(ResultSize: Outs.size(), Subtarget); |
| 1557 | } |
| 1558 | |
| 1559 | SDValue WebAssemblyTargetLowering::LowerReturn( |
| 1560 | SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/, |
| 1561 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 1562 | const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, |
| 1563 | SelectionDAG &DAG) const { |
| 1564 | assert(WebAssembly::canLowerReturn(Outs.size(), Subtarget) && |
| 1565 | "MVP WebAssembly can only return up to one value" ); |
| 1566 | if (!callingConvSupported(CallConv)) |
| 1567 | fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions" ); |
| 1568 | |
| 1569 | SmallVector<SDValue, 4> RetOps(1, Chain); |
| 1570 | RetOps.append(in_start: OutVals.begin(), in_end: OutVals.end()); |
| 1571 | Chain = DAG.getNode(Opcode: WebAssemblyISD::RETURN, DL, VT: MVT::Other, Ops: RetOps); |
| 1572 | |
| 1573 | // Record the number and types of the return values. |
| 1574 | for (const ISD::OutputArg &Out : Outs) { |
| 1575 | assert(!Out.Flags.isByVal() && "byval is not valid for return values" ); |
| 1576 | assert(!Out.Flags.isNest() && "nest is not valid for return values" ); |
| 1577 | assert(!Out.Flags.isVarArg() && "non-fixed return value is not valid" ); |
| 1578 | if (Out.Flags.isInAlloca()) |
| 1579 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca results" ); |
| 1580 | if (Out.Flags.isInConsecutiveRegs()) |
| 1581 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs results" ); |
| 1582 | if (Out.Flags.isInConsecutiveRegsLast()) |
| 1583 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last results" ); |
| 1584 | } |
| 1585 | |
| 1586 | return Chain; |
| 1587 | } |
| 1588 | |
| 1589 | SDValue WebAssemblyTargetLowering::LowerFormalArguments( |
| 1590 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 1591 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 1592 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 1593 | if (!callingConvSupported(CallConv)) |
| 1594 | fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions" ); |
| 1595 | |
| 1596 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1597 | auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>(); |
| 1598 | |
| 1599 | // Set up the incoming ARGUMENTS value, which serves to represent the liveness |
| 1600 | // of the incoming values before they're represented by virtual registers. |
| 1601 | MF.getRegInfo().addLiveIn(Reg: WebAssembly::ARGUMENTS); |
| 1602 | |
| 1603 | bool HasSwiftErrorArg = false; |
| 1604 | bool HasSwiftSelfArg = false; |
| 1605 | for (const ISD::InputArg &In : Ins) { |
| 1606 | HasSwiftSelfArg |= In.Flags.isSwiftSelf(); |
| 1607 | HasSwiftErrorArg |= In.Flags.isSwiftError(); |
| 1608 | if (In.Flags.isInAlloca()) |
| 1609 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments" ); |
| 1610 | if (In.Flags.isNest()) |
| 1611 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments" ); |
| 1612 | if (In.Flags.isInConsecutiveRegs()) |
| 1613 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments" ); |
| 1614 | if (In.Flags.isInConsecutiveRegsLast()) |
| 1615 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments" ); |
| 1616 | // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in |
| 1617 | // registers. |
| 1618 | InVals.push_back(Elt: In.Used ? DAG.getNode(Opcode: WebAssemblyISD::ARGUMENT, DL, VT: In.VT, |
| 1619 | Operand: DAG.getTargetConstant(Val: InVals.size(), |
| 1620 | DL, VT: MVT::i32)) |
| 1621 | : DAG.getUNDEF(VT: In.VT)); |
| 1622 | |
| 1623 | // Record the number and types of arguments. |
| 1624 | MFI->addParam(VT: In.VT); |
| 1625 | } |
| 1626 | |
| 1627 | // For swiftcc, emit additional swiftself and swifterror arguments |
| 1628 | // if there aren't. These additional arguments are also added for callee |
| 1629 | // signature They are necessary to match callee and caller signature for |
| 1630 | // indirect call. |
| 1631 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 1632 | if (CallConv == CallingConv::Swift) { |
| 1633 | if (!HasSwiftSelfArg) { |
| 1634 | MFI->addParam(VT: PtrVT); |
| 1635 | } |
| 1636 | if (!HasSwiftErrorArg) { |
| 1637 | MFI->addParam(VT: PtrVT); |
| 1638 | } |
| 1639 | } |
| 1640 | // Varargs are copied into a buffer allocated by the caller, and a pointer to |
| 1641 | // the buffer is passed as an argument. |
| 1642 | if (IsVarArg) { |
| 1643 | MVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 1644 | Register VarargVreg = |
| 1645 | MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: PtrVT)); |
| 1646 | MFI->setVarargBufferVreg(VarargVreg); |
| 1647 | Chain = DAG.getCopyToReg( |
| 1648 | Chain, dl: DL, Reg: VarargVreg, |
| 1649 | N: DAG.getNode(Opcode: WebAssemblyISD::ARGUMENT, DL, VT: PtrVT, |
| 1650 | Operand: DAG.getTargetConstant(Val: Ins.size(), DL, VT: MVT::i32))); |
| 1651 | MFI->addParam(VT: PtrVT); |
| 1652 | } |
| 1653 | |
| 1654 | // Record the number and types of arguments and results. |
| 1655 | SmallVector<MVT, 4> Params; |
| 1656 | SmallVector<MVT, 4> Results; |
| 1657 | computeSignatureVTs(Ty: MF.getFunction().getFunctionType(), TargetFunc: &MF.getFunction(), |
| 1658 | ContextFunc: MF.getFunction(), TM: DAG.getTarget(), Params, Results); |
| 1659 | for (MVT VT : Results) |
| 1660 | MFI->addResult(VT); |
| 1661 | // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify |
| 1662 | // the param logic here with ComputeSignatureVTs |
| 1663 | assert(MFI->getParams().size() == Params.size() && |
| 1664 | std::equal(MFI->getParams().begin(), MFI->getParams().end(), |
| 1665 | Params.begin())); |
| 1666 | |
| 1667 | return Chain; |
| 1668 | } |
| 1669 | |
| 1670 | void WebAssemblyTargetLowering::ReplaceNodeResults( |
| 1671 | SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { |
| 1672 | switch (N->getOpcode()) { |
| 1673 | case ISD::SIGN_EXTEND_INREG: |
| 1674 | // Do not add any results, signifying that N should not be custom lowered |
| 1675 | // after all. This happens because simd128 turns on custom lowering for |
| 1676 | // SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an |
| 1677 | // illegal type. |
| 1678 | break; |
| 1679 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 1680 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 1681 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 1682 | // Do not add any results, signifying that N should not be custom lowered. |
| 1683 | // EXTEND_VECTOR_INREG is implemented for some vectors, but not all. |
| 1684 | break; |
| 1685 | case ISD::ADD: |
| 1686 | case ISD::SUB: |
| 1687 | Results.push_back(Elt: Replace128Op(N, DAG)); |
| 1688 | break; |
| 1689 | default: |
| 1690 | llvm_unreachable( |
| 1691 | "ReplaceNodeResults not implemented for this op for WebAssembly!" ); |
| 1692 | } |
| 1693 | } |
| 1694 | |
| 1695 | //===----------------------------------------------------------------------===// |
| 1696 | // Custom lowering hooks. |
| 1697 | //===----------------------------------------------------------------------===// |
| 1698 | |
| 1699 | SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op, |
| 1700 | SelectionDAG &DAG) const { |
| 1701 | SDLoc DL(Op); |
| 1702 | switch (Op.getOpcode()) { |
| 1703 | default: |
| 1704 | llvm_unreachable("unimplemented operation lowering" ); |
| 1705 | return SDValue(); |
| 1706 | case ISD::FrameIndex: |
| 1707 | return LowerFrameIndex(Op, DAG); |
| 1708 | case ISD::GlobalAddress: |
| 1709 | return LowerGlobalAddress(Op, DAG); |
| 1710 | case ISD::GlobalTLSAddress: |
| 1711 | return LowerGlobalTLSAddress(Op, DAG); |
| 1712 | case ISD::ExternalSymbol: |
| 1713 | return LowerExternalSymbol(Op, DAG); |
| 1714 | case ISD::JumpTable: |
| 1715 | return LowerJumpTable(Op, DAG); |
| 1716 | case ISD::BR_JT: |
| 1717 | return LowerBR_JT(Op, DAG); |
| 1718 | case ISD::VASTART: |
| 1719 | return LowerVASTART(Op, DAG); |
| 1720 | case ISD::BlockAddress: |
| 1721 | case ISD::BRIND: |
| 1722 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented computed gotos" ); |
| 1723 | return SDValue(); |
| 1724 | case ISD::RETURNADDR: |
| 1725 | return LowerRETURNADDR(Op, DAG); |
| 1726 | case ISD::FRAMEADDR: |
| 1727 | return LowerFRAMEADDR(Op, DAG); |
| 1728 | case ISD::CopyToReg: |
| 1729 | return LowerCopyToReg(Op, DAG); |
| 1730 | case ISD::EXTRACT_VECTOR_ELT: |
| 1731 | case ISD::INSERT_VECTOR_ELT: |
| 1732 | return LowerAccessVectorElement(Op, DAG); |
| 1733 | case ISD::INTRINSIC_VOID: |
| 1734 | case ISD::INTRINSIC_WO_CHAIN: |
| 1735 | case ISD::INTRINSIC_W_CHAIN: |
| 1736 | return LowerIntrinsic(Op, DAG); |
| 1737 | case ISD::SIGN_EXTEND_INREG: |
| 1738 | return LowerSIGN_EXTEND_INREG(Op, DAG); |
| 1739 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 1740 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 1741 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 1742 | return LowerEXTEND_VECTOR_INREG(Op, DAG); |
| 1743 | case ISD::BUILD_VECTOR: |
| 1744 | return LowerBUILD_VECTOR(Op, DAG); |
| 1745 | case ISD::VECTOR_SHUFFLE: |
| 1746 | return LowerVECTOR_SHUFFLE(Op, DAG); |
| 1747 | case ISD::SETCC: |
| 1748 | return LowerSETCC(Op, DAG); |
| 1749 | case ISD::SHL: |
| 1750 | case ISD::SRA: |
| 1751 | case ISD::SRL: |
| 1752 | return LowerShift(Op, DAG); |
| 1753 | case ISD::FP_TO_SINT_SAT: |
| 1754 | case ISD::FP_TO_UINT_SAT: |
| 1755 | return LowerFP_TO_INT_SAT(Op, DAG); |
| 1756 | case ISD::FMINNUM: |
| 1757 | case ISD::FMINIMUMNUM: |
| 1758 | return LowerFMIN(Op, DAG); |
| 1759 | case ISD::FMAXNUM: |
| 1760 | case ISD::FMAXIMUMNUM: |
| 1761 | return LowerFMAX(Op, DAG); |
| 1762 | case ISD::LOAD: |
| 1763 | return LowerLoad(Op, DAG); |
| 1764 | case ISD::STORE: |
| 1765 | return LowerStore(Op, DAG); |
| 1766 | case ISD::CTPOP: |
| 1767 | case ISD::CTLZ: |
| 1768 | case ISD::CTTZ: |
| 1769 | return DAG.UnrollVectorOp(N: Op.getNode()); |
| 1770 | case ISD::CLEAR_CACHE: |
| 1771 | report_fatal_error(reason: "llvm.clear_cache is not supported on wasm" ); |
| 1772 | case ISD::SMUL_LOHI: |
| 1773 | case ISD::UMUL_LOHI: |
| 1774 | return LowerMUL_LOHI(Op, DAG); |
| 1775 | case ISD::UADDO: |
| 1776 | return LowerUADDO(Op, DAG); |
| 1777 | } |
| 1778 | } |
| 1779 | |
| 1780 | static bool IsWebAssemblyGlobal(SDValue Op) { |
| 1781 | if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op)) |
| 1782 | return WebAssembly::isWasmVarAddressSpace(AS: GA->getAddressSpace()); |
| 1783 | |
| 1784 | return false; |
| 1785 | } |
| 1786 | |
| 1787 | static std::optional<unsigned> IsWebAssemblyLocal(SDValue Op, |
| 1788 | SelectionDAG &DAG) { |
| 1789 | const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Op); |
| 1790 | if (!FI) |
| 1791 | return std::nullopt; |
| 1792 | |
| 1793 | auto &MF = DAG.getMachineFunction(); |
| 1794 | return WebAssemblyFrameLowering::getLocalForStackObject(MF, FrameIndex: FI->getIndex()); |
| 1795 | } |
| 1796 | |
| 1797 | SDValue WebAssemblyTargetLowering::LowerStore(SDValue Op, |
| 1798 | SelectionDAG &DAG) const { |
| 1799 | SDLoc DL(Op); |
| 1800 | StoreSDNode *SN = cast<StoreSDNode>(Val: Op.getNode()); |
| 1801 | const SDValue &Value = SN->getValue(); |
| 1802 | const SDValue &Base = SN->getBasePtr(); |
| 1803 | const SDValue &Offset = SN->getOffset(); |
| 1804 | |
| 1805 | if (IsWebAssemblyGlobal(Op: Base)) { |
| 1806 | if (!Offset->isUndef()) |
| 1807 | report_fatal_error(reason: "unexpected offset when storing to webassembly global" , |
| 1808 | gen_crash_diag: false); |
| 1809 | |
| 1810 | SDVTList Tys = DAG.getVTList(VT: MVT::Other); |
| 1811 | SDValue Ops[] = {SN->getChain(), Value, Base}; |
| 1812 | return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_SET, dl: DL, VTList: Tys, Ops, |
| 1813 | MemVT: SN->getMemoryVT(), MMO: SN->getMemOperand()); |
| 1814 | } |
| 1815 | |
| 1816 | if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) { |
| 1817 | if (!Offset->isUndef()) |
| 1818 | report_fatal_error(reason: "unexpected offset when storing to webassembly local" , |
| 1819 | gen_crash_diag: false); |
| 1820 | |
| 1821 | SDValue Idx = DAG.getTargetConstant(Val: *Local, DL: Base, VT: MVT::i32); |
| 1822 | SDVTList Tys = DAG.getVTList(VT: MVT::Other); // The chain. |
| 1823 | SDValue Ops[] = {SN->getChain(), Idx, Value}; |
| 1824 | return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_SET, DL, VTList: Tys, Ops); |
| 1825 | } |
| 1826 | |
| 1827 | if (WebAssembly::isWasmVarAddressSpace(AS: SN->getAddressSpace())) |
| 1828 | report_fatal_error( |
| 1829 | reason: "Encountered an unlowerable store to the wasm_var address space" , |
| 1830 | gen_crash_diag: false); |
| 1831 | |
| 1832 | return Op; |
| 1833 | } |
| 1834 | |
| 1835 | SDValue WebAssemblyTargetLowering::LowerLoad(SDValue Op, |
| 1836 | SelectionDAG &DAG) const { |
| 1837 | SDLoc DL(Op); |
| 1838 | LoadSDNode *LN = cast<LoadSDNode>(Val: Op.getNode()); |
| 1839 | const SDValue &Base = LN->getBasePtr(); |
| 1840 | const SDValue &Offset = LN->getOffset(); |
| 1841 | |
| 1842 | if (IsWebAssemblyGlobal(Op: Base)) { |
| 1843 | if (!Offset->isUndef()) |
| 1844 | report_fatal_error( |
| 1845 | reason: "unexpected offset when loading from webassembly global" , gen_crash_diag: false); |
| 1846 | |
| 1847 | SDVTList Tys = DAG.getVTList(VT1: LN->getValueType(ResNo: 0), VT2: MVT::Other); |
| 1848 | SDValue Ops[] = {LN->getChain(), Base}; |
| 1849 | return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_GET, dl: DL, VTList: Tys, Ops, |
| 1850 | MemVT: LN->getMemoryVT(), MMO: LN->getMemOperand()); |
| 1851 | } |
| 1852 | |
| 1853 | if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) { |
| 1854 | if (!Offset->isUndef()) |
| 1855 | report_fatal_error( |
| 1856 | reason: "unexpected offset when loading from webassembly local" , gen_crash_diag: false); |
| 1857 | |
| 1858 | SDValue Idx = DAG.getTargetConstant(Val: *Local, DL: Base, VT: MVT::i32); |
| 1859 | EVT LocalVT = LN->getValueType(ResNo: 0); |
| 1860 | return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_GET, DL, ResultTys: {LocalVT, MVT::Other}, |
| 1861 | Ops: {LN->getChain(), Idx}); |
| 1862 | } |
| 1863 | |
| 1864 | if (WebAssembly::isWasmVarAddressSpace(AS: LN->getAddressSpace())) |
| 1865 | report_fatal_error( |
| 1866 | reason: "Encountered an unlowerable load from the wasm_var address space" , |
| 1867 | gen_crash_diag: false); |
| 1868 | |
| 1869 | return Op; |
| 1870 | } |
| 1871 | |
| 1872 | SDValue WebAssemblyTargetLowering::LowerMUL_LOHI(SDValue Op, |
| 1873 | SelectionDAG &DAG) const { |
| 1874 | assert(Subtarget->hasWideArithmetic()); |
| 1875 | assert(Op.getValueType() == MVT::i64); |
| 1876 | SDLoc DL(Op); |
| 1877 | unsigned Opcode; |
| 1878 | switch (Op.getOpcode()) { |
| 1879 | case ISD::UMUL_LOHI: |
| 1880 | Opcode = WebAssemblyISD::I64_MUL_WIDE_U; |
| 1881 | break; |
| 1882 | case ISD::SMUL_LOHI: |
| 1883 | Opcode = WebAssemblyISD::I64_MUL_WIDE_S; |
| 1884 | break; |
| 1885 | default: |
| 1886 | llvm_unreachable("unexpected opcode" ); |
| 1887 | } |
| 1888 | SDValue LHS = Op.getOperand(i: 0); |
| 1889 | SDValue RHS = Op.getOperand(i: 1); |
| 1890 | SDValue Lo = |
| 1891 | DAG.getNode(Opcode, DL, VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), N1: LHS, N2: RHS); |
| 1892 | SDValue Hi(Lo.getNode(), 1); |
| 1893 | SDValue Ops[] = {Lo, Hi}; |
| 1894 | return DAG.getMergeValues(Ops, dl: DL); |
| 1895 | } |
| 1896 | |
| 1897 | // Lowers `UADDO` intrinsics to an `i64.add128` instruction when it's enabled. |
| 1898 | // |
| 1899 | // This enables generating a single wasm instruction for this operation where |
| 1900 | // the upper half of both operands are constant zeros. The upper half of the |
| 1901 | // result is then whether the overflow happened. |
| 1902 | SDValue WebAssemblyTargetLowering::LowerUADDO(SDValue Op, |
| 1903 | SelectionDAG &DAG) const { |
| 1904 | assert(Subtarget->hasWideArithmetic()); |
| 1905 | assert(Op.getValueType() == MVT::i64); |
| 1906 | assert(Op.getOpcode() == ISD::UADDO); |
| 1907 | SDLoc DL(Op); |
| 1908 | SDValue LHS = Op.getOperand(i: 0); |
| 1909 | SDValue RHS = Op.getOperand(i: 1); |
| 1910 | SDValue Zero = DAG.getConstant(Val: 0, DL, VT: MVT::i64); |
| 1911 | SDValue Result = |
| 1912 | DAG.getNode(Opcode: WebAssemblyISD::I64_ADD128, DL, |
| 1913 | VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), N1: LHS, N2: Zero, N3: RHS, N4: Zero); |
| 1914 | SDValue CarryI64(Result.getNode(), 1); |
| 1915 | SDValue CarryI32 = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: CarryI64); |
| 1916 | SDValue Ops[] = {Result, CarryI32}; |
| 1917 | return DAG.getMergeValues(Ops, dl: DL); |
| 1918 | } |
| 1919 | |
| 1920 | SDValue WebAssemblyTargetLowering::Replace128Op(SDNode *N, |
| 1921 | SelectionDAG &DAG) const { |
| 1922 | assert(Subtarget->hasWideArithmetic()); |
| 1923 | assert(N->getValueType(0) == MVT::i128); |
| 1924 | SDLoc DL(N); |
| 1925 | unsigned Opcode; |
| 1926 | switch (N->getOpcode()) { |
| 1927 | case ISD::ADD: |
| 1928 | Opcode = WebAssemblyISD::I64_ADD128; |
| 1929 | break; |
| 1930 | case ISD::SUB: |
| 1931 | Opcode = WebAssemblyISD::I64_SUB128; |
| 1932 | break; |
| 1933 | default: |
| 1934 | llvm_unreachable("unexpected opcode" ); |
| 1935 | } |
| 1936 | SDValue LHS = N->getOperand(Num: 0); |
| 1937 | SDValue RHS = N->getOperand(Num: 1); |
| 1938 | |
| 1939 | SDValue C0 = DAG.getConstant(Val: 0, DL, VT: MVT::i64); |
| 1940 | SDValue C1 = DAG.getConstant(Val: 1, DL, VT: MVT::i64); |
| 1941 | SDValue LHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: LHS, N2: C0); |
| 1942 | SDValue LHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: LHS, N2: C1); |
| 1943 | SDValue RHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: RHS, N2: C0); |
| 1944 | SDValue RHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: RHS, N2: C1); |
| 1945 | SDValue Result_LO = DAG.getNode(Opcode, DL, VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), |
| 1946 | N1: LHS_0, N2: LHS_1, N3: RHS_0, N4: RHS_1); |
| 1947 | SDValue Result_HI(Result_LO.getNode(), 1); |
| 1948 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VTList: N->getVTList(), N1: Result_LO, N2: Result_HI); |
| 1949 | } |
| 1950 | |
| 1951 | SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op, |
| 1952 | SelectionDAG &DAG) const { |
| 1953 | SDValue Src = Op.getOperand(i: 2); |
| 1954 | if (isa<FrameIndexSDNode>(Val: Src.getNode())) { |
| 1955 | // CopyToReg nodes don't support FrameIndex operands. Other targets select |
| 1956 | // the FI to some LEA-like instruction, but since we don't have that, we |
| 1957 | // need to insert some kind of instruction that can take an FI operand and |
| 1958 | // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy |
| 1959 | // local.copy between Op and its FI operand. |
| 1960 | SDValue Chain = Op.getOperand(i: 0); |
| 1961 | SDLoc DL(Op); |
| 1962 | Register Reg = cast<RegisterSDNode>(Val: Op.getOperand(i: 1))->getReg(); |
| 1963 | EVT VT = Src.getValueType(); |
| 1964 | SDValue Copy(DAG.getMachineNode(Opcode: VT == MVT::i32 ? WebAssembly::COPY_I32 |
| 1965 | : WebAssembly::COPY_I64, |
| 1966 | dl: DL, VT, Op1: Src), |
| 1967 | 0); |
| 1968 | return Op.getNode()->getNumValues() == 1 |
| 1969 | ? DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy) |
| 1970 | : DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy, |
| 1971 | Glue: Op.getNumOperands() == 4 ? Op.getOperand(i: 3) |
| 1972 | : SDValue()); |
| 1973 | } |
| 1974 | return SDValue(); |
| 1975 | } |
| 1976 | |
| 1977 | SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op, |
| 1978 | SelectionDAG &DAG) const { |
| 1979 | int FI = cast<FrameIndexSDNode>(Val&: Op)->getIndex(); |
| 1980 | return DAG.getTargetFrameIndex(FI, VT: Op.getValueType()); |
| 1981 | } |
| 1982 | |
| 1983 | SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op, |
| 1984 | SelectionDAG &DAG) const { |
| 1985 | SDLoc DL(Op); |
| 1986 | |
| 1987 | if (!Subtarget->getTargetTriple().isOSEmscripten()) { |
| 1988 | fail(DL, DAG, |
| 1989 | Msg: "Non-Emscripten WebAssembly hasn't implemented " |
| 1990 | "__builtin_return_address" ); |
| 1991 | return SDValue(); |
| 1992 | } |
| 1993 | |
| 1994 | unsigned Depth = Op.getConstantOperandVal(i: 0); |
| 1995 | MakeLibCallOptions CallOptions; |
| 1996 | return makeLibCall(DAG, LC: RTLIB::RETURN_ADDRESS, RetVT: Op.getValueType(), |
| 1997 | Ops: {DAG.getConstant(Val: Depth, DL, VT: MVT::i32)}, CallOptions, dl: DL) |
| 1998 | .first; |
| 1999 | } |
| 2000 | |
| 2001 | SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op, |
| 2002 | SelectionDAG &DAG) const { |
| 2003 | // Non-zero depths are not supported by WebAssembly currently. Use the |
| 2004 | // legalizer's default expansion, which is to return 0 (what this function is |
| 2005 | // documented to do). |
| 2006 | if (Op.getConstantOperandVal(i: 0) > 0) |
| 2007 | return SDValue(); |
| 2008 | |
| 2009 | DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true); |
| 2010 | EVT VT = Op.getValueType(); |
| 2011 | Register FP = |
| 2012 | Subtarget->getRegisterInfo()->getFrameRegister(MF: DAG.getMachineFunction()); |
| 2013 | return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), Reg: FP, VT); |
| 2014 | } |
| 2015 | |
| 2016 | SDValue |
| 2017 | WebAssemblyTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
| 2018 | SelectionDAG &DAG) const { |
| 2019 | SDLoc DL(Op); |
| 2020 | const auto *GA = cast<GlobalAddressSDNode>(Val&: Op); |
| 2021 | |
| 2022 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2023 | if (!MF.getSubtarget<WebAssemblySubtarget>().hasBulkMemory()) |
| 2024 | report_fatal_error(reason: "cannot use thread-local storage without bulk memory" , |
| 2025 | gen_crash_diag: false); |
| 2026 | |
| 2027 | const GlobalValue *GV = GA->getGlobal(); |
| 2028 | |
| 2029 | // Currently only Emscripten supports dynamic linking with threads. Therefore, |
| 2030 | // on other targets, if we have thread-local storage, only the local-exec |
| 2031 | // model is possible. |
| 2032 | auto model = Subtarget->getTargetTriple().isOSEmscripten() |
| 2033 | ? GV->getThreadLocalMode() |
| 2034 | : GlobalValue::LocalExecTLSModel; |
| 2035 | |
| 2036 | // Unsupported TLS modes |
| 2037 | assert(model != GlobalValue::NotThreadLocal); |
| 2038 | assert(model != GlobalValue::InitialExecTLSModel); |
| 2039 | |
| 2040 | if (model == GlobalValue::LocalExecTLSModel || |
| 2041 | model == GlobalValue::LocalDynamicTLSModel || |
| 2042 | (model == GlobalValue::GeneralDynamicTLSModel && |
| 2043 | getTargetMachine().shouldAssumeDSOLocal(GV))) { |
| 2044 | // For DSO-local TLS variables we use offset from __tls_base |
| 2045 | |
| 2046 | MVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2047 | auto GlobalGet = PtrVT == MVT::i64 ? WebAssembly::GLOBAL_GET_I64 |
| 2048 | : WebAssembly::GLOBAL_GET_I32; |
| 2049 | const char *BaseName = MF.createExternalSymbolName(Name: "__tls_base" ); |
| 2050 | |
| 2051 | SDValue BaseAddr( |
| 2052 | DAG.getMachineNode(Opcode: GlobalGet, dl: DL, VT: PtrVT, |
| 2053 | Op1: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)), |
| 2054 | 0); |
| 2055 | |
| 2056 | SDValue TLSOffset = DAG.getTargetGlobalAddress( |
| 2057 | GV, DL, VT: PtrVT, offset: GA->getOffset(), TargetFlags: WebAssemblyII::MO_TLS_BASE_REL); |
| 2058 | SDValue SymOffset = |
| 2059 | DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: TLSOffset); |
| 2060 | |
| 2061 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymOffset); |
| 2062 | } |
| 2063 | |
| 2064 | assert(model == GlobalValue::GeneralDynamicTLSModel); |
| 2065 | |
| 2066 | EVT VT = Op.getValueType(); |
| 2067 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
| 2068 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, |
| 2069 | offset: GA->getOffset(), |
| 2070 | TargetFlags: WebAssemblyII::MO_GOT_TLS)); |
| 2071 | } |
| 2072 | |
| 2073 | SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op, |
| 2074 | SelectionDAG &DAG) const { |
| 2075 | SDLoc DL(Op); |
| 2076 | const auto *GA = cast<GlobalAddressSDNode>(Val&: Op); |
| 2077 | EVT VT = Op.getValueType(); |
| 2078 | assert(GA->getTargetFlags() == 0 && |
| 2079 | "Unexpected target flags on generic GlobalAddressSDNode" ); |
| 2080 | if (!WebAssembly::isValidAddressSpace(AS: GA->getAddressSpace())) |
| 2081 | fail(DL, DAG, Msg: "Invalid address space for WebAssembly target" ); |
| 2082 | |
| 2083 | unsigned OperandFlags = 0; |
| 2084 | const GlobalValue *GV = GA->getGlobal(); |
| 2085 | // Since WebAssembly tables cannot yet be shared accross modules, we don't |
| 2086 | // need special treatment for tables in PIC mode. |
| 2087 | if (isPositionIndependent() && |
| 2088 | !WebAssembly::isWebAssemblyTableType(Ty: GV->getValueType())) { |
| 2089 | if (getTargetMachine().shouldAssumeDSOLocal(GV)) { |
| 2090 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2091 | MVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 2092 | const char *BaseName; |
| 2093 | if (GV->getValueType()->isFunctionTy()) { |
| 2094 | BaseName = MF.createExternalSymbolName(Name: "__table_base" ); |
| 2095 | OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL; |
| 2096 | } else { |
| 2097 | BaseName = MF.createExternalSymbolName(Name: "__memory_base" ); |
| 2098 | OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL; |
| 2099 | } |
| 2100 | SDValue BaseAddr = |
| 2101 | DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, |
| 2102 | Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)); |
| 2103 | |
| 2104 | SDValue SymAddr = DAG.getNode( |
| 2105 | Opcode: WebAssemblyISD::WrapperREL, DL, VT, |
| 2106 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, offset: GA->getOffset(), |
| 2107 | TargetFlags: OperandFlags)); |
| 2108 | |
| 2109 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: BaseAddr, N2: SymAddr); |
| 2110 | } |
| 2111 | OperandFlags = WebAssemblyII::MO_GOT; |
| 2112 | } |
| 2113 | |
| 2114 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
| 2115 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, |
| 2116 | offset: GA->getOffset(), TargetFlags: OperandFlags)); |
| 2117 | } |
| 2118 | |
| 2119 | SDValue |
| 2120 | WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op, |
| 2121 | SelectionDAG &DAG) const { |
| 2122 | SDLoc DL(Op); |
| 2123 | const auto *ES = cast<ExternalSymbolSDNode>(Val&: Op); |
| 2124 | EVT VT = Op.getValueType(); |
| 2125 | assert(ES->getTargetFlags() == 0 && |
| 2126 | "Unexpected target flags on generic ExternalSymbolSDNode" ); |
| 2127 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
| 2128 | Operand: DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), VT)); |
| 2129 | } |
| 2130 | |
| 2131 | SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op, |
| 2132 | SelectionDAG &DAG) const { |
| 2133 | // There's no need for a Wrapper node because we always incorporate a jump |
| 2134 | // table operand into a BR_TABLE instruction, rather than ever |
| 2135 | // materializing it in a register. |
| 2136 | const JumpTableSDNode *JT = cast<JumpTableSDNode>(Val&: Op); |
| 2137 | return DAG.getTargetJumpTable(JTI: JT->getIndex(), VT: Op.getValueType(), |
| 2138 | TargetFlags: JT->getTargetFlags()); |
| 2139 | } |
| 2140 | |
| 2141 | SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op, |
| 2142 | SelectionDAG &DAG) const { |
| 2143 | SDLoc DL(Op); |
| 2144 | SDValue Chain = Op.getOperand(i: 0); |
| 2145 | const auto *JT = cast<JumpTableSDNode>(Val: Op.getOperand(i: 1)); |
| 2146 | SDValue Index = Op.getOperand(i: 2); |
| 2147 | assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags" ); |
| 2148 | |
| 2149 | SmallVector<SDValue, 8> Ops; |
| 2150 | Ops.push_back(Elt: Chain); |
| 2151 | Ops.push_back(Elt: Index); |
| 2152 | |
| 2153 | MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo(); |
| 2154 | const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs; |
| 2155 | |
| 2156 | // Add an operand for each case. |
| 2157 | for (auto *MBB : MBBs) |
| 2158 | Ops.push_back(Elt: DAG.getBasicBlock(MBB)); |
| 2159 | |
| 2160 | // Add the first MBB as a dummy default target for now. This will be replaced |
| 2161 | // with the proper default target (and the preceding range check eliminated) |
| 2162 | // if possible by WebAssemblyFixBrTableDefaults. |
| 2163 | Ops.push_back(Elt: DAG.getBasicBlock(MBB: *MBBs.begin())); |
| 2164 | return DAG.getNode(Opcode: WebAssemblyISD::BR_TABLE, DL, VT: MVT::Other, Ops); |
| 2165 | } |
| 2166 | |
| 2167 | SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op, |
| 2168 | SelectionDAG &DAG) const { |
| 2169 | SDLoc DL(Op); |
| 2170 | EVT PtrVT = getPointerTy(DL: DAG.getMachineFunction().getDataLayout()); |
| 2171 | |
| 2172 | auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>(); |
| 2173 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
| 2174 | |
| 2175 | SDValue ArgN = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: DL, |
| 2176 | Reg: MFI->getVarargBufferVreg(), VT: PtrVT); |
| 2177 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: ArgN, Ptr: Op.getOperand(i: 1), |
| 2178 | PtrInfo: MachinePointerInfo(SV)); |
| 2179 | } |
| 2180 | |
| 2181 | SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op, |
| 2182 | SelectionDAG &DAG) const { |
| 2183 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2184 | unsigned IntNo; |
| 2185 | switch (Op.getOpcode()) { |
| 2186 | case ISD::INTRINSIC_VOID: |
| 2187 | case ISD::INTRINSIC_W_CHAIN: |
| 2188 | IntNo = Op.getConstantOperandVal(i: 1); |
| 2189 | break; |
| 2190 | case ISD::INTRINSIC_WO_CHAIN: |
| 2191 | IntNo = Op.getConstantOperandVal(i: 0); |
| 2192 | break; |
| 2193 | default: |
| 2194 | llvm_unreachable("Invalid intrinsic" ); |
| 2195 | } |
| 2196 | SDLoc DL(Op); |
| 2197 | |
| 2198 | switch (IntNo) { |
| 2199 | default: |
| 2200 | return SDValue(); // Don't custom lower most intrinsics. |
| 2201 | |
| 2202 | case Intrinsic::wasm_lsda: { |
| 2203 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 2204 | const char *SymName = MF.createExternalSymbolName( |
| 2205 | Name: "GCC_except_table" + std::to_string(val: MF.getFunctionNumber())); |
| 2206 | if (isPositionIndependent()) { |
| 2207 | SDValue Node = DAG.getTargetExternalSymbol( |
| 2208 | Sym: SymName, VT: PtrVT, TargetFlags: WebAssemblyII::MO_MEMORY_BASE_REL); |
| 2209 | const char *BaseName = MF.createExternalSymbolName(Name: "__memory_base" ); |
| 2210 | SDValue BaseAddr = |
| 2211 | DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, |
| 2212 | Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)); |
| 2213 | SDValue SymAddr = |
| 2214 | DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: Node); |
| 2215 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymAddr); |
| 2216 | } |
| 2217 | SDValue Node = DAG.getTargetExternalSymbol(Sym: SymName, VT: PtrVT); |
| 2218 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, Operand: Node); |
| 2219 | } |
| 2220 | |
| 2221 | case Intrinsic::wasm_shuffle: { |
| 2222 | // Drop in-chain and replace undefs, but otherwise pass through unchanged |
| 2223 | SDValue Ops[18]; |
| 2224 | size_t OpIdx = 0; |
| 2225 | Ops[OpIdx++] = Op.getOperand(i: 1); |
| 2226 | Ops[OpIdx++] = Op.getOperand(i: 2); |
| 2227 | while (OpIdx < 18) { |
| 2228 | const SDValue &MaskIdx = Op.getOperand(i: OpIdx + 1); |
| 2229 | if (MaskIdx.isUndef() || MaskIdx.getNode()->getAsZExtVal() >= 32) { |
| 2230 | bool isTarget = MaskIdx.getNode()->getOpcode() == ISD::TargetConstant; |
| 2231 | Ops[OpIdx++] = DAG.getConstant(Val: 0, DL, VT: MVT::i32, isTarget); |
| 2232 | } else { |
| 2233 | Ops[OpIdx++] = MaskIdx; |
| 2234 | } |
| 2235 | } |
| 2236 | return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops); |
| 2237 | } |
| 2238 | |
| 2239 | case Intrinsic::thread_pointer: { |
| 2240 | MVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2241 | auto GlobalGet = PtrVT == MVT::i64 ? WebAssembly::GLOBAL_GET_I64 |
| 2242 | : WebAssembly::GLOBAL_GET_I32; |
| 2243 | const char *TlsBase = MF.createExternalSymbolName(Name: "__tls_base" ); |
| 2244 | return SDValue( |
| 2245 | DAG.getMachineNode(Opcode: GlobalGet, dl: DL, VT: PtrVT, |
| 2246 | Op1: DAG.getTargetExternalSymbol(Sym: TlsBase, VT: PtrVT)), |
| 2247 | 0); |
| 2248 | } |
| 2249 | } |
| 2250 | } |
| 2251 | |
| 2252 | SDValue |
| 2253 | WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, |
| 2254 | SelectionDAG &DAG) const { |
| 2255 | SDLoc DL(Op); |
| 2256 | // If sign extension operations are disabled, allow sext_inreg only if operand |
| 2257 | // is a vector extract of an i8 or i16 lane. SIMD does not depend on sign |
| 2258 | // extension operations, but allowing sext_inreg in this context lets us have |
| 2259 | // simple patterns to select extract_lane_s instructions. Expanding sext_inreg |
| 2260 | // everywhere would be simpler in this file, but would necessitate large and |
| 2261 | // brittle patterns to undo the expansion and select extract_lane_s |
| 2262 | // instructions. |
| 2263 | assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128()); |
| 2264 | if (Op.getOperand(i: 0).getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2265 | return SDValue(); |
| 2266 | |
| 2267 | const SDValue & = Op.getOperand(i: 0); |
| 2268 | MVT VecT = Extract.getOperand(i: 0).getSimpleValueType(); |
| 2269 | if (VecT.getVectorElementType().getSizeInBits() > 32) |
| 2270 | return SDValue(); |
| 2271 | MVT = |
| 2272 | cast<VTSDNode>(Val: Op.getOperand(i: 1).getNode())->getVT().getSimpleVT(); |
| 2273 | MVT = |
| 2274 | MVT::getVectorVT(VT: ExtractedLaneT, NumElements: 128 / ExtractedLaneT.getSizeInBits()); |
| 2275 | if (ExtractedVecT == VecT) |
| 2276 | return Op; |
| 2277 | |
| 2278 | // Bitcast vector to appropriate type to ensure ISel pattern coverage |
| 2279 | const SDNode *Index = Extract.getOperand(i: 1).getNode(); |
| 2280 | if (!isa<ConstantSDNode>(Val: Index)) |
| 2281 | return SDValue(); |
| 2282 | unsigned IndexVal = Index->getAsZExtVal(); |
| 2283 | unsigned Scale = |
| 2284 | ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements(); |
| 2285 | assert(Scale > 1); |
| 2286 | SDValue NewIndex = |
| 2287 | DAG.getConstant(Val: IndexVal * Scale, DL, VT: Index->getValueType(ResNo: 0)); |
| 2288 | SDValue = DAG.getNode( |
| 2289 | Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: Extract.getValueType(), |
| 2290 | N1: DAG.getBitcast(VT: ExtractedVecT, V: Extract.getOperand(i: 0)), N2: NewIndex); |
| 2291 | return DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: Op.getValueType(), N1: NewExtract, |
| 2292 | N2: Op.getOperand(i: 1)); |
| 2293 | } |
| 2294 | |
| 2295 | static SDValue GetExtendHigh(SDValue Op, unsigned UserOpc, EVT VT, |
| 2296 | SelectionDAG &DAG) { |
| 2297 | SDValue Source = peekThroughBitcasts(V: Op); |
| 2298 | if (Source.getOpcode() != ISD::VECTOR_SHUFFLE) |
| 2299 | return SDValue(); |
| 2300 | |
| 2301 | assert((UserOpc == WebAssemblyISD::EXTEND_LOW_U || |
| 2302 | UserOpc == WebAssemblyISD::EXTEND_LOW_S) && |
| 2303 | "expected extend_low" ); |
| 2304 | auto *Shuffle = cast<ShuffleVectorSDNode>(Val: Source.getNode()); |
| 2305 | |
| 2306 | ArrayRef<int> Mask = Shuffle->getMask(); |
| 2307 | // Look for a shuffle which moves from the high half to the low half. |
| 2308 | size_t FirstIdx = Mask.size() / 2; |
| 2309 | for (size_t i = 0; i < Mask.size() / 2; ++i) { |
| 2310 | if (Mask[i] != static_cast<int>(FirstIdx + i)) { |
| 2311 | return SDValue(); |
| 2312 | } |
| 2313 | } |
| 2314 | |
| 2315 | SDLoc DL(Op); |
| 2316 | unsigned Opc = UserOpc == WebAssemblyISD::EXTEND_LOW_S |
| 2317 | ? WebAssemblyISD::EXTEND_HIGH_S |
| 2318 | : WebAssemblyISD::EXTEND_HIGH_U; |
| 2319 | SDValue ShuffleSrc = Shuffle->getOperand(Num: 0); |
| 2320 | if (Op.getOpcode() == ISD::BITCAST) |
| 2321 | ShuffleSrc = DAG.getBitcast(VT: Op.getValueType(), V: ShuffleSrc); |
| 2322 | |
| 2323 | return DAG.getNode(Opcode: Opc, DL, VT, Operand: ShuffleSrc); |
| 2324 | } |
| 2325 | |
| 2326 | SDValue |
| 2327 | WebAssemblyTargetLowering::LowerEXTEND_VECTOR_INREG(SDValue Op, |
| 2328 | SelectionDAG &DAG) const { |
| 2329 | SDLoc DL(Op); |
| 2330 | EVT VT = Op.getValueType(); |
| 2331 | SDValue Src = Op.getOperand(i: 0); |
| 2332 | EVT SrcVT = Src.getValueType(); |
| 2333 | |
| 2334 | if (SrcVT.getVectorElementType() == MVT::i1 || |
| 2335 | SrcVT.getVectorElementType() == MVT::i64) |
| 2336 | return SDValue(); |
| 2337 | |
| 2338 | assert(VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits() == 0 && |
| 2339 | "Unexpected extension factor." ); |
| 2340 | unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits(); |
| 2341 | |
| 2342 | if (Scale != 2 && Scale != 4 && Scale != 8) |
| 2343 | return SDValue(); |
| 2344 | |
| 2345 | unsigned Ext; |
| 2346 | switch (Op.getOpcode()) { |
| 2347 | default: |
| 2348 | llvm_unreachable("unexpected opcode" ); |
| 2349 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 2350 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 2351 | Ext = WebAssemblyISD::EXTEND_LOW_U; |
| 2352 | break; |
| 2353 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 2354 | Ext = WebAssemblyISD::EXTEND_LOW_S; |
| 2355 | break; |
| 2356 | } |
| 2357 | |
| 2358 | if (Scale == 2) { |
| 2359 | // See if we can use EXTEND_HIGH. |
| 2360 | if (auto ExtendHigh = GetExtendHigh(Op: Op.getOperand(i: 0), UserOpc: Ext, VT, DAG)) |
| 2361 | return ExtendHigh; |
| 2362 | } |
| 2363 | |
| 2364 | SDValue Ret = Src; |
| 2365 | while (Scale != 1) { |
| 2366 | Ret = DAG.getNode(Opcode: Ext, DL, |
| 2367 | VT: Ret.getValueType() |
| 2368 | .widenIntegerVectorElementType(Context&: *DAG.getContext()) |
| 2369 | .getHalfNumVectorElementsVT(Context&: *DAG.getContext()), |
| 2370 | Operand: Ret); |
| 2371 | Scale /= 2; |
| 2372 | } |
| 2373 | assert(Ret.getValueType() == VT); |
| 2374 | return Ret; |
| 2375 | } |
| 2376 | |
| 2377 | static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) { |
| 2378 | SDLoc DL(Op); |
| 2379 | if (Op.getValueType() != MVT::v2f64) |
| 2380 | return SDValue(); |
| 2381 | |
| 2382 | auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec, |
| 2383 | unsigned &Index) -> bool { |
| 2384 | switch (Op.getOpcode()) { |
| 2385 | case ISD::SINT_TO_FP: |
| 2386 | Opcode = WebAssemblyISD::CONVERT_LOW_S; |
| 2387 | break; |
| 2388 | case ISD::UINT_TO_FP: |
| 2389 | Opcode = WebAssemblyISD::CONVERT_LOW_U; |
| 2390 | break; |
| 2391 | case ISD::FP_EXTEND: |
| 2392 | Opcode = WebAssemblyISD::PROMOTE_LOW; |
| 2393 | break; |
| 2394 | default: |
| 2395 | return false; |
| 2396 | } |
| 2397 | |
| 2398 | auto = Op.getOperand(i: 0); |
| 2399 | if (ExtractVector.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2400 | return false; |
| 2401 | |
| 2402 | if (!isa<ConstantSDNode>(Val: ExtractVector.getOperand(i: 1).getNode())) |
| 2403 | return false; |
| 2404 | |
| 2405 | SrcVec = ExtractVector.getOperand(i: 0); |
| 2406 | Index = ExtractVector.getConstantOperandVal(i: 1); |
| 2407 | return true; |
| 2408 | }; |
| 2409 | |
| 2410 | unsigned LHSOpcode, RHSOpcode, LHSIndex, RHSIndex; |
| 2411 | SDValue LHSSrcVec, RHSSrcVec; |
| 2412 | if (!GetConvertedLane(Op.getOperand(i: 0), LHSOpcode, LHSSrcVec, LHSIndex) || |
| 2413 | !GetConvertedLane(Op.getOperand(i: 1), RHSOpcode, RHSSrcVec, RHSIndex)) |
| 2414 | return SDValue(); |
| 2415 | |
| 2416 | if (LHSOpcode != RHSOpcode) |
| 2417 | return SDValue(); |
| 2418 | |
| 2419 | MVT ExpectedSrcVT; |
| 2420 | switch (LHSOpcode) { |
| 2421 | case WebAssemblyISD::CONVERT_LOW_S: |
| 2422 | case WebAssemblyISD::CONVERT_LOW_U: |
| 2423 | ExpectedSrcVT = MVT::v4i32; |
| 2424 | break; |
| 2425 | case WebAssemblyISD::PROMOTE_LOW: |
| 2426 | ExpectedSrcVT = MVT::v4f32; |
| 2427 | break; |
| 2428 | } |
| 2429 | if (LHSSrcVec.getValueType() != ExpectedSrcVT) |
| 2430 | return SDValue(); |
| 2431 | |
| 2432 | auto Src = LHSSrcVec; |
| 2433 | if (LHSIndex != 0 || RHSIndex != 1 || LHSSrcVec != RHSSrcVec) { |
| 2434 | // Shuffle the source vector so that the converted lanes are the low lanes. |
| 2435 | Src = DAG.getVectorShuffle( |
| 2436 | VT: ExpectedSrcVT, dl: DL, N1: LHSSrcVec, N2: RHSSrcVec, |
| 2437 | Mask: {static_cast<int>(LHSIndex), static_cast<int>(RHSIndex) + 4, -1, -1}); |
| 2438 | } |
| 2439 | return DAG.getNode(Opcode: LHSOpcode, DL, VT: MVT::v2f64, Operand: Src); |
| 2440 | } |
| 2441 | |
| 2442 | SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op, |
| 2443 | SelectionDAG &DAG) const { |
| 2444 | MVT VT = Op.getSimpleValueType(); |
| 2445 | if (VT == MVT::v8f16) { |
| 2446 | // BUILD_VECTOR can't handle FP16 operands since Wasm doesn't have a scaler |
| 2447 | // FP16 type, so cast them to I16s. |
| 2448 | MVT IVT = VT.changeVectorElementType(EltVT: MVT::i16); |
| 2449 | SmallVector<SDValue, 8> NewOps; |
| 2450 | for (unsigned I = 0, E = Op.getNumOperands(); I < E; ++I) |
| 2451 | NewOps.push_back(Elt: DAG.getBitcast(VT: MVT::i16, V: Op.getOperand(i: I))); |
| 2452 | SDValue Res = DAG.getNode(Opcode: ISD::BUILD_VECTOR, DL: SDLoc(), VT: IVT, Ops: NewOps); |
| 2453 | return DAG.getBitcast(VT, V: Res); |
| 2454 | } |
| 2455 | |
| 2456 | if (auto ConvertLow = LowerConvertLow(Op, DAG)) |
| 2457 | return ConvertLow; |
| 2458 | |
| 2459 | SDLoc DL(Op); |
| 2460 | const EVT VecT = Op.getValueType(); |
| 2461 | const EVT LaneT = Op.getOperand(i: 0).getValueType(); |
| 2462 | const size_t Lanes = Op.getNumOperands(); |
| 2463 | bool CanSwizzle = VecT == MVT::v16i8; |
| 2464 | |
| 2465 | // BUILD_VECTORs are lowered to the instruction that initializes the highest |
| 2466 | // possible number of lanes at once followed by a sequence of replace_lane |
| 2467 | // instructions to individually initialize any remaining lanes. |
| 2468 | |
| 2469 | // TODO: Tune this. For example, lanewise swizzling is very expensive, so |
| 2470 | // swizzled lanes should be given greater weight. |
| 2471 | |
| 2472 | // TODO: Investigate looping rather than always extracting/replacing specific |
| 2473 | // lanes to fill gaps. |
| 2474 | |
| 2475 | auto IsConstant = [](const SDValue &V) { |
| 2476 | return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP; |
| 2477 | }; |
| 2478 | |
| 2479 | // Returns the source vector and index vector pair if they exist. Checks for: |
| 2480 | // (extract_vector_elt |
| 2481 | // $src, |
| 2482 | // (sign_extend_inreg (extract_vector_elt $indices, $i)) |
| 2483 | // ) |
| 2484 | auto GetSwizzleSrcs = [](size_t I, const SDValue &Lane) { |
| 2485 | auto Bail = std::make_pair(x: SDValue(), y: SDValue()); |
| 2486 | if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2487 | return Bail; |
| 2488 | const SDValue &SwizzleSrc = Lane->getOperand(Num: 0); |
| 2489 | const SDValue &IndexExt = Lane->getOperand(Num: 1); |
| 2490 | if (IndexExt->getOpcode() != ISD::SIGN_EXTEND_INREG) |
| 2491 | return Bail; |
| 2492 | const SDValue &Index = IndexExt->getOperand(Num: 0); |
| 2493 | if (Index->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2494 | return Bail; |
| 2495 | const SDValue &SwizzleIndices = Index->getOperand(Num: 0); |
| 2496 | if (SwizzleSrc.getValueType() != MVT::v16i8 || |
| 2497 | SwizzleIndices.getValueType() != MVT::v16i8 || |
| 2498 | Index->getOperand(Num: 1)->getOpcode() != ISD::Constant || |
| 2499 | Index->getConstantOperandVal(Num: 1) != I) |
| 2500 | return Bail; |
| 2501 | return std::make_pair(x: SwizzleSrc, y: SwizzleIndices); |
| 2502 | }; |
| 2503 | |
| 2504 | // If the lane is extracted from another vector at a constant index, return |
| 2505 | // that vector. The source vector must not have more lanes than the dest |
| 2506 | // because the shufflevector indices are in terms of the destination lanes and |
| 2507 | // would not be able to address the smaller individual source lanes. |
| 2508 | auto GetShuffleSrc = [&](const SDValue &Lane) { |
| 2509 | if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2510 | return SDValue(); |
| 2511 | if (!isa<ConstantSDNode>(Val: Lane->getOperand(Num: 1).getNode())) |
| 2512 | return SDValue(); |
| 2513 | if (Lane->getOperand(Num: 0).getValueType().getVectorNumElements() > |
| 2514 | VecT.getVectorNumElements()) |
| 2515 | return SDValue(); |
| 2516 | return Lane->getOperand(Num: 0); |
| 2517 | }; |
| 2518 | |
| 2519 | using ValueEntry = std::pair<SDValue, size_t>; |
| 2520 | SmallVector<ValueEntry, 16> SplatValueCounts; |
| 2521 | |
| 2522 | using SwizzleEntry = std::pair<std::pair<SDValue, SDValue>, size_t>; |
| 2523 | SmallVector<SwizzleEntry, 16> SwizzleCounts; |
| 2524 | |
| 2525 | using ShuffleEntry = std::pair<SDValue, size_t>; |
| 2526 | SmallVector<ShuffleEntry, 16> ShuffleCounts; |
| 2527 | |
| 2528 | auto AddCount = [](auto &Counts, const auto &Val) { |
| 2529 | auto CountIt = |
| 2530 | llvm::find_if(Counts, [&Val](auto E) { return E.first == Val; }); |
| 2531 | if (CountIt == Counts.end()) { |
| 2532 | Counts.emplace_back(Val, 1); |
| 2533 | } else { |
| 2534 | CountIt->second++; |
| 2535 | } |
| 2536 | }; |
| 2537 | |
| 2538 | auto GetMostCommon = [](auto &Counts) { |
| 2539 | auto CommonIt = llvm::max_element(Counts, llvm::less_second()); |
| 2540 | assert(CommonIt != Counts.end() && "Unexpected all-undef build_vector" ); |
| 2541 | return *CommonIt; |
| 2542 | }; |
| 2543 | |
| 2544 | size_t NumConstantLanes = 0; |
| 2545 | |
| 2546 | // Count eligible lanes for each type of vector creation op |
| 2547 | for (size_t I = 0; I < Lanes; ++I) { |
| 2548 | const SDValue &Lane = Op->getOperand(Num: I); |
| 2549 | if (Lane.isUndef()) |
| 2550 | continue; |
| 2551 | |
| 2552 | AddCount(SplatValueCounts, Lane); |
| 2553 | |
| 2554 | if (IsConstant(Lane)) |
| 2555 | NumConstantLanes++; |
| 2556 | if (auto ShuffleSrc = GetShuffleSrc(Lane)) |
| 2557 | AddCount(ShuffleCounts, ShuffleSrc); |
| 2558 | if (CanSwizzle) { |
| 2559 | auto SwizzleSrcs = GetSwizzleSrcs(I, Lane); |
| 2560 | if (SwizzleSrcs.first) |
| 2561 | AddCount(SwizzleCounts, SwizzleSrcs); |
| 2562 | } |
| 2563 | } |
| 2564 | |
| 2565 | SDValue SplatValue; |
| 2566 | size_t NumSplatLanes; |
| 2567 | std::tie(args&: SplatValue, args&: NumSplatLanes) = GetMostCommon(SplatValueCounts); |
| 2568 | |
| 2569 | SDValue SwizzleSrc; |
| 2570 | SDValue SwizzleIndices; |
| 2571 | size_t NumSwizzleLanes = 0; |
| 2572 | if (SwizzleCounts.size()) |
| 2573 | std::forward_as_tuple(args: std::tie(args&: SwizzleSrc, args&: SwizzleIndices), |
| 2574 | args&: NumSwizzleLanes) = GetMostCommon(SwizzleCounts); |
| 2575 | |
| 2576 | // Shuffles can draw from up to two vectors, so find the two most common |
| 2577 | // sources. |
| 2578 | SDValue ShuffleSrc1, ShuffleSrc2; |
| 2579 | size_t NumShuffleLanes = 0; |
| 2580 | if (ShuffleCounts.size()) { |
| 2581 | std::tie(args&: ShuffleSrc1, args&: NumShuffleLanes) = GetMostCommon(ShuffleCounts); |
| 2582 | llvm::erase_if(C&: ShuffleCounts, |
| 2583 | P: [&](const auto &Pair) { return Pair.first == ShuffleSrc1; }); |
| 2584 | } |
| 2585 | if (ShuffleCounts.size()) { |
| 2586 | size_t AdditionalShuffleLanes; |
| 2587 | std::tie(args&: ShuffleSrc2, args&: AdditionalShuffleLanes) = |
| 2588 | GetMostCommon(ShuffleCounts); |
| 2589 | NumShuffleLanes += AdditionalShuffleLanes; |
| 2590 | } |
| 2591 | |
| 2592 | // Predicate returning true if the lane is properly initialized by the |
| 2593 | // original instruction |
| 2594 | std::function<bool(size_t, const SDValue &)> IsLaneConstructed; |
| 2595 | SDValue Result; |
| 2596 | // Prefer swizzles over shuffles over vector consts over splats |
| 2597 | if (NumSwizzleLanes >= NumShuffleLanes && |
| 2598 | NumSwizzleLanes >= NumConstantLanes && NumSwizzleLanes >= NumSplatLanes) { |
| 2599 | Result = DAG.getNode(Opcode: WebAssemblyISD::SWIZZLE, DL, VT: VecT, N1: SwizzleSrc, |
| 2600 | N2: SwizzleIndices); |
| 2601 | auto Swizzled = std::make_pair(x&: SwizzleSrc, y&: SwizzleIndices); |
| 2602 | IsLaneConstructed = [&, Swizzled](size_t I, const SDValue &Lane) { |
| 2603 | return Swizzled == GetSwizzleSrcs(I, Lane); |
| 2604 | }; |
| 2605 | } else if (NumShuffleLanes >= NumConstantLanes && |
| 2606 | NumShuffleLanes >= NumSplatLanes) { |
| 2607 | size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits() / 8; |
| 2608 | size_t DestLaneCount = VecT.getVectorNumElements(); |
| 2609 | size_t Scale1 = 1; |
| 2610 | size_t Scale2 = 1; |
| 2611 | SDValue Src1 = ShuffleSrc1; |
| 2612 | SDValue Src2 = ShuffleSrc2 ? ShuffleSrc2 : DAG.getUNDEF(VT: VecT); |
| 2613 | if (Src1.getValueType() != VecT) { |
| 2614 | size_t LaneSize = |
| 2615 | Src1.getValueType().getVectorElementType().getFixedSizeInBits() / 8; |
| 2616 | assert(LaneSize > DestLaneSize); |
| 2617 | Scale1 = LaneSize / DestLaneSize; |
| 2618 | Src1 = DAG.getBitcast(VT: VecT, V: Src1); |
| 2619 | } |
| 2620 | if (Src2.getValueType() != VecT) { |
| 2621 | size_t LaneSize = |
| 2622 | Src2.getValueType().getVectorElementType().getFixedSizeInBits() / 8; |
| 2623 | assert(LaneSize > DestLaneSize); |
| 2624 | Scale2 = LaneSize / DestLaneSize; |
| 2625 | Src2 = DAG.getBitcast(VT: VecT, V: Src2); |
| 2626 | } |
| 2627 | |
| 2628 | int Mask[16]; |
| 2629 | assert(DestLaneCount <= 16); |
| 2630 | for (size_t I = 0; I < DestLaneCount; ++I) { |
| 2631 | const SDValue &Lane = Op->getOperand(Num: I); |
| 2632 | SDValue Src = GetShuffleSrc(Lane); |
| 2633 | if (Src == ShuffleSrc1) { |
| 2634 | Mask[I] = Lane->getConstantOperandVal(Num: 1) * Scale1; |
| 2635 | } else if (Src && Src == ShuffleSrc2) { |
| 2636 | Mask[I] = DestLaneCount + Lane->getConstantOperandVal(Num: 1) * Scale2; |
| 2637 | } else { |
| 2638 | Mask[I] = -1; |
| 2639 | } |
| 2640 | } |
| 2641 | ArrayRef<int> MaskRef(Mask, DestLaneCount); |
| 2642 | Result = DAG.getVectorShuffle(VT: VecT, dl: DL, N1: Src1, N2: Src2, Mask: MaskRef); |
| 2643 | IsLaneConstructed = [&](size_t, const SDValue &Lane) { |
| 2644 | auto Src = GetShuffleSrc(Lane); |
| 2645 | return Src == ShuffleSrc1 || (Src && Src == ShuffleSrc2); |
| 2646 | }; |
| 2647 | } else if (NumConstantLanes >= NumSplatLanes) { |
| 2648 | SmallVector<SDValue, 16> ConstLanes; |
| 2649 | for (const SDValue &Lane : Op->op_values()) { |
| 2650 | if (IsConstant(Lane)) { |
| 2651 | // Values may need to be fixed so that they will sign extend to be |
| 2652 | // within the expected range during ISel. Check whether the value is in |
| 2653 | // bounds based on the lane bit width and if it is out of bounds, lop |
| 2654 | // off the extra bits. |
| 2655 | uint64_t LaneBits = 128 / Lanes; |
| 2656 | if (auto *Const = dyn_cast<ConstantSDNode>(Val: Lane.getNode())) { |
| 2657 | ConstLanes.push_back(Elt: DAG.getConstant( |
| 2658 | Val: Const->getAPIntValue().trunc(width: LaneBits).getZExtValue(), |
| 2659 | DL: SDLoc(Lane), VT: LaneT)); |
| 2660 | } else { |
| 2661 | ConstLanes.push_back(Elt: Lane); |
| 2662 | } |
| 2663 | } else if (LaneT.isFloatingPoint()) { |
| 2664 | ConstLanes.push_back(Elt: DAG.getConstantFP(Val: 0, DL, VT: LaneT)); |
| 2665 | } else { |
| 2666 | ConstLanes.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: LaneT)); |
| 2667 | } |
| 2668 | } |
| 2669 | Result = DAG.getBuildVector(VT: VecT, DL, Ops: ConstLanes); |
| 2670 | IsLaneConstructed = [&IsConstant](size_t _, const SDValue &Lane) { |
| 2671 | return IsConstant(Lane); |
| 2672 | }; |
| 2673 | } else { |
| 2674 | size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits(); |
| 2675 | if (NumSplatLanes == 1 && Op->getOperand(Num: 0) == SplatValue && |
| 2676 | (DestLaneSize == 32 || DestLaneSize == 64)) { |
| 2677 | // Could be selected to load_zero. |
| 2678 | Result = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL, VT: VecT, Operand: SplatValue); |
| 2679 | } else { |
| 2680 | // Use a splat (which might be selected as a load splat) |
| 2681 | Result = DAG.getSplatBuildVector(VT: VecT, DL, Op: SplatValue); |
| 2682 | } |
| 2683 | IsLaneConstructed = [&SplatValue](size_t _, const SDValue &Lane) { |
| 2684 | return Lane == SplatValue; |
| 2685 | }; |
| 2686 | } |
| 2687 | |
| 2688 | assert(Result); |
| 2689 | assert(IsLaneConstructed); |
| 2690 | |
| 2691 | // Add replace_lane instructions for any unhandled values |
| 2692 | for (size_t I = 0; I < Lanes; ++I) { |
| 2693 | const SDValue &Lane = Op->getOperand(Num: I); |
| 2694 | if (!Lane.isUndef() && !IsLaneConstructed(I, Lane)) |
| 2695 | Result = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL, VT: VecT, N1: Result, N2: Lane, |
| 2696 | N3: DAG.getConstant(Val: I, DL, VT: MVT::i32)); |
| 2697 | } |
| 2698 | |
| 2699 | return Result; |
| 2700 | } |
| 2701 | |
| 2702 | SDValue |
| 2703 | WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, |
| 2704 | SelectionDAG &DAG) const { |
| 2705 | SDLoc DL(Op); |
| 2706 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: Op.getNode())->getMask(); |
| 2707 | MVT VecType = Op.getOperand(i: 0).getSimpleValueType(); |
| 2708 | assert(VecType.is128BitVector() && "Unexpected shuffle vector type" ); |
| 2709 | size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8; |
| 2710 | |
| 2711 | // Space for two vector args and sixteen mask indices |
| 2712 | SDValue Ops[18]; |
| 2713 | size_t OpIdx = 0; |
| 2714 | Ops[OpIdx++] = Op.getOperand(i: 0); |
| 2715 | Ops[OpIdx++] = Op.getOperand(i: 1); |
| 2716 | |
| 2717 | // Expand mask indices to byte indices and materialize them as operands |
| 2718 | for (int M : Mask) { |
| 2719 | for (size_t J = 0; J < LaneBytes; ++J) { |
| 2720 | // Lower undefs (represented by -1 in mask) to {0..J}, which use a |
| 2721 | // whole lane of vector input, to allow further reduction at VM. E.g. |
| 2722 | // match an 8x16 byte shuffle to an equivalent cheaper 32x4 shuffle. |
| 2723 | uint64_t ByteIndex = M == -1 ? J : (uint64_t)M * LaneBytes + J; |
| 2724 | Ops[OpIdx++] = DAG.getConstant(Val: ByteIndex, DL, VT: MVT::i32); |
| 2725 | } |
| 2726 | } |
| 2727 | |
| 2728 | return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops); |
| 2729 | } |
| 2730 | |
| 2731 | SDValue WebAssemblyTargetLowering::LowerSETCC(SDValue Op, |
| 2732 | SelectionDAG &DAG) const { |
| 2733 | SDLoc DL(Op); |
| 2734 | // The legalizer does not know how to expand the unsupported comparison modes |
| 2735 | // of i64x2 vectors, so we manually unroll them here. |
| 2736 | assert(Op->getOperand(0)->getSimpleValueType(0) == MVT::v2i64); |
| 2737 | SmallVector<SDValue, 2> LHS, RHS; |
| 2738 | DAG.ExtractVectorElements(Op: Op->getOperand(Num: 0), Args&: LHS); |
| 2739 | DAG.ExtractVectorElements(Op: Op->getOperand(Num: 1), Args&: RHS); |
| 2740 | const SDValue &CC = Op->getOperand(Num: 2); |
| 2741 | auto MakeLane = [&](unsigned I) { |
| 2742 | return DAG.getNode(Opcode: ISD::SELECT_CC, DL, VT: MVT::i64, N1: LHS[I], N2: RHS[I], |
| 2743 | N3: DAG.getConstant(Val: uint64_t(-1), DL, VT: MVT::i64), |
| 2744 | N4: DAG.getConstant(Val: uint64_t(0), DL, VT: MVT::i64), N5: CC); |
| 2745 | }; |
| 2746 | return DAG.getBuildVector(VT: Op->getValueType(ResNo: 0), DL, |
| 2747 | Ops: {MakeLane(0), MakeLane(1)}); |
| 2748 | } |
| 2749 | |
| 2750 | SDValue |
| 2751 | WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op, |
| 2752 | SelectionDAG &DAG) const { |
| 2753 | // Allow constant lane indices, expand variable lane indices |
| 2754 | SDNode *IdxNode = Op.getOperand(i: Op.getNumOperands() - 1).getNode(); |
| 2755 | if (isa<ConstantSDNode>(Val: IdxNode)) { |
| 2756 | // Ensure the index type is i32 to match the tablegen patterns |
| 2757 | uint64_t Idx = IdxNode->getAsZExtVal(); |
| 2758 | SmallVector<SDValue, 3> Ops(Op.getNode()->ops()); |
| 2759 | Ops[Op.getNumOperands() - 1] = |
| 2760 | DAG.getConstant(Val: Idx, DL: SDLoc(IdxNode), VT: MVT::i32); |
| 2761 | return DAG.getNode(Opcode: Op.getOpcode(), DL: SDLoc(Op), VT: Op.getValueType(), Ops); |
| 2762 | } |
| 2763 | // Perform default expansion |
| 2764 | return SDValue(); |
| 2765 | } |
| 2766 | |
| 2767 | static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) { |
| 2768 | EVT LaneT = Op.getSimpleValueType().getVectorElementType(); |
| 2769 | // 32-bit and 64-bit unrolled shifts will have proper semantics |
| 2770 | if (LaneT.bitsGE(VT: MVT::i32)) |
| 2771 | return DAG.UnrollVectorOp(N: Op.getNode()); |
| 2772 | // Otherwise mask the shift value to get proper semantics from 32-bit shift |
| 2773 | SDLoc DL(Op); |
| 2774 | size_t NumLanes = Op.getSimpleValueType().getVectorNumElements(); |
| 2775 | SDValue Mask = DAG.getConstant(Val: LaneT.getSizeInBits() - 1, DL, VT: MVT::i32); |
| 2776 | unsigned ShiftOpcode = Op.getOpcode(); |
| 2777 | SmallVector<SDValue, 16> ShiftedElements; |
| 2778 | DAG.ExtractVectorElements(Op: Op.getOperand(i: 0), Args&: ShiftedElements, Start: 0, Count: 0, EltVT: MVT::i32); |
| 2779 | SmallVector<SDValue, 16> ShiftElements; |
| 2780 | DAG.ExtractVectorElements(Op: Op.getOperand(i: 1), Args&: ShiftElements, Start: 0, Count: 0, EltVT: MVT::i32); |
| 2781 | SmallVector<SDValue, 16> UnrolledOps; |
| 2782 | for (size_t i = 0; i < NumLanes; ++i) { |
| 2783 | SDValue MaskedShiftValue = |
| 2784 | DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: ShiftElements[i], N2: Mask); |
| 2785 | SDValue ShiftedValue = ShiftedElements[i]; |
| 2786 | if (ShiftOpcode == ISD::SRA) |
| 2787 | ShiftedValue = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: MVT::i32, |
| 2788 | N1: ShiftedValue, N2: DAG.getValueType(LaneT)); |
| 2789 | UnrolledOps.push_back( |
| 2790 | Elt: DAG.getNode(Opcode: ShiftOpcode, DL, VT: MVT::i32, N1: ShiftedValue, N2: MaskedShiftValue)); |
| 2791 | } |
| 2792 | return DAG.getBuildVector(VT: Op.getValueType(), DL, Ops: UnrolledOps); |
| 2793 | } |
| 2794 | |
| 2795 | SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op, |
| 2796 | SelectionDAG &DAG) const { |
| 2797 | SDLoc DL(Op); |
| 2798 | |
| 2799 | // Only manually lower vector shifts |
| 2800 | assert(Op.getSimpleValueType().isVector()); |
| 2801 | |
| 2802 | uint64_t LaneBits = Op.getValueType().getScalarSizeInBits(); |
| 2803 | auto ShiftVal = Op.getOperand(i: 1); |
| 2804 | |
| 2805 | // Try to skip bitmask operation since it is implied inside shift instruction |
| 2806 | auto SkipImpliedMask = [](SDValue MaskOp, uint64_t MaskBits) { |
| 2807 | if (MaskOp.getOpcode() != ISD::AND) |
| 2808 | return MaskOp; |
| 2809 | SDValue LHS = MaskOp.getOperand(i: 0); |
| 2810 | SDValue RHS = MaskOp.getOperand(i: 1); |
| 2811 | if (MaskOp.getValueType().isVector()) { |
| 2812 | APInt MaskVal; |
| 2813 | if (!ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal)) |
| 2814 | std::swap(a&: LHS, b&: RHS); |
| 2815 | |
| 2816 | if (ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal) && |
| 2817 | MaskVal == MaskBits) |
| 2818 | MaskOp = LHS; |
| 2819 | } else { |
| 2820 | if (!isa<ConstantSDNode>(Val: RHS.getNode())) |
| 2821 | std::swap(a&: LHS, b&: RHS); |
| 2822 | |
| 2823 | auto ConstantRHS = dyn_cast<ConstantSDNode>(Val: RHS.getNode()); |
| 2824 | if (ConstantRHS && ConstantRHS->getAPIntValue() == MaskBits) |
| 2825 | MaskOp = LHS; |
| 2826 | } |
| 2827 | |
| 2828 | return MaskOp; |
| 2829 | }; |
| 2830 | |
| 2831 | // Skip vector and operation |
| 2832 | ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1); |
| 2833 | ShiftVal = DAG.getSplatValue(V: ShiftVal); |
| 2834 | if (!ShiftVal) |
| 2835 | return unrollVectorShift(Op, DAG); |
| 2836 | |
| 2837 | // Skip scalar and operation |
| 2838 | ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1); |
| 2839 | // Use anyext because none of the high bits can affect the shift |
| 2840 | ShiftVal = DAG.getAnyExtOrTrunc(Op: ShiftVal, DL, VT: MVT::i32); |
| 2841 | |
| 2842 | unsigned Opcode; |
| 2843 | switch (Op.getOpcode()) { |
| 2844 | case ISD::SHL: |
| 2845 | Opcode = WebAssemblyISD::VEC_SHL; |
| 2846 | break; |
| 2847 | case ISD::SRA: |
| 2848 | Opcode = WebAssemblyISD::VEC_SHR_S; |
| 2849 | break; |
| 2850 | case ISD::SRL: |
| 2851 | Opcode = WebAssemblyISD::VEC_SHR_U; |
| 2852 | break; |
| 2853 | default: |
| 2854 | llvm_unreachable("unexpected opcode" ); |
| 2855 | } |
| 2856 | |
| 2857 | return DAG.getNode(Opcode, DL, VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: ShiftVal); |
| 2858 | } |
| 2859 | |
| 2860 | SDValue WebAssemblyTargetLowering::LowerFP_TO_INT_SAT(SDValue Op, |
| 2861 | SelectionDAG &DAG) const { |
| 2862 | EVT ResT = Op.getValueType(); |
| 2863 | EVT SatVT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT(); |
| 2864 | |
| 2865 | if ((ResT == MVT::i32 || ResT == MVT::i64) && |
| 2866 | (SatVT == MVT::i32 || SatVT == MVT::i64)) |
| 2867 | return Op; |
| 2868 | |
| 2869 | if (ResT == MVT::v4i32 && SatVT == MVT::i32) |
| 2870 | return Op; |
| 2871 | |
| 2872 | if (ResT == MVT::v8i16 && SatVT == MVT::i16) |
| 2873 | return Op; |
| 2874 | |
| 2875 | return SDValue(); |
| 2876 | } |
| 2877 | |
| 2878 | static bool HasNoSignedZerosOrNaNs(SDValue Op, SelectionDAG &DAG) { |
| 2879 | return (Op->getFlags().hasNoNaNs() || |
| 2880 | (DAG.isKnownNeverNaN(Op: Op->getOperand(Num: 0)) && |
| 2881 | DAG.isKnownNeverNaN(Op: Op->getOperand(Num: 1)))) && |
| 2882 | (Op->getFlags().hasNoSignedZeros() || |
| 2883 | DAG.isKnownNeverZeroFloat(Op: Op->getOperand(Num: 0)) || |
| 2884 | DAG.isKnownNeverZeroFloat(Op: Op->getOperand(Num: 1))); |
| 2885 | } |
| 2886 | |
| 2887 | SDValue WebAssemblyTargetLowering::LowerFMIN(SDValue Op, |
| 2888 | SelectionDAG &DAG) const { |
| 2889 | if (Subtarget->hasRelaxedSIMD() && HasNoSignedZerosOrNaNs(Op, DAG)) { |
| 2890 | return DAG.getNode(Opcode: WebAssemblyISD::RELAXED_FMIN, DL: SDLoc(Op), |
| 2891 | VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: Op.getOperand(i: 1)); |
| 2892 | } |
| 2893 | return SDValue(); |
| 2894 | } |
| 2895 | |
| 2896 | SDValue WebAssemblyTargetLowering::LowerFMAX(SDValue Op, |
| 2897 | SelectionDAG &DAG) const { |
| 2898 | if (Subtarget->hasRelaxedSIMD() && HasNoSignedZerosOrNaNs(Op, DAG)) { |
| 2899 | return DAG.getNode(Opcode: WebAssemblyISD::RELAXED_FMAX, DL: SDLoc(Op), |
| 2900 | VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: Op.getOperand(i: 1)); |
| 2901 | } |
| 2902 | return SDValue(); |
| 2903 | } |
| 2904 | |
| 2905 | //===----------------------------------------------------------------------===// |
| 2906 | // Custom DAG combine hooks |
| 2907 | //===----------------------------------------------------------------------===// |
| 2908 | static SDValue |
| 2909 | performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
| 2910 | auto &DAG = DCI.DAG; |
| 2911 | auto Shuffle = cast<ShuffleVectorSDNode>(Val: N); |
| 2912 | |
| 2913 | // Hoist vector bitcasts that don't change the number of lanes out of unary |
| 2914 | // shuffles, where they are less likely to get in the way of other combines. |
| 2915 | // (shuffle (vNxT1 (bitcast (vNxT0 x))), undef, mask) -> |
| 2916 | // (vNxT1 (bitcast (vNxT0 (shuffle x, undef, mask)))) |
| 2917 | SDValue Bitcast = N->getOperand(Num: 0); |
| 2918 | if (Bitcast.getOpcode() != ISD::BITCAST) |
| 2919 | return SDValue(); |
| 2920 | if (!N->getOperand(Num: 1).isUndef()) |
| 2921 | return SDValue(); |
| 2922 | SDValue CastOp = Bitcast.getOperand(i: 0); |
| 2923 | EVT SrcType = CastOp.getValueType(); |
| 2924 | EVT DstType = Bitcast.getValueType(); |
| 2925 | if (!SrcType.is128BitVector() || |
| 2926 | SrcType.getVectorNumElements() != DstType.getVectorNumElements()) |
| 2927 | return SDValue(); |
| 2928 | SDValue NewShuffle = DAG.getVectorShuffle( |
| 2929 | VT: SrcType, dl: SDLoc(N), N1: CastOp, N2: DAG.getUNDEF(VT: SrcType), Mask: Shuffle->getMask()); |
| 2930 | return DAG.getBitcast(VT: DstType, V: NewShuffle); |
| 2931 | } |
| 2932 | |
| 2933 | /// Convert ({u,s}itofp vec) --> ({u,s}itofp ({s,z}ext vec)) so it doesn't get |
| 2934 | /// split up into scalar instructions during legalization, and the vector |
| 2935 | /// extending instructions are selected in performVectorExtendCombine below. |
| 2936 | static SDValue |
| 2937 | performVectorExtendToFPCombine(SDNode *N, |
| 2938 | TargetLowering::DAGCombinerInfo &DCI) { |
| 2939 | auto &DAG = DCI.DAG; |
| 2940 | assert(N->getOpcode() == ISD::UINT_TO_FP || |
| 2941 | N->getOpcode() == ISD::SINT_TO_FP); |
| 2942 | |
| 2943 | EVT InVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
| 2944 | EVT ResVT = N->getValueType(ResNo: 0); |
| 2945 | MVT ExtVT; |
| 2946 | if (ResVT == MVT::v4f32 && (InVT == MVT::v4i16 || InVT == MVT::v4i8)) |
| 2947 | ExtVT = MVT::v4i32; |
| 2948 | else if (ResVT == MVT::v2f64 && (InVT == MVT::v2i16 || InVT == MVT::v2i8)) |
| 2949 | ExtVT = MVT::v2i32; |
| 2950 | else |
| 2951 | return SDValue(); |
| 2952 | |
| 2953 | unsigned Op = |
| 2954 | N->getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; |
| 2955 | SDValue Conv = DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ExtVT, Operand: N->getOperand(Num: 0)); |
| 2956 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: ResVT, Operand: Conv); |
| 2957 | } |
| 2958 | |
| 2959 | static SDValue |
| 2960 | performVectorNonNegToFPCombine(SDNode *N, |
| 2961 | TargetLowering::DAGCombinerInfo &DCI) { |
| 2962 | auto &DAG = DCI.DAG; |
| 2963 | |
| 2964 | SDNodeFlags Flags = N->getFlags(); |
| 2965 | SDValue Op0 = N->getOperand(Num: 0); |
| 2966 | EVT VT = N->getValueType(ResNo: 0); |
| 2967 | |
| 2968 | // Optimize uitofp to sitofp when the sign bit is known to be zero. |
| 2969 | // Depending on the target (runtime) backend, this might be performance |
| 2970 | // neutral (e.g. AArch64) or a significant improvement (e.g. x86_64). |
| 2971 | if (VT.isVector() && (Flags.hasNonNeg() || DAG.SignBitIsZero(Op: Op0))) { |
| 2972 | return DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: SDLoc(N), VT, Operand: Op0); |
| 2973 | } |
| 2974 | |
| 2975 | return SDValue(); |
| 2976 | } |
| 2977 | |
| 2978 | static SDValue |
| 2979 | performVectorExtendCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
| 2980 | auto &DAG = DCI.DAG; |
| 2981 | assert(N->getOpcode() == ISD::SIGN_EXTEND || |
| 2982 | N->getOpcode() == ISD::ZERO_EXTEND); |
| 2983 | |
| 2984 | // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if |
| 2985 | // possible before the extract_subvector can be expanded. |
| 2986 | auto = N->getOperand(Num: 0); |
| 2987 | if (Extract.getOpcode() != ISD::EXTRACT_SUBVECTOR) |
| 2988 | return SDValue(); |
| 2989 | auto Source = Extract.getOperand(i: 0); |
| 2990 | auto *IndexNode = dyn_cast<ConstantSDNode>(Val: Extract.getOperand(i: 1)); |
| 2991 | if (IndexNode == nullptr) |
| 2992 | return SDValue(); |
| 2993 | auto Index = IndexNode->getZExtValue(); |
| 2994 | |
| 2995 | // Only v8i8, v4i16, and v2i32 extracts can be widened, and only if the |
| 2996 | // extracted subvector is the low or high half of its source. |
| 2997 | EVT ResVT = N->getValueType(ResNo: 0); |
| 2998 | if (ResVT == MVT::v8i16) { |
| 2999 | if (Extract.getValueType() != MVT::v8i8 || |
| 3000 | Source.getValueType() != MVT::v16i8 || (Index != 0 && Index != 8)) |
| 3001 | return SDValue(); |
| 3002 | } else if (ResVT == MVT::v4i32) { |
| 3003 | if (Extract.getValueType() != MVT::v4i16 || |
| 3004 | Source.getValueType() != MVT::v8i16 || (Index != 0 && Index != 4)) |
| 3005 | return SDValue(); |
| 3006 | } else if (ResVT == MVT::v2i64) { |
| 3007 | if (Extract.getValueType() != MVT::v2i32 || |
| 3008 | Source.getValueType() != MVT::v4i32 || (Index != 0 && Index != 2)) |
| 3009 | return SDValue(); |
| 3010 | } else { |
| 3011 | return SDValue(); |
| 3012 | } |
| 3013 | |
| 3014 | bool IsSext = N->getOpcode() == ISD::SIGN_EXTEND; |
| 3015 | bool IsLow = Index == 0; |
| 3016 | |
| 3017 | unsigned Op = IsSext ? (IsLow ? WebAssemblyISD::EXTEND_LOW_S |
| 3018 | : WebAssemblyISD::EXTEND_HIGH_S) |
| 3019 | : (IsLow ? WebAssemblyISD::EXTEND_LOW_U |
| 3020 | : WebAssemblyISD::EXTEND_HIGH_U); |
| 3021 | |
| 3022 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
| 3023 | } |
| 3024 | |
| 3025 | static SDValue |
| 3026 | performVectorTruncZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
| 3027 | auto &DAG = DCI.DAG; |
| 3028 | |
| 3029 | auto GetWasmConversionOp = [](unsigned Op) { |
| 3030 | switch (Op) { |
| 3031 | case ISD::FP_TO_SINT_SAT: |
| 3032 | return WebAssemblyISD::TRUNC_SAT_ZERO_S; |
| 3033 | case ISD::FP_TO_UINT_SAT: |
| 3034 | return WebAssemblyISD::TRUNC_SAT_ZERO_U; |
| 3035 | case ISD::FP_ROUND: |
| 3036 | return WebAssemblyISD::DEMOTE_ZERO; |
| 3037 | } |
| 3038 | llvm_unreachable("unexpected op" ); |
| 3039 | }; |
| 3040 | |
| 3041 | auto IsZeroSplat = [](SDValue SplatVal) { |
| 3042 | auto *Splat = dyn_cast<BuildVectorSDNode>(Val: SplatVal.getNode()); |
| 3043 | APInt SplatValue, SplatUndef; |
| 3044 | unsigned SplatBitSize; |
| 3045 | bool HasAnyUndefs; |
| 3046 | // Endianness doesn't matter in this context because we are looking for |
| 3047 | // an all-zero value. |
| 3048 | return Splat && |
| 3049 | Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, |
| 3050 | HasAnyUndefs) && |
| 3051 | SplatValue == 0; |
| 3052 | }; |
| 3053 | |
| 3054 | if (N->getOpcode() == ISD::CONCAT_VECTORS) { |
| 3055 | // Combine this: |
| 3056 | // |
| 3057 | // (concat_vectors (v2i32 (fp_to_{s,u}int_sat $x, 32)), (v2i32 (splat 0))) |
| 3058 | // |
| 3059 | // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x). |
| 3060 | // |
| 3061 | // Or this: |
| 3062 | // |
| 3063 | // (concat_vectors (v2f32 (fp_round (v2f64 $x))), (v2f32 (splat 0))) |
| 3064 | // |
| 3065 | // into (f32x4.demote_zero_f64x2 $x). |
| 3066 | EVT ResVT; |
| 3067 | EVT ExpectedConversionType; |
| 3068 | auto Conversion = N->getOperand(Num: 0); |
| 3069 | auto ConversionOp = Conversion.getOpcode(); |
| 3070 | switch (ConversionOp) { |
| 3071 | case ISD::FP_TO_SINT_SAT: |
| 3072 | case ISD::FP_TO_UINT_SAT: |
| 3073 | ResVT = MVT::v4i32; |
| 3074 | ExpectedConversionType = MVT::v2i32; |
| 3075 | break; |
| 3076 | case ISD::FP_ROUND: |
| 3077 | ResVT = MVT::v4f32; |
| 3078 | ExpectedConversionType = MVT::v2f32; |
| 3079 | break; |
| 3080 | default: |
| 3081 | return SDValue(); |
| 3082 | } |
| 3083 | |
| 3084 | if (N->getValueType(ResNo: 0) != ResVT) |
| 3085 | return SDValue(); |
| 3086 | |
| 3087 | if (Conversion.getValueType() != ExpectedConversionType) |
| 3088 | return SDValue(); |
| 3089 | |
| 3090 | auto Source = Conversion.getOperand(i: 0); |
| 3091 | if (Source.getValueType() != MVT::v2f64) |
| 3092 | return SDValue(); |
| 3093 | |
| 3094 | if (!IsZeroSplat(N->getOperand(Num: 1)) || |
| 3095 | N->getOperand(Num: 1).getValueType() != ExpectedConversionType) |
| 3096 | return SDValue(); |
| 3097 | |
| 3098 | unsigned Op = GetWasmConversionOp(ConversionOp); |
| 3099 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
| 3100 | } |
| 3101 | |
| 3102 | // Combine this: |
| 3103 | // |
| 3104 | // (fp_to_{s,u}int_sat (concat_vectors $x, (v2f64 (splat 0))), 32) |
| 3105 | // |
| 3106 | // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x). |
| 3107 | // |
| 3108 | // Or this: |
| 3109 | // |
| 3110 | // (v4f32 (fp_round (concat_vectors $x, (v2f64 (splat 0))))) |
| 3111 | // |
| 3112 | // into (f32x4.demote_zero_f64x2 $x). |
| 3113 | EVT ResVT; |
| 3114 | auto ConversionOp = N->getOpcode(); |
| 3115 | switch (ConversionOp) { |
| 3116 | case ISD::FP_TO_SINT_SAT: |
| 3117 | case ISD::FP_TO_UINT_SAT: |
| 3118 | ResVT = MVT::v4i32; |
| 3119 | break; |
| 3120 | case ISD::FP_ROUND: |
| 3121 | ResVT = MVT::v4f32; |
| 3122 | break; |
| 3123 | default: |
| 3124 | llvm_unreachable("unexpected op" ); |
| 3125 | } |
| 3126 | |
| 3127 | if (N->getValueType(ResNo: 0) != ResVT) |
| 3128 | return SDValue(); |
| 3129 | |
| 3130 | auto Concat = N->getOperand(Num: 0); |
| 3131 | if (Concat.getValueType() != MVT::v4f64) |
| 3132 | return SDValue(); |
| 3133 | |
| 3134 | auto Source = Concat.getOperand(i: 0); |
| 3135 | if (Source.getValueType() != MVT::v2f64) |
| 3136 | return SDValue(); |
| 3137 | |
| 3138 | if (!IsZeroSplat(Concat.getOperand(i: 1)) || |
| 3139 | Concat.getOperand(i: 1).getValueType() != MVT::v2f64) |
| 3140 | return SDValue(); |
| 3141 | |
| 3142 | unsigned Op = GetWasmConversionOp(ConversionOp); |
| 3143 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
| 3144 | } |
| 3145 | |
| 3146 | // Helper to extract VectorWidth bits from Vec, starting from IdxVal. |
| 3147 | static SDValue (SDValue Vec, unsigned IdxVal, SelectionDAG &DAG, |
| 3148 | const SDLoc &DL, unsigned VectorWidth) { |
| 3149 | EVT VT = Vec.getValueType(); |
| 3150 | EVT ElVT = VT.getVectorElementType(); |
| 3151 | unsigned Factor = VT.getSizeInBits() / VectorWidth; |
| 3152 | EVT ResultVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ElVT, |
| 3153 | NumElements: VT.getVectorNumElements() / Factor); |
| 3154 | |
| 3155 | // Extract the relevant VectorWidth bits. Generate an EXTRACT_SUBVECTOR |
| 3156 | unsigned ElemsPerChunk = VectorWidth / ElVT.getSizeInBits(); |
| 3157 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2" ); |
| 3158 | |
| 3159 | // This is the index of the first element of the VectorWidth-bit chunk |
| 3160 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. |
| 3161 | IdxVal &= ~(ElemsPerChunk - 1); |
| 3162 | |
| 3163 | // If the input is a buildvector just emit a smaller one. |
| 3164 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) |
| 3165 | return DAG.getBuildVector(VT: ResultVT, DL, |
| 3166 | Ops: Vec->ops().slice(N: IdxVal, M: ElemsPerChunk)); |
| 3167 | |
| 3168 | SDValue VecIdx = DAG.getIntPtrConstant(Val: IdxVal, DL); |
| 3169 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: ResultVT, N1: Vec, N2: VecIdx); |
| 3170 | } |
| 3171 | |
| 3172 | // Helper to recursively truncate vector elements in half with NARROW_U. DstVT |
| 3173 | // is the expected destination value type after recursion. In is the initial |
| 3174 | // input. Note that the input should have enough leading zero bits to prevent |
| 3175 | // NARROW_U from saturating results. |
| 3176 | static SDValue truncateVectorWithNARROW(EVT DstVT, SDValue In, const SDLoc &DL, |
| 3177 | SelectionDAG &DAG) { |
| 3178 | EVT SrcVT = In.getValueType(); |
| 3179 | |
| 3180 | // No truncation required, we might get here due to recursive calls. |
| 3181 | if (SrcVT == DstVT) |
| 3182 | return In; |
| 3183 | |
| 3184 | unsigned SrcSizeInBits = SrcVT.getSizeInBits(); |
| 3185 | unsigned NumElems = SrcVT.getVectorNumElements(); |
| 3186 | if (!isPowerOf2_32(Value: NumElems)) |
| 3187 | return SDValue(); |
| 3188 | assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation" ); |
| 3189 | assert(SrcSizeInBits > DstVT.getSizeInBits() && "Illegal truncation" ); |
| 3190 | |
| 3191 | LLVMContext &Ctx = *DAG.getContext(); |
| 3192 | EVT PackedSVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: SrcVT.getScalarSizeInBits() / 2); |
| 3193 | |
| 3194 | // Narrow to the largest type possible: |
| 3195 | // vXi64/vXi32 -> i16x8.narrow_i32x4_u and vXi16 -> i8x16.narrow_i16x8_u. |
| 3196 | EVT InVT = MVT::i16, OutVT = MVT::i8; |
| 3197 | if (SrcVT.getScalarSizeInBits() > 16) { |
| 3198 | InVT = MVT::i32; |
| 3199 | OutVT = MVT::i16; |
| 3200 | } |
| 3201 | unsigned SubSizeInBits = SrcSizeInBits / 2; |
| 3202 | InVT = EVT::getVectorVT(Context&: Ctx, VT: InVT, NumElements: SubSizeInBits / InVT.getSizeInBits()); |
| 3203 | OutVT = EVT::getVectorVT(Context&: Ctx, VT: OutVT, NumElements: SubSizeInBits / OutVT.getSizeInBits()); |
| 3204 | |
| 3205 | // Split lower/upper subvectors. |
| 3206 | SDValue Lo = extractSubVector(Vec: In, IdxVal: 0, DAG, DL, VectorWidth: SubSizeInBits); |
| 3207 | SDValue Hi = extractSubVector(Vec: In, IdxVal: NumElems / 2, DAG, DL, VectorWidth: SubSizeInBits); |
| 3208 | |
| 3209 | // 256bit -> 128bit truncate - Narrow lower/upper 128-bit subvectors. |
| 3210 | if (SrcVT.is256BitVector() && DstVT.is128BitVector()) { |
| 3211 | Lo = DAG.getBitcast(VT: InVT, V: Lo); |
| 3212 | Hi = DAG.getBitcast(VT: InVT, V: Hi); |
| 3213 | SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::NARROW_U, DL, VT: OutVT, N1: Lo, N2: Hi); |
| 3214 | return DAG.getBitcast(VT: DstVT, V: Res); |
| 3215 | } |
| 3216 | |
| 3217 | // Recursively narrow lower/upper subvectors, concat result and narrow again. |
| 3218 | EVT PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems / 2); |
| 3219 | Lo = truncateVectorWithNARROW(DstVT: PackedVT, In: Lo, DL, DAG); |
| 3220 | Hi = truncateVectorWithNARROW(DstVT: PackedVT, In: Hi, DL, DAG); |
| 3221 | |
| 3222 | PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems); |
| 3223 | SDValue Res = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: PackedVT, N1: Lo, N2: Hi); |
| 3224 | return truncateVectorWithNARROW(DstVT, In: Res, DL, DAG); |
| 3225 | } |
| 3226 | |
| 3227 | static SDValue performTruncateCombine(SDNode *N, |
| 3228 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3229 | auto &DAG = DCI.DAG; |
| 3230 | |
| 3231 | SDValue In = N->getOperand(Num: 0); |
| 3232 | EVT InVT = In.getValueType(); |
| 3233 | if (!InVT.isSimple()) |
| 3234 | return SDValue(); |
| 3235 | |
| 3236 | EVT OutVT = N->getValueType(ResNo: 0); |
| 3237 | if (!OutVT.isVector()) |
| 3238 | return SDValue(); |
| 3239 | |
| 3240 | EVT OutSVT = OutVT.getVectorElementType(); |
| 3241 | EVT InSVT = InVT.getVectorElementType(); |
| 3242 | // Currently only cover truncate to v16i8 or v8i16. |
| 3243 | if (!((InSVT == MVT::i16 || InSVT == MVT::i32 || InSVT == MVT::i64) && |
| 3244 | (OutSVT == MVT::i8 || OutSVT == MVT::i16) && OutVT.is128BitVector())) |
| 3245 | return SDValue(); |
| 3246 | |
| 3247 | SDLoc DL(N); |
| 3248 | APInt Mask = APInt::getLowBitsSet(numBits: InVT.getScalarSizeInBits(), |
| 3249 | loBitsSet: OutVT.getScalarSizeInBits()); |
| 3250 | In = DAG.getNode(Opcode: ISD::AND, DL, VT: InVT, N1: In, N2: DAG.getConstant(Val: Mask, DL, VT: InVT)); |
| 3251 | return truncateVectorWithNARROW(DstVT: OutVT, In, DL, DAG); |
| 3252 | } |
| 3253 | |
| 3254 | static SDValue performBitcastCombine(SDNode *N, |
| 3255 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3256 | using namespace llvm::SDPatternMatch; |
| 3257 | auto &DAG = DCI.DAG; |
| 3258 | SDLoc DL(N); |
| 3259 | SDValue Src = N->getOperand(Num: 0); |
| 3260 | EVT VT = N->getValueType(ResNo: 0); |
| 3261 | EVT SrcVT = Src.getValueType(); |
| 3262 | |
| 3263 | if (!(DCI.isBeforeLegalize() && VT.isScalarInteger() && |
| 3264 | SrcVT.isFixedLengthVector() && SrcVT.getScalarType() == MVT::i1)) |
| 3265 | return SDValue(); |
| 3266 | |
| 3267 | unsigned NumElts = SrcVT.getVectorNumElements(); |
| 3268 | EVT Width = MVT::getIntegerVT(BitWidth: 128 / NumElts); |
| 3269 | |
| 3270 | // bitcast <N x i1> to iN, where N = 2, 4, 8, 16 (legal) |
| 3271 | // ==> bitmask |
| 3272 | if (NumElts == 2 || NumElts == 4 || NumElts == 8 || NumElts == 16) { |
| 3273 | return DAG.getZExtOrTrunc( |
| 3274 | Op: DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3275 | Ops: {DAG.getConstant(Val: Intrinsic::wasm_bitmask, DL, VT: MVT::i32), |
| 3276 | DAG.getSExtOrTrunc(Op: N->getOperand(Num: 0), DL, |
| 3277 | VT: SrcVT.changeVectorElementType( |
| 3278 | Context&: *DAG.getContext(), EltVT: Width))}), |
| 3279 | DL, VT); |
| 3280 | } |
| 3281 | |
| 3282 | // bitcast <N x i1>(setcc ...) to concat iN, where N = 32 and 64 (illegal) |
| 3283 | if (NumElts == 32 || NumElts == 64) { |
| 3284 | // Strategy: We will setcc them separately in v16i8 -> v16i1 |
| 3285 | // Bitcast them to i16, extend them to either i32 or i64. |
| 3286 | // Add them together, shifting left 1 by 1. |
| 3287 | SDValue Concat, SetCCVector; |
| 3288 | ISD::CondCode SetCond; |
| 3289 | |
| 3290 | if (!sd_match(N, P: m_BitCast(Op: m_c_SetCC(LHS: m_Value(N&: Concat), RHS: m_Value(N&: SetCCVector), |
| 3291 | CC: m_CondCode(CC&: SetCond))))) |
| 3292 | return SDValue(); |
| 3293 | if (Concat.getOpcode() != ISD::CONCAT_VECTORS) |
| 3294 | return SDValue(); |
| 3295 | |
| 3296 | uint64_t ElementWidth = |
| 3297 | SetCCVector.getValueType().getVectorElementType().getFixedSizeInBits(); |
| 3298 | |
| 3299 | SmallVector<SDValue> VectorsToShuffle; |
| 3300 | for (size_t I = 0; I < Concat->ops().size(); I++) { |
| 3301 | VectorsToShuffle.push_back(Elt: DAG.getBitcast( |
| 3302 | VT: MVT::i16, |
| 3303 | V: DAG.getSetCC(DL, VT: MVT::v16i1, LHS: Concat->ops()[I], |
| 3304 | RHS: extractSubVector(Vec: SetCCVector, IdxVal: I * (128 / ElementWidth), |
| 3305 | DAG, DL, VectorWidth: 128), |
| 3306 | Cond: SetCond))); |
| 3307 | } |
| 3308 | |
| 3309 | MVT ReturnType = VectorsToShuffle.size() == 2 ? MVT::i32 : MVT::i64; |
| 3310 | SDValue ReturningInteger = DAG.getConstant(Val: 0, DL, VT: ReturnType); |
| 3311 | |
| 3312 | for (SDValue V : VectorsToShuffle) { |
| 3313 | ReturningInteger = DAG.getNode( |
| 3314 | Opcode: ISD::SHL, DL, VT: ReturnType, |
| 3315 | Ops: {DAG.getShiftAmountConstant(Val: 16, VT: ReturnType, DL), ReturningInteger}); |
| 3316 | |
| 3317 | SDValue ExtendedV = DAG.getZExtOrTrunc(Op: V, DL, VT: ReturnType); |
| 3318 | ReturningInteger = |
| 3319 | DAG.getNode(Opcode: ISD::ADD, DL, VT: ReturnType, Ops: {ReturningInteger, ExtendedV}); |
| 3320 | } |
| 3321 | |
| 3322 | return ReturningInteger; |
| 3323 | } |
| 3324 | |
| 3325 | return SDValue(); |
| 3326 | } |
| 3327 | |
| 3328 | static SDValue performAnyAllCombine(SDNode *N, SelectionDAG &DAG) { |
| 3329 | // any_true (setcc <X>, 0, eq) => (not (all_true X)) |
| 3330 | // all_true (setcc <X>, 0, eq) => (not (any_true X)) |
| 3331 | // any_true (setcc <X>, 0, ne) => (any_true X) |
| 3332 | // all_true (setcc <X>, 0, ne) => (all_true X) |
| 3333 | assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN); |
| 3334 | using namespace llvm::SDPatternMatch; |
| 3335 | |
| 3336 | SDValue LHS; |
| 3337 | if (N->getNumOperands() < 2 || |
| 3338 | !sd_match(N: N->getOperand(Num: 1), |
| 3339 | P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(), CC: m_CondCode()))) |
| 3340 | return SDValue(); |
| 3341 | EVT LT = LHS.getValueType(); |
| 3342 | if (LT.getScalarSizeInBits() > 128 / LT.getVectorNumElements()) |
| 3343 | return SDValue(); |
| 3344 | |
| 3345 | auto CombineSetCC = [&N, &DAG](Intrinsic::WASMIntrinsics InPre, |
| 3346 | ISD::CondCode SetType, |
| 3347 | Intrinsic::WASMIntrinsics InPost) { |
| 3348 | if (N->getConstantOperandVal(Num: 0) != InPre) |
| 3349 | return SDValue(); |
| 3350 | |
| 3351 | SDValue LHS; |
| 3352 | if (!sd_match(N: N->getOperand(Num: 1), P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(), |
| 3353 | CC: m_SpecificCondCode(CC: SetType)))) |
| 3354 | return SDValue(); |
| 3355 | |
| 3356 | SDLoc DL(N); |
| 3357 | SDValue Ret = DAG.getZExtOrTrunc( |
| 3358 | Op: DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3359 | Ops: {DAG.getConstant(Val: InPost, DL, VT: MVT::i32), LHS}), |
| 3360 | DL, VT: MVT::i1); |
| 3361 | if (SetType == ISD::SETEQ) |
| 3362 | Ret = DAG.getNOT(DL, Val: Ret, VT: MVT::i1); |
| 3363 | return DAG.getZExtOrTrunc(Op: Ret, DL, VT: N->getValueType(ResNo: 0)); |
| 3364 | }; |
| 3365 | |
| 3366 | if (SDValue AnyTrueEQ = CombineSetCC(Intrinsic::wasm_anytrue, ISD::SETEQ, |
| 3367 | Intrinsic::wasm_alltrue)) |
| 3368 | return AnyTrueEQ; |
| 3369 | if (SDValue AllTrueEQ = CombineSetCC(Intrinsic::wasm_alltrue, ISD::SETEQ, |
| 3370 | Intrinsic::wasm_anytrue)) |
| 3371 | return AllTrueEQ; |
| 3372 | if (SDValue AnyTrueNE = CombineSetCC(Intrinsic::wasm_anytrue, ISD::SETNE, |
| 3373 | Intrinsic::wasm_anytrue)) |
| 3374 | return AnyTrueNE; |
| 3375 | if (SDValue AllTrueNE = CombineSetCC(Intrinsic::wasm_alltrue, ISD::SETNE, |
| 3376 | Intrinsic::wasm_alltrue)) |
| 3377 | return AllTrueNE; |
| 3378 | |
| 3379 | return SDValue(); |
| 3380 | } |
| 3381 | |
| 3382 | template <int MatchRHS, ISD::CondCode MatchCond, bool RequiresNegate, |
| 3383 | Intrinsic::ID Intrin> |
| 3384 | static SDValue TryMatchTrue(SDNode *N, EVT VecVT, SelectionDAG &DAG) { |
| 3385 | SDValue LHS = N->getOperand(Num: 0); |
| 3386 | SDValue RHS = N->getOperand(Num: 1); |
| 3387 | SDValue Cond = N->getOperand(Num: 2); |
| 3388 | if (MatchCond != cast<CondCodeSDNode>(Val&: Cond)->get()) |
| 3389 | return SDValue(); |
| 3390 | |
| 3391 | if (MatchRHS != cast<ConstantSDNode>(Val&: RHS)->getSExtValue()) |
| 3392 | return SDValue(); |
| 3393 | |
| 3394 | SDLoc DL(N); |
| 3395 | SDValue Ret = DAG.getZExtOrTrunc( |
| 3396 | Op: DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3397 | Ops: {DAG.getConstant(Val: Intrin, DL, VT: MVT::i32), |
| 3398 | DAG.getSExtOrTrunc(Op: LHS->getOperand(Num: 0), DL, VT: VecVT)}), |
| 3399 | DL, VT: MVT::i1); |
| 3400 | if (RequiresNegate) |
| 3401 | Ret = DAG.getNOT(DL, Val: Ret, VT: MVT::i1); |
| 3402 | return DAG.getZExtOrTrunc(Op: Ret, DL, VT: N->getValueType(ResNo: 0)); |
| 3403 | } |
| 3404 | |
| 3405 | /// Try to convert a i128 comparison to a v16i8 comparison before type |
| 3406 | /// legalization splits it up into chunks |
| 3407 | static SDValue |
| 3408 | combineVectorSizedSetCCEquality(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, |
| 3409 | const WebAssemblySubtarget *Subtarget) { |
| 3410 | |
| 3411 | SDLoc DL(N); |
| 3412 | SDValue X = N->getOperand(Num: 0); |
| 3413 | SDValue Y = N->getOperand(Num: 1); |
| 3414 | EVT VT = N->getValueType(ResNo: 0); |
| 3415 | EVT OpVT = X.getValueType(); |
| 3416 | |
| 3417 | SelectionDAG &DAG = DCI.DAG; |
| 3418 | if (DCI.DAG.getMachineFunction().getFunction().hasFnAttribute( |
| 3419 | Kind: Attribute::NoImplicitFloat)) |
| 3420 | return SDValue(); |
| 3421 | |
| 3422 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get(); |
| 3423 | // We're looking for an oversized integer equality comparison with SIMD |
| 3424 | if (!OpVT.isScalarInteger() || !OpVT.isByteSized() || OpVT != MVT::i128 || |
| 3425 | !Subtarget->hasSIMD128() || !isIntEqualitySetCC(Code: CC)) |
| 3426 | return SDValue(); |
| 3427 | |
| 3428 | // Don't perform this combine if constructing the vector will be expensive. |
| 3429 | auto IsVectorBitCastCheap = [](SDValue X) { |
| 3430 | X = peekThroughBitcasts(V: X); |
| 3431 | return isa<ConstantSDNode>(Val: X) || X.getOpcode() == ISD::LOAD; |
| 3432 | }; |
| 3433 | |
| 3434 | if (!IsVectorBitCastCheap(X) || !IsVectorBitCastCheap(Y)) |
| 3435 | return SDValue(); |
| 3436 | |
| 3437 | SDValue VecX = DAG.getBitcast(VT: MVT::v16i8, V: X); |
| 3438 | SDValue VecY = DAG.getBitcast(VT: MVT::v16i8, V: Y); |
| 3439 | SDValue Cmp = DAG.getSetCC(DL, VT: MVT::v16i8, LHS: VecX, RHS: VecY, Cond: CC); |
| 3440 | |
| 3441 | SDValue Intr = |
| 3442 | DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3443 | Ops: {DAG.getConstant(Val: CC == ISD::SETEQ ? Intrinsic::wasm_alltrue |
| 3444 | : Intrinsic::wasm_anytrue, |
| 3445 | DL, VT: MVT::i32), |
| 3446 | Cmp}); |
| 3447 | |
| 3448 | return DAG.getSetCC(DL, VT, LHS: Intr, RHS: DAG.getConstant(Val: 0, DL, VT: MVT::i32), |
| 3449 | Cond: ISD::SETNE); |
| 3450 | } |
| 3451 | |
| 3452 | static SDValue performSETCCCombine(SDNode *N, |
| 3453 | TargetLowering::DAGCombinerInfo &DCI, |
| 3454 | const WebAssemblySubtarget *Subtarget) { |
| 3455 | if (!DCI.isBeforeLegalize()) |
| 3456 | return SDValue(); |
| 3457 | |
| 3458 | EVT VT = N->getValueType(ResNo: 0); |
| 3459 | if (!VT.isScalarInteger()) |
| 3460 | return SDValue(); |
| 3461 | |
| 3462 | if (SDValue V = combineVectorSizedSetCCEquality(N, DCI, Subtarget)) |
| 3463 | return V; |
| 3464 | |
| 3465 | SDValue LHS = N->getOperand(Num: 0); |
| 3466 | if (LHS->getOpcode() != ISD::BITCAST) |
| 3467 | return SDValue(); |
| 3468 | |
| 3469 | EVT FromVT = LHS->getOperand(Num: 0).getValueType(); |
| 3470 | if (!FromVT.isFixedLengthVector() || FromVT.getVectorElementType() != MVT::i1) |
| 3471 | return SDValue(); |
| 3472 | |
| 3473 | unsigned NumElts = FromVT.getVectorNumElements(); |
| 3474 | if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16) |
| 3475 | return SDValue(); |
| 3476 | |
| 3477 | if (!cast<ConstantSDNode>(Val: N->getOperand(Num: 1))) |
| 3478 | return SDValue(); |
| 3479 | |
| 3480 | auto &DAG = DCI.DAG; |
| 3481 | EVT VecVT = FromVT.changeVectorElementType(Context&: *DAG.getContext(), |
| 3482 | EltVT: MVT::getIntegerVT(BitWidth: 128 / NumElts)); |
| 3483 | // setcc (iN (bitcast (vNi1 X))), 0, ne |
| 3484 | // ==> any_true (vNi1 X) |
| 3485 | if (auto Match = TryMatchTrue<0, ISD::SETNE, false, Intrinsic::wasm_anytrue>( |
| 3486 | N, VecVT, DAG)) { |
| 3487 | return Match; |
| 3488 | } |
| 3489 | // setcc (iN (bitcast (vNi1 X))), 0, eq |
| 3490 | // ==> xor (any_true (vNi1 X)), -1 |
| 3491 | if (auto Match = TryMatchTrue<0, ISD::SETEQ, true, Intrinsic::wasm_anytrue>( |
| 3492 | N, VecVT, DAG)) { |
| 3493 | return Match; |
| 3494 | } |
| 3495 | // setcc (iN (bitcast (vNi1 X))), -1, eq |
| 3496 | // ==> all_true (vNi1 X) |
| 3497 | if (auto Match = TryMatchTrue<-1, ISD::SETEQ, false, Intrinsic::wasm_alltrue>( |
| 3498 | N, VecVT, DAG)) { |
| 3499 | return Match; |
| 3500 | } |
| 3501 | // setcc (iN (bitcast (vNi1 X))), -1, ne |
| 3502 | // ==> xor (all_true (vNi1 X)), -1 |
| 3503 | if (auto Match = TryMatchTrue<-1, ISD::SETNE, true, Intrinsic::wasm_alltrue>( |
| 3504 | N, VecVT, DAG)) { |
| 3505 | return Match; |
| 3506 | } |
| 3507 | return SDValue(); |
| 3508 | } |
| 3509 | |
| 3510 | static SDValue TryWideExtMulCombine(SDNode *N, SelectionDAG &DAG) { |
| 3511 | EVT VT = N->getValueType(ResNo: 0); |
| 3512 | if (VT != MVT::v8i32 && VT != MVT::v16i32) |
| 3513 | return SDValue(); |
| 3514 | |
| 3515 | // Mul with extending inputs. |
| 3516 | SDValue LHS = N->getOperand(Num: 0); |
| 3517 | SDValue RHS = N->getOperand(Num: 1); |
| 3518 | if (LHS.getOpcode() != RHS.getOpcode()) |
| 3519 | return SDValue(); |
| 3520 | |
| 3521 | if (LHS.getOpcode() != ISD::SIGN_EXTEND && |
| 3522 | LHS.getOpcode() != ISD::ZERO_EXTEND) |
| 3523 | return SDValue(); |
| 3524 | |
| 3525 | if (LHS->getOperand(Num: 0).getValueType() != RHS->getOperand(Num: 0).getValueType()) |
| 3526 | return SDValue(); |
| 3527 | |
| 3528 | EVT FromVT = LHS->getOperand(Num: 0).getValueType(); |
| 3529 | EVT EltTy = FromVT.getVectorElementType(); |
| 3530 | if (EltTy != MVT::i8) |
| 3531 | return SDValue(); |
| 3532 | |
| 3533 | // For an input DAG that looks like this |
| 3534 | // %a = input_type |
| 3535 | // %b = input_type |
| 3536 | // %lhs = extend %a to output_type |
| 3537 | // %rhs = extend %b to output_type |
| 3538 | // %mul = mul %lhs, %rhs |
| 3539 | |
| 3540 | // input_type | output_type | instructions |
| 3541 | // v16i8 | v16i32 | %low = i16x8.extmul_low_i8x16_ %a, %b |
| 3542 | // | | %high = i16x8.extmul_high_i8x16_, %a, %b |
| 3543 | // | | %low_low = i32x4.ext_low_i16x8_ %low |
| 3544 | // | | %low_high = i32x4.ext_high_i16x8_ %low |
| 3545 | // | | %high_low = i32x4.ext_low_i16x8_ %high |
| 3546 | // | | %high_high = i32x4.ext_high_i16x8_ %high |
| 3547 | // | | %res = concat_vector(...) |
| 3548 | // v8i8 | v8i32 | %low = i16x8.extmul_low_i8x16_ %a, %b |
| 3549 | // | | %low_low = i32x4.ext_low_i16x8_ %low |
| 3550 | // | | %low_high = i32x4.ext_high_i16x8_ %low |
| 3551 | // | | %res = concat_vector(%low_low, %low_high) |
| 3552 | |
| 3553 | SDLoc DL(N); |
| 3554 | unsigned NumElts = VT.getVectorNumElements(); |
| 3555 | SDValue ExtendInLHS = LHS->getOperand(Num: 0); |
| 3556 | SDValue ExtendInRHS = RHS->getOperand(Num: 0); |
| 3557 | bool IsSigned = LHS->getOpcode() == ISD::SIGN_EXTEND; |
| 3558 | unsigned ExtendLowOpc = |
| 3559 | IsSigned ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U; |
| 3560 | unsigned ExtendHighOpc = |
| 3561 | IsSigned ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U; |
| 3562 | |
| 3563 | auto GetExtendLow = [&DAG, &DL, &ExtendLowOpc](EVT VT, SDValue Op) { |
| 3564 | return DAG.getNode(Opcode: ExtendLowOpc, DL, VT, Operand: Op); |
| 3565 | }; |
| 3566 | auto GetExtendHigh = [&DAG, &DL, &ExtendHighOpc](EVT VT, SDValue Op) { |
| 3567 | return DAG.getNode(Opcode: ExtendHighOpc, DL, VT, Operand: Op); |
| 3568 | }; |
| 3569 | |
| 3570 | if (NumElts == 16) { |
| 3571 | SDValue LowLHS = GetExtendLow(MVT::v8i16, ExtendInLHS); |
| 3572 | SDValue LowRHS = GetExtendLow(MVT::v8i16, ExtendInRHS); |
| 3573 | SDValue MulLow = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: LowLHS, N2: LowRHS); |
| 3574 | SDValue HighLHS = GetExtendHigh(MVT::v8i16, ExtendInLHS); |
| 3575 | SDValue HighRHS = GetExtendHigh(MVT::v8i16, ExtendInRHS); |
| 3576 | SDValue MulHigh = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: HighLHS, N2: HighRHS); |
| 3577 | SDValue SubVectors[] = { |
| 3578 | GetExtendLow(MVT::v4i32, MulLow), |
| 3579 | GetExtendHigh(MVT::v4i32, MulLow), |
| 3580 | GetExtendLow(MVT::v4i32, MulHigh), |
| 3581 | GetExtendHigh(MVT::v4i32, MulHigh), |
| 3582 | }; |
| 3583 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, Ops: SubVectors); |
| 3584 | } else { |
| 3585 | assert(NumElts == 8); |
| 3586 | SDValue LowLHS = DAG.getNode(Opcode: LHS->getOpcode(), DL, VT: MVT::v8i16, Operand: ExtendInLHS); |
| 3587 | SDValue LowRHS = DAG.getNode(Opcode: RHS->getOpcode(), DL, VT: MVT::v8i16, Operand: ExtendInRHS); |
| 3588 | SDValue MulLow = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: LowLHS, N2: LowRHS); |
| 3589 | SDValue Lo = GetExtendLow(MVT::v4i32, MulLow); |
| 3590 | SDValue Hi = GetExtendHigh(MVT::v4i32, MulLow); |
| 3591 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, N1: Lo, N2: Hi); |
| 3592 | } |
| 3593 | return SDValue(); |
| 3594 | } |
| 3595 | |
| 3596 | static SDValue performMulCombine(SDNode *N, |
| 3597 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3598 | assert(N->getOpcode() == ISD::MUL); |
| 3599 | EVT VT = N->getValueType(ResNo: 0); |
| 3600 | if (!VT.isVector()) |
| 3601 | return SDValue(); |
| 3602 | |
| 3603 | if (auto Res = TryWideExtMulCombine(N, DAG&: DCI.DAG)) |
| 3604 | return Res; |
| 3605 | |
| 3606 | // We don't natively support v16i8 or v8i8 mul, but we do support v8i16. So, |
| 3607 | // extend them to v8i16. |
| 3608 | if (VT != MVT::v8i8 && VT != MVT::v16i8) |
| 3609 | return SDValue(); |
| 3610 | |
| 3611 | SDLoc DL(N); |
| 3612 | SelectionDAG &DAG = DCI.DAG; |
| 3613 | SDValue LHS = N->getOperand(Num: 0); |
| 3614 | SDValue RHS = N->getOperand(Num: 1); |
| 3615 | EVT MulVT = MVT::v8i16; |
| 3616 | |
| 3617 | if (VT == MVT::v8i8) { |
| 3618 | SDValue PromotedLHS = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: MVT::v16i8, N1: LHS, |
| 3619 | N2: DAG.getUNDEF(VT: MVT::v8i8)); |
| 3620 | SDValue PromotedRHS = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: MVT::v16i8, N1: RHS, |
| 3621 | N2: DAG.getUNDEF(VT: MVT::v8i8)); |
| 3622 | SDValue LowLHS = |
| 3623 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: PromotedLHS); |
| 3624 | SDValue LowRHS = |
| 3625 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: PromotedRHS); |
| 3626 | SDValue MulLow = DAG.getBitcast( |
| 3627 | VT: MVT::v16i8, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: LowLHS, N2: LowRHS)); |
| 3628 | // Take the low byte of each lane. |
| 3629 | SDValue Shuffle = DAG.getVectorShuffle( |
| 3630 | VT: MVT::v16i8, dl: DL, N1: MulLow, N2: DAG.getUNDEF(VT: MVT::v16i8), |
| 3631 | Mask: {0, 2, 4, 6, 8, 10, 12, 14, -1, -1, -1, -1, -1, -1, -1, -1}); |
| 3632 | return extractSubVector(Vec: Shuffle, IdxVal: 0, DAG, DL, VectorWidth: 64); |
| 3633 | } else { |
| 3634 | assert(VT == MVT::v16i8 && "Expected v16i8" ); |
| 3635 | SDValue LowLHS = DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: LHS); |
| 3636 | SDValue LowRHS = DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: RHS); |
| 3637 | SDValue HighLHS = |
| 3638 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_HIGH_U, DL, VT: MulVT, Operand: LHS); |
| 3639 | SDValue HighRHS = |
| 3640 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_HIGH_U, DL, VT: MulVT, Operand: RHS); |
| 3641 | |
| 3642 | SDValue MulLow = |
| 3643 | DAG.getBitcast(VT, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: LowLHS, N2: LowRHS)); |
| 3644 | SDValue MulHigh = |
| 3645 | DAG.getBitcast(VT, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: HighLHS, N2: HighRHS)); |
| 3646 | |
| 3647 | // Take the low byte of each lane. |
| 3648 | return DAG.getVectorShuffle( |
| 3649 | VT, dl: DL, N1: MulLow, N2: MulHigh, |
| 3650 | Mask: {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30}); |
| 3651 | } |
| 3652 | } |
| 3653 | |
| 3654 | SDValue DoubleVectorWidth(SDValue In, unsigned RequiredNumElems, |
| 3655 | SelectionDAG &DAG) { |
| 3656 | SDLoc DL(In); |
| 3657 | LLVMContext &Ctx = *DAG.getContext(); |
| 3658 | EVT InVT = In.getValueType(); |
| 3659 | unsigned NumElems = InVT.getVectorNumElements() * 2; |
| 3660 | EVT OutVT = EVT::getVectorVT(Context&: Ctx, VT: InVT.getVectorElementType(), NumElements: NumElems); |
| 3661 | SDValue Concat = |
| 3662 | DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: OutVT, N1: In, N2: DAG.getPOISON(VT: InVT)); |
| 3663 | if (NumElems < RequiredNumElems) { |
| 3664 | return DoubleVectorWidth(In: Concat, RequiredNumElems, DAG); |
| 3665 | } |
| 3666 | return Concat; |
| 3667 | } |
| 3668 | |
| 3669 | SDValue performConvertFPCombine(SDNode *N, SelectionDAG &DAG) { |
| 3670 | EVT OutVT = N->getValueType(ResNo: 0); |
| 3671 | if (!OutVT.isVector()) |
| 3672 | return SDValue(); |
| 3673 | |
| 3674 | EVT OutElTy = OutVT.getVectorElementType(); |
| 3675 | if (OutElTy != MVT::i8 && OutElTy != MVT::i16) |
| 3676 | return SDValue(); |
| 3677 | |
| 3678 | unsigned NumElems = OutVT.getVectorNumElements(); |
| 3679 | if (!isPowerOf2_32(Value: NumElems)) |
| 3680 | return SDValue(); |
| 3681 | |
| 3682 | EVT FPVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
| 3683 | if (FPVT.getVectorElementType() != MVT::f32) |
| 3684 | return SDValue(); |
| 3685 | |
| 3686 | SDLoc DL(N); |
| 3687 | |
| 3688 | // First, convert to i32. |
| 3689 | LLVMContext &Ctx = *DAG.getContext(); |
| 3690 | EVT IntVT = EVT::getVectorVT(Context&: Ctx, VT: MVT::i32, NumElements: NumElems); |
| 3691 | SDValue ToInt = DAG.getNode(Opcode: N->getOpcode(), DL, VT: IntVT, Operand: N->getOperand(Num: 0)); |
| 3692 | APInt Mask = APInt::getLowBitsSet(numBits: IntVT.getScalarSizeInBits(), |
| 3693 | loBitsSet: OutVT.getScalarSizeInBits()); |
| 3694 | // Mask out the top MSBs. |
| 3695 | SDValue Masked = |
| 3696 | DAG.getNode(Opcode: ISD::AND, DL, VT: IntVT, N1: ToInt, N2: DAG.getConstant(Val: Mask, DL, VT: IntVT)); |
| 3697 | |
| 3698 | if (OutVT.getSizeInBits() < 128) { |
| 3699 | // Create a wide enough vector that we can use narrow. |
| 3700 | EVT NarrowedVT = OutElTy == MVT::i8 ? MVT::v16i8 : MVT::v8i16; |
| 3701 | unsigned NumRequiredElems = NarrowedVT.getVectorNumElements(); |
| 3702 | SDValue WideVector = DoubleVectorWidth(In: Masked, RequiredNumElems: NumRequiredElems, DAG); |
| 3703 | SDValue Trunc = truncateVectorWithNARROW(DstVT: NarrowedVT, In: WideVector, DL, DAG); |
| 3704 | return DAG.getBitcast( |
| 3705 | VT: OutVT, V: extractSubVector(Vec: Trunc, IdxVal: 0, DAG, DL, VectorWidth: OutVT.getSizeInBits())); |
| 3706 | } else { |
| 3707 | return truncateVectorWithNARROW(DstVT: OutVT, In: Masked, DL, DAG); |
| 3708 | } |
| 3709 | return SDValue(); |
| 3710 | } |
| 3711 | |
| 3712 | SDValue |
| 3713 | WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N, |
| 3714 | DAGCombinerInfo &DCI) const { |
| 3715 | switch (N->getOpcode()) { |
| 3716 | default: |
| 3717 | return SDValue(); |
| 3718 | case ISD::BITCAST: |
| 3719 | return performBitcastCombine(N, DCI); |
| 3720 | case ISD::SETCC: |
| 3721 | return performSETCCCombine(N, DCI, Subtarget); |
| 3722 | case ISD::VECTOR_SHUFFLE: |
| 3723 | return performVECTOR_SHUFFLECombine(N, DCI); |
| 3724 | case ISD::SIGN_EXTEND: |
| 3725 | case ISD::ZERO_EXTEND: |
| 3726 | return performVectorExtendCombine(N, DCI); |
| 3727 | case ISD::UINT_TO_FP: |
| 3728 | if (auto ExtCombine = performVectorExtendToFPCombine(N, DCI)) |
| 3729 | return ExtCombine; |
| 3730 | return performVectorNonNegToFPCombine(N, DCI); |
| 3731 | case ISD::SINT_TO_FP: |
| 3732 | return performVectorExtendToFPCombine(N, DCI); |
| 3733 | case ISD::FP_TO_SINT_SAT: |
| 3734 | case ISD::FP_TO_UINT_SAT: |
| 3735 | case ISD::FP_ROUND: |
| 3736 | case ISD::CONCAT_VECTORS: |
| 3737 | return performVectorTruncZeroCombine(N, DCI); |
| 3738 | case ISD::FP_TO_SINT: |
| 3739 | case ISD::FP_TO_UINT: |
| 3740 | return performConvertFPCombine(N, DAG&: DCI.DAG); |
| 3741 | case ISD::TRUNCATE: |
| 3742 | return performTruncateCombine(N, DCI); |
| 3743 | case ISD::INTRINSIC_WO_CHAIN: |
| 3744 | return performAnyAllCombine(N, DAG&: DCI.DAG); |
| 3745 | case ISD::MUL: |
| 3746 | return performMulCombine(N, DCI); |
| 3747 | } |
| 3748 | } |
| 3749 | |