| 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 | if (Subtarget->hasSIMD128() && MVT(T).isVector()) { |
| 156 | setOperationAction(Op: ISD::PSEUDO_FMIN, VT: T, Action: Legal); |
| 157 | setOperationAction(Op: ISD::PSEUDO_FMAX, VT: T, Action: Legal); |
| 158 | } |
| 159 | // When experimental v8f16 support is enabled these instructions don't need |
| 160 | // to be expanded. |
| 161 | if (T != MVT::v8f16) { |
| 162 | setOperationAction(Op: ISD::FP16_TO_FP, VT: T, Action: Expand); |
| 163 | setOperationAction(Op: ISD::FP_TO_FP16, VT: T, Action: Expand); |
| 164 | } |
| 165 | if (Subtarget->hasFP16() && T == MVT::f32) { |
| 166 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: T, MemVT: MVT::f16, Action: Legal); |
| 167 | setTruncStoreAction(ValVT: T, MemVT: MVT::f16, Action: Legal); |
| 168 | } else { |
| 169 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: T, MemVT: MVT::f16, Action: Expand); |
| 170 | setTruncStoreAction(ValVT: T, MemVT: MVT::f16, Action: Expand); |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | // Expand unavailable integer operations. |
| 175 | for (auto Op : |
| 176 | {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU, |
| 177 | ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS, |
| 178 | ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) { |
| 179 | for (auto T : {MVT::i32, MVT::i64}) |
| 180 | setOperationAction(Op, VT: T, Action: Expand); |
| 181 | if (Subtarget->hasSIMD128()) |
| 182 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 183 | setOperationAction(Op, VT: T, Action: Expand); |
| 184 | } |
| 185 | |
| 186 | if (Subtarget->hasWideArithmetic()) { |
| 187 | setOperationAction(Op: ISD::ADD, VT: MVT::i128, Action: Custom); |
| 188 | setOperationAction(Op: ISD::SUB, VT: MVT::i128, Action: Custom); |
| 189 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i64, Action: Custom); |
| 190 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i64, Action: Custom); |
| 191 | setOperationAction(Op: ISD::UADDO, VT: MVT::i64, Action: Custom); |
| 192 | } |
| 193 | |
| 194 | if (Subtarget->hasNontrappingFPToInt()) |
| 195 | for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) |
| 196 | for (auto T : {MVT::i32, MVT::i64}) |
| 197 | setOperationAction(Op, VT: T, Action: Custom); |
| 198 | |
| 199 | if (Subtarget->hasRelaxedSIMD()) { |
| 200 | setOperationAction( |
| 201 | Ops: {ISD::FMINNUM, ISD::FMINIMUMNUM, ISD::FMAXNUM, ISD::FMAXIMUMNUM}, |
| 202 | VTs: {MVT::v4f32, MVT::v2f64}, Action: Custom); |
| 203 | } |
| 204 | |
| 205 | // Combine expands these operations, because wasi-libc and emscripten do not |
| 206 | // yet have the dedicated libcalls. |
| 207 | setTargetDAGCombine( |
| 208 | {ISD::FMINIMUM, ISD::FMAXIMUM, ISD::FMINIMUMNUM, ISD::FMAXIMUMNUM}); |
| 209 | |
| 210 | // SIMD-specific configuration |
| 211 | if (Subtarget->hasSIMD128()) { |
| 212 | |
| 213 | setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); |
| 214 | |
| 215 | // Combine wide-vector muls, with extend inputs, to extmul_half. |
| 216 | setTargetDAGCombine(ISD::MUL); |
| 217 | setTargetDAGCombine(ISD::SHL); |
| 218 | |
| 219 | // Combine vector mask reductions into alltrue/anytrue |
| 220 | setTargetDAGCombine(ISD::SETCC); |
| 221 | |
| 222 | // Convert vector to integer bitcasts to bitmask |
| 223 | setTargetDAGCombine(ISD::BITCAST); |
| 224 | |
| 225 | // Hoist bitcasts out of shuffles |
| 226 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); |
| 227 | |
| 228 | // Combine extends of extract_subvectors into widening ops |
| 229 | setTargetDAGCombine({ISD::SIGN_EXTEND, ISD::ZERO_EXTEND}); |
| 230 | |
| 231 | // Combine int_to_fp or fp_extend of extract_vectors and vice versa into |
| 232 | // conversions ops |
| 233 | setTargetDAGCombine({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_EXTEND, |
| 234 | ISD::EXTRACT_SUBVECTOR}); |
| 235 | |
| 236 | // Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa |
| 237 | // into conversion ops |
| 238 | setTargetDAGCombine({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT, |
| 239 | ISD::FP_TO_SINT, ISD::FP_TO_UINT, ISD::FP_ROUND, |
| 240 | ISD::CONCAT_VECTORS}); |
| 241 | |
| 242 | setTargetDAGCombine(ISD::TRUNCATE); |
| 243 | |
| 244 | // Support saturating add/sub for i8x16 and i16x8 |
| 245 | for (auto Op : {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT}) |
| 246 | for (auto T : {MVT::v16i8, MVT::v8i16}) |
| 247 | setOperationAction(Op, VT: T, Action: Legal); |
| 248 | |
| 249 | // Support integer abs |
| 250 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 251 | setOperationAction(Op: ISD::ABS, VT: T, Action: Legal); |
| 252 | |
| 253 | // Custom lower BUILD_VECTORs to minimize number of replace_lanes |
| 254 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 255 | MVT::v2f64}) |
| 256 | setOperationAction(Op: ISD::BUILD_VECTOR, VT: T, Action: Custom); |
| 257 | |
| 258 | if (Subtarget->hasFP16()) { |
| 259 | setOperationAction(Op: ISD::BUILD_VECTOR, VT: MVT::f16, Action: Custom); |
| 260 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::v4f16, Action: Custom); |
| 261 | } |
| 262 | |
| 263 | // We have custom shuffle lowering to expose the shuffle mask |
| 264 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 265 | MVT::v2f64}) |
| 266 | setOperationAction(Op: ISD::VECTOR_SHUFFLE, VT: T, Action: Custom); |
| 267 | |
| 268 | if (Subtarget->hasFP16()) |
| 269 | setOperationAction(Op: ISD::VECTOR_SHUFFLE, VT: MVT::v8f16, Action: Custom); |
| 270 | |
| 271 | // Support splatting |
| 272 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 273 | MVT::v2f64}) |
| 274 | setOperationAction(Op: ISD::SPLAT_VECTOR, VT: T, Action: Legal); |
| 275 | |
| 276 | setOperationAction(Ops: ISD::AVGCEILU, VTs: {MVT::v8i16, MVT::v16i8}, Action: Legal); |
| 277 | |
| 278 | // Custom lowering since wasm shifts must have a scalar shift amount |
| 279 | for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL}) |
| 280 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 281 | setOperationAction(Op, VT: T, Action: Custom); |
| 282 | |
| 283 | // Custom lower lane accesses to expand out variable indices |
| 284 | for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT}) |
| 285 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
| 286 | MVT::v2f64}) |
| 287 | setOperationAction(Op, VT: T, Action: Custom); |
| 288 | |
| 289 | // There is no i8x16.mul instruction |
| 290 | setOperationAction(Op: ISD::MUL, VT: MVT::v16i8, Action: Expand); |
| 291 | |
| 292 | // Expand integer operations supported for scalars but not SIMD |
| 293 | for (auto Op : |
| 294 | {ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR}) |
| 295 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 296 | setOperationAction(Op, VT: T, Action: Expand); |
| 297 | |
| 298 | // But we do have integer min and max operations |
| 299 | for (auto Op : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}) |
| 300 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32}) |
| 301 | setOperationAction(Op, VT: T, Action: Legal); |
| 302 | |
| 303 | // And we have popcnt for i8x16. It can be used to expand ctlz/cttz. |
| 304 | setOperationAction(Op: ISD::CTPOP, VT: MVT::v16i8, Action: Legal); |
| 305 | setOperationAction(Op: ISD::CTLZ, VT: MVT::v16i8, Action: Expand); |
| 306 | setOperationAction(Op: ISD::CTTZ, VT: MVT::v16i8, Action: Expand); |
| 307 | |
| 308 | // Custom lower bit counting operations for other types to scalarize them. |
| 309 | for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP}) |
| 310 | for (auto T : {MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
| 311 | setOperationAction(Op, VT: T, Action: Custom); |
| 312 | |
| 313 | // Expand float operations supported for scalars but not SIMD |
| 314 | for (auto Op : {ISD::FCOPYSIGN, ISD::FLOG, ISD::FLOG2, ISD::FLOG10, |
| 315 | ISD::FEXP, ISD::FEXP2, ISD::FEXP10}) |
| 316 | for (auto T : {MVT::v4f32, MVT::v2f64}) |
| 317 | setOperationAction(Op, VT: T, Action: Expand); |
| 318 | |
| 319 | // Unsigned comparison operations are unavailable for i64x2 vectors. |
| 320 | for (auto CC : {ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE}) |
| 321 | setCondCodeAction(CCs: CC, VT: MVT::v2i64, Action: Custom); |
| 322 | |
| 323 | // 64x2 conversions are not in the spec |
| 324 | for (auto Op : |
| 325 | {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT}) |
| 326 | for (auto T : {MVT::v2i64, MVT::v2f64}) |
| 327 | setOperationAction(Op, VT: T, Action: Expand); |
| 328 | |
| 329 | // But saturating fp_to_int converstions are |
| 330 | for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) { |
| 331 | setOperationAction(Op, VT: MVT::v4i32, Action: Custom); |
| 332 | if (Subtarget->hasFP16()) { |
| 333 | setOperationAction(Op, VT: MVT::v8i16, Action: Custom); |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | // Support vector extending |
| 338 | for (auto T : MVT::integer_fixedlen_vector_valuetypes()) { |
| 339 | setOperationAction(Op: ISD::ANY_EXTEND_VECTOR_INREG, VT: T, Action: Custom); |
| 340 | setOperationAction(Op: ISD::SIGN_EXTEND_VECTOR_INREG, VT: T, Action: Custom); |
| 341 | setOperationAction(Op: ISD::ZERO_EXTEND_VECTOR_INREG, VT: T, Action: Custom); |
| 342 | } |
| 343 | |
| 344 | if (Subtarget->hasFP16()) { |
| 345 | setOperationAction(Op: ISD::FMA, VT: MVT::v8f16, Action: Legal); |
| 346 | } |
| 347 | |
| 348 | if (Subtarget->hasRelaxedSIMD()) { |
| 349 | setOperationAction(Op: ISD::FMULADD, VT: MVT::v4f32, Action: Legal); |
| 350 | setOperationAction(Op: ISD::FMULADD, VT: MVT::v2f64, Action: Legal); |
| 351 | } |
| 352 | |
| 353 | // Partial MLA reductions. |
| 354 | for (auto Op : {ISD::PARTIAL_REDUCE_SMLA, ISD::PARTIAL_REDUCE_UMLA}) { |
| 355 | setPartialReduceMLAAction(Opc: Op, AccVT: MVT::v4i32, InputVT: MVT::v16i8, Action: Legal); |
| 356 | setPartialReduceMLAAction(Opc: Op, AccVT: MVT::v4i32, InputVT: MVT::v8i16, Action: Legal); |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | // As a special case, these operators use the type to mean the type to |
| 361 | // sign-extend from. |
| 362 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1, Action: Expand); |
| 363 | if (!Subtarget->hasSignExt()) { |
| 364 | // Sign extends are legal only when extending a vector extract |
| 365 | auto Action = Subtarget->hasSIMD128() ? Custom : Expand; |
| 366 | for (auto T : {MVT::i8, MVT::i16, MVT::i32}) |
| 367 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: T, Action); |
| 368 | } |
| 369 | for (auto T : MVT::integer_fixedlen_vector_valuetypes()) |
| 370 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: T, Action: Expand); |
| 371 | |
| 372 | // Dynamic stack allocation: use the default expansion. |
| 373 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
| 374 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand); |
| 375 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVTPtr, Action: Expand); |
| 376 | |
| 377 | setOperationAction(Op: ISD::FrameIndex, VT: MVT::i32, Action: Custom); |
| 378 | setOperationAction(Op: ISD::FrameIndex, VT: MVT::i64, Action: Custom); |
| 379 | setOperationAction(Op: ISD::CopyToReg, VT: MVT::Other, Action: Custom); |
| 380 | |
| 381 | // Expand these forms; we pattern-match the forms that we can handle in isel. |
| 382 | for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) |
| 383 | for (auto Op : {ISD::BR_CC, ISD::SELECT_CC}) |
| 384 | setOperationAction(Op, VT: T, Action: Expand); |
| 385 | |
| 386 | if (Subtarget->hasReferenceTypes()) |
| 387 | for (auto Op : {ISD::BR_CC, ISD::SELECT_CC}) |
| 388 | for (auto T : {MVT::externref, MVT::funcref}) |
| 389 | setOperationAction(Op, VT: T, Action: Expand); |
| 390 | |
| 391 | // There is no vector conditional select instruction |
| 392 | for (auto T : |
| 393 | {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, MVT::v2f64}) |
| 394 | setOperationAction(Op: ISD::SELECT_CC, VT: T, Action: Expand); |
| 395 | |
| 396 | // We have custom switch handling. |
| 397 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Custom); |
| 398 | |
| 399 | // WebAssembly doesn't have: |
| 400 | // - Floating-point extending loads. |
| 401 | // - Floating-point truncating stores. |
| 402 | // - i1 extending loads. |
| 403 | // - truncating SIMD stores and most extending loads |
| 404 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
| 405 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
| 406 | for (auto T : MVT::integer_valuetypes()) |
| 407 | for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD}) |
| 408 | setLoadExtAction(ExtType: Ext, ValVT: T, MemVT: MVT::i1, Action: Promote); |
| 409 | if (Subtarget->hasSIMD128()) { |
| 410 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32, |
| 411 | MVT::v2f64}) { |
| 412 | for (auto MemT : MVT::fixedlen_vector_valuetypes()) { |
| 413 | if (MVT(T) != MemT) { |
| 414 | setTruncStoreAction(ValVT: T, MemVT: MemT, Action: Expand); |
| 415 | for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD}) |
| 416 | setLoadExtAction(ExtType: Ext, ValVT: T, MemVT: MemT, Action: Expand); |
| 417 | } |
| 418 | } |
| 419 | } |
| 420 | // But some vector extending loads are legal |
| 421 | for (auto Ext : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) { |
| 422 | setLoadExtAction(ExtType: Ext, ValVT: MVT::v8i16, MemVT: MVT::v8i8, Action: Legal); |
| 423 | setLoadExtAction(ExtType: Ext, ValVT: MVT::v4i32, MemVT: MVT::v4i16, Action: Legal); |
| 424 | setLoadExtAction(ExtType: Ext, ValVT: MVT::v2i64, MemVT: MVT::v2i32, Action: Legal); |
| 425 | } |
| 426 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::v2f64, MemVT: MVT::v2f32, Action: Legal); |
| 427 | } |
| 428 | |
| 429 | // Don't do anything clever with build_pairs |
| 430 | setOperationAction(Op: ISD::BUILD_PAIR, VT: MVT::i64, Action: Expand); |
| 431 | |
| 432 | // Trap lowers to wasm unreachable |
| 433 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Legal); |
| 434 | setOperationAction(Op: ISD::DEBUGTRAP, VT: MVT::Other, Action: Legal); |
| 435 | |
| 436 | // Exception handling intrinsics |
| 437 | setOperationAction(Op: ISD::INTRINSIC_WO_CHAIN, VT: MVT::Other, Action: Custom); |
| 438 | setOperationAction(Op: ISD::INTRINSIC_W_CHAIN, VT: MVT::Other, Action: Custom); |
| 439 | setOperationAction(Op: ISD::INTRINSIC_VOID, VT: MVT::Other, Action: Custom); |
| 440 | |
| 441 | setMaxAtomicSizeInBitsSupported(64); |
| 442 | |
| 443 | // Always convert switches to br_tables unless there is only one case, which |
| 444 | // is equivalent to a simple branch. This reduces code size for wasm, and we |
| 445 | // defer possible jump table optimizations to the VM. |
| 446 | setMinimumJumpTableEntries(2); |
| 447 | } |
| 448 | |
| 449 | TargetLowering::AtomicExpansionKind |
| 450 | WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR( |
| 451 | const AtomicRMWInst *AI) const { |
| 452 | // We have wasm instructions for these |
| 453 | switch (AI->getOperation()) { |
| 454 | case AtomicRMWInst::Add: |
| 455 | case AtomicRMWInst::Sub: |
| 456 | case AtomicRMWInst::And: |
| 457 | case AtomicRMWInst::Or: |
| 458 | case AtomicRMWInst::Xor: |
| 459 | case AtomicRMWInst::Xchg: |
| 460 | return AtomicExpansionKind::None; |
| 461 | default: |
| 462 | break; |
| 463 | } |
| 464 | return AtomicExpansionKind::CmpXChg; |
| 465 | } |
| 466 | |
| 467 | bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const { |
| 468 | // Implementation copied from X86TargetLowering. |
| 469 | unsigned Opc = VecOp.getOpcode(); |
| 470 | |
| 471 | // Assume target opcodes can't be scalarized. |
| 472 | // TODO - do we have any exceptions? |
| 473 | if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opcode: Opc)) |
| 474 | return false; |
| 475 | |
| 476 | // If the vector op is not supported, try to convert to scalar. |
| 477 | EVT VecVT = VecOp.getValueType(); |
| 478 | if (!isOperationLegalOrCustomOrPromote(Op: Opc, VT: VecVT)) |
| 479 | return true; |
| 480 | |
| 481 | // If the vector op is supported, but the scalar op is not, the transform may |
| 482 | // not be worthwhile. |
| 483 | EVT ScalarVT = VecVT.getScalarType(); |
| 484 | return isOperationLegalOrCustomOrPromote(Op: Opc, VT: ScalarVT); |
| 485 | } |
| 486 | |
| 487 | FastISel *WebAssemblyTargetLowering::createFastISel( |
| 488 | FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo, |
| 489 | const LibcallLoweringInfo *LibcallLowering) const { |
| 490 | return WebAssembly::createFastISel(funcInfo&: FuncInfo, libInfo: LibInfo, libcallLowering: LibcallLowering); |
| 491 | } |
| 492 | |
| 493 | MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/, |
| 494 | EVT VT) const { |
| 495 | unsigned BitWidth = NextPowerOf2(A: VT.getSizeInBits() - 1); |
| 496 | if (BitWidth > 1 && BitWidth < 8) |
| 497 | BitWidth = 8; |
| 498 | |
| 499 | if (BitWidth > 64) { |
| 500 | // The shift will be lowered to a libcall, and compiler-rt libcalls expect |
| 501 | // the count to be an i32. |
| 502 | BitWidth = 32; |
| 503 | assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) && |
| 504 | "32-bit shift counts ought to be enough for anyone" ); |
| 505 | } |
| 506 | |
| 507 | MVT Result = MVT::getIntegerVT(BitWidth); |
| 508 | assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE && |
| 509 | "Unable to represent scalar shift amount type" ); |
| 510 | return Result; |
| 511 | } |
| 512 | |
| 513 | // Lower an fp-to-int conversion operator from the LLVM opcode, which has an |
| 514 | // undefined result on invalid/overflow, to the WebAssembly opcode, which |
| 515 | // traps on invalid/overflow. |
| 516 | static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL, |
| 517 | MachineBasicBlock *BB, |
| 518 | const TargetInstrInfo &TII, |
| 519 | bool IsUnsigned, bool Int64, |
| 520 | bool Float64, unsigned LoweredOpcode) { |
| 521 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
| 522 | |
| 523 | Register OutReg = MI.getOperand(i: 0).getReg(); |
| 524 | Register InReg = MI.getOperand(i: 1).getReg(); |
| 525 | |
| 526 | unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32; |
| 527 | unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32; |
| 528 | unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32; |
| 529 | unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32; |
| 530 | unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32; |
| 531 | unsigned Eqz = WebAssembly::EQZ_I32; |
| 532 | unsigned And = WebAssembly::AND_I32; |
| 533 | int64_t Limit = Int64 ? INT64_MIN : INT32_MIN; |
| 534 | int64_t Substitute = IsUnsigned ? 0 : Limit; |
| 535 | double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit; |
| 536 | auto &Context = BB->getParent()->getFunction().getContext(); |
| 537 | Type *Ty = Float64 ? Type::getDoubleTy(C&: Context) : Type::getFloatTy(C&: Context); |
| 538 | |
| 539 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
| 540 | MachineFunction *F = BB->getParent(); |
| 541 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 542 | MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 543 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 544 | |
| 545 | MachineFunction::iterator It = ++BB->getIterator(); |
| 546 | F->insert(MBBI: It, MBB: FalseMBB); |
| 547 | F->insert(MBBI: It, MBB: TrueMBB); |
| 548 | F->insert(MBBI: It, MBB: DoneMBB); |
| 549 | |
| 550 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
| 551 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
| 552 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 553 | |
| 554 | BB->addSuccessor(Succ: TrueMBB); |
| 555 | BB->addSuccessor(Succ: FalseMBB); |
| 556 | TrueMBB->addSuccessor(Succ: DoneMBB); |
| 557 | FalseMBB->addSuccessor(Succ: DoneMBB); |
| 558 | |
| 559 | unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg; |
| 560 | Tmp0 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
| 561 | Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
| 562 | CmpReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 563 | EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 564 | FalseReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg)); |
| 565 | TrueReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg)); |
| 566 | |
| 567 | MI.eraseFromParent(); |
| 568 | // For signed numbers, we can do a single comparison to determine whether |
| 569 | // fabs(x) is within range. |
| 570 | if (IsUnsigned) { |
| 571 | Tmp0 = InReg; |
| 572 | } else { |
| 573 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Abs), DestReg: Tmp0).addReg(RegNo: InReg); |
| 574 | } |
| 575 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1) |
| 576 | .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: CmpVal))); |
| 577 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: LT), DestReg: CmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1); |
| 578 | |
| 579 | // For unsigned numbers, we have to do a separate comparison with zero. |
| 580 | if (IsUnsigned) { |
| 581 | Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
| 582 | Register SecondCmpReg = |
| 583 | MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 584 | Register AndReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 585 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1) |
| 586 | .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: 0.0))); |
| 587 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: GE), DestReg: SecondCmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1); |
| 588 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: And), DestReg: AndReg).addReg(RegNo: CmpReg).addReg(RegNo: SecondCmpReg); |
| 589 | CmpReg = AndReg; |
| 590 | } |
| 591 | |
| 592 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).addReg(RegNo: CmpReg); |
| 593 | |
| 594 | // Create the CFG diamond to select between doing the conversion or using |
| 595 | // the substitute value. |
| 596 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: TrueMBB).addReg(RegNo: EqzReg); |
| 597 | BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: LoweredOpcode), DestReg: FalseReg).addReg(RegNo: InReg); |
| 598 | BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB); |
| 599 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: IConst), DestReg: TrueReg).addImm(Val: Substitute); |
| 600 | BuildMI(BB&: *DoneMBB, I: DoneMBB->begin(), MIMD: DL, MCID: TII.get(Opcode: TargetOpcode::PHI), DestReg: OutReg) |
| 601 | .addReg(RegNo: FalseReg) |
| 602 | .addMBB(MBB: FalseMBB) |
| 603 | .addReg(RegNo: TrueReg) |
| 604 | .addMBB(MBB: TrueMBB); |
| 605 | |
| 606 | return DoneMBB; |
| 607 | } |
| 608 | |
| 609 | // Lower a `MEMCPY` instruction into a CFG triangle around a `MEMORY_COPY` |
| 610 | // instuction to handle the zero-length case. |
| 611 | static MachineBasicBlock *LowerMemcpy(MachineInstr &MI, DebugLoc DL, |
| 612 | MachineBasicBlock *BB, |
| 613 | const TargetInstrInfo &TII, bool Int64) { |
| 614 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
| 615 | |
| 616 | MachineOperand DstMem = MI.getOperand(i: 0); |
| 617 | MachineOperand SrcMem = MI.getOperand(i: 1); |
| 618 | MachineOperand Dst = MI.getOperand(i: 2); |
| 619 | MachineOperand Src = MI.getOperand(i: 3); |
| 620 | MachineOperand Len = MI.getOperand(i: 4); |
| 621 | |
| 622 | // If the length is a constant, we don't actually need the check. |
| 623 | if (MachineInstr *Def = MRI.getVRegDef(Reg: Len.getReg())) { |
| 624 | if (Def->getOpcode() == WebAssembly::CONST_I32 || |
| 625 | Def->getOpcode() == WebAssembly::CONST_I64) { |
| 626 | if (Def->getOperand(i: 1).getImm() == 0) { |
| 627 | // A zero-length memcpy is a no-op. |
| 628 | MI.eraseFromParent(); |
| 629 | return BB; |
| 630 | } |
| 631 | // A non-zero-length memcpy doesn't need a zero check. |
| 632 | unsigned MemoryCopy = |
| 633 | Int64 ? WebAssembly::MEMORY_COPY_A64 : WebAssembly::MEMORY_COPY_A32; |
| 634 | BuildMI(BB&: *BB, I&: MI, MIMD: DL, MCID: TII.get(Opcode: MemoryCopy)) |
| 635 | .add(MO: DstMem) |
| 636 | .add(MO: SrcMem) |
| 637 | .add(MO: Dst) |
| 638 | .add(MO: Src) |
| 639 | .add(MO: Len); |
| 640 | MI.eraseFromParent(); |
| 641 | return BB; |
| 642 | } |
| 643 | } |
| 644 | |
| 645 | // We're going to add an extra use to `Len` to test if it's zero; that |
| 646 | // use shouldn't be a kill, even if the original use is. |
| 647 | MachineOperand NoKillLen = Len; |
| 648 | NoKillLen.setIsKill(false); |
| 649 | |
| 650 | // Decide on which `MachineInstr` opcode we're going to use. |
| 651 | unsigned Eqz = Int64 ? WebAssembly::EQZ_I64 : WebAssembly::EQZ_I32; |
| 652 | unsigned MemoryCopy = |
| 653 | Int64 ? WebAssembly::MEMORY_COPY_A64 : WebAssembly::MEMORY_COPY_A32; |
| 654 | |
| 655 | // Create two new basic blocks; one for the new `memory.fill` that we can |
| 656 | // branch over, and one for the rest of the instructions after the original |
| 657 | // `memory.fill`. |
| 658 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
| 659 | MachineFunction *F = BB->getParent(); |
| 660 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 661 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 662 | |
| 663 | MachineFunction::iterator It = ++BB->getIterator(); |
| 664 | F->insert(MBBI: It, MBB: TrueMBB); |
| 665 | F->insert(MBBI: It, MBB: DoneMBB); |
| 666 | |
| 667 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
| 668 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
| 669 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 670 | |
| 671 | // Connect the CFG edges. |
| 672 | BB->addSuccessor(Succ: TrueMBB); |
| 673 | BB->addSuccessor(Succ: DoneMBB); |
| 674 | TrueMBB->addSuccessor(Succ: DoneMBB); |
| 675 | |
| 676 | // Create a virtual register for the `Eqz` result. |
| 677 | unsigned EqzReg; |
| 678 | EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 679 | |
| 680 | // Erase the original `memory.copy`. |
| 681 | MI.eraseFromParent(); |
| 682 | |
| 683 | // Test if `Len` is zero. |
| 684 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).add(MO: NoKillLen); |
| 685 | |
| 686 | // Insert a new `memory.copy`. |
| 687 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: MemoryCopy)) |
| 688 | .add(MO: DstMem) |
| 689 | .add(MO: SrcMem) |
| 690 | .add(MO: Dst) |
| 691 | .add(MO: Src) |
| 692 | .add(MO: Len); |
| 693 | |
| 694 | // Create the CFG triangle. |
| 695 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: DoneMBB).addReg(RegNo: EqzReg); |
| 696 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB); |
| 697 | |
| 698 | return DoneMBB; |
| 699 | } |
| 700 | |
| 701 | // Lower a `MEMSET` instruction into a CFG triangle around a `MEMORY_FILL` |
| 702 | // instuction to handle the zero-length case. |
| 703 | static MachineBasicBlock *LowerMemset(MachineInstr &MI, DebugLoc DL, |
| 704 | MachineBasicBlock *BB, |
| 705 | const TargetInstrInfo &TII, bool Int64) { |
| 706 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
| 707 | |
| 708 | MachineOperand Mem = MI.getOperand(i: 0); |
| 709 | MachineOperand Dst = MI.getOperand(i: 1); |
| 710 | MachineOperand Val = MI.getOperand(i: 2); |
| 711 | MachineOperand Len = MI.getOperand(i: 3); |
| 712 | |
| 713 | // If the length is a constant, we don't actually need the check. |
| 714 | if (MachineInstr *Def = MRI.getVRegDef(Reg: Len.getReg())) { |
| 715 | if (Def->getOpcode() == WebAssembly::CONST_I32 || |
| 716 | Def->getOpcode() == WebAssembly::CONST_I64) { |
| 717 | if (Def->getOperand(i: 1).getImm() == 0) { |
| 718 | // A zero-length memset is a no-op. |
| 719 | MI.eraseFromParent(); |
| 720 | return BB; |
| 721 | } |
| 722 | // A non-zero-length memset doesn't need a zero check. |
| 723 | unsigned MemoryFill = |
| 724 | Int64 ? WebAssembly::MEMORY_FILL_A64 : WebAssembly::MEMORY_FILL_A32; |
| 725 | BuildMI(BB&: *BB, I&: MI, MIMD: DL, MCID: TII.get(Opcode: MemoryFill)) |
| 726 | .add(MO: Mem) |
| 727 | .add(MO: Dst) |
| 728 | .add(MO: Val) |
| 729 | .add(MO: Len); |
| 730 | MI.eraseFromParent(); |
| 731 | return BB; |
| 732 | } |
| 733 | } |
| 734 | |
| 735 | // We're going to add an extra use to `Len` to test if it's zero; that |
| 736 | // use shouldn't be a kill, even if the original use is. |
| 737 | MachineOperand NoKillLen = Len; |
| 738 | NoKillLen.setIsKill(false); |
| 739 | |
| 740 | // Decide on which `MachineInstr` opcode we're going to use. |
| 741 | unsigned Eqz = Int64 ? WebAssembly::EQZ_I64 : WebAssembly::EQZ_I32; |
| 742 | unsigned MemoryFill = |
| 743 | Int64 ? WebAssembly::MEMORY_FILL_A64 : WebAssembly::MEMORY_FILL_A32; |
| 744 | |
| 745 | // Create two new basic blocks; one for the new `memory.fill` that we can |
| 746 | // branch over, and one for the rest of the instructions after the original |
| 747 | // `memory.fill`. |
| 748 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
| 749 | MachineFunction *F = BB->getParent(); |
| 750 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 751 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
| 752 | |
| 753 | MachineFunction::iterator It = ++BB->getIterator(); |
| 754 | F->insert(MBBI: It, MBB: TrueMBB); |
| 755 | F->insert(MBBI: It, MBB: DoneMBB); |
| 756 | |
| 757 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
| 758 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
| 759 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 760 | |
| 761 | // Connect the CFG edges. |
| 762 | BB->addSuccessor(Succ: TrueMBB); |
| 763 | BB->addSuccessor(Succ: DoneMBB); |
| 764 | TrueMBB->addSuccessor(Succ: DoneMBB); |
| 765 | |
| 766 | // Create a virtual register for the `Eqz` result. |
| 767 | unsigned EqzReg; |
| 768 | EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 769 | |
| 770 | // Erase the original `memory.fill`. |
| 771 | MI.eraseFromParent(); |
| 772 | |
| 773 | // Test if `Len` is zero. |
| 774 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).add(MO: NoKillLen); |
| 775 | |
| 776 | // Insert a new `memory.copy`. |
| 777 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: MemoryFill)).add(MO: Mem).add(MO: Dst).add(MO: Val).add(MO: Len); |
| 778 | |
| 779 | // Create the CFG triangle. |
| 780 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: DoneMBB).addReg(RegNo: EqzReg); |
| 781 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB); |
| 782 | |
| 783 | return DoneMBB; |
| 784 | } |
| 785 | |
| 786 | static MachineBasicBlock * |
| 787 | LowerCallResults(MachineInstr &CallResults, DebugLoc DL, MachineBasicBlock *BB, |
| 788 | const WebAssemblySubtarget *Subtarget, |
| 789 | const TargetInstrInfo &TII) { |
| 790 | MachineInstr &CallParams = *CallResults.getPrevNode(); |
| 791 | assert(CallParams.getOpcode() == WebAssembly::CALL_PARAMS); |
| 792 | assert(CallResults.getOpcode() == WebAssembly::CALL_RESULTS || |
| 793 | CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS); |
| 794 | |
| 795 | bool IsIndirect = |
| 796 | CallParams.getOperand(i: 0).isReg() || CallParams.getOperand(i: 0).isFI(); |
| 797 | bool IsRetCall = CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS; |
| 798 | |
| 799 | bool IsFuncrefCall = false; |
| 800 | if (IsIndirect && CallParams.getOperand(i: 0).isReg()) { |
| 801 | Register Reg = CallParams.getOperand(i: 0).getReg(); |
| 802 | const MachineFunction *MF = BB->getParent(); |
| 803 | const MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 804 | const TargetRegisterClass *TRC = MRI.getRegClass(Reg); |
| 805 | IsFuncrefCall = (TRC == &WebAssembly::FUNCREFRegClass); |
| 806 | assert(!IsFuncrefCall || Subtarget->hasReferenceTypes()); |
| 807 | } |
| 808 | |
| 809 | unsigned CallOp; |
| 810 | if (IsIndirect && IsRetCall) { |
| 811 | CallOp = WebAssembly::RET_CALL_INDIRECT; |
| 812 | } else if (IsIndirect) { |
| 813 | CallOp = WebAssembly::CALL_INDIRECT; |
| 814 | } else if (IsRetCall) { |
| 815 | CallOp = WebAssembly::RET_CALL; |
| 816 | } else { |
| 817 | CallOp = WebAssembly::CALL; |
| 818 | } |
| 819 | |
| 820 | MachineFunction &MF = *BB->getParent(); |
| 821 | const MCInstrDesc &MCID = TII.get(Opcode: CallOp); |
| 822 | MachineInstrBuilder MIB(MF, MF.CreateMachineInstr(MCID, DL)); |
| 823 | |
| 824 | // Move the function pointer to the end of the arguments for indirect calls |
| 825 | if (IsIndirect) { |
| 826 | auto FnPtr = CallParams.getOperand(i: 0); |
| 827 | CallParams.removeOperand(OpNo: 0); |
| 828 | |
| 829 | // For funcrefs, call_indirect is done through __funcref_call_table and the |
| 830 | // funcref is always installed in slot 0 of the table, therefore instead of |
| 831 | // having the function pointer added at the end of the params list, a zero |
| 832 | // (the index in |
| 833 | // __funcref_call_table is added). |
| 834 | if (IsFuncrefCall) { |
| 835 | Register RegZero = |
| 836 | MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 837 | MachineInstrBuilder MIBC0 = |
| 838 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::CONST_I32), DestReg: RegZero).addImm(Val: 0); |
| 839 | |
| 840 | BB->insert(I: CallResults.getIterator(), M: MIBC0); |
| 841 | MachineInstrBuilder(MF, CallParams).addReg(RegNo: RegZero); |
| 842 | } else |
| 843 | CallParams.addOperand(Op: FnPtr); |
| 844 | } |
| 845 | |
| 846 | for (auto Def : CallResults.defs()) |
| 847 | MIB.add(MO: Def); |
| 848 | |
| 849 | if (IsIndirect) { |
| 850 | // Placeholder for the type index. |
| 851 | // This gets replaced with the correct value in WebAssemblyMCInstLower.cpp |
| 852 | MIB.addImm(Val: 0); |
| 853 | // The table into which this call_indirect indexes. |
| 854 | MCSymbolWasm *Table = IsFuncrefCall |
| 855 | ? WebAssembly::getOrCreateFuncrefCallTableSymbol( |
| 856 | Ctx&: MF.getContext(), Subtarget) |
| 857 | : WebAssembly::getOrCreateFunctionTableSymbol( |
| 858 | Ctx&: MF.getContext(), Subtarget); |
| 859 | if (Subtarget->hasCallIndirectOverlong()) { |
| 860 | MIB.addSym(Sym: Table); |
| 861 | } else { |
| 862 | // For the MVP there is at most one table whose number is 0, but we can't |
| 863 | // write a table symbol or issue relocations. Instead we just ensure the |
| 864 | // table is live and write a zero. |
| 865 | Table->setNoStrip(); |
| 866 | MIB.addImm(Val: 0); |
| 867 | } |
| 868 | } |
| 869 | |
| 870 | for (auto Use : CallParams.uses()) |
| 871 | MIB.add(MO: Use); |
| 872 | |
| 873 | BB->insert(I: CallResults.getIterator(), M: MIB); |
| 874 | CallParams.eraseFromParent(); |
| 875 | CallResults.eraseFromParent(); |
| 876 | |
| 877 | // If this is a funcref call, to avoid hidden GC roots, we need to clear the |
| 878 | // table slot with ref.null upon call_indirect return. |
| 879 | // |
| 880 | // This generates the following code, which comes right after a call_indirect |
| 881 | // of a funcref: |
| 882 | // |
| 883 | // i32.const 0 |
| 884 | // ref.null func |
| 885 | // table.set __funcref_call_table |
| 886 | if (IsIndirect && IsFuncrefCall) { |
| 887 | MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol( |
| 888 | Ctx&: MF.getContext(), Subtarget); |
| 889 | Register RegZero = |
| 890 | MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::I32RegClass); |
| 891 | MachineInstr *Const0 = |
| 892 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::CONST_I32), DestReg: RegZero).addImm(Val: 0); |
| 893 | BB->insertAfter(I: MIB.getInstr()->getIterator(), MI: Const0); |
| 894 | |
| 895 | Register RegFuncref = |
| 896 | MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::FUNCREFRegClass); |
| 897 | MachineInstr *RefNull = |
| 898 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::REF_NULL_FUNCREF), DestReg: RegFuncref); |
| 899 | BB->insertAfter(I: Const0->getIterator(), MI: RefNull); |
| 900 | |
| 901 | MachineInstr *TableSet = |
| 902 | BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::TABLE_SET_FUNCREF)) |
| 903 | .addSym(Sym: Table) |
| 904 | .addReg(RegNo: RegZero) |
| 905 | .addReg(RegNo: RegFuncref); |
| 906 | BB->insertAfter(I: RefNull->getIterator(), MI: TableSet); |
| 907 | } |
| 908 | |
| 909 | return BB; |
| 910 | } |
| 911 | |
| 912 | MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter( |
| 913 | MachineInstr &MI, MachineBasicBlock *BB) const { |
| 914 | const TargetInstrInfo &TII = *Subtarget->getInstrInfo(); |
| 915 | DebugLoc DL = MI.getDebugLoc(); |
| 916 | |
| 917 | switch (MI.getOpcode()) { |
| 918 | default: |
| 919 | llvm_unreachable("Unexpected instr type to insert" ); |
| 920 | case WebAssembly::FP_TO_SINT_I32_F32: |
| 921 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: false, Float64: false, |
| 922 | LoweredOpcode: WebAssembly::I32_TRUNC_S_F32); |
| 923 | case WebAssembly::FP_TO_UINT_I32_F32: |
| 924 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: false, Float64: false, |
| 925 | LoweredOpcode: WebAssembly::I32_TRUNC_U_F32); |
| 926 | case WebAssembly::FP_TO_SINT_I64_F32: |
| 927 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: true, Float64: false, |
| 928 | LoweredOpcode: WebAssembly::I64_TRUNC_S_F32); |
| 929 | case WebAssembly::FP_TO_UINT_I64_F32: |
| 930 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: true, Float64: false, |
| 931 | LoweredOpcode: WebAssembly::I64_TRUNC_U_F32); |
| 932 | case WebAssembly::FP_TO_SINT_I32_F64: |
| 933 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: false, Float64: true, |
| 934 | LoweredOpcode: WebAssembly::I32_TRUNC_S_F64); |
| 935 | case WebAssembly::FP_TO_UINT_I32_F64: |
| 936 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: false, Float64: true, |
| 937 | LoweredOpcode: WebAssembly::I32_TRUNC_U_F64); |
| 938 | case WebAssembly::FP_TO_SINT_I64_F64: |
| 939 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: true, Float64: true, |
| 940 | LoweredOpcode: WebAssembly::I64_TRUNC_S_F64); |
| 941 | case WebAssembly::FP_TO_UINT_I64_F64: |
| 942 | return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: true, Float64: true, |
| 943 | LoweredOpcode: WebAssembly::I64_TRUNC_U_F64); |
| 944 | case WebAssembly::MEMCPY_A32: |
| 945 | return LowerMemcpy(MI, DL, BB, TII, Int64: false); |
| 946 | case WebAssembly::MEMCPY_A64: |
| 947 | return LowerMemcpy(MI, DL, BB, TII, Int64: true); |
| 948 | case WebAssembly::MEMSET_A32: |
| 949 | return LowerMemset(MI, DL, BB, TII, Int64: false); |
| 950 | case WebAssembly::MEMSET_A64: |
| 951 | return LowerMemset(MI, DL, BB, TII, Int64: true); |
| 952 | case WebAssembly::CALL_RESULTS: |
| 953 | case WebAssembly::RET_CALL_RESULTS: |
| 954 | return LowerCallResults(CallResults&: MI, DL, BB, Subtarget, TII); |
| 955 | } |
| 956 | } |
| 957 | |
| 958 | std::pair<unsigned, const TargetRegisterClass *> |
| 959 | WebAssemblyTargetLowering::getRegForInlineAsmConstraint( |
| 960 | const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const { |
| 961 | // First, see if this is a constraint that directly corresponds to a |
| 962 | // WebAssembly register class. |
| 963 | if (Constraint.size() == 1) { |
| 964 | switch (Constraint[0]) { |
| 965 | case 'r': |
| 966 | assert(VT != MVT::iPTR && "Pointer MVT not expected here" ); |
| 967 | if (Subtarget->hasSIMD128() && VT.isVector()) { |
| 968 | if (VT.getSizeInBits() == 128) |
| 969 | return std::make_pair(x: 0U, y: &WebAssembly::V128RegClass); |
| 970 | } |
| 971 | if (VT.isInteger() && !VT.isVector()) { |
| 972 | if (VT.getSizeInBits() <= 32) |
| 973 | return std::make_pair(x: 0U, y: &WebAssembly::I32RegClass); |
| 974 | if (VT.getSizeInBits() <= 64) |
| 975 | return std::make_pair(x: 0U, y: &WebAssembly::I64RegClass); |
| 976 | } |
| 977 | if (VT.isFloatingPoint() && !VT.isVector()) { |
| 978 | switch (VT.getSizeInBits()) { |
| 979 | case 32: |
| 980 | return std::make_pair(x: 0U, y: &WebAssembly::F32RegClass); |
| 981 | case 64: |
| 982 | return std::make_pair(x: 0U, y: &WebAssembly::F64RegClass); |
| 983 | default: |
| 984 | break; |
| 985 | } |
| 986 | } |
| 987 | break; |
| 988 | default: |
| 989 | break; |
| 990 | } |
| 991 | } |
| 992 | |
| 993 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
| 994 | } |
| 995 | |
| 996 | bool WebAssemblyTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
| 997 | // Assume ctz is a relatively cheap operation. |
| 998 | return true; |
| 999 | } |
| 1000 | |
| 1001 | bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const { |
| 1002 | // Assume clz is a relatively cheap operation. |
| 1003 | return true; |
| 1004 | } |
| 1005 | |
| 1006 | bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
| 1007 | const AddrMode &AM, |
| 1008 | Type *Ty, unsigned AS, |
| 1009 | Instruction *I) const { |
| 1010 | // WebAssembly offsets are added as unsigned without wrapping. The |
| 1011 | // isLegalAddressingMode gives us no way to determine if wrapping could be |
| 1012 | // happening, so we approximate this by accepting only non-negative offsets. |
| 1013 | if (AM.BaseOffs < 0) |
| 1014 | return false; |
| 1015 | |
| 1016 | // WebAssembly has no scale register operands. |
| 1017 | if (AM.Scale != 0) |
| 1018 | return false; |
| 1019 | |
| 1020 | // Everything else is legal. |
| 1021 | return true; |
| 1022 | } |
| 1023 | |
| 1024 | bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses( |
| 1025 | EVT /*VT*/, unsigned /*AddrSpace*/, Align /*Align*/, |
| 1026 | MachineMemOperand::Flags /*Flags*/, unsigned *Fast) const { |
| 1027 | // WebAssembly supports unaligned accesses, though it should be declared |
| 1028 | // with the p2align attribute on loads and stores which do so, and there |
| 1029 | // may be a performance impact. We tell LLVM they're "fast" because |
| 1030 | // for the kinds of things that LLVM uses this for (merging adjacent stores |
| 1031 | // of constants, etc.), WebAssembly implementations will either want the |
| 1032 | // unaligned access or they'll split anyway. |
| 1033 | if (Fast) |
| 1034 | *Fast = 1; |
| 1035 | return true; |
| 1036 | } |
| 1037 | |
| 1038 | bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT, |
| 1039 | AttributeList Attr) const { |
| 1040 | // The current thinking is that wasm engines will perform this optimization, |
| 1041 | // so we can save on code size. |
| 1042 | return true; |
| 1043 | } |
| 1044 | |
| 1045 | bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const { |
| 1046 | EVT ExtT = ExtVal.getValueType(); |
| 1047 | SDValue N0 = peekThroughFreeze(V: ExtVal->getOperand(Num: 0)); |
| 1048 | auto *Load = dyn_cast<LoadSDNode>(Val&: N0); |
| 1049 | if (!Load) |
| 1050 | return false; |
| 1051 | EVT MemT = Load->getValueType(ResNo: 0); |
| 1052 | return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) || |
| 1053 | (ExtT == MVT::v4i32 && MemT == MVT::v4i16) || |
| 1054 | (ExtT == MVT::v2i64 && MemT == MVT::v2i32); |
| 1055 | } |
| 1056 | |
| 1057 | bool WebAssemblyTargetLowering::isOffsetFoldingLegal( |
| 1058 | const GlobalAddressSDNode *GA) const { |
| 1059 | // Wasm doesn't support function addresses with offsets |
| 1060 | const GlobalValue *GV = GA->getGlobal(); |
| 1061 | return isa<Function>(Val: GV) ? false : TargetLowering::isOffsetFoldingLegal(GA); |
| 1062 | } |
| 1063 | |
| 1064 | EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL, |
| 1065 | LLVMContext &C, |
| 1066 | EVT VT) const { |
| 1067 | if (VT.isVector()) { |
| 1068 | if (VT.getVectorElementType() == MVT::f16 && !Subtarget->hasFP16()) |
| 1069 | return VT.changeElementType(Context&: C, EltVT: MVT::i1); |
| 1070 | |
| 1071 | return VT.changeVectorElementTypeToInteger(); |
| 1072 | } |
| 1073 | |
| 1074 | // So far, all branch instructions in Wasm take an I32 condition. |
| 1075 | // The default TargetLowering::getSetCCResultType returns the pointer size, |
| 1076 | // which would be useful to reduce instruction counts when testing |
| 1077 | // against 64-bit pointers/values if at some point Wasm supports that. |
| 1078 | return EVT::getIntegerVT(Context&: C, BitWidth: 32); |
| 1079 | } |
| 1080 | |
| 1081 | void WebAssemblyTargetLowering::getTgtMemIntrinsic( |
| 1082 | SmallVectorImpl<IntrinsicInfo> &Infos, const CallBase &I, |
| 1083 | MachineFunction &MF, unsigned Intrinsic) const { |
| 1084 | IntrinsicInfo Info; |
| 1085 | switch (Intrinsic) { |
| 1086 | case Intrinsic::wasm_memory_atomic_notify: |
| 1087 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1088 | Info.memVT = MVT::i32; |
| 1089 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1090 | Info.offset = 0; |
| 1091 | Info.align = Align(4); |
| 1092 | // atomic.notify instruction does not really load the memory specified with |
| 1093 | // this argument, but MachineMemOperand should either be load or store, so |
| 1094 | // we set this to a load. |
| 1095 | // FIXME Volatile isn't really correct, but currently all LLVM atomic |
| 1096 | // instructions are treated as volatiles in the backend, so we should be |
| 1097 | // consistent. The same applies for wasm_atomic_wait intrinsics too. |
| 1098 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
| 1099 | Infos.push_back(Elt: Info); |
| 1100 | return; |
| 1101 | case Intrinsic::wasm_memory_atomic_wait32: |
| 1102 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1103 | Info.memVT = MVT::i32; |
| 1104 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1105 | Info.offset = 0; |
| 1106 | Info.align = Align(4); |
| 1107 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
| 1108 | Infos.push_back(Elt: Info); |
| 1109 | return; |
| 1110 | case Intrinsic::wasm_memory_atomic_wait64: |
| 1111 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1112 | Info.memVT = MVT::i64; |
| 1113 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1114 | Info.offset = 0; |
| 1115 | Info.align = Align(8); |
| 1116 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
| 1117 | Infos.push_back(Elt: Info); |
| 1118 | return; |
| 1119 | case Intrinsic::wasm_loadf16_f32: |
| 1120 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 1121 | Info.memVT = MVT::f16; |
| 1122 | Info.ptrVal = I.getArgOperand(i: 0); |
| 1123 | Info.offset = 0; |
| 1124 | Info.align = Align(2); |
| 1125 | Info.flags = MachineMemOperand::MOLoad; |
| 1126 | Infos.push_back(Elt: Info); |
| 1127 | return; |
| 1128 | case Intrinsic::wasm_storef16_f32: |
| 1129 | Info.opc = ISD::INTRINSIC_VOID; |
| 1130 | Info.memVT = MVT::f16; |
| 1131 | Info.ptrVal = I.getArgOperand(i: 1); |
| 1132 | Info.offset = 0; |
| 1133 | Info.align = Align(2); |
| 1134 | Info.flags = MachineMemOperand::MOStore; |
| 1135 | Infos.push_back(Elt: Info); |
| 1136 | return; |
| 1137 | default: |
| 1138 | return; |
| 1139 | } |
| 1140 | } |
| 1141 | |
| 1142 | void WebAssemblyTargetLowering::computeKnownBitsForTargetNode( |
| 1143 | const SDValue Op, KnownBits &Known, const APInt &DemandedElts, |
| 1144 | const SelectionDAG &DAG, unsigned Depth) const { |
| 1145 | switch (Op.getOpcode()) { |
| 1146 | default: |
| 1147 | break; |
| 1148 | case ISD::INTRINSIC_WO_CHAIN: { |
| 1149 | unsigned IntNo = Op.getConstantOperandVal(i: 0); |
| 1150 | switch (IntNo) { |
| 1151 | default: |
| 1152 | break; |
| 1153 | case Intrinsic::wasm_bitmask: { |
| 1154 | unsigned BitWidth = Known.getBitWidth(); |
| 1155 | EVT VT = Op.getOperand(i: 1).getSimpleValueType(); |
| 1156 | unsigned PossibleBits = VT.getVectorNumElements(); |
| 1157 | APInt ZeroMask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - PossibleBits); |
| 1158 | Known.Zero |= ZeroMask; |
| 1159 | break; |
| 1160 | } |
| 1161 | } |
| 1162 | break; |
| 1163 | } |
| 1164 | case WebAssemblyISD::EXTEND_LOW_U: |
| 1165 | case WebAssemblyISD::EXTEND_HIGH_U: { |
| 1166 | // We know the high half, of each destination vector element, will be zero. |
| 1167 | SDValue SrcOp = Op.getOperand(i: 0); |
| 1168 | EVT VT = SrcOp.getSimpleValueType(); |
| 1169 | unsigned BitWidth = Known.getBitWidth(); |
| 1170 | if (VT == MVT::v8i8 || VT == MVT::v16i8) { |
| 1171 | assert(BitWidth >= 8 && "Unexpected width!" ); |
| 1172 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 8); |
| 1173 | Known.Zero |= Mask; |
| 1174 | } else if (VT == MVT::v4i16 || VT == MVT::v8i16) { |
| 1175 | assert(BitWidth >= 16 && "Unexpected width!" ); |
| 1176 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 16); |
| 1177 | Known.Zero |= Mask; |
| 1178 | } else if (VT == MVT::v2i32 || VT == MVT::v4i32) { |
| 1179 | assert(BitWidth >= 32 && "Unexpected width!" ); |
| 1180 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 32); |
| 1181 | Known.Zero |= Mask; |
| 1182 | } |
| 1183 | break; |
| 1184 | } |
| 1185 | // For 128-bit addition if the upper bits are all zero then it's known that |
| 1186 | // the upper bits of the result will have all bits guaranteed zero except the |
| 1187 | // first. |
| 1188 | case WebAssemblyISD::I64_ADD128: |
| 1189 | if (Op.getResNo() == 1) { |
| 1190 | SDValue LHS_HI = Op.getOperand(i: 1); |
| 1191 | SDValue RHS_HI = Op.getOperand(i: 3); |
| 1192 | if (isNullConstant(V: LHS_HI) && isNullConstant(V: RHS_HI)) |
| 1193 | Known.Zero.setBitsFrom(1); |
| 1194 | } |
| 1195 | break; |
| 1196 | } |
| 1197 | } |
| 1198 | |
| 1199 | TargetLoweringBase::LegalizeTypeAction |
| 1200 | WebAssemblyTargetLowering::getPreferredVectorAction(MVT VT) const { |
| 1201 | if (VT.isFixedLengthVector()) { |
| 1202 | MVT EltVT = VT.getVectorElementType(); |
| 1203 | // We have legal vector types with these lane types, so widening the |
| 1204 | // vector would let us use some of the lanes directly without having to |
| 1205 | // extend or truncate values. |
| 1206 | if (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 || |
| 1207 | EltVT == MVT::i64 || EltVT == MVT::f32 || EltVT == MVT::f64) |
| 1208 | return TypeWidenVector; |
| 1209 | } |
| 1210 | |
| 1211 | return TargetLoweringBase::getPreferredVectorAction(VT); |
| 1212 | } |
| 1213 | |
| 1214 | bool WebAssemblyTargetLowering::isFMAFasterThanFMulAndFAdd( |
| 1215 | const MachineFunction &MF, EVT VT) const { |
| 1216 | if (!Subtarget->hasFP16() || !VT.isVector()) |
| 1217 | return false; |
| 1218 | |
| 1219 | EVT ScalarVT = VT.getScalarType(); |
| 1220 | if (!ScalarVT.isSimple()) |
| 1221 | return false; |
| 1222 | |
| 1223 | return ScalarVT.getSimpleVT().SimpleTy == MVT::f16; |
| 1224 | } |
| 1225 | |
| 1226 | bool WebAssemblyTargetLowering::shouldSimplifyDemandedVectorElts( |
| 1227 | SDValue Op, const TargetLoweringOpt &TLO) const { |
| 1228 | // ISel process runs DAGCombiner after legalization; this step is called |
| 1229 | // SelectionDAG optimization phase. This post-legalization combining process |
| 1230 | // runs DAGCombiner on each node, and if there was a change to be made, |
| 1231 | // re-runs legalization again on it and its user nodes to make sure |
| 1232 | // everythiing is in a legalized state. |
| 1233 | // |
| 1234 | // The legalization calls lowering routines, and we do our custom lowering for |
| 1235 | // build_vectors (LowerBUILD_VECTOR), which converts undef vector elements |
| 1236 | // into zeros. But there is a set of routines in DAGCombiner that turns unused |
| 1237 | // (= not demanded) nodes into undef, among which SimplifyDemandedVectorElts |
| 1238 | // turns unused vector elements into undefs. But this routine does not work |
| 1239 | // with our custom LowerBUILD_VECTOR, which turns undefs into zeros. This |
| 1240 | // combination can result in a infinite loop, in which undefs are converted to |
| 1241 | // zeros in legalization and back to undefs in combining. |
| 1242 | // |
| 1243 | // So after DAG is legalized, we prevent SimplifyDemandedVectorElts from |
| 1244 | // running for build_vectors. |
| 1245 | if (Op.getOpcode() == ISD::BUILD_VECTOR && TLO.LegalOps && TLO.LegalTys) |
| 1246 | return false; |
| 1247 | return true; |
| 1248 | } |
| 1249 | |
| 1250 | //===----------------------------------------------------------------------===// |
| 1251 | // WebAssembly Lowering private implementation. |
| 1252 | //===----------------------------------------------------------------------===// |
| 1253 | |
| 1254 | //===----------------------------------------------------------------------===// |
| 1255 | // Lowering Code |
| 1256 | //===----------------------------------------------------------------------===// |
| 1257 | |
| 1258 | static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) { |
| 1259 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1260 | DAG.getContext()->diagnose( |
| 1261 | DI: DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc())); |
| 1262 | } |
| 1263 | |
| 1264 | // Test whether the given calling convention is supported. |
| 1265 | static bool callingConvSupported(CallingConv::ID CallConv) { |
| 1266 | // We currently support the language-independent target-independent |
| 1267 | // conventions. We don't yet have a way to annotate calls with properties like |
| 1268 | // "cold", and we don't have any call-clobbered registers, so these are mostly |
| 1269 | // all handled the same. |
| 1270 | return CallConv == CallingConv::C || CallConv == CallingConv::Fast || |
| 1271 | CallConv == CallingConv::Cold || |
| 1272 | CallConv == CallingConv::PreserveMost || |
| 1273 | CallConv == CallingConv::PreserveAll || |
| 1274 | CallConv == CallingConv::CXX_FAST_TLS || |
| 1275 | CallConv == CallingConv::WASM_EmscriptenInvoke || |
| 1276 | CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail; |
| 1277 | } |
| 1278 | |
| 1279 | SDValue |
| 1280 | WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI, |
| 1281 | SmallVectorImpl<SDValue> &InVals) const { |
| 1282 | SelectionDAG &DAG = CLI.DAG; |
| 1283 | SDLoc DL = CLI.DL; |
| 1284 | SDValue Chain = CLI.Chain; |
| 1285 | SDValue Callee = CLI.Callee; |
| 1286 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1287 | auto Layout = MF.getDataLayout(); |
| 1288 | |
| 1289 | // A call through a funcref is expressed in IR as a call through the pointer |
| 1290 | // produced by the llvm.wasm.funcref.to_ptr intrinsic. Detect this here and |
| 1291 | // recover the underlying funcref value so the call can be lowered to a |
| 1292 | // table.set + call_indirect through the dedicated __funcref_call_table. |
| 1293 | bool IsFuncrefCall = false; |
| 1294 | if (Callee.getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
| 1295 | Callee.getConstantOperandVal(i: 0) == Intrinsic::wasm_funcref_to_ptr) { |
| 1296 | Callee = Callee.getOperand(i: 1); |
| 1297 | IsFuncrefCall = true; |
| 1298 | } |
| 1299 | |
| 1300 | CallingConv::ID CallConv = CLI.CallConv; |
| 1301 | if (!callingConvSupported(CallConv)) |
| 1302 | fail(DL, DAG, |
| 1303 | Msg: "WebAssembly doesn't support language-specific or target-specific " |
| 1304 | "calling conventions yet" ); |
| 1305 | if (CLI.IsPatchPoint) |
| 1306 | fail(DL, DAG, Msg: "WebAssembly doesn't support patch point yet" ); |
| 1307 | |
| 1308 | if (CLI.IsTailCall) { |
| 1309 | auto NoTail = [&](const char *Msg) { |
| 1310 | if (CLI.CB && CLI.CB->isMustTailCall()) |
| 1311 | fail(DL, DAG, Msg); |
| 1312 | CLI.IsTailCall = false; |
| 1313 | }; |
| 1314 | |
| 1315 | if (!Subtarget->hasTailCall()) |
| 1316 | NoTail("WebAssembly 'tail-call' feature not enabled" ); |
| 1317 | |
| 1318 | // Varargs calls cannot be tail calls because the buffer is on the stack |
| 1319 | if (CLI.IsVarArg) |
| 1320 | NoTail("WebAssembly does not support varargs tail calls" ); |
| 1321 | |
| 1322 | // Do not tail call unless caller and callee return types match |
| 1323 | const Function &F = MF.getFunction(); |
| 1324 | const TargetMachine &TM = getTargetMachine(); |
| 1325 | Type *RetTy = F.getReturnType(); |
| 1326 | SmallVector<MVT, 4> CallerRetTys; |
| 1327 | SmallVector<MVT, 4> CalleeRetTys; |
| 1328 | computeLegalValueVTs(F, TM, Ty: RetTy, ValueVTs&: CallerRetTys); |
| 1329 | computeLegalValueVTs(F, TM, Ty: CLI.RetTy, ValueVTs&: CalleeRetTys); |
| 1330 | bool TypesMatch = CallerRetTys.size() == CalleeRetTys.size() && |
| 1331 | std::equal(first1: CallerRetTys.begin(), last1: CallerRetTys.end(), |
| 1332 | first2: CalleeRetTys.begin()); |
| 1333 | if (!TypesMatch) |
| 1334 | NoTail("WebAssembly tail call requires caller and callee return types to " |
| 1335 | "match" ); |
| 1336 | |
| 1337 | // If pointers to local stack values are passed, we cannot tail call |
| 1338 | if (CLI.CB) { |
| 1339 | for (auto &Arg : CLI.CB->args()) { |
| 1340 | Value *Val = Arg.get(); |
| 1341 | // Trace the value back through pointer operations |
| 1342 | while (true) { |
| 1343 | Value *Src = Val->stripPointerCastsAndAliases(); |
| 1344 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: Src)) |
| 1345 | Src = GEP->getPointerOperand(); |
| 1346 | if (Val == Src) |
| 1347 | break; |
| 1348 | Val = Src; |
| 1349 | } |
| 1350 | if (isa<AllocaInst>(Val)) { |
| 1351 | NoTail( |
| 1352 | "WebAssembly does not support tail calling with stack arguments" ); |
| 1353 | break; |
| 1354 | } |
| 1355 | } |
| 1356 | } |
| 1357 | } |
| 1358 | |
| 1359 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
| 1360 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
| 1361 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
| 1362 | |
| 1363 | // The generic code may have added an sret argument. If we're lowering an |
| 1364 | // invoke function, the ABI requires that the function pointer be the first |
| 1365 | // argument, so we may have to swap the arguments. |
| 1366 | if (CallConv == CallingConv::WASM_EmscriptenInvoke && Outs.size() >= 2 && |
| 1367 | Outs[0].Flags.isSRet()) { |
| 1368 | std::swap(a&: Outs[0], b&: Outs[1]); |
| 1369 | std::swap(a&: OutVals[0], b&: OutVals[1]); |
| 1370 | } |
| 1371 | |
| 1372 | bool HasSwiftSelfArg = false; |
| 1373 | bool HasSwiftErrorArg = false; |
| 1374 | bool HasSwiftAsyncArg = false; |
| 1375 | unsigned NumFixedArgs = 0; |
| 1376 | for (unsigned I = 0; I < Outs.size(); ++I) { |
| 1377 | const ISD::OutputArg &Out = Outs[I]; |
| 1378 | SDValue &OutVal = OutVals[I]; |
| 1379 | HasSwiftSelfArg |= Out.Flags.isSwiftSelf(); |
| 1380 | HasSwiftErrorArg |= Out.Flags.isSwiftError(); |
| 1381 | HasSwiftAsyncArg |= Out.Flags.isSwiftAsync(); |
| 1382 | if (Out.Flags.isNest()) |
| 1383 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments" ); |
| 1384 | if (Out.Flags.isInAlloca()) |
| 1385 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments" ); |
| 1386 | if (Out.Flags.isInConsecutiveRegs()) |
| 1387 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments" ); |
| 1388 | if (Out.Flags.isInConsecutiveRegsLast()) |
| 1389 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments" ); |
| 1390 | if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) { |
| 1391 | auto &MFI = MF.getFrameInfo(); |
| 1392 | int FI = MFI.CreateStackObject(Size: Out.Flags.getByValSize(), |
| 1393 | Alignment: Out.Flags.getNonZeroByValAlign(), |
| 1394 | /*isSS=*/isSpillSlot: false); |
| 1395 | SDValue SizeNode = |
| 1396 | DAG.getConstant(Val: Out.Flags.getByValSize(), DL, VT: MVT::i32); |
| 1397 | SDValue FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout)); |
| 1398 | Align Alignment = Out.Flags.getNonZeroByValAlign(); |
| 1399 | Chain = DAG.getMemcpy(Chain, dl: DL, Dst: FINode, Src: OutVal, Size: SizeNode, DstAlign: Alignment, |
| 1400 | SrcAlign: Alignment, |
| 1401 | /*isVolatile*/ isVol: false, /*AlwaysInline=*/false, |
| 1402 | /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(), |
| 1403 | SrcPtrInfo: MachinePointerInfo()); |
| 1404 | OutVal = FINode; |
| 1405 | } |
| 1406 | // Count the number of fixed args *after* legalization. |
| 1407 | NumFixedArgs += !Out.Flags.isVarArg(); |
| 1408 | } |
| 1409 | |
| 1410 | bool IsVarArg = CLI.IsVarArg; |
| 1411 | auto PtrVT = getPointerTy(DL: Layout); |
| 1412 | |
| 1413 | // For swiftcc and swifttailcc, emit additional swiftself, swifterror, and |
| 1414 | // (for swifttailcc) swiftasync arguments if there aren't. These additional |
| 1415 | // arguments are also added for callee signature. They are necessary to match |
| 1416 | // callee and caller signature for indirect call. |
| 1417 | if (CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail) { |
| 1418 | Type *PtrTy = PointerType::getUnqual(C&: *DAG.getContext()); |
| 1419 | if (!HasSwiftSelfArg) { |
| 1420 | NumFixedArgs++; |
| 1421 | ISD::ArgFlagsTy Flags; |
| 1422 | Flags.setSwiftSelf(); |
| 1423 | ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0); |
| 1424 | CLI.Outs.push_back(Elt: Arg); |
| 1425 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
| 1426 | CLI.OutVals.push_back(Elt: ArgVal); |
| 1427 | } |
| 1428 | if (!HasSwiftErrorArg) { |
| 1429 | NumFixedArgs++; |
| 1430 | ISD::ArgFlagsTy Flags; |
| 1431 | Flags.setSwiftError(); |
| 1432 | ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0); |
| 1433 | CLI.Outs.push_back(Elt: Arg); |
| 1434 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
| 1435 | CLI.OutVals.push_back(Elt: ArgVal); |
| 1436 | } |
| 1437 | if (CallConv == CallingConv::SwiftTail && !HasSwiftAsyncArg) { |
| 1438 | NumFixedArgs++; |
| 1439 | ISD::ArgFlagsTy Flags; |
| 1440 | Flags.setSwiftAsync(); |
| 1441 | ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0); |
| 1442 | CLI.Outs.push_back(Elt: Arg); |
| 1443 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
| 1444 | CLI.OutVals.push_back(Elt: ArgVal); |
| 1445 | } |
| 1446 | } |
| 1447 | |
| 1448 | // Analyze operands of the call, assigning locations to each operand. |
| 1449 | SmallVector<CCValAssign, 16> ArgLocs; |
| 1450 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
| 1451 | |
| 1452 | if (IsVarArg) { |
| 1453 | // Outgoing non-fixed arguments are placed in a buffer. First |
| 1454 | // compute their offsets and the total amount of buffer space needed. |
| 1455 | for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) { |
| 1456 | const ISD::OutputArg &Out = Outs[I]; |
| 1457 | SDValue &Arg = OutVals[I]; |
| 1458 | EVT VT = Arg.getValueType(); |
| 1459 | assert(VT != MVT::iPTR && "Legalized args should be concrete" ); |
| 1460 | Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext()); |
| 1461 | Align Alignment = |
| 1462 | std::max(a: Out.Flags.getNonZeroOrigAlign(), b: Layout.getABITypeAlign(Ty)); |
| 1463 | unsigned Offset = |
| 1464 | CCInfo.AllocateStack(Size: Layout.getTypeAllocSize(Ty), Alignment); |
| 1465 | CCInfo.addLoc(V: CCValAssign::getMem(ValNo: ArgLocs.size(), ValVT: VT.getSimpleVT(), |
| 1466 | Offset, LocVT: VT.getSimpleVT(), |
| 1467 | HTP: CCValAssign::Full)); |
| 1468 | } |
| 1469 | } |
| 1470 | |
| 1471 | unsigned NumBytes = CCInfo.getAlignedCallFrameSize(); |
| 1472 | |
| 1473 | SDValue FINode; |
| 1474 | if (IsVarArg && NumBytes) { |
| 1475 | // For non-fixed arguments, next emit stores to store the argument values |
| 1476 | // to the stack buffer at the offsets computed above. |
| 1477 | MaybeAlign StackAlign = Layout.getStackAlignment(); |
| 1478 | assert(StackAlign && "data layout string is missing stack alignment" ); |
| 1479 | int FI = MF.getFrameInfo().CreateStackObject(Size: NumBytes, Alignment: *StackAlign, |
| 1480 | /*isSS=*/isSpillSlot: false); |
| 1481 | unsigned ValNo = 0; |
| 1482 | SmallVector<SDValue, 8> Chains; |
| 1483 | for (SDValue Arg : drop_begin(RangeOrContainer&: OutVals, N: NumFixedArgs)) { |
| 1484 | assert(ArgLocs[ValNo].getValNo() == ValNo && |
| 1485 | "ArgLocs should remain in order and only hold varargs args" ); |
| 1486 | unsigned Offset = ArgLocs[ValNo++].getLocMemOffset(); |
| 1487 | FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout)); |
| 1488 | SDValue Add = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: FINode, |
| 1489 | N2: DAG.getConstant(Val: Offset, DL, VT: PtrVT)); |
| 1490 | Chains.push_back( |
| 1491 | Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: Add, |
| 1492 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI, Offset))); |
| 1493 | } |
| 1494 | if (!Chains.empty()) |
| 1495 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: Chains); |
| 1496 | } else if (IsVarArg) { |
| 1497 | FINode = DAG.getIntPtrConstant(Val: 0, DL); |
| 1498 | } |
| 1499 | |
| 1500 | if (Callee->getOpcode() == ISD::GlobalAddress) { |
| 1501 | // If the callee is a GlobalAddress node (quite common, every direct call |
| 1502 | // is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress |
| 1503 | // doesn't at MO_GOT which is not needed for direct calls. |
| 1504 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Callee); |
| 1505 | Callee = DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, |
| 1506 | VT: getPointerTy(DL: DAG.getDataLayout()), |
| 1507 | offset: GA->getOffset()); |
| 1508 | Callee = DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, |
| 1509 | VT: getPointerTy(DL: DAG.getDataLayout()), Operand: Callee); |
| 1510 | } |
| 1511 | |
| 1512 | // Compute the operands for the CALLn node. |
| 1513 | SmallVector<SDValue, 16> Ops; |
| 1514 | Ops.push_back(Elt: Chain); |
| 1515 | Ops.push_back(Elt: Callee); |
| 1516 | |
| 1517 | // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs |
| 1518 | // isn't reliable. |
| 1519 | Ops.append(in_start: OutVals.begin(), |
| 1520 | in_end: IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end()); |
| 1521 | // Add a pointer to the vararg buffer. |
| 1522 | if (IsVarArg) |
| 1523 | Ops.push_back(Elt: FINode); |
| 1524 | |
| 1525 | SmallVector<EVT, 8> InTys; |
| 1526 | for (const auto &In : Ins) { |
| 1527 | assert(!In.Flags.isByVal() && "byval is not valid for return values" ); |
| 1528 | assert(!In.Flags.isNest() && "nest is not valid for return values" ); |
| 1529 | if (In.Flags.isInAlloca()) |
| 1530 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca return values" ); |
| 1531 | if (In.Flags.isInConsecutiveRegs()) |
| 1532 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs return values" ); |
| 1533 | if (In.Flags.isInConsecutiveRegsLast()) |
| 1534 | fail(DL, DAG, |
| 1535 | Msg: "WebAssembly hasn't implemented cons regs last return values" ); |
| 1536 | // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in |
| 1537 | // registers. |
| 1538 | InTys.push_back(Elt: In.VT); |
| 1539 | } |
| 1540 | |
| 1541 | // Lastly, if this is a call to a funcref we need to add an instruction |
| 1542 | // table.set to the chain and transform the call. |
| 1543 | if (IsFuncrefCall) { |
| 1544 | // In the absence of function references proposal where a funcref call is |
| 1545 | // lowered to call_ref, using reference types we generate a table.set to set |
| 1546 | // the funcref to a special table used solely for this purpose, followed by |
| 1547 | // a call_indirect. Here we just generate the table set, and return the |
| 1548 | // SDValue of the table.set so that LowerCall can finalize the lowering by |
| 1549 | // generating the call_indirect. |
| 1550 | SDValue Chain = Ops[0]; |
| 1551 | |
| 1552 | MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol( |
| 1553 | Ctx&: MF.getContext(), Subtarget); |
| 1554 | SDValue Sym = DAG.getMCSymbol(Sym: Table, VT: PtrVT); |
| 1555 | SDValue TableSlot = DAG.getConstant(Val: 0, DL, VT: MVT::i32); |
| 1556 | SDValue TableSetOps[] = {Chain, Sym, TableSlot, Callee}; |
| 1557 | SDValue TableSet = DAG.getMemIntrinsicNode( |
| 1558 | Opcode: WebAssemblyISD::TABLE_SET, dl: DL, VTList: DAG.getVTList(VT: MVT::Other), Ops: TableSetOps, |
| 1559 | MemVT: MVT::funcref, PtrInfo: MachinePointerInfo(), Alignment: Align(1), |
| 1560 | Flags: MachineMemOperand::MOStore); |
| 1561 | |
| 1562 | Ops[0] = TableSet; // The new chain is the TableSet itself |
| 1563 | } |
| 1564 | |
| 1565 | if (CLI.IsTailCall) { |
| 1566 | // ret_calls do not return values to the current frame |
| 1567 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
| 1568 | return DAG.getNode(Opcode: WebAssemblyISD::RET_CALL, DL, VTList: NodeTys, Ops); |
| 1569 | } |
| 1570 | |
| 1571 | InTys.push_back(Elt: MVT::Other); |
| 1572 | SDVTList InTyList = DAG.getVTList(VTs: InTys); |
| 1573 | SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::CALL, DL, VTList: InTyList, Ops); |
| 1574 | |
| 1575 | for (size_t I = 0; I < Ins.size(); ++I) |
| 1576 | InVals.push_back(Elt: Res.getValue(R: I)); |
| 1577 | |
| 1578 | // Return the chain |
| 1579 | return Res.getValue(R: Ins.size()); |
| 1580 | } |
| 1581 | |
| 1582 | bool WebAssemblyTargetLowering::CanLowerReturn( |
| 1583 | CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/, |
| 1584 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext & /*Context*/, |
| 1585 | const Type *RetTy) const { |
| 1586 | // WebAssembly can only handle returning tuples with multivalue enabled |
| 1587 | return WebAssembly::canLowerReturn(ResultSize: Outs.size(), Subtarget); |
| 1588 | } |
| 1589 | |
| 1590 | SDValue WebAssemblyTargetLowering::LowerReturn( |
| 1591 | SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/, |
| 1592 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 1593 | const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, |
| 1594 | SelectionDAG &DAG) const { |
| 1595 | assert(WebAssembly::canLowerReturn(Outs.size(), Subtarget) && |
| 1596 | "MVP WebAssembly can only return up to one value" ); |
| 1597 | if (!callingConvSupported(CallConv)) |
| 1598 | fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions" ); |
| 1599 | |
| 1600 | SmallVector<SDValue, 4> RetOps(1, Chain); |
| 1601 | RetOps.append(in_start: OutVals.begin(), in_end: OutVals.end()); |
| 1602 | Chain = DAG.getNode(Opcode: WebAssemblyISD::RETURN, DL, VT: MVT::Other, Ops: RetOps); |
| 1603 | |
| 1604 | // Record the number and types of the return values. |
| 1605 | for (const ISD::OutputArg &Out : Outs) { |
| 1606 | assert(!Out.Flags.isByVal() && "byval is not valid for return values" ); |
| 1607 | assert(!Out.Flags.isNest() && "nest is not valid for return values" ); |
| 1608 | assert(!Out.Flags.isVarArg() && "non-fixed return value is not valid" ); |
| 1609 | if (Out.Flags.isInAlloca()) |
| 1610 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca results" ); |
| 1611 | if (Out.Flags.isInConsecutiveRegs()) |
| 1612 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs results" ); |
| 1613 | if (Out.Flags.isInConsecutiveRegsLast()) |
| 1614 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last results" ); |
| 1615 | } |
| 1616 | |
| 1617 | return Chain; |
| 1618 | } |
| 1619 | |
| 1620 | SDValue WebAssemblyTargetLowering::LowerFormalArguments( |
| 1621 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 1622 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 1623 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 1624 | if (!callingConvSupported(CallConv)) |
| 1625 | fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions" ); |
| 1626 | |
| 1627 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1628 | auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>(); |
| 1629 | |
| 1630 | // Set up the incoming ARGUMENTS value, which serves to represent the liveness |
| 1631 | // of the incoming values before they're represented by virtual registers. |
| 1632 | MF.getRegInfo().addLiveIn(Reg: WebAssembly::ARGUMENTS); |
| 1633 | |
| 1634 | bool HasSwiftErrorArg = false; |
| 1635 | bool HasSwiftSelfArg = false; |
| 1636 | bool HasSwiftAsyncArg = false; |
| 1637 | for (const ISD::InputArg &In : Ins) { |
| 1638 | HasSwiftSelfArg |= In.Flags.isSwiftSelf(); |
| 1639 | HasSwiftErrorArg |= In.Flags.isSwiftError(); |
| 1640 | HasSwiftAsyncArg |= In.Flags.isSwiftAsync(); |
| 1641 | if (In.Flags.isInAlloca()) |
| 1642 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments" ); |
| 1643 | if (In.Flags.isNest()) |
| 1644 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments" ); |
| 1645 | if (In.Flags.isInConsecutiveRegs()) |
| 1646 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments" ); |
| 1647 | if (In.Flags.isInConsecutiveRegsLast()) |
| 1648 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments" ); |
| 1649 | // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in |
| 1650 | // registers. |
| 1651 | InVals.push_back(Elt: In.Used ? DAG.getNode(Opcode: WebAssemblyISD::ARGUMENT, DL, VT: In.VT, |
| 1652 | Operand: DAG.getTargetConstant(Val: InVals.size(), |
| 1653 | DL, VT: MVT::i32)) |
| 1654 | : DAG.getUNDEF(VT: In.VT)); |
| 1655 | |
| 1656 | // Record the number and types of arguments. |
| 1657 | MFI->addParam(VT: In.VT); |
| 1658 | } |
| 1659 | |
| 1660 | // For swiftcc and swifttailcc, emit additional swiftself, swifterror, and |
| 1661 | // (for swifttailcc) swiftasync arguments if there aren't. These additional |
| 1662 | // arguments are also added for callee signature. They are necessary to match |
| 1663 | // callee and caller signature for indirect call. |
| 1664 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 1665 | if (CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail) { |
| 1666 | if (!HasSwiftSelfArg) { |
| 1667 | MFI->addParam(VT: PtrVT); |
| 1668 | } |
| 1669 | if (!HasSwiftErrorArg) { |
| 1670 | MFI->addParam(VT: PtrVT); |
| 1671 | } |
| 1672 | if (CallConv == CallingConv::SwiftTail && !HasSwiftAsyncArg) { |
| 1673 | MFI->addParam(VT: PtrVT); |
| 1674 | } |
| 1675 | } |
| 1676 | // Varargs are copied into a buffer allocated by the caller, and a pointer to |
| 1677 | // the buffer is passed as an argument. |
| 1678 | if (IsVarArg) { |
| 1679 | MVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 1680 | Register VarargVreg = |
| 1681 | MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: PtrVT)); |
| 1682 | MFI->setVarargBufferVreg(VarargVreg); |
| 1683 | Chain = DAG.getCopyToReg( |
| 1684 | Chain, dl: DL, Reg: VarargVreg, |
| 1685 | N: DAG.getNode(Opcode: WebAssemblyISD::ARGUMENT, DL, VT: PtrVT, |
| 1686 | Operand: DAG.getTargetConstant(Val: Ins.size(), DL, VT: MVT::i32))); |
| 1687 | MFI->addParam(VT: PtrVT); |
| 1688 | } |
| 1689 | |
| 1690 | // Record the number and types of arguments and results. |
| 1691 | SmallVector<MVT, 4> Params; |
| 1692 | SmallVector<MVT, 4> Results; |
| 1693 | computeSignatureVTs(Ty: MF.getFunction().getFunctionType(), TargetFunc: &MF.getFunction(), |
| 1694 | ContextFunc: MF.getFunction(), TM: DAG.getTarget(), Params, Results); |
| 1695 | for (MVT VT : Results) |
| 1696 | MFI->addResult(VT); |
| 1697 | // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify |
| 1698 | // the param logic here with ComputeSignatureVTs |
| 1699 | assert(MFI->getParams().size() == Params.size() && |
| 1700 | std::equal(MFI->getParams().begin(), MFI->getParams().end(), |
| 1701 | Params.begin())); |
| 1702 | |
| 1703 | return Chain; |
| 1704 | } |
| 1705 | |
| 1706 | void WebAssemblyTargetLowering::ReplaceNodeResults( |
| 1707 | SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { |
| 1708 | switch (N->getOpcode()) { |
| 1709 | case ISD::SIGN_EXTEND_INREG: |
| 1710 | // Do not add any results, signifying that N should not be custom lowered |
| 1711 | // after all. This happens because simd128 turns on custom lowering for |
| 1712 | // SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an |
| 1713 | // illegal type. |
| 1714 | break; |
| 1715 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 1716 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 1717 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 1718 | // Do not add any results, signifying that N should not be custom lowered. |
| 1719 | // EXTEND_VECTOR_INREG is implemented for some vectors, but not all. |
| 1720 | break; |
| 1721 | case ISD::FP_ROUND: { |
| 1722 | EVT VT = N->getValueType(ResNo: 0); |
| 1723 | SDValue Src = N->getOperand(Num: 0); |
| 1724 | if (VT == MVT::v4f16 && Src.getValueType() == MVT::v4f32) { |
| 1725 | Results.push_back( |
| 1726 | Elt: DAG.getNode(Opcode: WebAssemblyISD::DEMOTE_ZERO, DL: SDLoc(N), VT: MVT::v8f16, Operand: Src)); |
| 1727 | } |
| 1728 | break; |
| 1729 | } |
| 1730 | case ISD::ADD: |
| 1731 | case ISD::SUB: |
| 1732 | Results.push_back(Elt: Replace128Op(N, DAG)); |
| 1733 | break; |
| 1734 | default: |
| 1735 | llvm_unreachable( |
| 1736 | "ReplaceNodeResults not implemented for this op for WebAssembly!" ); |
| 1737 | } |
| 1738 | } |
| 1739 | |
| 1740 | //===----------------------------------------------------------------------===// |
| 1741 | // Custom lowering hooks. |
| 1742 | //===----------------------------------------------------------------------===// |
| 1743 | |
| 1744 | SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op, |
| 1745 | SelectionDAG &DAG) const { |
| 1746 | SDLoc DL(Op); |
| 1747 | switch (Op.getOpcode()) { |
| 1748 | default: |
| 1749 | llvm_unreachable("unimplemented operation lowering" ); |
| 1750 | return SDValue(); |
| 1751 | case ISD::FrameIndex: |
| 1752 | return LowerFrameIndex(Op, DAG); |
| 1753 | case ISD::GlobalAddress: |
| 1754 | return LowerGlobalAddress(Op, DAG); |
| 1755 | case ISD::GlobalTLSAddress: |
| 1756 | return LowerGlobalTLSAddress(Op, DAG); |
| 1757 | case ISD::ExternalSymbol: |
| 1758 | return LowerExternalSymbol(Op, DAG); |
| 1759 | case ISD::JumpTable: |
| 1760 | return LowerJumpTable(Op, DAG); |
| 1761 | case ISD::BR_JT: |
| 1762 | return LowerBR_JT(Op, DAG); |
| 1763 | case ISD::VASTART: |
| 1764 | return LowerVASTART(Op, DAG); |
| 1765 | case ISD::BlockAddress: |
| 1766 | case ISD::BRIND: |
| 1767 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented computed gotos" ); |
| 1768 | return SDValue(); |
| 1769 | case ISD::RETURNADDR: |
| 1770 | return LowerRETURNADDR(Op, DAG); |
| 1771 | case ISD::FRAMEADDR: |
| 1772 | return LowerFRAMEADDR(Op, DAG); |
| 1773 | case ISD::CopyToReg: |
| 1774 | return LowerCopyToReg(Op, DAG); |
| 1775 | case ISD::EXTRACT_VECTOR_ELT: |
| 1776 | case ISD::INSERT_VECTOR_ELT: |
| 1777 | return LowerAccessVectorElement(Op, DAG); |
| 1778 | case ISD::INTRINSIC_VOID: |
| 1779 | case ISD::INTRINSIC_WO_CHAIN: |
| 1780 | case ISD::INTRINSIC_W_CHAIN: |
| 1781 | return LowerIntrinsic(Op, DAG); |
| 1782 | case ISD::SIGN_EXTEND_INREG: |
| 1783 | return LowerSIGN_EXTEND_INREG(Op, DAG); |
| 1784 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 1785 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 1786 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 1787 | return LowerEXTEND_VECTOR_INREG(Op, DAG); |
| 1788 | case ISD::BUILD_VECTOR: |
| 1789 | return LowerBUILD_VECTOR(Op, DAG); |
| 1790 | case ISD::VECTOR_SHUFFLE: |
| 1791 | return LowerVECTOR_SHUFFLE(Op, DAG); |
| 1792 | case ISD::SETCC: |
| 1793 | return LowerSETCC(Op, DAG); |
| 1794 | case ISD::SHL: |
| 1795 | case ISD::SRA: |
| 1796 | case ISD::SRL: |
| 1797 | return LowerShift(Op, DAG); |
| 1798 | case ISD::FP_TO_SINT_SAT: |
| 1799 | case ISD::FP_TO_UINT_SAT: |
| 1800 | return LowerFP_TO_INT_SAT(Op, DAG); |
| 1801 | case ISD::FMINNUM: |
| 1802 | case ISD::FMINIMUMNUM: |
| 1803 | return LowerFMIN(Op, DAG); |
| 1804 | case ISD::FMAXNUM: |
| 1805 | case ISD::FMAXIMUMNUM: |
| 1806 | return LowerFMAX(Op, DAG); |
| 1807 | case ISD::LOAD: |
| 1808 | return LowerLoad(Op, DAG); |
| 1809 | case ISD::STORE: |
| 1810 | return LowerStore(Op, DAG); |
| 1811 | case ISD::CTPOP: |
| 1812 | case ISD::CTLZ: |
| 1813 | case ISD::CTTZ: |
| 1814 | return DAG.UnrollVectorOp(N: Op.getNode()); |
| 1815 | case ISD::CLEAR_CACHE: |
| 1816 | report_fatal_error(reason: "llvm.clear_cache is not supported on wasm" ); |
| 1817 | case ISD::SMUL_LOHI: |
| 1818 | case ISD::UMUL_LOHI: |
| 1819 | return LowerMUL_LOHI(Op, DAG); |
| 1820 | case ISD::UADDO: |
| 1821 | return LowerUADDO(Op, DAG); |
| 1822 | } |
| 1823 | } |
| 1824 | |
| 1825 | static bool IsWebAssemblyGlobal(SDValue Op) { |
| 1826 | if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op)) |
| 1827 | return WebAssembly::isWasmVarAddressSpace(AS: GA->getAddressSpace()); |
| 1828 | |
| 1829 | return false; |
| 1830 | } |
| 1831 | |
| 1832 | static std::optional<unsigned> IsWebAssemblyLocal(SDValue Op, |
| 1833 | SelectionDAG &DAG) { |
| 1834 | const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Op); |
| 1835 | if (!FI) |
| 1836 | return std::nullopt; |
| 1837 | |
| 1838 | auto &MF = DAG.getMachineFunction(); |
| 1839 | return WebAssemblyFrameLowering::getLocalForStackObject(MF, FrameIndex: FI->getIndex()); |
| 1840 | } |
| 1841 | |
| 1842 | SDValue WebAssemblyTargetLowering::LowerStore(SDValue Op, |
| 1843 | SelectionDAG &DAG) const { |
| 1844 | SDLoc DL(Op); |
| 1845 | StoreSDNode *SN = cast<StoreSDNode>(Val: Op.getNode()); |
| 1846 | const SDValue &Value = SN->getValue(); |
| 1847 | const SDValue &Base = SN->getBasePtr(); |
| 1848 | const SDValue &Offset = SN->getOffset(); |
| 1849 | |
| 1850 | if (IsWebAssemblyGlobal(Op: Base)) { |
| 1851 | if (!Offset->isUndef()) |
| 1852 | report_fatal_error(reason: "unexpected offset when storing to webassembly global" , |
| 1853 | gen_crash_diag: false); |
| 1854 | |
| 1855 | SDVTList Tys = DAG.getVTList(VT: MVT::Other); |
| 1856 | SDValue Ops[] = {SN->getChain(), Value, Base}; |
| 1857 | return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_SET, dl: DL, VTList: Tys, Ops, |
| 1858 | MemVT: SN->getMemoryVT(), MMO: SN->getMemOperand()); |
| 1859 | } |
| 1860 | |
| 1861 | if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) { |
| 1862 | if (!Offset->isUndef()) |
| 1863 | report_fatal_error(reason: "unexpected offset when storing to webassembly local" , |
| 1864 | gen_crash_diag: false); |
| 1865 | |
| 1866 | SDValue Idx = DAG.getTargetConstant(Val: *Local, DL: Base, VT: MVT::i32); |
| 1867 | SDVTList Tys = DAG.getVTList(VT: MVT::Other); // The chain. |
| 1868 | SDValue Ops[] = {SN->getChain(), Idx, Value}; |
| 1869 | return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_SET, DL, VTList: Tys, Ops); |
| 1870 | } |
| 1871 | |
| 1872 | if (WebAssembly::isWasmVarAddressSpace(AS: SN->getAddressSpace())) |
| 1873 | report_fatal_error( |
| 1874 | reason: "Encountered an unlowerable store to the wasm_var address space" , |
| 1875 | gen_crash_diag: false); |
| 1876 | |
| 1877 | return Op; |
| 1878 | } |
| 1879 | |
| 1880 | SDValue WebAssemblyTargetLowering::LowerLoad(SDValue Op, |
| 1881 | SelectionDAG &DAG) const { |
| 1882 | SDLoc DL(Op); |
| 1883 | LoadSDNode *LN = cast<LoadSDNode>(Val: Op.getNode()); |
| 1884 | const SDValue &Base = LN->getBasePtr(); |
| 1885 | const SDValue &Offset = LN->getOffset(); |
| 1886 | |
| 1887 | if (IsWebAssemblyGlobal(Op: Base)) { |
| 1888 | if (!Offset->isUndef()) |
| 1889 | report_fatal_error( |
| 1890 | reason: "unexpected offset when loading from webassembly global" , gen_crash_diag: false); |
| 1891 | |
| 1892 | SDVTList Tys = DAG.getVTList(VT1: LN->getValueType(ResNo: 0), VT2: MVT::Other); |
| 1893 | SDValue Ops[] = {LN->getChain(), Base}; |
| 1894 | return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_GET, dl: DL, VTList: Tys, Ops, |
| 1895 | MemVT: LN->getMemoryVT(), MMO: LN->getMemOperand()); |
| 1896 | } |
| 1897 | |
| 1898 | if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) { |
| 1899 | if (!Offset->isUndef()) |
| 1900 | report_fatal_error( |
| 1901 | reason: "unexpected offset when loading from webassembly local" , gen_crash_diag: false); |
| 1902 | |
| 1903 | SDValue Idx = DAG.getTargetConstant(Val: *Local, DL: Base, VT: MVT::i32); |
| 1904 | EVT LocalVT = LN->getValueType(ResNo: 0); |
| 1905 | return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_GET, DL, ResultTys: {LocalVT, MVT::Other}, |
| 1906 | Ops: {LN->getChain(), Idx}); |
| 1907 | } |
| 1908 | |
| 1909 | if (WebAssembly::isWasmVarAddressSpace(AS: LN->getAddressSpace())) |
| 1910 | report_fatal_error( |
| 1911 | reason: "Encountered an unlowerable load from the wasm_var address space" , |
| 1912 | gen_crash_diag: false); |
| 1913 | |
| 1914 | return Op; |
| 1915 | } |
| 1916 | |
| 1917 | SDValue WebAssemblyTargetLowering::LowerMUL_LOHI(SDValue Op, |
| 1918 | SelectionDAG &DAG) const { |
| 1919 | assert(Subtarget->hasWideArithmetic()); |
| 1920 | assert(Op.getValueType() == MVT::i64); |
| 1921 | SDLoc DL(Op); |
| 1922 | unsigned Opcode; |
| 1923 | switch (Op.getOpcode()) { |
| 1924 | case ISD::UMUL_LOHI: |
| 1925 | Opcode = WebAssemblyISD::I64_MUL_WIDE_U; |
| 1926 | break; |
| 1927 | case ISD::SMUL_LOHI: |
| 1928 | Opcode = WebAssemblyISD::I64_MUL_WIDE_S; |
| 1929 | break; |
| 1930 | default: |
| 1931 | llvm_unreachable("unexpected opcode" ); |
| 1932 | } |
| 1933 | SDValue LHS = Op.getOperand(i: 0); |
| 1934 | SDValue RHS = Op.getOperand(i: 1); |
| 1935 | SDValue Lo = |
| 1936 | DAG.getNode(Opcode, DL, VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), N1: LHS, N2: RHS); |
| 1937 | SDValue Hi(Lo.getNode(), 1); |
| 1938 | SDValue Ops[] = {Lo, Hi}; |
| 1939 | return DAG.getMergeValues(Ops, dl: DL); |
| 1940 | } |
| 1941 | |
| 1942 | // Lowers `UADDO` intrinsics to an `i64.add128` instruction when it's enabled. |
| 1943 | // |
| 1944 | // This enables generating a single wasm instruction for this operation where |
| 1945 | // the upper half of both operands are constant zeros. The upper half of the |
| 1946 | // result is then whether the overflow happened. |
| 1947 | SDValue WebAssemblyTargetLowering::LowerUADDO(SDValue Op, |
| 1948 | SelectionDAG &DAG) const { |
| 1949 | assert(Subtarget->hasWideArithmetic()); |
| 1950 | assert(Op.getValueType() == MVT::i64); |
| 1951 | assert(Op.getOpcode() == ISD::UADDO); |
| 1952 | SDLoc DL(Op); |
| 1953 | SDValue LHS = Op.getOperand(i: 0); |
| 1954 | SDValue RHS = Op.getOperand(i: 1); |
| 1955 | SDValue Zero = DAG.getConstant(Val: 0, DL, VT: MVT::i64); |
| 1956 | SDValue Result = |
| 1957 | DAG.getNode(Opcode: WebAssemblyISD::I64_ADD128, DL, |
| 1958 | VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), N1: LHS, N2: Zero, N3: RHS, N4: Zero); |
| 1959 | SDValue CarryI64(Result.getNode(), 1); |
| 1960 | SDValue CarryI32 = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: CarryI64); |
| 1961 | SDValue Ops[] = {Result, CarryI32}; |
| 1962 | return DAG.getMergeValues(Ops, dl: DL); |
| 1963 | } |
| 1964 | |
| 1965 | SDValue WebAssemblyTargetLowering::Replace128Op(SDNode *N, |
| 1966 | SelectionDAG &DAG) const { |
| 1967 | assert(Subtarget->hasWideArithmetic()); |
| 1968 | assert(N->getValueType(0) == MVT::i128); |
| 1969 | SDLoc DL(N); |
| 1970 | unsigned Opcode; |
| 1971 | switch (N->getOpcode()) { |
| 1972 | case ISD::ADD: |
| 1973 | Opcode = WebAssemblyISD::I64_ADD128; |
| 1974 | break; |
| 1975 | case ISD::SUB: |
| 1976 | Opcode = WebAssemblyISD::I64_SUB128; |
| 1977 | break; |
| 1978 | default: |
| 1979 | llvm_unreachable("unexpected opcode" ); |
| 1980 | } |
| 1981 | SDValue LHS = N->getOperand(Num: 0); |
| 1982 | SDValue RHS = N->getOperand(Num: 1); |
| 1983 | |
| 1984 | SDValue C0 = DAG.getConstant(Val: 0, DL, VT: MVT::i64); |
| 1985 | SDValue C1 = DAG.getConstant(Val: 1, DL, VT: MVT::i64); |
| 1986 | SDValue LHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: LHS, N2: C0); |
| 1987 | SDValue LHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: LHS, N2: C1); |
| 1988 | SDValue RHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: RHS, N2: C0); |
| 1989 | SDValue RHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: RHS, N2: C1); |
| 1990 | SDValue Result_LO = DAG.getNode(Opcode, DL, VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), |
| 1991 | N1: LHS_0, N2: LHS_1, N3: RHS_0, N4: RHS_1); |
| 1992 | SDValue Result_HI(Result_LO.getNode(), 1); |
| 1993 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VTList: N->getVTList(), N1: Result_LO, N2: Result_HI); |
| 1994 | } |
| 1995 | |
| 1996 | SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op, |
| 1997 | SelectionDAG &DAG) const { |
| 1998 | SDValue Src = Op.getOperand(i: 2); |
| 1999 | if (isa<FrameIndexSDNode>(Val: Src.getNode())) { |
| 2000 | // CopyToReg nodes don't support FrameIndex operands. Other targets select |
| 2001 | // the FI to some LEA-like instruction, but since we don't have that, we |
| 2002 | // need to insert some kind of instruction that can take an FI operand and |
| 2003 | // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy |
| 2004 | // local.copy between Op and its FI operand. |
| 2005 | SDValue Chain = Op.getOperand(i: 0); |
| 2006 | SDLoc DL(Op); |
| 2007 | Register Reg = cast<RegisterSDNode>(Val: Op.getOperand(i: 1))->getReg(); |
| 2008 | EVT VT = Src.getValueType(); |
| 2009 | SDValue Copy(DAG.getMachineNode(Opcode: VT == MVT::i32 ? WebAssembly::COPY_I32 |
| 2010 | : WebAssembly::COPY_I64, |
| 2011 | dl: DL, VT, Op1: Src), |
| 2012 | 0); |
| 2013 | return Op.getNode()->getNumValues() == 1 |
| 2014 | ? DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy) |
| 2015 | : DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy, |
| 2016 | Glue: Op.getNumOperands() == 4 ? Op.getOperand(i: 3) |
| 2017 | : SDValue()); |
| 2018 | } |
| 2019 | return SDValue(); |
| 2020 | } |
| 2021 | |
| 2022 | SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op, |
| 2023 | SelectionDAG &DAG) const { |
| 2024 | int FI = cast<FrameIndexSDNode>(Val&: Op)->getIndex(); |
| 2025 | return DAG.getTargetFrameIndex(FI, VT: Op.getValueType()); |
| 2026 | } |
| 2027 | |
| 2028 | SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op, |
| 2029 | SelectionDAG &DAG) const { |
| 2030 | SDLoc DL(Op); |
| 2031 | |
| 2032 | if (!Subtarget->getTargetTriple().isOSEmscripten()) { |
| 2033 | fail(DL, DAG, |
| 2034 | Msg: "Non-Emscripten WebAssembly hasn't implemented " |
| 2035 | "__builtin_return_address" ); |
| 2036 | return SDValue(); |
| 2037 | } |
| 2038 | |
| 2039 | unsigned Depth = Op.getConstantOperandVal(i: 0); |
| 2040 | MakeLibCallOptions CallOptions; |
| 2041 | return makeLibCall(DAG, LC: RTLIB::RETURN_ADDRESS, RetVT: Op.getValueType(), |
| 2042 | Ops: {DAG.getConstant(Val: Depth, DL, VT: MVT::i32)}, CallOptions, dl: DL) |
| 2043 | .first; |
| 2044 | } |
| 2045 | |
| 2046 | SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op, |
| 2047 | SelectionDAG &DAG) const { |
| 2048 | // Non-zero depths are not supported by WebAssembly currently. Use the |
| 2049 | // legalizer's default expansion, which is to return 0 (what this function is |
| 2050 | // documented to do). |
| 2051 | if (Op.getConstantOperandVal(i: 0) > 0) |
| 2052 | return SDValue(); |
| 2053 | |
| 2054 | DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true); |
| 2055 | EVT VT = Op.getValueType(); |
| 2056 | Register FP = |
| 2057 | Subtarget->getRegisterInfo()->getFrameRegister(MF: DAG.getMachineFunction()); |
| 2058 | return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), Reg: FP, VT); |
| 2059 | } |
| 2060 | |
| 2061 | SDValue |
| 2062 | WebAssemblyTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
| 2063 | SelectionDAG &DAG) const { |
| 2064 | SDLoc DL(Op); |
| 2065 | const auto *GA = cast<GlobalAddressSDNode>(Val&: Op); |
| 2066 | |
| 2067 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2068 | if (!MF.getSubtarget<WebAssemblySubtarget>().hasBulkMemory()) |
| 2069 | report_fatal_error(reason: "cannot use thread-local storage without bulk memory" , |
| 2070 | gen_crash_diag: false); |
| 2071 | |
| 2072 | const GlobalValue *GV = GA->getGlobal(); |
| 2073 | |
| 2074 | // Currently only Emscripten supports dynamic linking with threads. Therefore, |
| 2075 | // on other targets, if we have thread-local storage, only the local-exec |
| 2076 | // model is possible. |
| 2077 | auto model = Subtarget->getTargetTriple().isOSEmscripten() |
| 2078 | ? GV->getThreadLocalMode() |
| 2079 | : GlobalValue::LocalExecTLSModel; |
| 2080 | |
| 2081 | // Unsupported TLS modes |
| 2082 | assert(model != GlobalValue::NotThreadLocal); |
| 2083 | assert(model != GlobalValue::InitialExecTLSModel); |
| 2084 | |
| 2085 | if (model == GlobalValue::LocalExecTLSModel || |
| 2086 | model == GlobalValue::LocalDynamicTLSModel || |
| 2087 | (model == GlobalValue::GeneralDynamicTLSModel && |
| 2088 | getTargetMachine().shouldAssumeDSOLocal(GV))) { |
| 2089 | // For DSO-local TLS variables we use offset from __tls_base, or |
| 2090 | // __wasm_get_tls_base() if using libcall thread context. |
| 2091 | |
| 2092 | MVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2093 | SDValue BaseAddr(WebAssembly::getTLSBase(DAG, DL, Subtarget), 0); |
| 2094 | |
| 2095 | SDValue TLSOffset = DAG.getTargetGlobalAddress( |
| 2096 | GV, DL, VT: PtrVT, offset: GA->getOffset(), TargetFlags: WebAssemblyII::MO_TLS_BASE_REL); |
| 2097 | SDValue SymOffset = |
| 2098 | DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: TLSOffset); |
| 2099 | |
| 2100 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymOffset); |
| 2101 | } |
| 2102 | |
| 2103 | assert(model == GlobalValue::GeneralDynamicTLSModel); |
| 2104 | |
| 2105 | EVT VT = Op.getValueType(); |
| 2106 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
| 2107 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, |
| 2108 | offset: GA->getOffset(), |
| 2109 | TargetFlags: WebAssemblyII::MO_GOT_TLS)); |
| 2110 | } |
| 2111 | |
| 2112 | SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op, |
| 2113 | SelectionDAG &DAG) const { |
| 2114 | SDLoc DL(Op); |
| 2115 | const auto *GA = cast<GlobalAddressSDNode>(Val&: Op); |
| 2116 | EVT VT = Op.getValueType(); |
| 2117 | assert(GA->getTargetFlags() == 0 && |
| 2118 | "Unexpected target flags on generic GlobalAddressSDNode" ); |
| 2119 | if (!WebAssembly::isValidAddressSpace(AS: GA->getAddressSpace())) |
| 2120 | fail(DL, DAG, Msg: "Invalid address space for WebAssembly target" ); |
| 2121 | |
| 2122 | unsigned OperandFlags = 0; |
| 2123 | const GlobalValue *GV = GA->getGlobal(); |
| 2124 | // Since WebAssembly tables cannot yet be shared accross modules, we don't |
| 2125 | // need special treatment for tables in PIC mode. |
| 2126 | if (isPositionIndependent() && |
| 2127 | !WebAssembly::isWebAssemblyTableType(Ty: GV->getValueType())) { |
| 2128 | if (getTargetMachine().shouldAssumeDSOLocal(GV)) { |
| 2129 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2130 | MVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 2131 | const char *BaseName; |
| 2132 | if (GV->getValueType()->isFunctionTy()) { |
| 2133 | BaseName = MF.createExternalSymbolName(Name: "__table_base" ); |
| 2134 | OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL; |
| 2135 | } else { |
| 2136 | BaseName = MF.createExternalSymbolName(Name: "__memory_base" ); |
| 2137 | OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL; |
| 2138 | } |
| 2139 | SDValue BaseAddr = |
| 2140 | DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, |
| 2141 | Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)); |
| 2142 | |
| 2143 | SDValue SymAddr = DAG.getNode( |
| 2144 | Opcode: WebAssemblyISD::WrapperREL, DL, VT, |
| 2145 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, offset: GA->getOffset(), |
| 2146 | TargetFlags: OperandFlags)); |
| 2147 | |
| 2148 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: BaseAddr, N2: SymAddr); |
| 2149 | } |
| 2150 | OperandFlags = WebAssemblyII::MO_GOT; |
| 2151 | } |
| 2152 | |
| 2153 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
| 2154 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, |
| 2155 | offset: GA->getOffset(), TargetFlags: OperandFlags)); |
| 2156 | } |
| 2157 | |
| 2158 | SDValue |
| 2159 | WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op, |
| 2160 | SelectionDAG &DAG) const { |
| 2161 | SDLoc DL(Op); |
| 2162 | const auto *ES = cast<ExternalSymbolSDNode>(Val&: Op); |
| 2163 | EVT VT = Op.getValueType(); |
| 2164 | assert(ES->getTargetFlags() == 0 && |
| 2165 | "Unexpected target flags on generic ExternalSymbolSDNode" ); |
| 2166 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
| 2167 | Operand: DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), VT)); |
| 2168 | } |
| 2169 | |
| 2170 | SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op, |
| 2171 | SelectionDAG &DAG) const { |
| 2172 | // There's no need for a Wrapper node because we always incorporate a jump |
| 2173 | // table operand into a BR_TABLE instruction, rather than ever |
| 2174 | // materializing it in a register. |
| 2175 | const JumpTableSDNode *JT = cast<JumpTableSDNode>(Val&: Op); |
| 2176 | return DAG.getTargetJumpTable(JTI: JT->getIndex(), VT: Op.getValueType(), |
| 2177 | TargetFlags: JT->getTargetFlags()); |
| 2178 | } |
| 2179 | |
| 2180 | SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op, |
| 2181 | SelectionDAG &DAG) const { |
| 2182 | SDLoc DL(Op); |
| 2183 | SDValue Chain = Op.getOperand(i: 0); |
| 2184 | const auto *JT = cast<JumpTableSDNode>(Val: Op.getOperand(i: 1)); |
| 2185 | SDValue Index = Op.getOperand(i: 2); |
| 2186 | assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags" ); |
| 2187 | |
| 2188 | SmallVector<SDValue, 8> Ops; |
| 2189 | Ops.push_back(Elt: Chain); |
| 2190 | Ops.push_back(Elt: Index); |
| 2191 | |
| 2192 | MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo(); |
| 2193 | const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs; |
| 2194 | |
| 2195 | // Add an operand for each case. |
| 2196 | for (auto *MBB : MBBs) |
| 2197 | Ops.push_back(Elt: DAG.getBasicBlock(MBB)); |
| 2198 | |
| 2199 | // Add the first MBB as a dummy default target for now. This will be replaced |
| 2200 | // with the proper default target (and the preceding range check eliminated) |
| 2201 | // if possible by WebAssemblyFixBrTableDefaults. |
| 2202 | Ops.push_back(Elt: DAG.getBasicBlock(MBB: *MBBs.begin())); |
| 2203 | return DAG.getNode(Opcode: WebAssemblyISD::BR_TABLE, DL, VT: MVT::Other, Ops); |
| 2204 | } |
| 2205 | |
| 2206 | SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op, |
| 2207 | SelectionDAG &DAG) const { |
| 2208 | SDLoc DL(Op); |
| 2209 | EVT PtrVT = getPointerTy(DL: DAG.getMachineFunction().getDataLayout()); |
| 2210 | |
| 2211 | auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>(); |
| 2212 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
| 2213 | |
| 2214 | SDValue ArgN = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: DL, |
| 2215 | Reg: MFI->getVarargBufferVreg(), VT: PtrVT); |
| 2216 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: ArgN, Ptr: Op.getOperand(i: 1), |
| 2217 | PtrInfo: MachinePointerInfo(SV)); |
| 2218 | } |
| 2219 | |
| 2220 | SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op, |
| 2221 | SelectionDAG &DAG) const { |
| 2222 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2223 | unsigned IntNo; |
| 2224 | switch (Op.getOpcode()) { |
| 2225 | case ISD::INTRINSIC_VOID: |
| 2226 | case ISD::INTRINSIC_W_CHAIN: |
| 2227 | IntNo = Op.getConstantOperandVal(i: 1); |
| 2228 | break; |
| 2229 | case ISD::INTRINSIC_WO_CHAIN: |
| 2230 | IntNo = Op.getConstantOperandVal(i: 0); |
| 2231 | break; |
| 2232 | default: |
| 2233 | llvm_unreachable("Invalid intrinsic" ); |
| 2234 | } |
| 2235 | SDLoc DL(Op); |
| 2236 | |
| 2237 | switch (IntNo) { |
| 2238 | default: |
| 2239 | return SDValue(); // Don't custom lower most intrinsics. |
| 2240 | |
| 2241 | case Intrinsic::wasm_lsda: { |
| 2242 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 2243 | const char *SymName = MF.createExternalSymbolName( |
| 2244 | Name: "GCC_except_table" + std::to_string(val: MF.getFunctionNumber())); |
| 2245 | if (isPositionIndependent()) { |
| 2246 | SDValue Node = DAG.getTargetExternalSymbol( |
| 2247 | Sym: SymName, VT: PtrVT, TargetFlags: WebAssemblyII::MO_MEMORY_BASE_REL); |
| 2248 | const char *BaseName = MF.createExternalSymbolName(Name: "__memory_base" ); |
| 2249 | SDValue BaseAddr = |
| 2250 | DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, |
| 2251 | Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)); |
| 2252 | SDValue SymAddr = |
| 2253 | DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: Node); |
| 2254 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymAddr); |
| 2255 | } |
| 2256 | SDValue Node = DAG.getTargetExternalSymbol(Sym: SymName, VT: PtrVT); |
| 2257 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, Operand: Node); |
| 2258 | } |
| 2259 | |
| 2260 | case Intrinsic::wasm_shuffle: { |
| 2261 | // Drop in-chain and replace undefs, but otherwise pass through unchanged |
| 2262 | SDValue Ops[18]; |
| 2263 | size_t OpIdx = 0; |
| 2264 | Ops[OpIdx++] = Op.getOperand(i: 1); |
| 2265 | Ops[OpIdx++] = Op.getOperand(i: 2); |
| 2266 | while (OpIdx < 18) { |
| 2267 | const SDValue &MaskIdx = Op.getOperand(i: OpIdx + 1); |
| 2268 | if (MaskIdx.isUndef() || MaskIdx.getNode()->getAsZExtVal() >= 32) { |
| 2269 | bool isTarget = MaskIdx.getNode()->getOpcode() == ISD::TargetConstant; |
| 2270 | Ops[OpIdx++] = DAG.getConstant(Val: 0, DL, VT: MVT::i32, isTarget); |
| 2271 | } else { |
| 2272 | Ops[OpIdx++] = MaskIdx; |
| 2273 | } |
| 2274 | } |
| 2275 | return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops); |
| 2276 | } |
| 2277 | |
| 2278 | case Intrinsic::wasm_funcref_to_ptr: { |
| 2279 | // llvm.wasm.funcref.to_ptr only has a defined lowering when its result |
| 2280 | // feeds directly into an indirect call. Reaching here means the pointer |
| 2281 | // escapes a direct call. We haven't implemented conversion of a funcref |
| 2282 | // into a real function pointer so we crash if we get here. |
| 2283 | fail(DL, DAG, |
| 2284 | Msg: "a funcref can only be converted to a pointer to be directly called; " |
| 2285 | "the resulting pointer cannot otherwise be used" ); |
| 2286 | return DAG.getPOISON(VT: Op.getValueType()); |
| 2287 | } |
| 2288 | |
| 2289 | case Intrinsic::thread_pointer: { |
| 2290 | return SDValue(WebAssembly::getTLSBase(DAG, DL, Subtarget), 0); |
| 2291 | } |
| 2292 | } |
| 2293 | } |
| 2294 | |
| 2295 | SDValue |
| 2296 | WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, |
| 2297 | SelectionDAG &DAG) const { |
| 2298 | SDLoc DL(Op); |
| 2299 | // If sign extension operations are disabled, allow sext_inreg only if operand |
| 2300 | // is a vector extract of an i8 or i16 lane. SIMD does not depend on sign |
| 2301 | // extension operations, but allowing sext_inreg in this context lets us have |
| 2302 | // simple patterns to select extract_lane_s instructions. Expanding sext_inreg |
| 2303 | // everywhere would be simpler in this file, but would necessitate large and |
| 2304 | // brittle patterns to undo the expansion and select extract_lane_s |
| 2305 | // instructions. |
| 2306 | assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128()); |
| 2307 | if (Op.getOperand(i: 0).getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2308 | return SDValue(); |
| 2309 | |
| 2310 | const SDValue & = Op.getOperand(i: 0); |
| 2311 | MVT VecT = Extract.getOperand(i: 0).getSimpleValueType(); |
| 2312 | if (VecT.getVectorElementType().getSizeInBits() > 32) |
| 2313 | return SDValue(); |
| 2314 | MVT = |
| 2315 | cast<VTSDNode>(Val: Op.getOperand(i: 1).getNode())->getVT().getSimpleVT(); |
| 2316 | MVT = |
| 2317 | MVT::getVectorVT(VT: ExtractedLaneT, NumElements: 128 / ExtractedLaneT.getSizeInBits()); |
| 2318 | if (ExtractedVecT == VecT) |
| 2319 | return Op; |
| 2320 | |
| 2321 | // Bitcast vector to appropriate type to ensure ISel pattern coverage |
| 2322 | const SDNode *Index = Extract.getOperand(i: 1).getNode(); |
| 2323 | if (!isa<ConstantSDNode>(Val: Index)) |
| 2324 | return SDValue(); |
| 2325 | unsigned IndexVal = Index->getAsZExtVal(); |
| 2326 | unsigned Scale = |
| 2327 | ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements(); |
| 2328 | assert(Scale > 1); |
| 2329 | SDValue NewIndex = |
| 2330 | DAG.getConstant(Val: IndexVal * Scale, DL, VT: Index->getValueType(ResNo: 0)); |
| 2331 | SDValue = DAG.getNode( |
| 2332 | Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: Extract.getValueType(), |
| 2333 | N1: DAG.getBitcast(VT: ExtractedVecT, V: Extract.getOperand(i: 0)), N2: NewIndex); |
| 2334 | return DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: Op.getValueType(), N1: NewExtract, |
| 2335 | N2: Op.getOperand(i: 1)); |
| 2336 | } |
| 2337 | |
| 2338 | static SDValue GetExtendHigh(SDValue Op, unsigned UserOpc, EVT VT, |
| 2339 | SelectionDAG &DAG) { |
| 2340 | SDValue Source = peekThroughBitcasts(V: Op); |
| 2341 | if (Source.getOpcode() != ISD::VECTOR_SHUFFLE) |
| 2342 | return SDValue(); |
| 2343 | |
| 2344 | assert((UserOpc == WebAssemblyISD::EXTEND_LOW_U || |
| 2345 | UserOpc == WebAssemblyISD::EXTEND_LOW_S) && |
| 2346 | "expected extend_low" ); |
| 2347 | auto *Shuffle = cast<ShuffleVectorSDNode>(Val: Source.getNode()); |
| 2348 | |
| 2349 | ArrayRef<int> Mask = Shuffle->getMask(); |
| 2350 | // Look for a shuffle which moves from the high half to the low half. |
| 2351 | size_t FirstIdx = Mask.size() / 2; |
| 2352 | for (size_t i = 0; i < Mask.size() / 2; ++i) { |
| 2353 | if (Mask[i] != static_cast<int>(FirstIdx + i)) { |
| 2354 | return SDValue(); |
| 2355 | } |
| 2356 | } |
| 2357 | |
| 2358 | SDLoc DL(Op); |
| 2359 | unsigned Opc = UserOpc == WebAssemblyISD::EXTEND_LOW_S |
| 2360 | ? WebAssemblyISD::EXTEND_HIGH_S |
| 2361 | : WebAssemblyISD::EXTEND_HIGH_U; |
| 2362 | SDValue ShuffleSrc = Shuffle->getOperand(Num: 0); |
| 2363 | if (Op.getOpcode() == ISD::BITCAST) |
| 2364 | ShuffleSrc = DAG.getBitcast(VT: Op.getValueType(), V: ShuffleSrc); |
| 2365 | |
| 2366 | return DAG.getNode(Opcode: Opc, DL, VT, Operand: ShuffleSrc); |
| 2367 | } |
| 2368 | |
| 2369 | SDValue |
| 2370 | WebAssemblyTargetLowering::LowerEXTEND_VECTOR_INREG(SDValue Op, |
| 2371 | SelectionDAG &DAG) const { |
| 2372 | SDLoc DL(Op); |
| 2373 | EVT VT = Op.getValueType(); |
| 2374 | SDValue Src = Op.getOperand(i: 0); |
| 2375 | EVT SrcVT = Src.getValueType(); |
| 2376 | |
| 2377 | if (SrcVT.getVectorElementType() == MVT::i1 || |
| 2378 | SrcVT.getVectorElementType() == MVT::i64) |
| 2379 | return SDValue(); |
| 2380 | |
| 2381 | assert(VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits() == 0 && |
| 2382 | "Unexpected extension factor." ); |
| 2383 | unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits(); |
| 2384 | |
| 2385 | if (Scale != 2 && Scale != 4 && Scale != 8) |
| 2386 | return SDValue(); |
| 2387 | |
| 2388 | unsigned Ext; |
| 2389 | switch (Op.getOpcode()) { |
| 2390 | default: |
| 2391 | llvm_unreachable("unexpected opcode" ); |
| 2392 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 2393 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 2394 | Ext = WebAssemblyISD::EXTEND_LOW_U; |
| 2395 | break; |
| 2396 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 2397 | Ext = WebAssemblyISD::EXTEND_LOW_S; |
| 2398 | break; |
| 2399 | } |
| 2400 | |
| 2401 | if (Scale == 2) { |
| 2402 | // See if we can use EXTEND_HIGH. |
| 2403 | if (auto ExtendHigh = GetExtendHigh(Op: Op.getOperand(i: 0), UserOpc: Ext, VT, DAG)) |
| 2404 | return ExtendHigh; |
| 2405 | } |
| 2406 | |
| 2407 | SDValue Ret = Src; |
| 2408 | while (Scale != 1) { |
| 2409 | Ret = DAG.getNode(Opcode: Ext, DL, |
| 2410 | VT: Ret.getValueType() |
| 2411 | .widenIntegerVectorElementType(Context&: *DAG.getContext()) |
| 2412 | .getHalfNumVectorElementsVT(Context&: *DAG.getContext()), |
| 2413 | Operand: Ret); |
| 2414 | Scale /= 2; |
| 2415 | } |
| 2416 | assert(Ret.getValueType() == VT); |
| 2417 | return Ret; |
| 2418 | } |
| 2419 | |
| 2420 | static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) { |
| 2421 | SDLoc DL(Op); |
| 2422 | if (Op.getValueType() != MVT::v2f64 && Op.getValueType() != MVT::v4f32) |
| 2423 | return SDValue(); |
| 2424 | |
| 2425 | auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec, |
| 2426 | unsigned &Index) -> bool { |
| 2427 | switch (Op.getOpcode()) { |
| 2428 | case ISD::SINT_TO_FP: |
| 2429 | Opcode = WebAssemblyISD::CONVERT_LOW_S; |
| 2430 | break; |
| 2431 | case ISD::UINT_TO_FP: |
| 2432 | Opcode = WebAssemblyISD::CONVERT_LOW_U; |
| 2433 | break; |
| 2434 | case ISD::FP_EXTEND: |
| 2435 | case ISD::FP16_TO_FP: |
| 2436 | Opcode = WebAssemblyISD::PROMOTE_LOW; |
| 2437 | break; |
| 2438 | default: |
| 2439 | return false; |
| 2440 | } |
| 2441 | |
| 2442 | auto = Op.getOperand(i: 0); |
| 2443 | if (ExtractVector.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2444 | return false; |
| 2445 | |
| 2446 | if (!isa<ConstantSDNode>(Val: ExtractVector.getOperand(i: 1).getNode())) |
| 2447 | return false; |
| 2448 | |
| 2449 | SrcVec = ExtractVector.getOperand(i: 0); |
| 2450 | Index = ExtractVector.getConstantOperandVal(i: 1); |
| 2451 | return true; |
| 2452 | }; |
| 2453 | |
| 2454 | unsigned NumLanes = Op.getValueType() == MVT::v2f64 ? 2 : 4; |
| 2455 | unsigned FirstOpcode = 0, SecondOpcode = 0, ThirdOpcode = 0, FourthOpcode = 0; |
| 2456 | unsigned FirstIndex = 0, SecondIndex = 0, ThirdIndex = 0, FourthIndex = 0; |
| 2457 | SDValue FirstSrcVec, SecondSrcVec, ThirdSrcVec, FourthSrcVec; |
| 2458 | |
| 2459 | if (!GetConvertedLane(Op.getOperand(i: 0), FirstOpcode, FirstSrcVec, |
| 2460 | FirstIndex) || |
| 2461 | !GetConvertedLane(Op.getOperand(i: 1), SecondOpcode, SecondSrcVec, |
| 2462 | SecondIndex)) |
| 2463 | return SDValue(); |
| 2464 | |
| 2465 | // If we're converting to v4f32, check the third and fourth lanes, too. |
| 2466 | if (NumLanes == 4 && (!GetConvertedLane(Op.getOperand(i: 2), ThirdOpcode, |
| 2467 | ThirdSrcVec, ThirdIndex) || |
| 2468 | !GetConvertedLane(Op.getOperand(i: 3), FourthOpcode, |
| 2469 | FourthSrcVec, FourthIndex))) |
| 2470 | return SDValue(); |
| 2471 | |
| 2472 | if (FirstOpcode != SecondOpcode) |
| 2473 | return SDValue(); |
| 2474 | |
| 2475 | // TODO Add an optimization similar to the v2f64 below for shuffling the |
| 2476 | // vectors when the lanes are in the wrong order or come from different src |
| 2477 | // vectors. |
| 2478 | if (NumLanes == 4 && |
| 2479 | (FirstOpcode != ThirdOpcode || FirstOpcode != FourthOpcode || |
| 2480 | FirstSrcVec != SecondSrcVec || FirstSrcVec != ThirdSrcVec || |
| 2481 | FirstSrcVec != FourthSrcVec || FirstIndex != 0 || SecondIndex != 1 || |
| 2482 | ThirdIndex != 2 || FourthIndex != 3)) |
| 2483 | return SDValue(); |
| 2484 | |
| 2485 | MVT ExpectedSrcVT; |
| 2486 | switch (FirstOpcode) { |
| 2487 | case WebAssemblyISD::CONVERT_LOW_S: |
| 2488 | case WebAssemblyISD::CONVERT_LOW_U: |
| 2489 | ExpectedSrcVT = MVT::v4i32; |
| 2490 | break; |
| 2491 | case WebAssemblyISD::PROMOTE_LOW: |
| 2492 | ExpectedSrcVT = NumLanes == 2 ? MVT::v4f32 : MVT::v8i16; |
| 2493 | break; |
| 2494 | } |
| 2495 | if (FirstSrcVec.getValueType() != ExpectedSrcVT) |
| 2496 | return SDValue(); |
| 2497 | |
| 2498 | auto Src = FirstSrcVec; |
| 2499 | if (NumLanes == 2 && |
| 2500 | (FirstIndex != 0 || SecondIndex != 1 || FirstSrcVec != SecondSrcVec)) { |
| 2501 | // Shuffle the source vector so that the converted lanes are the low lanes. |
| 2502 | Src = DAG.getVectorShuffle(VT: ExpectedSrcVT, dl: DL, N1: FirstSrcVec, N2: SecondSrcVec, |
| 2503 | Mask: {static_cast<int>(FirstIndex), |
| 2504 | static_cast<int>(SecondIndex) + 4, -1, -1}); |
| 2505 | } |
| 2506 | return DAG.getNode(Opcode: FirstOpcode, DL, VT: NumLanes == 2 ? MVT::v2f64 : MVT::v4f32, |
| 2507 | Operand: Src); |
| 2508 | } |
| 2509 | |
| 2510 | SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op, |
| 2511 | SelectionDAG &DAG) const { |
| 2512 | MVT VT = Op.getSimpleValueType(); |
| 2513 | if (VT == MVT::v8f16) { |
| 2514 | // BUILD_VECTOR can't handle FP16 operands since Wasm doesn't have a scaler |
| 2515 | // FP16 type, so cast them to I16s. |
| 2516 | MVT IVT = VT.changeVectorElementType(EltVT: MVT::i16); |
| 2517 | SmallVector<SDValue, 8> NewOps; |
| 2518 | for (unsigned I = 0, E = Op.getNumOperands(); I < E; ++I) |
| 2519 | NewOps.push_back(Elt: DAG.getBitcast(VT: MVT::i16, V: Op.getOperand(i: I))); |
| 2520 | SDValue Res = DAG.getNode(Opcode: ISD::BUILD_VECTOR, DL: SDLoc(), VT: IVT, Ops: NewOps); |
| 2521 | return DAG.getBitcast(VT, V: Res); |
| 2522 | } |
| 2523 | |
| 2524 | if (auto ConvertLow = LowerConvertLow(Op, DAG)) |
| 2525 | return ConvertLow; |
| 2526 | |
| 2527 | SDLoc DL(Op); |
| 2528 | const EVT VecT = Op.getValueType(); |
| 2529 | const EVT LaneT = Op.getOperand(i: 0).getValueType(); |
| 2530 | const size_t Lanes = Op.getNumOperands(); |
| 2531 | bool CanSwizzle = VecT == MVT::v16i8; |
| 2532 | |
| 2533 | // BUILD_VECTORs are lowered to the instruction that initializes the highest |
| 2534 | // possible number of lanes at once followed by a sequence of replace_lane |
| 2535 | // instructions to individually initialize any remaining lanes. |
| 2536 | |
| 2537 | // TODO: Tune this. For example, lanewise swizzling is very expensive, so |
| 2538 | // swizzled lanes should be given greater weight. |
| 2539 | |
| 2540 | // TODO: Investigate looping rather than always extracting/replacing specific |
| 2541 | // lanes to fill gaps. |
| 2542 | |
| 2543 | auto IsConstant = [](const SDValue &V) { |
| 2544 | return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP; |
| 2545 | }; |
| 2546 | |
| 2547 | // Returns the source vector and index vector pair if they exist. Checks for: |
| 2548 | // (extract_vector_elt |
| 2549 | // $src, |
| 2550 | // (sign_extend_inreg (extract_vector_elt $indices, $i)) |
| 2551 | // ) |
| 2552 | auto GetSwizzleSrcs = [](size_t I, const SDValue &Lane) { |
| 2553 | auto Bail = std::make_pair(x: SDValue(), y: SDValue()); |
| 2554 | if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2555 | return Bail; |
| 2556 | const SDValue &SwizzleSrc = Lane->getOperand(Num: 0); |
| 2557 | const SDValue &IndexExt = Lane->getOperand(Num: 1); |
| 2558 | if (IndexExt->getOpcode() != ISD::SIGN_EXTEND_INREG) |
| 2559 | return Bail; |
| 2560 | const SDValue &Index = IndexExt->getOperand(Num: 0); |
| 2561 | if (Index->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2562 | return Bail; |
| 2563 | const SDValue &SwizzleIndices = Index->getOperand(Num: 0); |
| 2564 | if (SwizzleSrc.getValueType() != MVT::v16i8 || |
| 2565 | SwizzleIndices.getValueType() != MVT::v16i8 || |
| 2566 | Index->getOperand(Num: 1)->getOpcode() != ISD::Constant || |
| 2567 | Index->getConstantOperandVal(Num: 1) != I) |
| 2568 | return Bail; |
| 2569 | return std::make_pair(x: SwizzleSrc, y: SwizzleIndices); |
| 2570 | }; |
| 2571 | |
| 2572 | // If the lane is extracted from another vector at a constant index, return |
| 2573 | // that vector. The source vector must not have more lanes than the dest |
| 2574 | // because the shufflevector indices are in terms of the destination lanes and |
| 2575 | // would not be able to address the smaller individual source lanes. |
| 2576 | auto GetShuffleSrc = [&](const SDValue &Lane) { |
| 2577 | if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
| 2578 | return SDValue(); |
| 2579 | if (!isa<ConstantSDNode>(Val: Lane->getOperand(Num: 1).getNode())) |
| 2580 | return SDValue(); |
| 2581 | if (Lane->getOperand(Num: 0).getValueType().getVectorNumElements() > |
| 2582 | VecT.getVectorNumElements()) |
| 2583 | return SDValue(); |
| 2584 | return Lane->getOperand(Num: 0); |
| 2585 | }; |
| 2586 | |
| 2587 | using ValueEntry = std::pair<SDValue, size_t>; |
| 2588 | SmallVector<ValueEntry, 16> SplatValueCounts; |
| 2589 | |
| 2590 | using SwizzleEntry = std::pair<std::pair<SDValue, SDValue>, size_t>; |
| 2591 | SmallVector<SwizzleEntry, 16> SwizzleCounts; |
| 2592 | |
| 2593 | using ShuffleEntry = std::pair<SDValue, size_t>; |
| 2594 | SmallVector<ShuffleEntry, 16> ShuffleCounts; |
| 2595 | |
| 2596 | auto AddCount = [](auto &Counts, const auto &Val) { |
| 2597 | auto CountIt = |
| 2598 | llvm::find_if(Counts, [&Val](auto E) { return E.first == Val; }); |
| 2599 | if (CountIt == Counts.end()) { |
| 2600 | Counts.emplace_back(Val, 1); |
| 2601 | } else { |
| 2602 | CountIt->second++; |
| 2603 | } |
| 2604 | }; |
| 2605 | |
| 2606 | auto GetMostCommon = [](auto &Counts) { |
| 2607 | auto CommonIt = llvm::max_element(Counts, llvm::less_second()); |
| 2608 | assert(CommonIt != Counts.end() && "Unexpected all-undef build_vector" ); |
| 2609 | return *CommonIt; |
| 2610 | }; |
| 2611 | |
| 2612 | size_t NumConstantLanes = 0; |
| 2613 | |
| 2614 | // Count eligible lanes for each type of vector creation op |
| 2615 | for (size_t I = 0; I < Lanes; ++I) { |
| 2616 | const SDValue &Lane = Op->getOperand(Num: I); |
| 2617 | if (Lane.isUndef()) |
| 2618 | continue; |
| 2619 | |
| 2620 | AddCount(SplatValueCounts, Lane); |
| 2621 | |
| 2622 | if (IsConstant(Lane)) |
| 2623 | NumConstantLanes++; |
| 2624 | if (auto ShuffleSrc = GetShuffleSrc(Lane)) |
| 2625 | AddCount(ShuffleCounts, ShuffleSrc); |
| 2626 | if (CanSwizzle) { |
| 2627 | auto SwizzleSrcs = GetSwizzleSrcs(I, Lane); |
| 2628 | if (SwizzleSrcs.first) |
| 2629 | AddCount(SwizzleCounts, SwizzleSrcs); |
| 2630 | } |
| 2631 | } |
| 2632 | |
| 2633 | SDValue SplatValue; |
| 2634 | size_t NumSplatLanes; |
| 2635 | std::tie(args&: SplatValue, args&: NumSplatLanes) = GetMostCommon(SplatValueCounts); |
| 2636 | |
| 2637 | SDValue SwizzleSrc; |
| 2638 | SDValue SwizzleIndices; |
| 2639 | size_t NumSwizzleLanes = 0; |
| 2640 | if (SwizzleCounts.size()) |
| 2641 | std::forward_as_tuple(args: std::tie(args&: SwizzleSrc, args&: SwizzleIndices), |
| 2642 | args&: NumSwizzleLanes) = GetMostCommon(SwizzleCounts); |
| 2643 | |
| 2644 | // Shuffles can draw from up to two vectors, so find the two most common |
| 2645 | // sources. |
| 2646 | SDValue ShuffleSrc1, ShuffleSrc2; |
| 2647 | size_t NumShuffleLanes = 0; |
| 2648 | if (ShuffleCounts.size()) { |
| 2649 | std::tie(args&: ShuffleSrc1, args&: NumShuffleLanes) = GetMostCommon(ShuffleCounts); |
| 2650 | llvm::erase_if(C&: ShuffleCounts, |
| 2651 | P: [&](const auto &Pair) { return Pair.first == ShuffleSrc1; }); |
| 2652 | } |
| 2653 | if (ShuffleCounts.size()) { |
| 2654 | size_t AdditionalShuffleLanes; |
| 2655 | std::tie(args&: ShuffleSrc2, args&: AdditionalShuffleLanes) = |
| 2656 | GetMostCommon(ShuffleCounts); |
| 2657 | NumShuffleLanes += AdditionalShuffleLanes; |
| 2658 | } |
| 2659 | |
| 2660 | // Predicate returning true if the lane is properly initialized by the |
| 2661 | // original instruction |
| 2662 | std::function<bool(size_t, const SDValue &)> IsLaneConstructed; |
| 2663 | SDValue Result; |
| 2664 | // Prefer swizzles over shuffles over vector consts over splats |
| 2665 | if (NumSwizzleLanes >= NumShuffleLanes && |
| 2666 | NumSwizzleLanes >= NumConstantLanes && NumSwizzleLanes >= NumSplatLanes) { |
| 2667 | Result = DAG.getNode(Opcode: WebAssemblyISD::SWIZZLE, DL, VT: VecT, N1: SwizzleSrc, |
| 2668 | N2: SwizzleIndices); |
| 2669 | auto Swizzled = std::make_pair(x&: SwizzleSrc, y&: SwizzleIndices); |
| 2670 | IsLaneConstructed = [&, Swizzled](size_t I, const SDValue &Lane) { |
| 2671 | return Swizzled == GetSwizzleSrcs(I, Lane); |
| 2672 | }; |
| 2673 | } else if (NumShuffleLanes >= NumConstantLanes && |
| 2674 | NumShuffleLanes >= NumSplatLanes) { |
| 2675 | size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits() / 8; |
| 2676 | size_t DestLaneCount = VecT.getVectorNumElements(); |
| 2677 | size_t Scale1 = 1; |
| 2678 | size_t Scale2 = 1; |
| 2679 | SDValue Src1 = ShuffleSrc1; |
| 2680 | SDValue Src2 = ShuffleSrc2 ? ShuffleSrc2 : DAG.getUNDEF(VT: VecT); |
| 2681 | if (Src1.getValueType() != VecT) { |
| 2682 | size_t LaneSize = |
| 2683 | Src1.getValueType().getVectorElementType().getFixedSizeInBits() / 8; |
| 2684 | assert(LaneSize > DestLaneSize); |
| 2685 | Scale1 = LaneSize / DestLaneSize; |
| 2686 | Src1 = DAG.getBitcast(VT: VecT, V: Src1); |
| 2687 | } |
| 2688 | if (Src2.getValueType() != VecT) { |
| 2689 | size_t LaneSize = |
| 2690 | Src2.getValueType().getVectorElementType().getFixedSizeInBits() / 8; |
| 2691 | assert(LaneSize > DestLaneSize); |
| 2692 | Scale2 = LaneSize / DestLaneSize; |
| 2693 | Src2 = DAG.getBitcast(VT: VecT, V: Src2); |
| 2694 | } |
| 2695 | |
| 2696 | int Mask[16]; |
| 2697 | assert(DestLaneCount <= 16); |
| 2698 | for (size_t I = 0; I < DestLaneCount; ++I) { |
| 2699 | const SDValue &Lane = Op->getOperand(Num: I); |
| 2700 | SDValue Src = GetShuffleSrc(Lane); |
| 2701 | if (Src == ShuffleSrc1) { |
| 2702 | Mask[I] = Lane->getConstantOperandVal(Num: 1) * Scale1; |
| 2703 | } else if (Src && Src == ShuffleSrc2) { |
| 2704 | Mask[I] = DestLaneCount + Lane->getConstantOperandVal(Num: 1) * Scale2; |
| 2705 | } else { |
| 2706 | Mask[I] = -1; |
| 2707 | } |
| 2708 | } |
| 2709 | ArrayRef<int> MaskRef(Mask, DestLaneCount); |
| 2710 | Result = DAG.getVectorShuffle(VT: VecT, dl: DL, N1: Src1, N2: Src2, Mask: MaskRef); |
| 2711 | IsLaneConstructed = [&](size_t, const SDValue &Lane) { |
| 2712 | auto Src = GetShuffleSrc(Lane); |
| 2713 | return Src == ShuffleSrc1 || (Src && Src == ShuffleSrc2); |
| 2714 | }; |
| 2715 | } else if (NumConstantLanes >= NumSplatLanes) { |
| 2716 | SmallVector<SDValue, 16> ConstLanes; |
| 2717 | for (const SDValue &Lane : Op->op_values()) { |
| 2718 | if (IsConstant(Lane)) { |
| 2719 | // Values may need to be fixed so that they will sign extend to be |
| 2720 | // within the expected range during ISel. Check whether the value is in |
| 2721 | // bounds based on the lane bit width and if it is out of bounds, lop |
| 2722 | // off the extra bits. |
| 2723 | uint64_t LaneBits = 128 / Lanes; |
| 2724 | if (auto *Const = dyn_cast<ConstantSDNode>(Val: Lane.getNode())) { |
| 2725 | ConstLanes.push_back(Elt: DAG.getConstant( |
| 2726 | Val: Const->getAPIntValue().trunc(width: LaneBits).getZExtValue(), |
| 2727 | DL: SDLoc(Lane), VT: LaneT)); |
| 2728 | } else { |
| 2729 | ConstLanes.push_back(Elt: Lane); |
| 2730 | } |
| 2731 | } else if (LaneT.isFloatingPoint()) { |
| 2732 | ConstLanes.push_back(Elt: DAG.getConstantFP(Val: 0, DL, VT: LaneT)); |
| 2733 | } else { |
| 2734 | ConstLanes.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: LaneT)); |
| 2735 | } |
| 2736 | } |
| 2737 | Result = DAG.getBuildVector(VT: VecT, DL, Ops: ConstLanes); |
| 2738 | IsLaneConstructed = [&IsConstant](size_t _, const SDValue &Lane) { |
| 2739 | return IsConstant(Lane); |
| 2740 | }; |
| 2741 | } else { |
| 2742 | size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits(); |
| 2743 | if (NumSplatLanes == 1 && Op->getOperand(Num: 0) == SplatValue && |
| 2744 | (DestLaneSize == 32 || DestLaneSize == 64)) { |
| 2745 | // Could be selected to load_zero. |
| 2746 | Result = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL, VT: VecT, Operand: SplatValue); |
| 2747 | } else { |
| 2748 | // Use a splat (which might be selected as a load splat) |
| 2749 | Result = DAG.getSplatBuildVector(VT: VecT, DL, Op: SplatValue); |
| 2750 | } |
| 2751 | IsLaneConstructed = [&SplatValue](size_t _, const SDValue &Lane) { |
| 2752 | return Lane == SplatValue; |
| 2753 | }; |
| 2754 | } |
| 2755 | |
| 2756 | assert(Result); |
| 2757 | assert(IsLaneConstructed); |
| 2758 | |
| 2759 | // Add replace_lane instructions for any unhandled values |
| 2760 | for (size_t I = 0; I < Lanes; ++I) { |
| 2761 | const SDValue &Lane = Op->getOperand(Num: I); |
| 2762 | if (!Lane.isUndef() && !IsLaneConstructed(I, Lane)) |
| 2763 | Result = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL, VT: VecT, N1: Result, N2: Lane, |
| 2764 | N3: DAG.getConstant(Val: I, DL, VT: MVT::i32)); |
| 2765 | } |
| 2766 | |
| 2767 | return Result; |
| 2768 | } |
| 2769 | |
| 2770 | SDValue |
| 2771 | WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, |
| 2772 | SelectionDAG &DAG) const { |
| 2773 | SDLoc DL(Op); |
| 2774 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: Op.getNode())->getMask(); |
| 2775 | MVT VecType = Op.getOperand(i: 0).getSimpleValueType(); |
| 2776 | assert(VecType.is128BitVector() && "Unexpected shuffle vector type" ); |
| 2777 | size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8; |
| 2778 | |
| 2779 | // Space for two vector args and sixteen mask indices |
| 2780 | SDValue Ops[18]; |
| 2781 | size_t OpIdx = 0; |
| 2782 | Ops[OpIdx++] = Op.getOperand(i: 0); |
| 2783 | Ops[OpIdx++] = Op.getOperand(i: 1); |
| 2784 | |
| 2785 | // Expand mask indices to byte indices and materialize them as operands |
| 2786 | for (int M : Mask) { |
| 2787 | for (size_t J = 0; J < LaneBytes; ++J) { |
| 2788 | // Lower undefs (represented by -1 in mask) to {0..J}, which use a |
| 2789 | // whole lane of vector input, to allow further reduction at VM. E.g. |
| 2790 | // match an 8x16 byte shuffle to an equivalent cheaper 32x4 shuffle. |
| 2791 | uint64_t ByteIndex = M == -1 ? J : (uint64_t)M * LaneBytes + J; |
| 2792 | Ops[OpIdx++] = DAG.getConstant(Val: ByteIndex, DL, VT: MVT::i32); |
| 2793 | } |
| 2794 | } |
| 2795 | |
| 2796 | return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops); |
| 2797 | } |
| 2798 | |
| 2799 | SDValue WebAssemblyTargetLowering::LowerSETCC(SDValue Op, |
| 2800 | SelectionDAG &DAG) const { |
| 2801 | SDLoc DL(Op); |
| 2802 | // The legalizer does not know how to expand the unsupported comparison modes |
| 2803 | // of i64x2 vectors, so we manually unroll them here. |
| 2804 | assert(Op->getOperand(0)->getSimpleValueType(0) == MVT::v2i64); |
| 2805 | SmallVector<SDValue, 2> LHS, RHS; |
| 2806 | DAG.ExtractVectorElements(Op: Op->getOperand(Num: 0), Args&: LHS); |
| 2807 | DAG.ExtractVectorElements(Op: Op->getOperand(Num: 1), Args&: RHS); |
| 2808 | const SDValue &CC = Op->getOperand(Num: 2); |
| 2809 | auto MakeLane = [&](unsigned I) { |
| 2810 | return DAG.getNode(Opcode: ISD::SELECT_CC, DL, VT: MVT::i64, N1: LHS[I], N2: RHS[I], |
| 2811 | N3: DAG.getConstant(Val: uint64_t(-1), DL, VT: MVT::i64), |
| 2812 | N4: DAG.getConstant(Val: uint64_t(0), DL, VT: MVT::i64), N5: CC); |
| 2813 | }; |
| 2814 | return DAG.getBuildVector(VT: Op->getValueType(ResNo: 0), DL, |
| 2815 | Ops: {MakeLane(0), MakeLane(1)}); |
| 2816 | } |
| 2817 | |
| 2818 | SDValue |
| 2819 | WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op, |
| 2820 | SelectionDAG &DAG) const { |
| 2821 | // Allow constant lane indices, expand variable lane indices |
| 2822 | SDNode *IdxNode = Op.getOperand(i: Op.getNumOperands() - 1).getNode(); |
| 2823 | if (isa<ConstantSDNode>(Val: IdxNode)) { |
| 2824 | // Ensure the index type is i32 to match the tablegen patterns |
| 2825 | uint64_t Idx = IdxNode->getAsZExtVal(); |
| 2826 | SmallVector<SDValue, 3> Ops(Op.getNode()->ops()); |
| 2827 | Ops[Op.getNumOperands() - 1] = |
| 2828 | DAG.getConstant(Val: Idx, DL: SDLoc(IdxNode), VT: MVT::i32); |
| 2829 | return DAG.getNode(Opcode: Op.getOpcode(), DL: SDLoc(Op), VT: Op.getValueType(), Ops); |
| 2830 | } |
| 2831 | // Perform default expansion |
| 2832 | return SDValue(); |
| 2833 | } |
| 2834 | |
| 2835 | static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) { |
| 2836 | EVT LaneT = Op.getSimpleValueType().getVectorElementType(); |
| 2837 | // 32-bit and 64-bit unrolled shifts will have proper semantics |
| 2838 | if (LaneT.bitsGE(VT: MVT::i32)) |
| 2839 | return DAG.UnrollVectorOp(N: Op.getNode()); |
| 2840 | // Otherwise mask the shift value to get proper semantics from 32-bit shift |
| 2841 | SDLoc DL(Op); |
| 2842 | size_t NumLanes = Op.getSimpleValueType().getVectorNumElements(); |
| 2843 | SDValue Mask = DAG.getConstant(Val: LaneT.getSizeInBits() - 1, DL, VT: MVT::i32); |
| 2844 | unsigned ShiftOpcode = Op.getOpcode(); |
| 2845 | SmallVector<SDValue, 16> ShiftedElements; |
| 2846 | DAG.ExtractVectorElements(Op: Op.getOperand(i: 0), Args&: ShiftedElements, Start: 0, Count: 0, EltVT: MVT::i32); |
| 2847 | SmallVector<SDValue, 16> ShiftElements; |
| 2848 | DAG.ExtractVectorElements(Op: Op.getOperand(i: 1), Args&: ShiftElements, Start: 0, Count: 0, EltVT: MVT::i32); |
| 2849 | SmallVector<SDValue, 16> UnrolledOps; |
| 2850 | for (size_t i = 0; i < NumLanes; ++i) { |
| 2851 | SDValue MaskedShiftValue = |
| 2852 | DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: ShiftElements[i], N2: Mask); |
| 2853 | SDValue ShiftedValue = ShiftedElements[i]; |
| 2854 | if (ShiftOpcode == ISD::SRA) |
| 2855 | ShiftedValue = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: MVT::i32, |
| 2856 | N1: ShiftedValue, N2: DAG.getValueType(LaneT)); |
| 2857 | UnrolledOps.push_back( |
| 2858 | Elt: DAG.getNode(Opcode: ShiftOpcode, DL, VT: MVT::i32, N1: ShiftedValue, N2: MaskedShiftValue)); |
| 2859 | } |
| 2860 | return DAG.getBuildVector(VT: Op.getValueType(), DL, Ops: UnrolledOps); |
| 2861 | } |
| 2862 | |
| 2863 | SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op, |
| 2864 | SelectionDAG &DAG) const { |
| 2865 | SDLoc DL(Op); |
| 2866 | // Only manually lower vector shifts |
| 2867 | assert(Op.getSimpleValueType().isVector()); |
| 2868 | |
| 2869 | uint64_t LaneBits = Op.getValueType().getScalarSizeInBits(); |
| 2870 | auto ShiftVal = Op.getOperand(i: 1); |
| 2871 | |
| 2872 | // Try to skip bitmask operation since it is implied inside shift instruction |
| 2873 | auto SkipImpliedMask = [](SDValue MaskOp, uint64_t MaskBits) { |
| 2874 | if (MaskOp.getOpcode() != ISD::AND) |
| 2875 | return MaskOp; |
| 2876 | SDValue LHS = MaskOp.getOperand(i: 0); |
| 2877 | SDValue RHS = MaskOp.getOperand(i: 1); |
| 2878 | if (MaskOp.getValueType().isVector()) { |
| 2879 | APInt MaskVal; |
| 2880 | if (!ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal)) |
| 2881 | std::swap(a&: LHS, b&: RHS); |
| 2882 | |
| 2883 | if (ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal) && |
| 2884 | MaskVal == MaskBits) |
| 2885 | MaskOp = LHS; |
| 2886 | } else { |
| 2887 | if (!isa<ConstantSDNode>(Val: RHS.getNode())) |
| 2888 | std::swap(a&: LHS, b&: RHS); |
| 2889 | |
| 2890 | auto ConstantRHS = dyn_cast<ConstantSDNode>(Val: RHS.getNode()); |
| 2891 | if (ConstantRHS && ConstantRHS->getAPIntValue() == MaskBits) |
| 2892 | MaskOp = LHS; |
| 2893 | } |
| 2894 | |
| 2895 | return MaskOp; |
| 2896 | }; |
| 2897 | |
| 2898 | // Skip vector and operation |
| 2899 | ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1); |
| 2900 | ShiftVal = DAG.getSplatValue(V: ShiftVal); |
| 2901 | if (!ShiftVal) |
| 2902 | return unrollVectorShift(Op, DAG); |
| 2903 | |
| 2904 | // Skip scalar and operation |
| 2905 | ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1); |
| 2906 | // Use anyext because none of the high bits can affect the shift |
| 2907 | ShiftVal = DAG.getAnyExtOrTrunc(Op: ShiftVal, DL, VT: MVT::i32); |
| 2908 | |
| 2909 | unsigned Opcode; |
| 2910 | switch (Op.getOpcode()) { |
| 2911 | case ISD::SHL: |
| 2912 | Opcode = WebAssemblyISD::VEC_SHL; |
| 2913 | break; |
| 2914 | case ISD::SRA: |
| 2915 | Opcode = WebAssemblyISD::VEC_SHR_S; |
| 2916 | break; |
| 2917 | case ISD::SRL: |
| 2918 | Opcode = WebAssemblyISD::VEC_SHR_U; |
| 2919 | break; |
| 2920 | default: |
| 2921 | llvm_unreachable("unexpected opcode" ); |
| 2922 | } |
| 2923 | |
| 2924 | return DAG.getNode(Opcode, DL, VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: ShiftVal); |
| 2925 | } |
| 2926 | |
| 2927 | SDValue WebAssemblyTargetLowering::LowerFP_TO_INT_SAT(SDValue Op, |
| 2928 | SelectionDAG &DAG) const { |
| 2929 | EVT ResT = Op.getValueType(); |
| 2930 | EVT SatVT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT(); |
| 2931 | |
| 2932 | if ((ResT == MVT::i32 || ResT == MVT::i64) && |
| 2933 | (SatVT == MVT::i32 || SatVT == MVT::i64)) |
| 2934 | return Op; |
| 2935 | |
| 2936 | if (ResT == MVT::v4i32 && SatVT == MVT::i32) |
| 2937 | return Op; |
| 2938 | |
| 2939 | if (ResT == MVT::v8i16 && SatVT == MVT::i16) |
| 2940 | return Op; |
| 2941 | |
| 2942 | return SDValue(); |
| 2943 | } |
| 2944 | |
| 2945 | static bool HasNoSignedZerosOrNaNs(SDValue Op, SelectionDAG &DAG) { |
| 2946 | return (Op->getFlags().hasNoNaNs() || |
| 2947 | (DAG.isKnownNeverNaN(Op: Op->getOperand(Num: 0)) && |
| 2948 | DAG.isKnownNeverNaN(Op: Op->getOperand(Num: 1)))) && |
| 2949 | (Op->getFlags().hasNoSignedZeros() || |
| 2950 | DAG.isKnownNeverLogicalZero(Op: Op->getOperand(Num: 0)) || |
| 2951 | DAG.isKnownNeverLogicalZero(Op: Op->getOperand(Num: 1))); |
| 2952 | } |
| 2953 | |
| 2954 | SDValue WebAssemblyTargetLowering::LowerFMIN(SDValue Op, |
| 2955 | SelectionDAG &DAG) const { |
| 2956 | if (Subtarget->hasRelaxedSIMD() && HasNoSignedZerosOrNaNs(Op, DAG)) { |
| 2957 | return DAG.getNode(Opcode: WebAssemblyISD::RELAXED_FMIN, DL: SDLoc(Op), |
| 2958 | VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: Op.getOperand(i: 1)); |
| 2959 | } |
| 2960 | return SDValue(); |
| 2961 | } |
| 2962 | |
| 2963 | SDValue WebAssemblyTargetLowering::LowerFMAX(SDValue Op, |
| 2964 | SelectionDAG &DAG) const { |
| 2965 | if (Subtarget->hasRelaxedSIMD() && HasNoSignedZerosOrNaNs(Op, DAG)) { |
| 2966 | return DAG.getNode(Opcode: WebAssemblyISD::RELAXED_FMAX, DL: SDLoc(Op), |
| 2967 | VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: Op.getOperand(i: 1)); |
| 2968 | } |
| 2969 | return SDValue(); |
| 2970 | } |
| 2971 | |
| 2972 | //===----------------------------------------------------------------------===// |
| 2973 | // Custom DAG combine hooks |
| 2974 | //===----------------------------------------------------------------------===// |
| 2975 | static SDValue |
| 2976 | performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
| 2977 | auto &DAG = DCI.DAG; |
| 2978 | auto Shuffle = cast<ShuffleVectorSDNode>(Val: N); |
| 2979 | |
| 2980 | // Hoist vector bitcasts that don't change the number of lanes out of unary |
| 2981 | // shuffles, where they are less likely to get in the way of other combines. |
| 2982 | // (shuffle (vNxT1 (bitcast (vNxT0 x))), undef, mask) -> |
| 2983 | // (vNxT1 (bitcast (vNxT0 (shuffle x, undef, mask)))) |
| 2984 | SDValue Bitcast = N->getOperand(Num: 0); |
| 2985 | if (Bitcast.getOpcode() != ISD::BITCAST) |
| 2986 | return SDValue(); |
| 2987 | if (!N->getOperand(Num: 1).isUndef()) |
| 2988 | return SDValue(); |
| 2989 | SDValue CastOp = Bitcast.getOperand(i: 0); |
| 2990 | EVT SrcType = CastOp.getValueType(); |
| 2991 | EVT DstType = Bitcast.getValueType(); |
| 2992 | if (!SrcType.is128BitVector() || |
| 2993 | SrcType.getVectorNumElements() != DstType.getVectorNumElements()) |
| 2994 | return SDValue(); |
| 2995 | SDValue NewShuffle = DAG.getVectorShuffle( |
| 2996 | VT: SrcType, dl: SDLoc(N), N1: CastOp, N2: DAG.getUNDEF(VT: SrcType), Mask: Shuffle->getMask()); |
| 2997 | return DAG.getBitcast(VT: DstType, V: NewShuffle); |
| 2998 | } |
| 2999 | |
| 3000 | /// Convert ({u,s}itofp vec) --> ({u,s}itofp ({s,z}ext vec)) so it doesn't get |
| 3001 | /// split up into scalar instructions during legalization, and the vector |
| 3002 | /// extending instructions are selected in performVectorExtendCombine below. |
| 3003 | static SDValue |
| 3004 | performVectorExtendToFPCombine(SDNode *N, |
| 3005 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3006 | auto &DAG = DCI.DAG; |
| 3007 | assert(N->getOpcode() == ISD::UINT_TO_FP || |
| 3008 | N->getOpcode() == ISD::SINT_TO_FP); |
| 3009 | |
| 3010 | EVT InVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
| 3011 | EVT ResVT = N->getValueType(ResNo: 0); |
| 3012 | MVT ExtVT; |
| 3013 | if (ResVT == MVT::v4f32 && (InVT == MVT::v4i16 || InVT == MVT::v4i8)) |
| 3014 | ExtVT = MVT::v4i32; |
| 3015 | else if (ResVT == MVT::v2f64 && (InVT == MVT::v2i16 || InVT == MVT::v2i8)) |
| 3016 | ExtVT = MVT::v2i32; |
| 3017 | else |
| 3018 | return SDValue(); |
| 3019 | |
| 3020 | unsigned Op = |
| 3021 | N->getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; |
| 3022 | SDValue Conv = DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ExtVT, Operand: N->getOperand(Num: 0)); |
| 3023 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: ResVT, Operand: Conv); |
| 3024 | } |
| 3025 | |
| 3026 | static SDValue |
| 3027 | performVectorNonNegToFPCombine(SDNode *N, |
| 3028 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3029 | auto &DAG = DCI.DAG; |
| 3030 | |
| 3031 | SDNodeFlags Flags = N->getFlags(); |
| 3032 | SDValue Op0 = N->getOperand(Num: 0); |
| 3033 | EVT VT = N->getValueType(ResNo: 0); |
| 3034 | |
| 3035 | // Optimize uitofp to sitofp when the sign bit is known to be zero. |
| 3036 | // Depending on the target (runtime) backend, this might be performance |
| 3037 | // neutral (e.g. AArch64) or a significant improvement (e.g. x86_64). |
| 3038 | if (VT.isVector() && (Flags.hasNonNeg() || DAG.SignBitIsZero(Op: Op0))) { |
| 3039 | return DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: SDLoc(N), VT, Operand: Op0); |
| 3040 | } |
| 3041 | |
| 3042 | return SDValue(); |
| 3043 | } |
| 3044 | |
| 3045 | static SDValue |
| 3046 | performVectorExtendCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
| 3047 | auto &DAG = DCI.DAG; |
| 3048 | assert(N->getOpcode() == ISD::SIGN_EXTEND || |
| 3049 | N->getOpcode() == ISD::ZERO_EXTEND); |
| 3050 | |
| 3051 | EVT ResVT = N->getValueType(ResNo: 0); |
| 3052 | bool IsSext = N->getOpcode() == ISD::SIGN_EXTEND; |
| 3053 | SDLoc DL(N); |
| 3054 | |
| 3055 | if (ResVT == MVT::v16i32 && N->getOperand(Num: 0)->getValueType(ResNo: 0) == MVT::v16i8) { |
| 3056 | // Use a tree of extend low/high to split and extend the input in two |
| 3057 | // layers to avoid doing several shuffles and even more extends. |
| 3058 | unsigned LowOp = |
| 3059 | IsSext ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U; |
| 3060 | unsigned HighOp = |
| 3061 | IsSext ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U; |
| 3062 | SDValue Input = N->getOperand(Num: 0); |
| 3063 | SDValue LowHalf = DAG.getNode(Opcode: LowOp, DL, VT: MVT::v8i16, Operand: Input); |
| 3064 | SDValue HighHalf = DAG.getNode(Opcode: HighOp, DL, VT: MVT::v8i16, Operand: Input); |
| 3065 | SDValue Subvectors[] = { |
| 3066 | DAG.getNode(Opcode: LowOp, DL, VT: MVT::v4i32, Operand: LowHalf), |
| 3067 | DAG.getNode(Opcode: HighOp, DL, VT: MVT::v4i32, Operand: LowHalf), |
| 3068 | DAG.getNode(Opcode: LowOp, DL, VT: MVT::v4i32, Operand: HighHalf), |
| 3069 | DAG.getNode(Opcode: HighOp, DL, VT: MVT::v4i32, Operand: HighHalf), |
| 3070 | }; |
| 3071 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: ResVT, Ops: Subvectors); |
| 3072 | } |
| 3073 | |
| 3074 | // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if |
| 3075 | // possible before the extract_subvector can be expanded. |
| 3076 | auto = N->getOperand(Num: 0); |
| 3077 | if (Extract.getOpcode() != ISD::EXTRACT_SUBVECTOR) |
| 3078 | return SDValue(); |
| 3079 | auto Source = Extract.getOperand(i: 0); |
| 3080 | auto *IndexNode = dyn_cast<ConstantSDNode>(Val: Extract.getOperand(i: 1)); |
| 3081 | if (IndexNode == nullptr) |
| 3082 | return SDValue(); |
| 3083 | auto Index = IndexNode->getZExtValue(); |
| 3084 | |
| 3085 | // Only v8i8, v4i16, and v2i32 extracts can be widened, and only if the |
| 3086 | // extracted subvector is the low or high half of its source. |
| 3087 | if (ResVT == MVT::v8i16) { |
| 3088 | if (Extract.getValueType() != MVT::v8i8 || |
| 3089 | Source.getValueType() != MVT::v16i8 || (Index != 0 && Index != 8)) |
| 3090 | return SDValue(); |
| 3091 | } else if (ResVT == MVT::v4i32) { |
| 3092 | if (Extract.getValueType() != MVT::v4i16 || |
| 3093 | Source.getValueType() != MVT::v8i16 || (Index != 0 && Index != 4)) |
| 3094 | return SDValue(); |
| 3095 | } else if (ResVT == MVT::v2i64) { |
| 3096 | if (Extract.getValueType() != MVT::v2i32 || |
| 3097 | Source.getValueType() != MVT::v4i32 || (Index != 0 && Index != 2)) |
| 3098 | return SDValue(); |
| 3099 | } else { |
| 3100 | return SDValue(); |
| 3101 | } |
| 3102 | |
| 3103 | bool IsLow = Index == 0; |
| 3104 | |
| 3105 | unsigned Op = IsSext ? (IsLow ? WebAssemblyISD::EXTEND_LOW_S |
| 3106 | : WebAssemblyISD::EXTEND_HIGH_S) |
| 3107 | : (IsLow ? WebAssemblyISD::EXTEND_LOW_U |
| 3108 | : WebAssemblyISD::EXTEND_HIGH_U); |
| 3109 | |
| 3110 | return DAG.getNode(Opcode: Op, DL, VT: ResVT, Operand: Source); |
| 3111 | } |
| 3112 | |
| 3113 | static SDValue |
| 3114 | performVectorTruncZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
| 3115 | auto &DAG = DCI.DAG; |
| 3116 | |
| 3117 | auto GetWasmConversionOp = [](unsigned Op) { |
| 3118 | switch (Op) { |
| 3119 | case ISD::FP_TO_SINT_SAT: |
| 3120 | return WebAssemblyISD::TRUNC_SAT_ZERO_S; |
| 3121 | case ISD::FP_TO_UINT_SAT: |
| 3122 | return WebAssemblyISD::TRUNC_SAT_ZERO_U; |
| 3123 | case ISD::FP_ROUND: |
| 3124 | return WebAssemblyISD::DEMOTE_ZERO; |
| 3125 | } |
| 3126 | llvm_unreachable("unexpected op" ); |
| 3127 | }; |
| 3128 | |
| 3129 | auto IsZeroSplat = [](SDValue SplatVal) { |
| 3130 | auto *Splat = dyn_cast<BuildVectorSDNode>(Val: SplatVal.getNode()); |
| 3131 | APInt SplatValue, SplatUndef; |
| 3132 | unsigned SplatBitSize; |
| 3133 | bool HasAnyUndefs; |
| 3134 | // Endianness doesn't matter in this context because we are looking for |
| 3135 | // an all-zero value. |
| 3136 | return Splat && |
| 3137 | Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, |
| 3138 | HasAnyUndefs) && |
| 3139 | SplatValue == 0; |
| 3140 | }; |
| 3141 | |
| 3142 | if (N->getOpcode() == ISD::CONCAT_VECTORS) { |
| 3143 | // Combine this: |
| 3144 | // |
| 3145 | // (concat_vectors (v2i32 (fp_to_{s,u}int_sat $x, 32)), (v2i32 (splat 0))) |
| 3146 | // |
| 3147 | // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x). |
| 3148 | // |
| 3149 | // Or this: |
| 3150 | // |
| 3151 | // (concat_vectors ({v2f32, v4f16} (fp_round ({v2f64, v4f32} $x))), |
| 3152 | // ({v2f32, v4f16} (splat 0))) |
| 3153 | // |
| 3154 | // into ({f32x4, f16x8}.demote_zero_{f64x2, f32x4} $x). |
| 3155 | EVT ResVT; |
| 3156 | EVT ExpectedConversionType; |
| 3157 | auto Conversion = N->getOperand(Num: 0); |
| 3158 | auto ConversionOp = Conversion.getOpcode(); |
| 3159 | switch (ConversionOp) { |
| 3160 | case ISD::FP_TO_SINT_SAT: |
| 3161 | case ISD::FP_TO_UINT_SAT: |
| 3162 | ResVT = MVT::v4i32; |
| 3163 | ExpectedConversionType = MVT::v2i32; |
| 3164 | break; |
| 3165 | case ISD::FP_ROUND: |
| 3166 | if (Conversion.getValueType() == MVT::v2f32) { |
| 3167 | ResVT = MVT::v4f32; |
| 3168 | ExpectedConversionType = MVT::v2f32; |
| 3169 | } else if (Conversion.getValueType() == MVT::v4f16) { |
| 3170 | ResVT = MVT::v8f16; |
| 3171 | ExpectedConversionType = MVT::v4f16; |
| 3172 | } else { |
| 3173 | return SDValue(); |
| 3174 | } |
| 3175 | break; |
| 3176 | default: |
| 3177 | return SDValue(); |
| 3178 | } |
| 3179 | |
| 3180 | if (N->getValueType(ResNo: 0) != ResVT) |
| 3181 | return SDValue(); |
| 3182 | |
| 3183 | if (Conversion.getValueType() != ExpectedConversionType) |
| 3184 | return SDValue(); |
| 3185 | |
| 3186 | auto Source = Conversion.getOperand(i: 0); |
| 3187 | if (!((Source.getValueType() == MVT::v2f64 && ResVT == MVT::v4f32) || |
| 3188 | (Source.getValueType() == MVT::v2f64 && ResVT == MVT::v4i32) || |
| 3189 | (Source.getValueType() == MVT::v4f32 && ResVT == MVT::v8f16))) |
| 3190 | return SDValue(); |
| 3191 | |
| 3192 | if (!IsZeroSplat(N->getOperand(Num: 1)) || |
| 3193 | N->getOperand(Num: 1).getValueType() != ExpectedConversionType) |
| 3194 | return SDValue(); |
| 3195 | |
| 3196 | unsigned Op = GetWasmConversionOp(ConversionOp); |
| 3197 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
| 3198 | } |
| 3199 | |
| 3200 | // Combine this: |
| 3201 | // |
| 3202 | // (fp_to_{s,u}int_sat (concat_vectors $x, (v2f64 (splat 0))), 32) |
| 3203 | // |
| 3204 | // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x). |
| 3205 | // |
| 3206 | // Or this: |
| 3207 | // |
| 3208 | // ({v4f32, v8f16} (fp_round (concat_vectors $x, |
| 3209 | // ({v2f64, v4f32} (splat 0))))) |
| 3210 | // |
| 3211 | // into ({f32x4, f16x8}.demote_zero_{f64x2, f32x4} $x). |
| 3212 | EVT ResVT; |
| 3213 | auto ConversionOp = N->getOpcode(); |
| 3214 | switch (ConversionOp) { |
| 3215 | case ISD::FP_TO_SINT_SAT: |
| 3216 | case ISD::FP_TO_UINT_SAT: |
| 3217 | ResVT = MVT::v4i32; |
| 3218 | break; |
| 3219 | case ISD::FP_ROUND: |
| 3220 | ResVT = N->getValueType(ResNo: 0); |
| 3221 | break; |
| 3222 | default: |
| 3223 | llvm_unreachable("unexpected op" ); |
| 3224 | } |
| 3225 | |
| 3226 | if (N->getValueType(ResNo: 0) != ResVT) |
| 3227 | return SDValue(); |
| 3228 | |
| 3229 | auto Concat = N->getOperand(Num: 0); |
| 3230 | if (Concat.getOpcode() != ISD::CONCAT_VECTORS) |
| 3231 | return SDValue(); |
| 3232 | EVT ConcatVT = Concat.getValueType(); |
| 3233 | EVT SourceVT = Concat.getOperand(i: 0).getValueType(); |
| 3234 | |
| 3235 | if (!IsZeroSplat(Concat.getOperand(i: 1))) |
| 3236 | return SDValue(); |
| 3237 | |
| 3238 | if (ConversionOp == ISD::FP_ROUND) { |
| 3239 | bool IsF64ToF32 = |
| 3240 | ConcatVT == MVT::v4f64 && SourceVT == MVT::v2f64 && ResVT == MVT::v4f32; |
| 3241 | bool IsF32ToF16 = |
| 3242 | ConcatVT == MVT::v8f32 && SourceVT == MVT::v4f32 && ResVT == MVT::v8f16; |
| 3243 | if (!(IsF64ToF32 || IsF32ToF16)) |
| 3244 | return SDValue(); |
| 3245 | } else { |
| 3246 | if (ConcatVT != MVT::v4f64 || SourceVT != MVT::v2f64 || ResVT != MVT::v4i32) |
| 3247 | return SDValue(); |
| 3248 | } |
| 3249 | |
| 3250 | unsigned Op = GetWasmConversionOp(ConversionOp); |
| 3251 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Concat.getOperand(i: 0)); |
| 3252 | } |
| 3253 | |
| 3254 | // Helper to extract VectorWidth bits from Vec, starting from IdxVal. |
| 3255 | static SDValue (SDValue Vec, unsigned IdxVal, SelectionDAG &DAG, |
| 3256 | const SDLoc &DL, unsigned VectorWidth) { |
| 3257 | EVT VT = Vec.getValueType(); |
| 3258 | EVT ElVT = VT.getVectorElementType(); |
| 3259 | unsigned Factor = VT.getSizeInBits() / VectorWidth; |
| 3260 | EVT ResultVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ElVT, |
| 3261 | NumElements: VT.getVectorNumElements() / Factor); |
| 3262 | |
| 3263 | // Extract the relevant VectorWidth bits. Generate an EXTRACT_SUBVECTOR |
| 3264 | unsigned ElemsPerChunk = VectorWidth / ElVT.getSizeInBits(); |
| 3265 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2" ); |
| 3266 | |
| 3267 | // This is the index of the first element of the VectorWidth-bit chunk |
| 3268 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. |
| 3269 | IdxVal &= ~(ElemsPerChunk - 1); |
| 3270 | |
| 3271 | // If the input is a buildvector just emit a smaller one. |
| 3272 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) |
| 3273 | return DAG.getBuildVector(VT: ResultVT, DL, |
| 3274 | Ops: Vec->ops().slice(N: IdxVal, M: ElemsPerChunk)); |
| 3275 | |
| 3276 | SDValue VecIdx = DAG.getIntPtrConstant(Val: IdxVal, DL); |
| 3277 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: ResultVT, N1: Vec, N2: VecIdx); |
| 3278 | } |
| 3279 | |
| 3280 | // Helper to recursively truncate vector elements in half with NARROW_U. DstVT |
| 3281 | // is the expected destination value type after recursion. In is the initial |
| 3282 | // input. Note that the input should have enough leading zero bits to prevent |
| 3283 | // NARROW_U from saturating results. |
| 3284 | static SDValue truncateVectorWithNARROW(EVT DstVT, SDValue In, const SDLoc &DL, |
| 3285 | SelectionDAG &DAG) { |
| 3286 | EVT SrcVT = In.getValueType(); |
| 3287 | |
| 3288 | // No truncation required, we might get here due to recursive calls. |
| 3289 | if (SrcVT == DstVT) |
| 3290 | return In; |
| 3291 | |
| 3292 | unsigned SrcSizeInBits = SrcVT.getSizeInBits(); |
| 3293 | unsigned NumElems = SrcVT.getVectorNumElements(); |
| 3294 | if (!isPowerOf2_32(Value: NumElems)) |
| 3295 | return SDValue(); |
| 3296 | assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation" ); |
| 3297 | assert(SrcSizeInBits > DstVT.getSizeInBits() && "Illegal truncation" ); |
| 3298 | |
| 3299 | LLVMContext &Ctx = *DAG.getContext(); |
| 3300 | EVT PackedSVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: SrcVT.getScalarSizeInBits() / 2); |
| 3301 | |
| 3302 | // Narrow to the largest type possible: |
| 3303 | // vXi64/vXi32 -> i16x8.narrow_i32x4_u and vXi16 -> i8x16.narrow_i16x8_u. |
| 3304 | EVT InVT = MVT::i16, OutVT = MVT::i8; |
| 3305 | if (SrcVT.getScalarSizeInBits() > 16) { |
| 3306 | InVT = MVT::i32; |
| 3307 | OutVT = MVT::i16; |
| 3308 | } |
| 3309 | unsigned SubSizeInBits = SrcSizeInBits / 2; |
| 3310 | InVT = EVT::getVectorVT(Context&: Ctx, VT: InVT, NumElements: SubSizeInBits / InVT.getSizeInBits()); |
| 3311 | OutVT = EVT::getVectorVT(Context&: Ctx, VT: OutVT, NumElements: SubSizeInBits / OutVT.getSizeInBits()); |
| 3312 | |
| 3313 | // Split lower/upper subvectors. |
| 3314 | SDValue Lo = extractSubVector(Vec: In, IdxVal: 0, DAG, DL, VectorWidth: SubSizeInBits); |
| 3315 | SDValue Hi = extractSubVector(Vec: In, IdxVal: NumElems / 2, DAG, DL, VectorWidth: SubSizeInBits); |
| 3316 | |
| 3317 | // 256bit -> 128bit truncate - Narrow lower/upper 128-bit subvectors. |
| 3318 | if (SrcVT.is256BitVector() && DstVT.is128BitVector()) { |
| 3319 | Lo = DAG.getBitcast(VT: InVT, V: Lo); |
| 3320 | Hi = DAG.getBitcast(VT: InVT, V: Hi); |
| 3321 | SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::NARROW_U, DL, VT: OutVT, N1: Lo, N2: Hi); |
| 3322 | return DAG.getBitcast(VT: DstVT, V: Res); |
| 3323 | } |
| 3324 | |
| 3325 | // Recursively narrow lower/upper subvectors, concat result and narrow again. |
| 3326 | EVT PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems / 2); |
| 3327 | Lo = truncateVectorWithNARROW(DstVT: PackedVT, In: Lo, DL, DAG); |
| 3328 | Hi = truncateVectorWithNARROW(DstVT: PackedVT, In: Hi, DL, DAG); |
| 3329 | |
| 3330 | PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems); |
| 3331 | SDValue Res = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: PackedVT, N1: Lo, N2: Hi); |
| 3332 | return truncateVectorWithNARROW(DstVT, In: Res, DL, DAG); |
| 3333 | } |
| 3334 | |
| 3335 | static SDValue performTruncateCombine(SDNode *N, |
| 3336 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3337 | auto &DAG = DCI.DAG; |
| 3338 | |
| 3339 | SDValue In = N->getOperand(Num: 0); |
| 3340 | EVT InVT = In.getValueType(); |
| 3341 | if (!InVT.isSimple()) |
| 3342 | return SDValue(); |
| 3343 | |
| 3344 | EVT OutVT = N->getValueType(ResNo: 0); |
| 3345 | if (!OutVT.isVector()) |
| 3346 | return SDValue(); |
| 3347 | |
| 3348 | EVT OutSVT = OutVT.getVectorElementType(); |
| 3349 | EVT InSVT = InVT.getVectorElementType(); |
| 3350 | // Currently only cover truncate to v16i8 or v8i16. |
| 3351 | if (!((InSVT == MVT::i16 || InSVT == MVT::i32 || InSVT == MVT::i64) && |
| 3352 | (OutSVT == MVT::i8 || OutSVT == MVT::i16) && OutVT.is128BitVector())) |
| 3353 | return SDValue(); |
| 3354 | |
| 3355 | SDLoc DL(N); |
| 3356 | APInt Mask = APInt::getLowBitsSet(numBits: InVT.getScalarSizeInBits(), |
| 3357 | loBitsSet: OutVT.getScalarSizeInBits()); |
| 3358 | In = DAG.getNode(Opcode: ISD::AND, DL, VT: InVT, N1: In, N2: DAG.getConstant(Val: Mask, DL, VT: InVT)); |
| 3359 | return truncateVectorWithNARROW(DstVT: OutVT, In, DL, DAG); |
| 3360 | } |
| 3361 | |
| 3362 | static SDValue performBitcastCombine(SDNode *N, |
| 3363 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3364 | using namespace llvm::SDPatternMatch; |
| 3365 | auto &DAG = DCI.DAG; |
| 3366 | SDLoc DL(N); |
| 3367 | SDValue Src = N->getOperand(Num: 0); |
| 3368 | EVT VT = N->getValueType(ResNo: 0); |
| 3369 | EVT SrcVT = Src.getValueType(); |
| 3370 | |
| 3371 | if (!(DCI.isBeforeLegalize() && VT.isScalarInteger() && |
| 3372 | SrcVT.isFixedLengthVectorOf(EltVT: MVT::i1))) |
| 3373 | return SDValue(); |
| 3374 | |
| 3375 | unsigned NumElts = SrcVT.getVectorNumElements(); |
| 3376 | EVT Width = MVT::getIntegerVT(BitWidth: 128 / NumElts); |
| 3377 | |
| 3378 | // bitcast <N x i1> to iN, where N = 2, 4, 8, 16 (legal) |
| 3379 | // ==> bitmask |
| 3380 | if (NumElts == 2 || NumElts == 4 || NumElts == 8 || NumElts == 16) { |
| 3381 | return DAG.getZExtOrTrunc( |
| 3382 | Op: DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3383 | Ops: {DAG.getConstant(Val: Intrinsic::wasm_bitmask, DL, VT: MVT::i32), |
| 3384 | DAG.getSExtOrTrunc(Op: N->getOperand(Num: 0), DL, |
| 3385 | VT: SrcVT.changeVectorElementType( |
| 3386 | Context&: *DAG.getContext(), EltVT: Width))}), |
| 3387 | DL, VT); |
| 3388 | } |
| 3389 | |
| 3390 | // bitcast <N x i1>(setcc ...) to concat iN, where N = 32 and 64 (illegal) |
| 3391 | if (NumElts == 32 || NumElts == 64) { |
| 3392 | SDValue Concat, SetCCVector; |
| 3393 | ISD::CondCode SetCond; |
| 3394 | |
| 3395 | if (!sd_match(N, P: m_BitCast(Op: m_c_SetCC(LHS: m_Value(N&: Concat), RHS: m_Value(N&: SetCCVector), |
| 3396 | CC: m_CondCode(CC&: SetCond))))) |
| 3397 | return SDValue(); |
| 3398 | if (Concat.getOpcode() != ISD::CONCAT_VECTORS) |
| 3399 | return SDValue(); |
| 3400 | |
| 3401 | // Reconstruct the wide bitmask from each CONCAT_VECTORS operand. |
| 3402 | // Derive the per-chunk mask/integer types from the actual operand type |
| 3403 | // instead of hardcoding v16i1 / i16 for every chunk. |
| 3404 | EVT ConcatOperandVT = Concat.getOperand(i: 0).getValueType(); |
| 3405 | unsigned ConcatOperandNumElts = ConcatOperandVT.getVectorNumElements(); |
| 3406 | |
| 3407 | EVT ConcatOperandMaskVT = |
| 3408 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 3409 | EC: ElementCount::getFixed(MinVal: ConcatOperandNumElts)); |
| 3410 | EVT ConcatOperandBitmaskVT = |
| 3411 | EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: ConcatOperandNumElts); |
| 3412 | EVT ReturnVT = N->getValueType(ResNo: 0); |
| 3413 | SDValue ReconstructedBitmask = DAG.getConstant(Val: 0, DL, VT: ReturnVT); |
| 3414 | // Example: |
| 3415 | // v32i16 = concat(v8i16, v8i16, v8i16, v8i16) |
| 3416 | // -> v8i1 + v8i1 + v8i1 + v8i1 |
| 3417 | // -> i8 + i8 + i8 + i8 |
| 3418 | // -> reconstructed i32 bitmask |
| 3419 | for (size_t I = 0; I < Concat->ops().size(); ++I) { |
| 3420 | SDValue ConcatOperand = Concat.getOperand(i: I); |
| 3421 | assert(ConcatOperand.getValueType() == ConcatOperandVT && |
| 3422 | "concat_vectors operands must have the same type" ); |
| 3423 | |
| 3424 | SDValue SetCCVectorOperand = |
| 3425 | extractSubVector(Vec: SetCCVector, IdxVal: I * ConcatOperandNumElts, DAG, DL, VectorWidth: 128); |
| 3426 | if (!SetCCVectorOperand || |
| 3427 | SetCCVectorOperand.getValueType() != ConcatOperandVT) |
| 3428 | return SDValue(); |
| 3429 | |
| 3430 | // Build the per-chunk mask using the correct chunk type: |
| 3431 | // v16i8 -> v16i1 -> i16 |
| 3432 | // v8i16 -> v8i1 -> i8 |
| 3433 | // v4i32 -> v4i1 -> i4 |
| 3434 | // v2i64 -> v2i1 -> i2 |
| 3435 | SDValue ConcatOperandMask = DAG.getSetCC( |
| 3436 | DL, VT: ConcatOperandMaskVT, LHS: ConcatOperand, RHS: SetCCVectorOperand, Cond: SetCond); |
| 3437 | SDValue ConcatOperandBitmask = |
| 3438 | DAG.getBitcast(VT: ConcatOperandBitmaskVT, V: ConcatOperandMask); |
| 3439 | SDValue ExtendedConcatOperandBitmask = |
| 3440 | DAG.getZExtOrTrunc(Op: ConcatOperandBitmask, DL, VT: ReturnVT); |
| 3441 | |
| 3442 | // Shift the previously reconstructed bits to make room for this chunk. |
| 3443 | if (I != 0) { |
| 3444 | ReconstructedBitmask = DAG.getNode( |
| 3445 | Opcode: ISD::SHL, DL, VT: ReturnVT, N1: ReconstructedBitmask, |
| 3446 | N2: DAG.getShiftAmountConstant(Val: ConcatOperandNumElts, VT: ReturnVT, DL)); |
| 3447 | } |
| 3448 | |
| 3449 | // Merge disjoint partial bitmasks with OR. |
| 3450 | ReconstructedBitmask = |
| 3451 | DAG.getNode(Opcode: ISD::OR, DL, VT: ReturnVT, N1: ReconstructedBitmask, |
| 3452 | N2: ExtendedConcatOperandBitmask); |
| 3453 | } |
| 3454 | |
| 3455 | return ReconstructedBitmask; |
| 3456 | } |
| 3457 | |
| 3458 | return SDValue(); |
| 3459 | } |
| 3460 | |
| 3461 | static SDValue performBitmaskCombine(SDNode *N, SelectionDAG &DAG) { |
| 3462 | // bitmask (setcc <X>, 0, setlt) => bitmask X |
| 3463 | assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN); |
| 3464 | using namespace llvm::SDPatternMatch; |
| 3465 | |
| 3466 | if (N->getConstantOperandVal(Num: 0) != Intrinsic::wasm_bitmask) |
| 3467 | return SDValue(); |
| 3468 | |
| 3469 | SDValue LHS; |
| 3470 | if (!sd_match(N: N->getOperand(Num: 1), P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(), |
| 3471 | CC: m_SpecificCondCode(CC: ISD::SETLT)))) |
| 3472 | return SDValue(); |
| 3473 | |
| 3474 | SDLoc DL(N); |
| 3475 | return DAG.getNode( |
| 3476 | Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: N->getValueType(ResNo: 0), |
| 3477 | Ops: {DAG.getConstant(Val: Intrinsic::wasm_bitmask, DL, VT: MVT::i32), LHS}); |
| 3478 | } |
| 3479 | |
| 3480 | static SDValue performAnyAllCombine(SDNode *N, SelectionDAG &DAG) { |
| 3481 | // any_true (setcc <X>, 0, eq) => (not (all_true X)) |
| 3482 | // all_true (setcc <X>, 0, eq) => (not (any_true X)) |
| 3483 | // any_true (setcc <X>, 0, ne) => (any_true X) |
| 3484 | // all_true (setcc <X>, 0, ne) => (all_true X) |
| 3485 | assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN); |
| 3486 | using namespace llvm::SDPatternMatch; |
| 3487 | |
| 3488 | SDValue LHS; |
| 3489 | if (N->getNumOperands() < 2 || |
| 3490 | !sd_match(N: N->getOperand(Num: 1), |
| 3491 | P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(), CC: m_CondCode()))) |
| 3492 | return SDValue(); |
| 3493 | EVT LT = LHS.getValueType(); |
| 3494 | if (LT.getScalarSizeInBits() > 128 / LT.getVectorNumElements()) |
| 3495 | return SDValue(); |
| 3496 | |
| 3497 | auto CombineSetCC = [&N, &DAG](Intrinsic::WASMIntrinsics InPre, |
| 3498 | ISD::CondCode SetType, |
| 3499 | Intrinsic::WASMIntrinsics InPost) { |
| 3500 | if (N->getConstantOperandVal(Num: 0) != InPre) |
| 3501 | return SDValue(); |
| 3502 | |
| 3503 | SDValue LHS; |
| 3504 | if (!sd_match(N: N->getOperand(Num: 1), P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(), |
| 3505 | CC: m_SpecificCondCode(CC: SetType)))) |
| 3506 | return SDValue(); |
| 3507 | |
| 3508 | SDLoc DL(N); |
| 3509 | SDValue Ret = DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3510 | Ops: {DAG.getConstant(Val: InPost, DL, VT: MVT::i32), LHS}); |
| 3511 | if (SetType == ISD::SETEQ) |
| 3512 | Ret = DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Ret, |
| 3513 | N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32)); |
| 3514 | return DAG.getZExtOrTrunc(Op: Ret, DL, VT: N->getValueType(ResNo: 0)); |
| 3515 | }; |
| 3516 | |
| 3517 | if (SDValue AnyTrueEQ = CombineSetCC(Intrinsic::wasm_anytrue, ISD::SETEQ, |
| 3518 | Intrinsic::wasm_alltrue)) |
| 3519 | return AnyTrueEQ; |
| 3520 | if (SDValue AllTrueEQ = CombineSetCC(Intrinsic::wasm_alltrue, ISD::SETEQ, |
| 3521 | Intrinsic::wasm_anytrue)) |
| 3522 | return AllTrueEQ; |
| 3523 | if (SDValue AnyTrueNE = CombineSetCC(Intrinsic::wasm_anytrue, ISD::SETNE, |
| 3524 | Intrinsic::wasm_anytrue)) |
| 3525 | return AnyTrueNE; |
| 3526 | if (SDValue AllTrueNE = CombineSetCC(Intrinsic::wasm_alltrue, ISD::SETNE, |
| 3527 | Intrinsic::wasm_alltrue)) |
| 3528 | return AllTrueNE; |
| 3529 | |
| 3530 | return SDValue(); |
| 3531 | } |
| 3532 | |
| 3533 | struct MaskReduceInfo { |
| 3534 | Intrinsic::ID IID; |
| 3535 | unsigned WideCombineOpcode; |
| 3536 | bool Invert; |
| 3537 | }; |
| 3538 | |
| 3539 | static SDValue combineSmallMaskReduction(SDNode *N, EVT FromVT, |
| 3540 | unsigned NumElts, |
| 3541 | const MaskReduceInfo &Info, |
| 3542 | SelectionDAG &DAG) { |
| 3543 | EVT VecVT = FromVT.changeVectorElementType(Context&: *DAG.getContext(), |
| 3544 | EltVT: MVT::getIntegerVT(BitWidth: 128 / NumElts)); |
| 3545 | assert(VecVT.getSizeInBits() == 128 && |
| 3546 | "mask reduction should be widened to a 128-bit vector" ); |
| 3547 | |
| 3548 | SDLoc DL(N); |
| 3549 | SDValue Mask = N->getOperand(Num: 0)->getOperand(Num: 0); |
| 3550 | SDValue Ret = DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3551 | Ops: {DAG.getConstant(Val: Info.IID, DL, VT: MVT::i32), |
| 3552 | DAG.getSExtOrTrunc(Op: Mask, DL, VT: VecVT)}); |
| 3553 | if (Info.Invert) |
| 3554 | Ret = DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Ret, |
| 3555 | N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32)); |
| 3556 | return DAG.getZExtOrTrunc(Op: Ret, DL, VT: N->getValueType(ResNo: 0)); |
| 3557 | } |
| 3558 | |
| 3559 | static SDValue combineWideMaskReduction(SDNode *N, SDValue Mask, EVT MaskVT, |
| 3560 | unsigned NumElts, |
| 3561 | const MaskReduceInfo &Info, |
| 3562 | SelectionDAG &DAG) { |
| 3563 | assert((NumElts == 32 || NumElts == 64) && |
| 3564 | "combineWideMaskReduction is only for wide masks" ); |
| 3565 | assert(MaskVT.isFixedLengthVector() && |
| 3566 | MaskVT.getVectorElementType() == MVT::i1); |
| 3567 | SDLoc DL(N); |
| 3568 | unsigned ChunkElts = 16; |
| 3569 | EVT ChunkMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 3570 | EC: ElementCount::getFixed(MinVal: ChunkElts)); |
| 3571 | EVT LegalVecVT = ChunkMaskVT.changeVectorElementType( |
| 3572 | Context&: *DAG.getContext(), EltVT: MVT::getIntegerVT(BitWidth: 128 / ChunkElts)); |
| 3573 | |
| 3574 | SmallVector<SDValue, 4> ChunkResults; |
| 3575 | // Split the wide mask into v16i1 chunks and reduce each chunk separately. |
| 3576 | // For example: |
| 3577 | // v32i1: [0..15] [16..31] |
| 3578 | // | | |
| 3579 | // v v |
| 3580 | // chunk0 chunk1 |
| 3581 | // |
| 3582 | // v64i1: [0..15] [16..31] [32..47] [48..63] |
| 3583 | // | | | | |
| 3584 | // v v v v |
| 3585 | // chunk0 chunk1 chunk2 chunk3 |
| 3586 | // |
| 3587 | // each chunk: |
| 3588 | // v16i1 -> v16i8 -> wasm_anytrue/alltrue -> i32 0/1 |
| 3589 | for (unsigned I = 0; I < NumElts; I += ChunkElts) { |
| 3590 | SDValue ChunkMask = DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: ChunkMaskVT, |
| 3591 | N1: Mask, N2: DAG.getVectorIdxConstant(Val: I, DL)); |
| 3592 | SDValue LegalMask = DAG.getSExtOrTrunc(Op: ChunkMask, DL, VT: LegalVecVT); |
| 3593 | SDValue Reduced = |
| 3594 | DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3595 | N1: DAG.getConstant(Val: Info.IID, DL, VT: MVT::i32), N2: LegalMask); |
| 3596 | ChunkResults.push_back(Elt: Reduced); |
| 3597 | } |
| 3598 | |
| 3599 | SDValue Acc = ChunkResults[0]; |
| 3600 | for (unsigned I = 1; I < ChunkResults.size(); ++I) |
| 3601 | Acc = |
| 3602 | DAG.getNode(Opcode: Info.WideCombineOpcode, DL, VT: MVT::i32, N1: Acc, N2: ChunkResults[I]); |
| 3603 | |
| 3604 | if (Info.Invert) |
| 3605 | Acc = DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Acc, |
| 3606 | N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32)); |
| 3607 | |
| 3608 | return DAG.getZExtOrTrunc(Op: Acc, DL, VT: N->getValueType(ResNo: 0)); |
| 3609 | } |
| 3610 | |
| 3611 | static std::optional<MaskReduceInfo> classifyMaskReduction(SDNode *N) { |
| 3612 | auto *C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1)); |
| 3613 | if (!C) |
| 3614 | return std::nullopt; |
| 3615 | |
| 3616 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get(); |
| 3617 | |
| 3618 | // setcc (bitcast mask), 0, ne -> any_true(mask) |
| 3619 | if (C->isZero() && CC == ISD::SETNE) |
| 3620 | return MaskReduceInfo{.IID: Intrinsic::wasm_anytrue, .WideCombineOpcode: ISD::OR, .Invert: false}; |
| 3621 | |
| 3622 | // setcc (bitcast mask), 0, eq -> !any_true(mask) |
| 3623 | if (C->isZero() && CC == ISD::SETEQ) |
| 3624 | return MaskReduceInfo{.IID: Intrinsic::wasm_anytrue, .WideCombineOpcode: ISD::OR, .Invert: true}; |
| 3625 | |
| 3626 | // setcc (bitcast mask), -1, eq -> all_true(mask) |
| 3627 | if (C->isAllOnes() && CC == ISD::SETEQ) |
| 3628 | return MaskReduceInfo{.IID: Intrinsic::wasm_alltrue, .WideCombineOpcode: ISD::AND, .Invert: false}; |
| 3629 | |
| 3630 | // setcc (bitcast mask), -1, ne -> !all_true(mask) |
| 3631 | if (C->isAllOnes() && CC == ISD::SETNE) |
| 3632 | return MaskReduceInfo{.IID: Intrinsic::wasm_alltrue, .WideCombineOpcode: ISD::AND, .Invert: true}; |
| 3633 | |
| 3634 | return std::nullopt; |
| 3635 | } |
| 3636 | |
| 3637 | /// Try to convert a i128 comparison to a v16i8 comparison before type |
| 3638 | /// legalization splits it up into chunks |
| 3639 | static SDValue |
| 3640 | combineVectorSizedSetCCEquality(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, |
| 3641 | const WebAssemblySubtarget *Subtarget) { |
| 3642 | |
| 3643 | SDLoc DL(N); |
| 3644 | SDValue X = N->getOperand(Num: 0); |
| 3645 | SDValue Y = N->getOperand(Num: 1); |
| 3646 | EVT VT = N->getValueType(ResNo: 0); |
| 3647 | EVT OpVT = X.getValueType(); |
| 3648 | |
| 3649 | SelectionDAG &DAG = DCI.DAG; |
| 3650 | if (DCI.DAG.getMachineFunction().getFunction().hasFnAttribute( |
| 3651 | Kind: Attribute::NoImplicitFloat)) |
| 3652 | return SDValue(); |
| 3653 | |
| 3654 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get(); |
| 3655 | // We're looking for an oversized integer equality comparison with SIMD |
| 3656 | if (!OpVT.isScalarInteger() || !OpVT.isByteSized() || OpVT != MVT::i128 || |
| 3657 | !Subtarget->hasSIMD128() || !isIntEqualitySetCC(Code: CC)) |
| 3658 | return SDValue(); |
| 3659 | |
| 3660 | // Don't perform this combine if constructing the vector will be expensive. |
| 3661 | auto IsVectorBitCastCheap = [](SDValue X) { |
| 3662 | X = peekThroughBitcasts(V: X); |
| 3663 | return isa<ConstantSDNode>(Val: X) || X.getOpcode() == ISD::LOAD; |
| 3664 | }; |
| 3665 | |
| 3666 | if (!IsVectorBitCastCheap(X) || !IsVectorBitCastCheap(Y)) |
| 3667 | return SDValue(); |
| 3668 | |
| 3669 | SDValue VecX = DAG.getBitcast(VT: MVT::v16i8, V: X); |
| 3670 | SDValue VecY = DAG.getBitcast(VT: MVT::v16i8, V: Y); |
| 3671 | SDValue Cmp = DAG.getSetCC(DL, VT: MVT::v16i8, LHS: VecX, RHS: VecY, Cond: CC); |
| 3672 | |
| 3673 | SDValue Intr = |
| 3674 | DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32, |
| 3675 | Ops: {DAG.getConstant(Val: CC == ISD::SETEQ ? Intrinsic::wasm_alltrue |
| 3676 | : Intrinsic::wasm_anytrue, |
| 3677 | DL, VT: MVT::i32), |
| 3678 | Cmp}); |
| 3679 | |
| 3680 | return DAG.getSetCC(DL, VT, LHS: Intr, RHS: DAG.getConstant(Val: 0, DL, VT: MVT::i32), |
| 3681 | Cond: ISD::SETNE); |
| 3682 | } |
| 3683 | |
| 3684 | static SDValue performSETCCCombine(SDNode *N, |
| 3685 | TargetLowering::DAGCombinerInfo &DCI, |
| 3686 | const WebAssemblySubtarget *Subtarget) { |
| 3687 | if (!DCI.isBeforeLegalize()) |
| 3688 | return SDValue(); |
| 3689 | |
| 3690 | EVT VT = N->getValueType(ResNo: 0); |
| 3691 | if (!VT.isScalarInteger()) |
| 3692 | return SDValue(); |
| 3693 | |
| 3694 | if (SDValue V = combineVectorSizedSetCCEquality(N, DCI, Subtarget)) |
| 3695 | return V; |
| 3696 | |
| 3697 | SDValue LHS = N->getOperand(Num: 0); |
| 3698 | if (LHS->getOpcode() != ISD::BITCAST) |
| 3699 | return SDValue(); |
| 3700 | |
| 3701 | EVT FromVT = LHS->getOperand(Num: 0).getValueType(); |
| 3702 | if (!FromVT.isFixedLengthVectorOf(EltVT: MVT::i1)) |
| 3703 | return SDValue(); |
| 3704 | |
| 3705 | unsigned NumElts = FromVT.getVectorNumElements(); |
| 3706 | auto Info = classifyMaskReduction(N); |
| 3707 | if (!Info) |
| 3708 | return SDValue(); |
| 3709 | |
| 3710 | auto &DAG = DCI.DAG; |
| 3711 | if (NumElts == 2 || NumElts == 4 || NumElts == 8 || NumElts == 16) |
| 3712 | return combineSmallMaskReduction(N, FromVT, NumElts, Info: *Info, DAG); |
| 3713 | |
| 3714 | if (NumElts == 32 || NumElts == 64) |
| 3715 | return combineWideMaskReduction(N, Mask: LHS.getOperand(i: 0), MaskVT: FromVT, NumElts, |
| 3716 | Info: *Info, DAG); |
| 3717 | |
| 3718 | return SDValue(); |
| 3719 | } |
| 3720 | |
| 3721 | static SDValue TryWideExtMulCombine(SDNode *N, SelectionDAG &DAG) { |
| 3722 | EVT VT = N->getValueType(ResNo: 0); |
| 3723 | if (VT != MVT::v8i32 && VT != MVT::v16i32) |
| 3724 | return SDValue(); |
| 3725 | |
| 3726 | // Mul with extending inputs. |
| 3727 | SDValue LHS = N->getOperand(Num: 0); |
| 3728 | SDValue RHS = N->getOperand(Num: 1); |
| 3729 | if (LHS.getOpcode() != RHS.getOpcode()) |
| 3730 | return SDValue(); |
| 3731 | |
| 3732 | if (LHS.getOpcode() != ISD::SIGN_EXTEND && |
| 3733 | LHS.getOpcode() != ISD::ZERO_EXTEND) |
| 3734 | return SDValue(); |
| 3735 | |
| 3736 | if (LHS->getOperand(Num: 0).getValueType() != RHS->getOperand(Num: 0).getValueType()) |
| 3737 | return SDValue(); |
| 3738 | |
| 3739 | EVT FromVT = LHS->getOperand(Num: 0).getValueType(); |
| 3740 | EVT EltTy = FromVT.getVectorElementType(); |
| 3741 | if (EltTy != MVT::i8) |
| 3742 | return SDValue(); |
| 3743 | |
| 3744 | // For an input DAG that looks like this |
| 3745 | // %a = input_type |
| 3746 | // %b = input_type |
| 3747 | // %lhs = extend %a to output_type |
| 3748 | // %rhs = extend %b to output_type |
| 3749 | // %mul = mul %lhs, %rhs |
| 3750 | |
| 3751 | // input_type | output_type | instructions |
| 3752 | // v16i8 | v16i32 | %low = i16x8.extmul_low_i8x16_ %a, %b |
| 3753 | // | | %high = i16x8.extmul_high_i8x16_, %a, %b |
| 3754 | // | | %low_low = i32x4.ext_low_i16x8_ %low |
| 3755 | // | | %low_high = i32x4.ext_high_i16x8_ %low |
| 3756 | // | | %high_low = i32x4.ext_low_i16x8_ %high |
| 3757 | // | | %high_high = i32x4.ext_high_i16x8_ %high |
| 3758 | // | | %res = concat_vector(...) |
| 3759 | // v8i8 | v8i32 | %low = i16x8.extmul_low_i8x16_ %a, %b |
| 3760 | // | | %low_low = i32x4.ext_low_i16x8_ %low |
| 3761 | // | | %low_high = i32x4.ext_high_i16x8_ %low |
| 3762 | // | | %res = concat_vector(%low_low, %low_high) |
| 3763 | |
| 3764 | SDLoc DL(N); |
| 3765 | unsigned NumElts = VT.getVectorNumElements(); |
| 3766 | SDValue ExtendInLHS = LHS->getOperand(Num: 0); |
| 3767 | SDValue ExtendInRHS = RHS->getOperand(Num: 0); |
| 3768 | bool IsSigned = LHS->getOpcode() == ISD::SIGN_EXTEND; |
| 3769 | unsigned ExtendLowOpc = |
| 3770 | IsSigned ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U; |
| 3771 | unsigned ExtendHighOpc = |
| 3772 | IsSigned ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U; |
| 3773 | |
| 3774 | auto GetExtendLow = [&DAG, &DL, &ExtendLowOpc](EVT VT, SDValue Op) { |
| 3775 | return DAG.getNode(Opcode: ExtendLowOpc, DL, VT, Operand: Op); |
| 3776 | }; |
| 3777 | auto GetExtendHigh = [&DAG, &DL, &ExtendHighOpc](EVT VT, SDValue Op) { |
| 3778 | return DAG.getNode(Opcode: ExtendHighOpc, DL, VT, Operand: Op); |
| 3779 | }; |
| 3780 | |
| 3781 | if (NumElts == 16) { |
| 3782 | SDValue LowLHS = GetExtendLow(MVT::v8i16, ExtendInLHS); |
| 3783 | SDValue LowRHS = GetExtendLow(MVT::v8i16, ExtendInRHS); |
| 3784 | SDValue MulLow = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: LowLHS, N2: LowRHS); |
| 3785 | SDValue HighLHS = GetExtendHigh(MVT::v8i16, ExtendInLHS); |
| 3786 | SDValue HighRHS = GetExtendHigh(MVT::v8i16, ExtendInRHS); |
| 3787 | SDValue MulHigh = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: HighLHS, N2: HighRHS); |
| 3788 | SDValue SubVectors[] = { |
| 3789 | GetExtendLow(MVT::v4i32, MulLow), |
| 3790 | GetExtendHigh(MVT::v4i32, MulLow), |
| 3791 | GetExtendLow(MVT::v4i32, MulHigh), |
| 3792 | GetExtendHigh(MVT::v4i32, MulHigh), |
| 3793 | }; |
| 3794 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, Ops: SubVectors); |
| 3795 | } else { |
| 3796 | assert(NumElts == 8); |
| 3797 | SDValue LowLHS = DAG.getNode(Opcode: LHS->getOpcode(), DL, VT: MVT::v8i16, Operand: ExtendInLHS); |
| 3798 | SDValue LowRHS = DAG.getNode(Opcode: RHS->getOpcode(), DL, VT: MVT::v8i16, Operand: ExtendInRHS); |
| 3799 | SDValue MulLow = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: LowLHS, N2: LowRHS); |
| 3800 | SDValue Lo = GetExtendLow(MVT::v4i32, MulLow); |
| 3801 | SDValue Hi = GetExtendHigh(MVT::v4i32, MulLow); |
| 3802 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, N1: Lo, N2: Hi); |
| 3803 | } |
| 3804 | return SDValue(); |
| 3805 | } |
| 3806 | |
| 3807 | static SDValue performMulCombine(SDNode *N, |
| 3808 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3809 | assert(N->getOpcode() == ISD::MUL); |
| 3810 | EVT VT = N->getValueType(ResNo: 0); |
| 3811 | if (!VT.isVector()) |
| 3812 | return SDValue(); |
| 3813 | |
| 3814 | if (auto Res = TryWideExtMulCombine(N, DAG&: DCI.DAG)) |
| 3815 | return Res; |
| 3816 | |
| 3817 | // We don't natively support v16i8 or v8i8 mul, but we do support v8i16. So, |
| 3818 | // extend them to v8i16. |
| 3819 | if (VT != MVT::v8i8 && VT != MVT::v16i8) |
| 3820 | return SDValue(); |
| 3821 | |
| 3822 | SDLoc DL(N); |
| 3823 | SelectionDAG &DAG = DCI.DAG; |
| 3824 | SDValue LHS = N->getOperand(Num: 0); |
| 3825 | SDValue RHS = N->getOperand(Num: 1); |
| 3826 | EVT MulVT = MVT::v8i16; |
| 3827 | |
| 3828 | if (VT == MVT::v8i8) { |
| 3829 | SDValue PromotedLHS = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: MVT::v16i8, N1: LHS, |
| 3830 | N2: DAG.getUNDEF(VT: MVT::v8i8)); |
| 3831 | SDValue PromotedRHS = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: MVT::v16i8, N1: RHS, |
| 3832 | N2: DAG.getUNDEF(VT: MVT::v8i8)); |
| 3833 | SDValue LowLHS = |
| 3834 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: PromotedLHS); |
| 3835 | SDValue LowRHS = |
| 3836 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: PromotedRHS); |
| 3837 | SDValue MulLow = DAG.getBitcast( |
| 3838 | VT: MVT::v16i8, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: LowLHS, N2: LowRHS)); |
| 3839 | // Take the low byte of each lane. |
| 3840 | SDValue Shuffle = DAG.getVectorShuffle( |
| 3841 | VT: MVT::v16i8, dl: DL, N1: MulLow, N2: DAG.getUNDEF(VT: MVT::v16i8), |
| 3842 | Mask: {0, 2, 4, 6, 8, 10, 12, 14, -1, -1, -1, -1, -1, -1, -1, -1}); |
| 3843 | return extractSubVector(Vec: Shuffle, IdxVal: 0, DAG, DL, VectorWidth: 64); |
| 3844 | } else { |
| 3845 | assert(VT == MVT::v16i8 && "Expected v16i8" ); |
| 3846 | SDValue LowLHS = DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: LHS); |
| 3847 | SDValue LowRHS = DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: RHS); |
| 3848 | SDValue HighLHS = |
| 3849 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_HIGH_U, DL, VT: MulVT, Operand: LHS); |
| 3850 | SDValue HighRHS = |
| 3851 | DAG.getNode(Opcode: WebAssemblyISD::EXTEND_HIGH_U, DL, VT: MulVT, Operand: RHS); |
| 3852 | |
| 3853 | SDValue MulLow = |
| 3854 | DAG.getBitcast(VT, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: LowLHS, N2: LowRHS)); |
| 3855 | SDValue MulHigh = |
| 3856 | DAG.getBitcast(VT, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: HighLHS, N2: HighRHS)); |
| 3857 | |
| 3858 | // Take the low byte of each lane. |
| 3859 | return DAG.getVectorShuffle( |
| 3860 | VT, dl: DL, N1: MulLow, N2: MulHigh, |
| 3861 | Mask: {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30}); |
| 3862 | } |
| 3863 | } |
| 3864 | |
| 3865 | SDValue DoubleVectorWidth(SDValue In, unsigned RequiredNumElems, |
| 3866 | SelectionDAG &DAG) { |
| 3867 | SDLoc DL(In); |
| 3868 | LLVMContext &Ctx = *DAG.getContext(); |
| 3869 | EVT InVT = In.getValueType(); |
| 3870 | unsigned NumElems = InVT.getVectorNumElements() * 2; |
| 3871 | EVT OutVT = EVT::getVectorVT(Context&: Ctx, VT: InVT.getVectorElementType(), NumElements: NumElems); |
| 3872 | SDValue Concat = |
| 3873 | DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: OutVT, N1: In, N2: DAG.getPOISON(VT: InVT)); |
| 3874 | if (NumElems < RequiredNumElems) { |
| 3875 | return DoubleVectorWidth(In: Concat, RequiredNumElems, DAG); |
| 3876 | } |
| 3877 | return Concat; |
| 3878 | } |
| 3879 | |
| 3880 | SDValue performConvertFPCombine(SDNode *N, SelectionDAG &DAG) { |
| 3881 | EVT OutVT = N->getValueType(ResNo: 0); |
| 3882 | if (!OutVT.isVector()) |
| 3883 | return SDValue(); |
| 3884 | |
| 3885 | EVT OutElTy = OutVT.getVectorElementType(); |
| 3886 | if (OutElTy != MVT::i8 && OutElTy != MVT::i16) |
| 3887 | return SDValue(); |
| 3888 | |
| 3889 | unsigned NumElems = OutVT.getVectorNumElements(); |
| 3890 | if (!isPowerOf2_32(Value: NumElems)) |
| 3891 | return SDValue(); |
| 3892 | |
| 3893 | EVT FPVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
| 3894 | if (FPVT.getVectorElementType() != MVT::f32) |
| 3895 | return SDValue(); |
| 3896 | |
| 3897 | SDLoc DL(N); |
| 3898 | |
| 3899 | // First, convert to i32. |
| 3900 | LLVMContext &Ctx = *DAG.getContext(); |
| 3901 | EVT IntVT = EVT::getVectorVT(Context&: Ctx, VT: MVT::i32, NumElements: NumElems); |
| 3902 | SDValue ToInt = DAG.getNode(Opcode: N->getOpcode(), DL, VT: IntVT, Operand: N->getOperand(Num: 0)); |
| 3903 | APInt Mask = APInt::getLowBitsSet(numBits: IntVT.getScalarSizeInBits(), |
| 3904 | loBitsSet: OutVT.getScalarSizeInBits()); |
| 3905 | // Mask out the top MSBs. |
| 3906 | SDValue Masked = |
| 3907 | DAG.getNode(Opcode: ISD::AND, DL, VT: IntVT, N1: ToInt, N2: DAG.getConstant(Val: Mask, DL, VT: IntVT)); |
| 3908 | |
| 3909 | if (OutVT.getSizeInBits() < 128) { |
| 3910 | // Create a wide enough vector that we can use narrow. |
| 3911 | EVT NarrowedVT = OutElTy == MVT::i8 ? MVT::v16i8 : MVT::v8i16; |
| 3912 | unsigned NumRequiredElems = NarrowedVT.getVectorNumElements(); |
| 3913 | SDValue WideVector = DoubleVectorWidth(In: Masked, RequiredNumElems: NumRequiredElems, DAG); |
| 3914 | SDValue Trunc = truncateVectorWithNARROW(DstVT: NarrowedVT, In: WideVector, DL, DAG); |
| 3915 | return DAG.getBitcast( |
| 3916 | VT: OutVT, V: extractSubVector(Vec: Trunc, IdxVal: 0, DAG, DL, VectorWidth: OutVT.getSizeInBits())); |
| 3917 | } else { |
| 3918 | return truncateVectorWithNARROW(DstVT: OutVT, In: Masked, DL, DAG); |
| 3919 | } |
| 3920 | return SDValue(); |
| 3921 | } |
| 3922 | |
| 3923 | // Wide vector shift operations such as v8i32 with sign-extended |
| 3924 | // operands cause Type Legalizer crashes because the target-specific |
| 3925 | // extension nodes cannot be directly mapped to the 256-bit size. |
| 3926 | // |
| 3927 | // To resolve the crash and optimize performance, we intercept the |
| 3928 | // illegal v8i32 shift in DAGCombine. We convert the shift amounts |
| 3929 | // into multipliers and manually split the vector into two v4i32 halves. |
| 3930 | // |
| 3931 | // Before: t1: v8i32 = shl (sign_extend v8i16), const_vec |
| 3932 | // After : t2: v4i32 = mul (ext_low_s v8i16), (ext_low_s narrow_vec) |
| 3933 | // t3: v4i32 = mul (ext_high_s v8i16), (ext_high_s narrow_vec) |
| 3934 | // t4: v8i32 = concat_vectors t2, t3 |
| 3935 | static SDValue performShiftCombine(SDNode *N, |
| 3936 | TargetLowering::DAGCombinerInfo &DCI) { |
| 3937 | SelectionDAG &DAG = DCI.DAG; |
| 3938 | assert(N->getOpcode() == ISD::SHL); |
| 3939 | EVT VT = N->getValueType(ResNo: 0); |
| 3940 | if (VT != MVT::v8i32) |
| 3941 | return SDValue(); |
| 3942 | |
| 3943 | SDValue LHS = N->getOperand(Num: 0); |
| 3944 | SDValue RHS = N->getOperand(Num: 1); |
| 3945 | unsigned ExtOpc = LHS.getOpcode(); |
| 3946 | if (ExtOpc != ISD::SIGN_EXTEND && ExtOpc != ISD::ZERO_EXTEND) |
| 3947 | return SDValue(); |
| 3948 | |
| 3949 | if (RHS.getOpcode() != ISD::BUILD_VECTOR) |
| 3950 | return SDValue(); |
| 3951 | |
| 3952 | SDLoc DL(N); |
| 3953 | SDValue ExtendIn = LHS.getOperand(i: 0); |
| 3954 | EVT FromVT = ExtendIn.getValueType(); |
| 3955 | if (FromVT != MVT::v8i16) |
| 3956 | return SDValue(); |
| 3957 | |
| 3958 | unsigned NumElts = VT.getVectorNumElements(); |
| 3959 | unsigned BitWidth = FromVT.getScalarSizeInBits(); |
| 3960 | bool IsSigned = (ExtOpc == ISD::SIGN_EXTEND); |
| 3961 | unsigned MaxValidShift = IsSigned ? (BitWidth - 1) : BitWidth; |
| 3962 | SmallVector<SDValue, 16> MulConsts; |
| 3963 | for (unsigned I = 0; I < NumElts; ++I) { |
| 3964 | auto *C = dyn_cast<ConstantSDNode>(Val: RHS.getOperand(i: I)); |
| 3965 | if (!C) |
| 3966 | return SDValue(); |
| 3967 | |
| 3968 | const APInt &ShiftAmt = C->getAPIntValue(); |
| 3969 | if (ShiftAmt.uge(RHS: MaxValidShift)) |
| 3970 | return SDValue(); |
| 3971 | |
| 3972 | APInt MulAmt = APInt::getOneBitSet(numBits: BitWidth, BitNo: ShiftAmt.getZExtValue()); |
| 3973 | MulConsts.push_back(Elt: DAG.getConstant(Val: MulAmt, DL, VT: FromVT.getScalarType(), |
| 3974 | /*isTarget=*/false, /*isOpaque=*/true)); |
| 3975 | } |
| 3976 | |
| 3977 | SDValue NarrowConst = DAG.getBuildVector(VT: FromVT, DL, Ops: MulConsts); |
| 3978 | unsigned ExtLowOpc = |
| 3979 | IsSigned ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U; |
| 3980 | unsigned ExtHighOpc = |
| 3981 | IsSigned ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U; |
| 3982 | |
| 3983 | EVT HalfVT = MVT::v4i32; |
| 3984 | SDValue LHSLo = DAG.getNode(Opcode: ExtLowOpc, DL, VT: HalfVT, Operand: ExtendIn); |
| 3985 | SDValue LHSHi = DAG.getNode(Opcode: ExtHighOpc, DL, VT: HalfVT, Operand: ExtendIn); |
| 3986 | SDValue RHSLo = DAG.getNode(Opcode: ExtLowOpc, DL, VT: HalfVT, Operand: NarrowConst); |
| 3987 | SDValue RHSHi = DAG.getNode(Opcode: ExtHighOpc, DL, VT: HalfVT, Operand: NarrowConst); |
| 3988 | SDValue MulLo = DAG.getNode(Opcode: ISD::MUL, DL, VT: HalfVT, N1: LHSLo, N2: RHSLo); |
| 3989 | SDValue MulHi = DAG.getNode(Opcode: ISD::MUL, DL, VT: HalfVT, N1: LHSHi, N2: RHSHi); |
| 3990 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, N1: MulLo, N2: MulHi); |
| 3991 | } |
| 3992 | |
| 3993 | static SDValue performMinMaxF128Combine(SDNode *N, SelectionDAG &DAG) { |
| 3994 | if (N->getValueType(ResNo: 0) != MVT::f128) |
| 3995 | return SDValue(); |
| 3996 | |
| 3997 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
| 3998 | switch (N->getOpcode()) { |
| 3999 | // wasi-libc and emscripten do not currently define fminimuml and fmaximuml. |
| 4000 | case ISD::FMINIMUM: |
| 4001 | case ISD::FMAXIMUM: |
| 4002 | return TLI.expandFMINIMUM_FMAXIMUM(N, DAG); |
| 4003 | |
| 4004 | // wasi-libc and emscripten do not currently define fminimum_numl and |
| 4005 | // fmaximum_numl. |
| 4006 | case ISD::FMINIMUMNUM: |
| 4007 | case ISD::FMAXIMUMNUM: |
| 4008 | return TLI.expandFMINIMUMNUM_FMAXIMUMNUM(N, DAG); |
| 4009 | |
| 4010 | default: |
| 4011 | return SDValue(); |
| 4012 | } |
| 4013 | } |
| 4014 | |
| 4015 | SDValue |
| 4016 | WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N, |
| 4017 | DAGCombinerInfo &DCI) const { |
| 4018 | switch (N->getOpcode()) { |
| 4019 | default: |
| 4020 | return SDValue(); |
| 4021 | case ISD::BITCAST: |
| 4022 | return performBitcastCombine(N, DCI); |
| 4023 | case ISD::SETCC: |
| 4024 | return performSETCCCombine(N, DCI, Subtarget); |
| 4025 | case ISD::VECTOR_SHUFFLE: |
| 4026 | return performVECTOR_SHUFFLECombine(N, DCI); |
| 4027 | case ISD::SIGN_EXTEND: |
| 4028 | case ISD::ZERO_EXTEND: |
| 4029 | return performVectorExtendCombine(N, DCI); |
| 4030 | case ISD::UINT_TO_FP: |
| 4031 | if (auto ExtCombine = performVectorExtendToFPCombine(N, DCI)) |
| 4032 | return ExtCombine; |
| 4033 | return performVectorNonNegToFPCombine(N, DCI); |
| 4034 | case ISD::SINT_TO_FP: |
| 4035 | return performVectorExtendToFPCombine(N, DCI); |
| 4036 | case ISD::FP_TO_SINT_SAT: |
| 4037 | case ISD::FP_TO_UINT_SAT: |
| 4038 | case ISD::FP_ROUND: |
| 4039 | case ISD::CONCAT_VECTORS: |
| 4040 | return performVectorTruncZeroCombine(N, DCI); |
| 4041 | case ISD::FP_TO_SINT: |
| 4042 | case ISD::FP_TO_UINT: |
| 4043 | return performConvertFPCombine(N, DAG&: DCI.DAG); |
| 4044 | case ISD::TRUNCATE: |
| 4045 | return performTruncateCombine(N, DCI); |
| 4046 | case ISD::INTRINSIC_WO_CHAIN: { |
| 4047 | if (SDValue V = performBitmaskCombine(N, DAG&: DCI.DAG)) |
| 4048 | return V; |
| 4049 | return performAnyAllCombine(N, DAG&: DCI.DAG); |
| 4050 | } |
| 4051 | case ISD::MUL: |
| 4052 | return performMulCombine(N, DCI); |
| 4053 | case ISD::SHL: |
| 4054 | return performShiftCombine(N, DCI); |
| 4055 | case ISD::FMINIMUM: |
| 4056 | case ISD::FMAXIMUM: |
| 4057 | case ISD::FMINIMUMNUM: |
| 4058 | case ISD::FMAXIMUMNUM: |
| 4059 | return performMinMaxF128Combine(N, DAG&: DCI.DAG); |
| 4060 | } |
| 4061 | } |
| 4062 | |