| OLD | NEW |
| 1 // Copyright 2016 the V8 project authors. All rights reserved. | 1 // Copyright 2016 the V8 project authors. All rights reserved. |
| 2 // Use of this source code is governed by a BSD-style license that can be | 2 // Use of this source code is governed by a BSD-style license that can be |
| 3 // found in the LICENSE file. | 3 // found in the LICENSE file. |
| 4 | 4 |
| 5 #include "src/compiler/simd-scalar-lowering.h" | 5 #include "src/compiler/simd-scalar-lowering.h" |
| 6 #include "src/compiler/diamond.h" | 6 #include "src/compiler/diamond.h" |
| 7 #include "src/compiler/linkage.h" | 7 #include "src/compiler/linkage.h" |
| 8 #include "src/compiler/node-matchers.h" | 8 #include "src/compiler/node-matchers.h" |
| 9 #include "src/compiler/node-properties.h" | 9 #include "src/compiler/node-properties.h" |
| 10 | 10 |
| (...skipping 62 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 73 | 73 |
| 74 #define FOREACH_INT32X4_OPCODE(V) \ | 74 #define FOREACH_INT32X4_OPCODE(V) \ |
| 75 V(Int32x4Splat) \ | 75 V(Int32x4Splat) \ |
| 76 V(Int32x4ExtractLane) \ | 76 V(Int32x4ExtractLane) \ |
| 77 V(Int32x4ReplaceLane) \ | 77 V(Int32x4ReplaceLane) \ |
| 78 V(Int32x4Neg) \ | 78 V(Int32x4Neg) \ |
| 79 V(Simd128Not) \ | 79 V(Simd128Not) \ |
| 80 V(Int32x4Add) \ | 80 V(Int32x4Add) \ |
| 81 V(Int32x4Sub) \ | 81 V(Int32x4Sub) \ |
| 82 V(Int32x4Mul) \ | 82 V(Int32x4Mul) \ |
| 83 V(Int32x4Min) \ | |
| 84 V(Int32x4Max) \ | |
| 85 V(Uint32x4Min) \ | |
| 86 V(Uint32x4Max) \ | |
| 87 V(Simd128And) \ | 83 V(Simd128And) \ |
| 88 V(Simd128Or) \ | 84 V(Simd128Or) \ |
| 89 V(Simd128Xor) \ | 85 V(Simd128Xor) |
| 90 V(Int32x4FromFloat32x4) \ | |
| 91 V(Uint32x4FromFloat32x4) | |
| 92 | 86 |
| 93 #define FOREACH_FLOAT32X4_OPCODE(V) \ | 87 #define FOREACH_FLOAT32X4_OPCODE(V) \ |
| 94 V(Float32x4Splat) \ | 88 V(Float32x4Splat) \ |
| 95 V(Float32x4ExtractLane) \ | 89 V(Float32x4ExtractLane) \ |
| 96 V(Float32x4ReplaceLane) \ | 90 V(Float32x4ReplaceLane) \ |
| 97 V(Float32x4Abs) \ | 91 V(Float32x4Abs) \ |
| 98 V(Float32x4Neg) \ | 92 V(Float32x4Neg) \ |
| 99 V(Float32x4Add) \ | 93 V(Float32x4Add) \ |
| 100 V(Float32x4Sub) \ | 94 V(Float32x4Sub) \ |
| 101 V(Float32x4Mul) \ | 95 V(Float32x4Mul) \ |
| (...skipping 12 matching lines...) Expand all Loading... |
| 114 case IrOpcode::kCall: { | 108 case IrOpcode::kCall: { |
| 115 replacements_[node->id()].type = SimdType::kInt32; | 109 replacements_[node->id()].type = SimdType::kInt32; |
| 116 break; | 110 break; |
| 117 } | 111 } |
| 118 FOREACH_FLOAT32X4_OPCODE(CASE_STMT) { | 112 FOREACH_FLOAT32X4_OPCODE(CASE_STMT) { |
| 119 replacements_[node->id()].type = SimdType::kFloat32; | 113 replacements_[node->id()].type = SimdType::kFloat32; |
| 120 break; | 114 break; |
| 121 } | 115 } |
| 122 #undef CASE_STMT | 116 #undef CASE_STMT |
| 123 default: { | 117 default: { |
| 124 switch (output->opcode()) { | 118 if (output->opcode() == IrOpcode::kFloat32x4FromInt32x4 || |
| 125 case IrOpcode::kFloat32x4FromInt32x4: | 119 output->opcode() == IrOpcode::kFloat32x4FromUint32x4) { |
| 126 case IrOpcode::kFloat32x4FromUint32x4: { | 120 replacements_[node->id()].type = SimdType::kInt32; |
| 127 replacements_[node->id()].type = SimdType::kInt32; | 121 } else { |
| 128 break; | 122 replacements_[node->id()].type = replacements_[output->id()].type; |
| 129 } | |
| 130 case IrOpcode::kInt32x4FromFloat32x4: | |
| 131 case IrOpcode::kUint32x4FromFloat32x4: { | |
| 132 replacements_[node->id()].type = SimdType::kFloat32; | |
| 133 break; | |
| 134 } | |
| 135 default: { | |
| 136 replacements_[node->id()].type = replacements_[output->id()].type; | |
| 137 } | |
| 138 } | 123 } |
| 139 } | 124 } |
| 140 } | 125 } |
| 141 } | 126 } |
| 142 | 127 |
| 143 static int GetParameterIndexAfterLowering( | 128 static int GetParameterIndexAfterLowering( |
| 144 Signature<MachineRepresentation>* signature, int old_index) { | 129 Signature<MachineRepresentation>* signature, int old_index) { |
| 145 // In function calls, the simd128 types are passed as 4 Int32 types. The | 130 // In function calls, the simd128 types are passed as 4 Int32 types. The |
| 146 // parameters are typecast to the types as needed for various operations. | 131 // parameters are typecast to the types as needed for various operations. |
| 147 int result = old_index; | 132 int result = old_index; |
| (...skipping 124 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 272 const Operator* op) { | 257 const Operator* op) { |
| 273 DCHECK(node->InputCount() == 1); | 258 DCHECK(node->InputCount() == 1); |
| 274 Node** rep = GetReplacementsWithType(node->InputAt(0), input_rep_type); | 259 Node** rep = GetReplacementsWithType(node->InputAt(0), input_rep_type); |
| 275 Node* rep_node[kMaxLanes]; | 260 Node* rep_node[kMaxLanes]; |
| 276 for (int i = 0; i < kMaxLanes; ++i) { | 261 for (int i = 0; i < kMaxLanes; ++i) { |
| 277 rep_node[i] = graph()->NewNode(op, rep[i]); | 262 rep_node[i] = graph()->NewNode(op, rep[i]); |
| 278 } | 263 } |
| 279 ReplaceNode(node, rep_node); | 264 ReplaceNode(node, rep_node); |
| 280 } | 265 } |
| 281 | 266 |
| 282 void SimdScalarLowering::LowerIntMinMax(Node* node, const Operator* op, | |
| 283 bool is_max) { | |
| 284 DCHECK(node->InputCount() == 2); | |
| 285 Node** rep_left = GetReplacementsWithType(node->InputAt(0), SimdType::kInt32); | |
| 286 Node** rep_right = | |
| 287 GetReplacementsWithType(node->InputAt(1), SimdType::kInt32); | |
| 288 Node* rep_node[kMaxLanes]; | |
| 289 for (int i = 0; i < kMaxLanes; ++i) { | |
| 290 Diamond d(graph(), common(), | |
| 291 graph()->NewNode(op, rep_left[i], rep_right[i])); | |
| 292 if (is_max) { | |
| 293 rep_node[i] = | |
| 294 d.Phi(MachineRepresentation::kWord32, rep_right[i], rep_left[i]); | |
| 295 } else { | |
| 296 rep_node[i] = | |
| 297 d.Phi(MachineRepresentation::kWord32, rep_left[i], rep_right[i]); | |
| 298 } | |
| 299 } | |
| 300 ReplaceNode(node, rep_node); | |
| 301 } | |
| 302 | |
| 303 void SimdScalarLowering::LowerConvertFromFloat(Node* node, bool is_signed) { | |
| 304 DCHECK(node->InputCount() == 1); | |
| 305 Node** rep = GetReplacementsWithType(node->InputAt(0), SimdType::kFloat32); | |
| 306 Node* rep_node[kMaxLanes]; | |
| 307 Node* double_zero = graph()->NewNode(common()->Float64Constant(0.0)); | |
| 308 Node* min = graph()->NewNode( | |
| 309 common()->Float64Constant(static_cast<double>(is_signed ? kMinInt : 0))); | |
| 310 Node* max = graph()->NewNode(common()->Float64Constant( | |
| 311 static_cast<double>(is_signed ? kMaxInt : 0xffffffffu))); | |
| 312 for (int i = 0; i < kMaxLanes; ++i) { | |
| 313 Node* double_rep = | |
| 314 graph()->NewNode(machine()->ChangeFloat32ToFloat64(), rep[i]); | |
| 315 Diamond nan_d(graph(), common(), graph()->NewNode(machine()->Float64Equal(), | |
| 316 double_rep, double_rep)); | |
| 317 Node* temp = | |
| 318 nan_d.Phi(MachineRepresentation::kFloat64, double_rep, double_zero); | |
| 319 Diamond min_d(graph(), common(), | |
| 320 graph()->NewNode(machine()->Float64LessThan(), temp, min)); | |
| 321 temp = min_d.Phi(MachineRepresentation::kFloat64, min, temp); | |
| 322 Diamond max_d(graph(), common(), | |
| 323 graph()->NewNode(machine()->Float64LessThan(), max, temp)); | |
| 324 temp = max_d.Phi(MachineRepresentation::kFloat64, max, temp); | |
| 325 Node* trunc = | |
| 326 graph()->NewNode(machine()->Float64RoundTruncate().op(), temp); | |
| 327 if (is_signed) { | |
| 328 rep_node[i] = graph()->NewNode(machine()->ChangeFloat64ToInt32(), trunc); | |
| 329 } else { | |
| 330 rep_node[i] = | |
| 331 graph()->NewNode(machine()->TruncateFloat64ToUint32(), trunc); | |
| 332 } | |
| 333 } | |
| 334 ReplaceNode(node, rep_node); | |
| 335 } | |
| 336 | |
| 337 void SimdScalarLowering::LowerShiftOp(Node* node, const Operator* op) { | |
| 338 static int32_t shift_mask = 0x1f; | |
| 339 DCHECK_EQ(1, node->InputCount()); | |
| 340 int32_t shift_amount = OpParameter<int32_t>(node); | |
| 341 Node* shift_node = | |
| 342 graph()->NewNode(common()->Int32Constant(shift_amount & shift_mask)); | |
| 343 Node** rep = GetReplacementsWithType(node->InputAt(0), SimdType::kInt32); | |
| 344 Node* rep_node[kMaxLanes]; | |
| 345 for (int i = 0; i < kMaxLanes; ++i) { | |
| 346 rep_node[i] = graph()->NewNode(op, rep[i], shift_node); | |
| 347 } | |
| 348 ReplaceNode(node, rep_node); | |
| 349 } | |
| 350 | |
| 351 void SimdScalarLowering::LowerNode(Node* node) { | 267 void SimdScalarLowering::LowerNode(Node* node) { |
| 352 SimdType rep_type = ReplacementType(node); | 268 SimdType rep_type = ReplacementType(node); |
| 353 switch (node->opcode()) { | 269 switch (node->opcode()) { |
| 354 case IrOpcode::kStart: { | 270 case IrOpcode::kStart: { |
| 355 int parameter_count = GetParameterCountAfterLowering(); | 271 int parameter_count = GetParameterCountAfterLowering(); |
| 356 // Only exchange the node if the parameter count actually changed. | 272 // Only exchange the node if the parameter count actually changed. |
| 357 if (parameter_count != static_cast<int>(signature()->parameter_count())) { | 273 if (parameter_count != static_cast<int>(signature()->parameter_count())) { |
| 358 int delta = | 274 int delta = |
| 359 parameter_count - static_cast<int>(signature()->parameter_count()); | 275 parameter_count - static_cast<int>(signature()->parameter_count()); |
| 360 int new_output_count = node->op()->ValueOutputCount() + delta; | 276 int new_output_count = node->op()->ValueOutputCount() + delta; |
| (...skipping 138 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 499 LowerBinaryOp(node, rep_type, machine()->instruction()); \ | 415 LowerBinaryOp(node, rep_type, machine()->instruction()); \ |
| 500 break; \ | 416 break; \ |
| 501 } | 417 } |
| 502 I32X4_BINOP_CASE(kInt32x4Add, Int32Add) | 418 I32X4_BINOP_CASE(kInt32x4Add, Int32Add) |
| 503 I32X4_BINOP_CASE(kInt32x4Sub, Int32Sub) | 419 I32X4_BINOP_CASE(kInt32x4Sub, Int32Sub) |
| 504 I32X4_BINOP_CASE(kInt32x4Mul, Int32Mul) | 420 I32X4_BINOP_CASE(kInt32x4Mul, Int32Mul) |
| 505 I32X4_BINOP_CASE(kSimd128And, Word32And) | 421 I32X4_BINOP_CASE(kSimd128And, Word32And) |
| 506 I32X4_BINOP_CASE(kSimd128Or, Word32Or) | 422 I32X4_BINOP_CASE(kSimd128Or, Word32Or) |
| 507 I32X4_BINOP_CASE(kSimd128Xor, Word32Xor) | 423 I32X4_BINOP_CASE(kSimd128Xor, Word32Xor) |
| 508 #undef I32X4_BINOP_CASE | 424 #undef I32X4_BINOP_CASE |
| 509 case IrOpcode::kInt32x4Max: { | |
| 510 LowerIntMinMax(node, machine()->Int32LessThan(), true); | |
| 511 break; | |
| 512 } | |
| 513 case IrOpcode::kInt32x4Min: { | |
| 514 LowerIntMinMax(node, machine()->Int32LessThan(), false); | |
| 515 break; | |
| 516 } | |
| 517 case IrOpcode::kUint32x4Max: { | |
| 518 LowerIntMinMax(node, machine()->Uint32LessThan(), true); | |
| 519 break; | |
| 520 } | |
| 521 case IrOpcode::kUint32x4Min: { | |
| 522 LowerIntMinMax(node, machine()->Uint32LessThan(), false); | |
| 523 break; | |
| 524 } | |
| 525 case IrOpcode::kInt32x4Neg: { | 425 case IrOpcode::kInt32x4Neg: { |
| 526 DCHECK(node->InputCount() == 1); | 426 DCHECK(node->InputCount() == 1); |
| 527 Node** rep = GetReplacementsWithType(node->InputAt(0), rep_type); | 427 Node** rep = GetReplacementsWithType(node->InputAt(0), rep_type); |
| 528 Node* rep_node[kMaxLanes]; | 428 Node* rep_node[kMaxLanes]; |
| 529 Node* zero = graph()->NewNode(common()->Int32Constant(0)); | 429 Node* zero = graph()->NewNode(common()->Int32Constant(0)); |
| 530 for (int i = 0; i < kMaxLanes; ++i) { | 430 for (int i = 0; i < kMaxLanes; ++i) { |
| 531 rep_node[i] = graph()->NewNode(machine()->Int32Sub(), zero, rep[i]); | 431 rep_node[i] = graph()->NewNode(machine()->Int32Sub(), zero, rep[i]); |
| 532 } | 432 } |
| 533 ReplaceNode(node, rep_node); | 433 ReplaceNode(node, rep_node); |
| 534 break; | 434 break; |
| 535 } | 435 } |
| 536 case IrOpcode::kSimd128Not: { | 436 case IrOpcode::kSimd128Not: { |
| 537 DCHECK(node->InputCount() == 1); | 437 DCHECK(node->InputCount() == 1); |
| 538 Node** rep = GetReplacementsWithType(node->InputAt(0), rep_type); | 438 Node** rep = GetReplacementsWithType(node->InputAt(0), rep_type); |
| 539 Node* rep_node[kMaxLanes]; | 439 Node* rep_node[kMaxLanes]; |
| 540 Node* mask = graph()->NewNode(common()->Int32Constant(0xffffffff)); | 440 Node* mask = graph()->NewNode(common()->Int32Constant(0xffffffff)); |
| 541 for (int i = 0; i < kMaxLanes; ++i) { | 441 for (int i = 0; i < kMaxLanes; ++i) { |
| 542 rep_node[i] = graph()->NewNode(machine()->Word32Xor(), rep[i], mask); | 442 rep_node[i] = graph()->NewNode(machine()->Word32Xor(), rep[i], mask); |
| 543 } | 443 } |
| 544 ReplaceNode(node, rep_node); | 444 ReplaceNode(node, rep_node); |
| 545 break; | 445 break; |
| 546 } | 446 } |
| 547 case IrOpcode::kInt32x4FromFloat32x4: { | |
| 548 LowerConvertFromFloat(node, true); | |
| 549 break; | |
| 550 } | |
| 551 case IrOpcode::kUint32x4FromFloat32x4: { | |
| 552 LowerConvertFromFloat(node, false); | |
| 553 break; | |
| 554 } | |
| 555 case IrOpcode::kInt32x4ShiftLeftByScalar: { | |
| 556 LowerShiftOp(node, machine()->Word32Shl()); | |
| 557 break; | |
| 558 } | |
| 559 case IrOpcode::kInt32x4ShiftRightByScalar: { | |
| 560 LowerShiftOp(node, machine()->Word32Sar()); | |
| 561 break; | |
| 562 } | |
| 563 case IrOpcode::kUint32x4ShiftRightByScalar: { | |
| 564 LowerShiftOp(node, machine()->Word32Shr()); | |
| 565 break; | |
| 566 } | |
| 567 #define F32X4_BINOP_CASE(name) \ | 447 #define F32X4_BINOP_CASE(name) \ |
| 568 case IrOpcode::kFloat32x4##name: { \ | 448 case IrOpcode::kFloat32x4##name: { \ |
| 569 LowerBinaryOp(node, rep_type, machine()->Float32##name()); \ | 449 LowerBinaryOp(node, rep_type, machine()->Float32##name()); \ |
| 570 break; \ | 450 break; \ |
| 571 } | 451 } |
| 572 F32X4_BINOP_CASE(Add) | 452 F32X4_BINOP_CASE(Add) |
| 573 F32X4_BINOP_CASE(Sub) | 453 F32X4_BINOP_CASE(Sub) |
| 574 F32X4_BINOP_CASE(Mul) | 454 F32X4_BINOP_CASE(Mul) |
| 575 F32X4_BINOP_CASE(Div) | 455 F32X4_BINOP_CASE(Div) |
| 576 F32X4_BINOP_CASE(Min) | 456 F32X4_BINOP_CASE(Min) |
| (...skipping 159 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
| 736 } else { | 616 } else { |
| 737 UNREACHABLE(); | 617 UNREACHABLE(); |
| 738 } | 618 } |
| 739 } | 619 } |
| 740 ReplaceNode(phi, rep_nodes); | 620 ReplaceNode(phi, rep_nodes); |
| 741 } | 621 } |
| 742 } | 622 } |
| 743 } // namespace compiler | 623 } // namespace compiler |
| 744 } // namespace internal | 624 } // namespace internal |
| 745 } // namespace v8 | 625 } // namespace v8 |
| OLD | NEW |