1 //=- WebAssemblyISelLowering.cpp - WebAssembly DAG Lowering Implementation -==//
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
7 //===----------------------------------------------------------------------===//
10 /// This file implements the WebAssemblyTargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "WebAssemblyISelLowering.h"
15 #include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
16 #include "WebAssemblyMachineFunctionInfo.h"
17 #include "WebAssemblySubtarget.h"
18 #include "WebAssemblyTargetMachine.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineJumpTableInfo.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/WasmEHFuncInfo.h"
27 #include "llvm/IR/DiagnosticInfo.h"
28 #include "llvm/IR/DiagnosticPrinter.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetOptions.h"
37 #define DEBUG_TYPE "wasm-lower"
39 WebAssemblyTargetLowering::WebAssemblyTargetLowering(
40 const TargetMachine &TM, const WebAssemblySubtarget &STI)
41 : TargetLowering(TM), Subtarget(&STI) {
42 auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32;
44 // Booleans always contain 0 or 1.
45 setBooleanContents(ZeroOrOneBooleanContent);
46 // Except in SIMD vectors
47 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
48 // WebAssembly does not produce floating-point exceptions on normal floating
50 setHasFloatingPointExceptions(false);
51 // We don't know the microarchitecture here, so just reduce register pressure.
52 setSchedulingPreference(Sched::RegPressure);
53 // Tell ISel that we have a stack pointer.
54 setStackPointerRegisterToSaveRestore(
55 Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32);
56 // Set up the register classes.
57 addRegisterClass(MVT::i32, &WebAssembly::I32RegClass);
58 addRegisterClass(MVT::i64, &WebAssembly::I64RegClass);
59 addRegisterClass(MVT::f32, &WebAssembly::F32RegClass);
60 addRegisterClass(MVT::f64, &WebAssembly::F64RegClass);
61 if (Subtarget->hasSIMD128()) {
62 addRegisterClass(MVT::v16i8, &WebAssembly::V128RegClass);
63 addRegisterClass(MVT::v8i16, &WebAssembly::V128RegClass);
64 addRegisterClass(MVT::v4i32, &WebAssembly::V128RegClass);
65 addRegisterClass(MVT::v4f32, &WebAssembly::V128RegClass);
67 if (Subtarget->hasUnimplementedSIMD128()) {
68 addRegisterClass(MVT::v2i64, &WebAssembly::V128RegClass);
69 addRegisterClass(MVT::v2f64, &WebAssembly::V128RegClass);
71 // Compute derived properties from the register classes.
72 computeRegisterProperties(Subtarget->getRegisterInfo());
74 setOperationAction(ISD::GlobalAddress, MVTPtr, Custom);
75 setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom);
76 setOperationAction(ISD::JumpTable, MVTPtr, Custom);
77 setOperationAction(ISD::BlockAddress, MVTPtr, Custom);
78 setOperationAction(ISD::BRIND, MVT::Other, Custom);
80 // Take the default expansion for va_arg, va_copy, and va_end. There is no
81 // default action for va_start, so we do that custom.
82 setOperationAction(ISD::VASTART, MVT::Other, Custom);
83 setOperationAction(ISD::VAARG, MVT::Other, Expand);
84 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
85 setOperationAction(ISD::VAEND, MVT::Other, Expand);
87 for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
88 // Don't expand the floating-point types to constant pools.
89 setOperationAction(ISD::ConstantFP, T, Legal);
90 // Expand floating-point comparisons.
91 for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
92 ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
93 setCondCodeAction(CC, T, Expand);
94 // Expand floating-point library function operators.
96 {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM, ISD::FMA})
97 setOperationAction(Op, T, Expand);
98 // Note supported floating-point library function operators that otherwise
101 {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT, ISD::FRINT})
102 setOperationAction(Op, T, Legal);
103 // Support minimum and maximum, which otherwise default to expand.
104 setOperationAction(ISD::FMINIMUM, T, Legal);
105 setOperationAction(ISD::FMAXIMUM, T, Legal);
106 // WebAssembly currently has no builtin f16 support.
107 setOperationAction(ISD::FP16_TO_FP, T, Expand);
108 setOperationAction(ISD::FP_TO_FP16, T, Expand);
109 setLoadExtAction(ISD::EXTLOAD, T, MVT::f16, Expand);
110 setTruncStoreAction(T, MVT::f16, Expand);
113 // Expand unavailable integer operations.
115 {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU,
116 ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS,
117 ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) {
118 for (auto T : {MVT::i32, MVT::i64})
119 setOperationAction(Op, T, Expand);
120 if (Subtarget->hasSIMD128())
121 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
122 setOperationAction(Op, T, Expand);
123 if (Subtarget->hasUnimplementedSIMD128())
124 setOperationAction(Op, MVT::v2i64, Expand);
127 // SIMD-specific configuration
128 if (Subtarget->hasSIMD128()) {
129 // Support saturating add for i8x16 and i16x8
130 for (auto Op : {ISD::SADDSAT, ISD::UADDSAT})
131 for (auto T : {MVT::v16i8, MVT::v8i16})
132 setOperationAction(Op, T, Legal);
134 // Custom lower BUILD_VECTORs to minimize number of replace_lanes
135 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
136 setOperationAction(ISD::BUILD_VECTOR, T, Custom);
137 if (Subtarget->hasUnimplementedSIMD128())
138 for (auto T : {MVT::v2i64, MVT::v2f64})
139 setOperationAction(ISD::BUILD_VECTOR, T, Custom);
141 // We have custom shuffle lowering to expose the shuffle mask
142 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
143 setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom);
144 if (Subtarget->hasUnimplementedSIMD128())
145 for (auto T: {MVT::v2i64, MVT::v2f64})
146 setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom);
148 // Custom lowering since wasm shifts must have a scalar shift amount
149 for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL}) {
150 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
151 setOperationAction(Op, T, Custom);
152 if (Subtarget->hasUnimplementedSIMD128())
153 setOperationAction(Op, MVT::v2i64, Custom);
156 // Custom lower lane accesses to expand out variable indices
157 for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT}) {
158 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
159 setOperationAction(Op, T, Custom);
160 if (Subtarget->hasUnimplementedSIMD128())
161 for (auto T : {MVT::v2i64, MVT::v2f64})
162 setOperationAction(Op, T, Custom);
165 // There is no i64x2.mul instruction
166 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
168 // There are no vector select instructions
169 for (auto Op : {ISD::VSELECT, ISD::SELECT_CC, ISD::SELECT}) {
170 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
171 setOperationAction(Op, T, Expand);
172 if (Subtarget->hasUnimplementedSIMD128())
173 for (auto T : {MVT::v2i64, MVT::v2f64})
174 setOperationAction(Op, T, Expand);
177 // Expand integer operations supported for scalars but not SIMD
178 for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP, ISD::SDIV, ISD::UDIV,
179 ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR}) {
180 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
181 setOperationAction(Op, T, Expand);
182 if (Subtarget->hasUnimplementedSIMD128())
183 setOperationAction(Op, MVT::v2i64, Expand);
186 // Expand float operations supported for scalars but not SIMD
187 for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT,
189 setOperationAction(Op, MVT::v4f32, Expand);
190 if (Subtarget->hasUnimplementedSIMD128())
191 setOperationAction(Op, MVT::v2f64, Expand);
194 // Expand additional SIMD ops that V8 hasn't implemented yet
195 if (!Subtarget->hasUnimplementedSIMD128()) {
196 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
197 setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
201 // As a special case, these operators use the type to mean the type to
203 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
204 if (!Subtarget->hasSignExt()) {
205 // Sign extends are legal only when extending a vector extract
206 auto Action = Subtarget->hasSIMD128() ? Custom : Expand;
207 for (auto T : {MVT::i8, MVT::i16, MVT::i32})
208 setOperationAction(ISD::SIGN_EXTEND_INREG, T, Action);
210 for (auto T : MVT::integer_vector_valuetypes())
211 setOperationAction(ISD::SIGN_EXTEND_INREG, T, Expand);
213 // Dynamic stack allocation: use the default expansion.
214 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
215 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
216 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVTPtr, Expand);
218 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
219 setOperationAction(ISD::CopyToReg, MVT::Other, Custom);
221 // Expand these forms; we pattern-match the forms that we can handle in isel.
222 for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64})
223 for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
224 setOperationAction(Op, T, Expand);
226 // We have custom switch handling.
227 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
229 // WebAssembly doesn't have:
230 // - Floating-point extending loads.
231 // - Floating-point truncating stores.
232 // - i1 extending loads.
233 // - extending/truncating SIMD loads/stores
234 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
235 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
236 for (auto T : MVT::integer_valuetypes())
237 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
238 setLoadExtAction(Ext, T, MVT::i1, Promote);
239 if (Subtarget->hasSIMD128()) {
240 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32,
242 for (auto MemT : MVT::vector_valuetypes()) {
243 if (MVT(T) != MemT) {
244 setTruncStoreAction(T, MemT, Expand);
245 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
246 setLoadExtAction(Ext, T, MemT, Expand);
252 // Don't do anything clever with build_pairs
253 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
255 // Trap lowers to wasm unreachable
256 setOperationAction(ISD::TRAP, MVT::Other, Legal);
258 // Exception handling intrinsics
259 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
260 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
262 setMaxAtomicSizeInBitsSupported(64);
264 if (Subtarget->hasBulkMemory()) {
265 // Use memory.copy and friends over multiple loads and stores
266 MaxStoresPerMemcpy = 1;
267 MaxStoresPerMemcpyOptSize = 1;
268 MaxStoresPerMemmove = 1;
269 MaxStoresPerMemmoveOptSize = 1;
270 MaxStoresPerMemset = 1;
271 MaxStoresPerMemsetOptSize = 1;
275 TargetLowering::AtomicExpansionKind
276 WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
277 // We have wasm instructions for these
278 switch (AI->getOperation()) {
279 case AtomicRMWInst::Add:
280 case AtomicRMWInst::Sub:
281 case AtomicRMWInst::And:
282 case AtomicRMWInst::Or:
283 case AtomicRMWInst::Xor:
284 case AtomicRMWInst::Xchg:
285 return AtomicExpansionKind::None;
289 return AtomicExpansionKind::CmpXChg;
292 FastISel *WebAssemblyTargetLowering::createFastISel(
293 FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo) const {
294 return WebAssembly::createFastISel(FuncInfo, LibInfo);
297 bool WebAssemblyTargetLowering::isOffsetFoldingLegal(
298 const GlobalAddressSDNode * /*GA*/) const {
299 // All offsets can be folded.
303 MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
305 unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
306 if (BitWidth > 1 && BitWidth < 8)
310 // The shift will be lowered to a libcall, and compiler-rt libcalls expect
311 // the count to be an i32.
313 assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
314 "32-bit shift counts ought to be enough for anyone");
317 MVT Result = MVT::getIntegerVT(BitWidth);
318 assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
319 "Unable to represent scalar shift amount type");
323 // Lower an fp-to-int conversion operator from the LLVM opcode, which has an
324 // undefined result on invalid/overflow, to the WebAssembly opcode, which
325 // traps on invalid/overflow.
326 static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL,
327 MachineBasicBlock *BB,
328 const TargetInstrInfo &TII,
329 bool IsUnsigned, bool Int64,
330 bool Float64, unsigned LoweredOpcode) {
331 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
333 unsigned OutReg = MI.getOperand(0).getReg();
334 unsigned InReg = MI.getOperand(1).getReg();
336 unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32;
337 unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32;
338 unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32;
339 unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32;
340 unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32;
341 unsigned Eqz = WebAssembly::EQZ_I32;
342 unsigned And = WebAssembly::AND_I32;
343 int64_t Limit = Int64 ? INT64_MIN : INT32_MIN;
344 int64_t Substitute = IsUnsigned ? 0 : Limit;
345 double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit;
346 auto &Context = BB->getParent()->getFunction().getContext();
347 Type *Ty = Float64 ? Type::getDoubleTy(Context) : Type::getFloatTy(Context);
349 const BasicBlock *LLVMBB = BB->getBasicBlock();
350 MachineFunction *F = BB->getParent();
351 MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(LLVMBB);
352 MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(LLVMBB);
353 MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(LLVMBB);
355 MachineFunction::iterator It = ++BB->getIterator();
356 F->insert(It, FalseMBB);
357 F->insert(It, TrueMBB);
358 F->insert(It, DoneMBB);
360 // Transfer the remainder of BB and its successor edges to DoneMBB.
361 DoneMBB->splice(DoneMBB->begin(), BB, std::next(MI.getIterator()), BB->end());
362 DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
364 BB->addSuccessor(TrueMBB);
365 BB->addSuccessor(FalseMBB);
366 TrueMBB->addSuccessor(DoneMBB);
367 FalseMBB->addSuccessor(DoneMBB);
369 unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg;
370 Tmp0 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
371 Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
372 CmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
373 EqzReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
374 FalseReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
375 TrueReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
377 MI.eraseFromParent();
378 // For signed numbers, we can do a single comparison to determine whether
379 // fabs(x) is within range.
383 BuildMI(BB, DL, TII.get(Abs), Tmp0).addReg(InReg);
385 BuildMI(BB, DL, TII.get(FConst), Tmp1)
386 .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, CmpVal)));
387 BuildMI(BB, DL, TII.get(LT), CmpReg).addReg(Tmp0).addReg(Tmp1);
389 // For unsigned numbers, we have to do a separate comparison with zero.
391 Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
392 unsigned SecondCmpReg =
393 MRI.createVirtualRegister(&WebAssembly::I32RegClass);
394 unsigned AndReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
395 BuildMI(BB, DL, TII.get(FConst), Tmp1)
396 .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, 0.0)));
397 BuildMI(BB, DL, TII.get(GE), SecondCmpReg).addReg(Tmp0).addReg(Tmp1);
398 BuildMI(BB, DL, TII.get(And), AndReg).addReg(CmpReg).addReg(SecondCmpReg);
402 BuildMI(BB, DL, TII.get(Eqz), EqzReg).addReg(CmpReg);
404 // Create the CFG diamond to select between doing the conversion or using
405 // the substitute value.
406 BuildMI(BB, DL, TII.get(WebAssembly::BR_IF)).addMBB(TrueMBB).addReg(EqzReg);
407 BuildMI(FalseMBB, DL, TII.get(LoweredOpcode), FalseReg).addReg(InReg);
408 BuildMI(FalseMBB, DL, TII.get(WebAssembly::BR)).addMBB(DoneMBB);
409 BuildMI(TrueMBB, DL, TII.get(IConst), TrueReg).addImm(Substitute);
410 BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(TargetOpcode::PHI), OutReg)
419 MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter(
420 MachineInstr &MI, MachineBasicBlock *BB) const {
421 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
422 DebugLoc DL = MI.getDebugLoc();
424 switch (MI.getOpcode()) {
426 llvm_unreachable("Unexpected instr type to insert");
427 case WebAssembly::FP_TO_SINT_I32_F32:
428 return LowerFPToInt(MI, DL, BB, TII, false, false, false,
429 WebAssembly::I32_TRUNC_S_F32);
430 case WebAssembly::FP_TO_UINT_I32_F32:
431 return LowerFPToInt(MI, DL, BB, TII, true, false, false,
432 WebAssembly::I32_TRUNC_U_F32);
433 case WebAssembly::FP_TO_SINT_I64_F32:
434 return LowerFPToInt(MI, DL, BB, TII, false, true, false,
435 WebAssembly::I64_TRUNC_S_F32);
436 case WebAssembly::FP_TO_UINT_I64_F32:
437 return LowerFPToInt(MI, DL, BB, TII, true, true, false,
438 WebAssembly::I64_TRUNC_U_F32);
439 case WebAssembly::FP_TO_SINT_I32_F64:
440 return LowerFPToInt(MI, DL, BB, TII, false, false, true,
441 WebAssembly::I32_TRUNC_S_F64);
442 case WebAssembly::FP_TO_UINT_I32_F64:
443 return LowerFPToInt(MI, DL, BB, TII, true, false, true,
444 WebAssembly::I32_TRUNC_U_F64);
445 case WebAssembly::FP_TO_SINT_I64_F64:
446 return LowerFPToInt(MI, DL, BB, TII, false, true, true,
447 WebAssembly::I64_TRUNC_S_F64);
448 case WebAssembly::FP_TO_UINT_I64_F64:
449 return LowerFPToInt(MI, DL, BB, TII, true, true, true,
450 WebAssembly::I64_TRUNC_U_F64);
451 llvm_unreachable("Unexpected instruction to emit with custom inserter");
456 WebAssemblyTargetLowering::getTargetNodeName(unsigned Opcode) const {
457 switch (static_cast<WebAssemblyISD::NodeType>(Opcode)) {
458 case WebAssemblyISD::FIRST_NUMBER:
460 #define HANDLE_NODETYPE(NODE) \
461 case WebAssemblyISD::NODE: \
462 return "WebAssemblyISD::" #NODE;
463 #include "WebAssemblyISD.def"
464 #undef HANDLE_NODETYPE
469 std::pair<unsigned, const TargetRegisterClass *>
470 WebAssemblyTargetLowering::getRegForInlineAsmConstraint(
471 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
472 // First, see if this is a constraint that directly corresponds to a
473 // WebAssembly register class.
474 if (Constraint.size() == 1) {
475 switch (Constraint[0]) {
477 assert(VT != MVT::iPTR && "Pointer MVT not expected here");
478 if (Subtarget->hasSIMD128() && VT.isVector()) {
479 if (VT.getSizeInBits() == 128)
480 return std::make_pair(0U, &WebAssembly::V128RegClass);
482 if (VT.isInteger() && !VT.isVector()) {
483 if (VT.getSizeInBits() <= 32)
484 return std::make_pair(0U, &WebAssembly::I32RegClass);
485 if (VT.getSizeInBits() <= 64)
486 return std::make_pair(0U, &WebAssembly::I64RegClass);
494 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
497 bool WebAssemblyTargetLowering::isCheapToSpeculateCttz() const {
498 // Assume ctz is a relatively cheap operation.
502 bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz() const {
503 // Assume clz is a relatively cheap operation.
507 bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL,
509 Type *Ty, unsigned AS,
510 Instruction *I) const {
511 // WebAssembly offsets are added as unsigned without wrapping. The
512 // isLegalAddressingMode gives us no way to determine if wrapping could be
513 // happening, so we approximate this by accepting only non-negative offsets.
517 // WebAssembly has no scale register operands.
521 // Everything else is legal.
525 bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses(
526 EVT /*VT*/, unsigned /*AddrSpace*/, unsigned /*Align*/, bool *Fast) const {
527 // WebAssembly supports unaligned accesses, though it should be declared
528 // with the p2align attribute on loads and stores which do so, and there
529 // may be a performance impact. We tell LLVM they're "fast" because
530 // for the kinds of things that LLVM uses this for (merging adjacent stores
531 // of constants, etc.), WebAssembly implementations will either want the
532 // unaligned access or they'll split anyway.
538 bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT,
539 AttributeList Attr) const {
540 // The current thinking is that wasm engines will perform this optimization,
541 // so we can save on code size.
545 EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
549 return VT.changeVectorElementTypeToInteger();
551 return TargetLowering::getSetCCResultType(DL, C, VT);
554 bool WebAssemblyTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
557 unsigned Intrinsic) const {
559 case Intrinsic::wasm_atomic_notify:
560 Info.opc = ISD::INTRINSIC_W_CHAIN;
561 Info.memVT = MVT::i32;
562 Info.ptrVal = I.getArgOperand(0);
565 // atomic.notify instruction does not really load the memory specified with
566 // this argument, but MachineMemOperand should either be load or store, so
567 // we set this to a load.
568 // FIXME Volatile isn't really correct, but currently all LLVM atomic
569 // instructions are treated as volatiles in the backend, so we should be
570 // consistent. The same applies for wasm_atomic_wait intrinsics too.
571 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
573 case Intrinsic::wasm_atomic_wait_i32:
574 Info.opc = ISD::INTRINSIC_W_CHAIN;
575 Info.memVT = MVT::i32;
576 Info.ptrVal = I.getArgOperand(0);
579 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
581 case Intrinsic::wasm_atomic_wait_i64:
582 Info.opc = ISD::INTRINSIC_W_CHAIN;
583 Info.memVT = MVT::i64;
584 Info.ptrVal = I.getArgOperand(0);
587 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
594 //===----------------------------------------------------------------------===//
595 // WebAssembly Lowering private implementation.
596 //===----------------------------------------------------------------------===//
598 //===----------------------------------------------------------------------===//
600 //===----------------------------------------------------------------------===//
602 static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) {
603 MachineFunction &MF = DAG.getMachineFunction();
604 DAG.getContext()->diagnose(
605 DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc()));
608 // Test whether the given calling convention is supported.
609 static bool callingConvSupported(CallingConv::ID CallConv) {
610 // We currently support the language-independent target-independent
611 // conventions. We don't yet have a way to annotate calls with properties like
612 // "cold", and we don't have any call-clobbered registers, so these are mostly
613 // all handled the same.
614 return CallConv == CallingConv::C || CallConv == CallingConv::Fast ||
615 CallConv == CallingConv::Cold ||
616 CallConv == CallingConv::PreserveMost ||
617 CallConv == CallingConv::PreserveAll ||
618 CallConv == CallingConv::CXX_FAST_TLS;
622 WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI,
623 SmallVectorImpl<SDValue> &InVals) const {
624 SelectionDAG &DAG = CLI.DAG;
626 SDValue Chain = CLI.Chain;
627 SDValue Callee = CLI.Callee;
628 MachineFunction &MF = DAG.getMachineFunction();
629 auto Layout = MF.getDataLayout();
631 CallingConv::ID CallConv = CLI.CallConv;
632 if (!callingConvSupported(CallConv))
634 "WebAssembly doesn't support language-specific or target-specific "
635 "calling conventions yet");
636 if (CLI.IsPatchPoint)
637 fail(DL, DAG, "WebAssembly doesn't support patch point yet");
639 // WebAssembly doesn't currently support explicit tail calls. If they are
640 // required, fail. Otherwise, just disable them.
641 if ((CallConv == CallingConv::Fast && CLI.IsTailCall &&
642 MF.getTarget().Options.GuaranteedTailCallOpt) ||
643 (CLI.CS && CLI.CS.isMustTailCall()))
644 fail(DL, DAG, "WebAssembly doesn't support tail call yet");
645 CLI.IsTailCall = false;
647 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
649 fail(DL, DAG, "WebAssembly doesn't support more than 1 returned value yet");
651 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
652 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
653 unsigned NumFixedArgs = 0;
654 for (unsigned I = 0; I < Outs.size(); ++I) {
655 const ISD::OutputArg &Out = Outs[I];
656 SDValue &OutVal = OutVals[I];
657 if (Out.Flags.isNest())
658 fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
659 if (Out.Flags.isInAlloca())
660 fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
661 if (Out.Flags.isInConsecutiveRegs())
662 fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
663 if (Out.Flags.isInConsecutiveRegsLast())
664 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
665 if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) {
666 auto &MFI = MF.getFrameInfo();
667 int FI = MFI.CreateStackObject(Out.Flags.getByValSize(),
668 Out.Flags.getByValAlign(),
671 DAG.getConstant(Out.Flags.getByValSize(), DL, MVT::i32);
672 SDValue FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
673 Chain = DAG.getMemcpy(
674 Chain, DL, FINode, OutVal, SizeNode, Out.Flags.getByValAlign(),
675 /*isVolatile*/ false, /*AlwaysInline=*/false,
676 /*isTailCall*/ false, MachinePointerInfo(), MachinePointerInfo());
679 // Count the number of fixed args *after* legalization.
680 NumFixedArgs += Out.IsFixed;
683 bool IsVarArg = CLI.IsVarArg;
684 auto PtrVT = getPointerTy(Layout);
686 // Analyze operands of the call, assigning locations to each operand.
687 SmallVector<CCValAssign, 16> ArgLocs;
688 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
691 // Outgoing non-fixed arguments are placed in a buffer. First
692 // compute their offsets and the total amount of buffer space needed.
693 for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) {
694 const ISD::OutputArg &Out = Outs[I];
695 SDValue &Arg = OutVals[I];
696 EVT VT = Arg.getValueType();
697 assert(VT != MVT::iPTR && "Legalized args should be concrete");
698 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
699 unsigned Align = std::max(Out.Flags.getOrigAlign(),
700 Layout.getABITypeAlignment(Ty));
701 unsigned Offset = CCInfo.AllocateStack(Layout.getTypeAllocSize(Ty),
703 CCInfo.addLoc(CCValAssign::getMem(ArgLocs.size(), VT.getSimpleVT(),
704 Offset, VT.getSimpleVT(),
709 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
712 if (IsVarArg && NumBytes) {
713 // For non-fixed arguments, next emit stores to store the argument values
714 // to the stack buffer at the offsets computed above.
715 int FI = MF.getFrameInfo().CreateStackObject(NumBytes,
716 Layout.getStackAlignment(),
719 SmallVector<SDValue, 8> Chains;
721 make_range(OutVals.begin() + NumFixedArgs, OutVals.end())) {
722 assert(ArgLocs[ValNo].getValNo() == ValNo &&
723 "ArgLocs should remain in order and only hold varargs args");
724 unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
725 FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
726 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, FINode,
727 DAG.getConstant(Offset, DL, PtrVT));
729 DAG.getStore(Chain, DL, Arg, Add,
730 MachinePointerInfo::getFixedStack(MF, FI, Offset), 0));
733 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
734 } else if (IsVarArg) {
735 FINode = DAG.getIntPtrConstant(0, DL);
738 // Compute the operands for the CALLn node.
739 SmallVector<SDValue, 16> Ops;
740 Ops.push_back(Chain);
741 Ops.push_back(Callee);
743 // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
745 Ops.append(OutVals.begin(),
746 IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
747 // Add a pointer to the vararg buffer.
749 Ops.push_back(FINode);
751 SmallVector<EVT, 8> InTys;
752 for (const auto &In : Ins) {
753 assert(!In.Flags.isByVal() && "byval is not valid for return values");
754 assert(!In.Flags.isNest() && "nest is not valid for return values");
755 if (In.Flags.isInAlloca())
756 fail(DL, DAG, "WebAssembly hasn't implemented inalloca return values");
757 if (In.Flags.isInConsecutiveRegs())
758 fail(DL, DAG, "WebAssembly hasn't implemented cons regs return values");
759 if (In.Flags.isInConsecutiveRegsLast())
761 "WebAssembly hasn't implemented cons regs last return values");
762 // Ignore In.getOrigAlign() because all our arguments are passed in
764 InTys.push_back(In.VT);
766 InTys.push_back(MVT::Other);
767 SDVTList InTyList = DAG.getVTList(InTys);
769 DAG.getNode(Ins.empty() ? WebAssemblyISD::CALL0 : WebAssemblyISD::CALL1,
774 InVals.push_back(Res);
775 Chain = Res.getValue(1);
781 bool WebAssemblyTargetLowering::CanLowerReturn(
782 CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
783 const SmallVectorImpl<ISD::OutputArg> &Outs,
784 LLVMContext & /*Context*/) const {
785 // WebAssembly can't currently handle returning tuples.
786 return Outs.size() <= 1;
789 SDValue WebAssemblyTargetLowering::LowerReturn(
790 SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
791 const SmallVectorImpl<ISD::OutputArg> &Outs,
792 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
793 SelectionDAG &DAG) const {
794 assert(Outs.size() <= 1 && "WebAssembly can only return up to one value");
795 if (!callingConvSupported(CallConv))
796 fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
798 SmallVector<SDValue, 4> RetOps(1, Chain);
799 RetOps.append(OutVals.begin(), OutVals.end());
800 Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps);
802 // Record the number and types of the return values.
803 for (const ISD::OutputArg &Out : Outs) {
804 assert(!Out.Flags.isByVal() && "byval is not valid for return values");
805 assert(!Out.Flags.isNest() && "nest is not valid for return values");
806 assert(Out.IsFixed && "non-fixed return value is not valid");
807 if (Out.Flags.isInAlloca())
808 fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
809 if (Out.Flags.isInConsecutiveRegs())
810 fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
811 if (Out.Flags.isInConsecutiveRegsLast())
812 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
818 SDValue WebAssemblyTargetLowering::LowerFormalArguments(
819 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
820 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
821 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
822 if (!callingConvSupported(CallConv))
823 fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
825 MachineFunction &MF = DAG.getMachineFunction();
826 auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>();
828 // Set up the incoming ARGUMENTS value, which serves to represent the liveness
829 // of the incoming values before they're represented by virtual registers.
830 MF.getRegInfo().addLiveIn(WebAssembly::ARGUMENTS);
832 for (const ISD::InputArg &In : Ins) {
833 if (In.Flags.isInAlloca())
834 fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
835 if (In.Flags.isNest())
836 fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
837 if (In.Flags.isInConsecutiveRegs())
838 fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
839 if (In.Flags.isInConsecutiveRegsLast())
840 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
841 // Ignore In.getOrigAlign() because all our arguments are passed in
843 InVals.push_back(In.Used ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT,
844 DAG.getTargetConstant(InVals.size(),
846 : DAG.getUNDEF(In.VT));
848 // Record the number and types of arguments.
849 MFI->addParam(In.VT);
852 // Varargs are copied into a buffer allocated by the caller, and a pointer to
853 // the buffer is passed as an argument.
855 MVT PtrVT = getPointerTy(MF.getDataLayout());
856 unsigned VarargVreg =
857 MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrVT));
858 MFI->setVarargBufferVreg(VarargVreg);
859 Chain = DAG.getCopyToReg(
860 Chain, DL, VarargVreg,
861 DAG.getNode(WebAssemblyISD::ARGUMENT, DL, PtrVT,
862 DAG.getTargetConstant(Ins.size(), DL, MVT::i32)));
863 MFI->addParam(PtrVT);
866 // Record the number and types of arguments and results.
867 SmallVector<MVT, 4> Params;
868 SmallVector<MVT, 4> Results;
869 computeSignatureVTs(MF.getFunction().getFunctionType(), MF.getFunction(),
870 DAG.getTarget(), Params, Results);
871 for (MVT VT : Results)
873 // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify
874 // the param logic here with ComputeSignatureVTs
875 assert(MFI->getParams().size() == Params.size() &&
876 std::equal(MFI->getParams().begin(), MFI->getParams().end(),
882 //===----------------------------------------------------------------------===//
883 // Custom lowering hooks.
884 //===----------------------------------------------------------------------===//
886 SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
887 SelectionDAG &DAG) const {
889 switch (Op.getOpcode()) {
891 llvm_unreachable("unimplemented operation lowering");
893 case ISD::FrameIndex:
894 return LowerFrameIndex(Op, DAG);
895 case ISD::GlobalAddress:
896 return LowerGlobalAddress(Op, DAG);
897 case ISD::ExternalSymbol:
898 return LowerExternalSymbol(Op, DAG);
900 return LowerJumpTable(Op, DAG);
902 return LowerBR_JT(Op, DAG);
904 return LowerVASTART(Op, DAG);
905 case ISD::BlockAddress:
907 fail(DL, DAG, "WebAssembly hasn't implemented computed gotos");
909 case ISD::RETURNADDR: // Probably nothing meaningful can be returned here.
910 fail(DL, DAG, "WebAssembly hasn't implemented __builtin_return_address");
913 return LowerFRAMEADDR(Op, DAG);
915 return LowerCopyToReg(Op, DAG);
916 case ISD::EXTRACT_VECTOR_ELT:
917 case ISD::INSERT_VECTOR_ELT:
918 return LowerAccessVectorElement(Op, DAG);
919 case ISD::INTRINSIC_VOID:
920 case ISD::INTRINSIC_WO_CHAIN:
921 case ISD::INTRINSIC_W_CHAIN:
922 return LowerIntrinsic(Op, DAG);
923 case ISD::SIGN_EXTEND_INREG:
924 return LowerSIGN_EXTEND_INREG(Op, DAG);
925 case ISD::BUILD_VECTOR:
926 return LowerBUILD_VECTOR(Op, DAG);
927 case ISD::VECTOR_SHUFFLE:
928 return LowerVECTOR_SHUFFLE(Op, DAG);
932 return LowerShift(Op, DAG);
936 SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op,
937 SelectionDAG &DAG) const {
938 SDValue Src = Op.getOperand(2);
939 if (isa<FrameIndexSDNode>(Src.getNode())) {
940 // CopyToReg nodes don't support FrameIndex operands. Other targets select
941 // the FI to some LEA-like instruction, but since we don't have that, we
942 // need to insert some kind of instruction that can take an FI operand and
943 // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
944 // local.copy between Op and its FI operand.
945 SDValue Chain = Op.getOperand(0);
947 unsigned Reg = cast<RegisterSDNode>(Op.getOperand(1))->getReg();
948 EVT VT = Src.getValueType();
949 SDValue Copy(DAG.getMachineNode(VT == MVT::i32 ? WebAssembly::COPY_I32
950 : WebAssembly::COPY_I64,
953 return Op.getNode()->getNumValues() == 1
954 ? DAG.getCopyToReg(Chain, DL, Reg, Copy)
955 : DAG.getCopyToReg(Chain, DL, Reg, Copy,
956 Op.getNumOperands() == 4 ? Op.getOperand(3)
962 SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op,
963 SelectionDAG &DAG) const {
964 int FI = cast<FrameIndexSDNode>(Op)->getIndex();
965 return DAG.getTargetFrameIndex(FI, Op.getValueType());
968 SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op,
969 SelectionDAG &DAG) const {
970 // Non-zero depths are not supported by WebAssembly currently. Use the
971 // legalizer's default expansion, which is to return 0 (what this function is
972 // documented to do).
973 if (Op.getConstantOperandVal(0) > 0)
976 DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true);
977 EVT VT = Op.getValueType();
979 Subtarget->getRegisterInfo()->getFrameRegister(DAG.getMachineFunction());
980 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), FP, VT);
983 SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
984 SelectionDAG &DAG) const {
986 const auto *GA = cast<GlobalAddressSDNode>(Op);
987 EVT VT = Op.getValueType();
988 assert(GA->getTargetFlags() == 0 &&
989 "Unexpected target flags on generic GlobalAddressSDNode");
990 if (GA->getAddressSpace() != 0)
991 fail(DL, DAG, "WebAssembly only expects the 0 address space");
993 WebAssemblyISD::Wrapper, DL, VT,
994 DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT, GA->getOffset()));
998 WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op,
999 SelectionDAG &DAG) const {
1001 const auto *ES = cast<ExternalSymbolSDNode>(Op);
1002 EVT VT = Op.getValueType();
1003 assert(ES->getTargetFlags() == 0 &&
1004 "Unexpected target flags on generic ExternalSymbolSDNode");
1005 // Set the TargetFlags to 0x1 which indicates that this is a "function"
1006 // symbol rather than a data symbol. We do this unconditionally even though
1007 // we don't know anything about the symbol other than its name, because all
1008 // external symbols used in target-independent SelectionDAG code are for
1011 WebAssemblyISD::Wrapper, DL, VT,
1012 DAG.getTargetExternalSymbol(ES->getSymbol(), VT,
1013 WebAssemblyII::MO_SYMBOL_FUNCTION));
1016 SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op,
1017 SelectionDAG &DAG) const {
1018 // There's no need for a Wrapper node because we always incorporate a jump
1019 // table operand into a BR_TABLE instruction, rather than ever
1020 // materializing it in a register.
1021 const JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1022 return DAG.getTargetJumpTable(JT->getIndex(), Op.getValueType(),
1023 JT->getTargetFlags());
1026 SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
1027 SelectionDAG &DAG) const {
1029 SDValue Chain = Op.getOperand(0);
1030 const auto *JT = cast<JumpTableSDNode>(Op.getOperand(1));
1031 SDValue Index = Op.getOperand(2);
1032 assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags");
1034 SmallVector<SDValue, 8> Ops;
1035 Ops.push_back(Chain);
1036 Ops.push_back(Index);
1038 MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
1039 const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs;
1041 // Add an operand for each case.
1042 for (auto MBB : MBBs)
1043 Ops.push_back(DAG.getBasicBlock(MBB));
1045 // TODO: For now, we just pick something arbitrary for a default case for now.
1046 // We really want to sniff out the guard and put in the real default case (and
1047 // delete the guard).
1048 Ops.push_back(DAG.getBasicBlock(MBBs[0]));
1050 return DAG.getNode(WebAssemblyISD::BR_TABLE, DL, MVT::Other, Ops);
1053 SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
1054 SelectionDAG &DAG) const {
1056 EVT PtrVT = getPointerTy(DAG.getMachineFunction().getDataLayout());
1058 auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>();
1059 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1061 SDValue ArgN = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1062 MFI->getVarargBufferVreg(), PtrVT);
1063 return DAG.getStore(Op.getOperand(0), DL, ArgN, Op.getOperand(1),
1064 MachinePointerInfo(SV), 0);
1067 SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op,
1068 SelectionDAG &DAG) const {
1069 MachineFunction &MF = DAG.getMachineFunction();
1071 switch (Op.getOpcode()) {
1072 case ISD::INTRINSIC_VOID:
1073 case ISD::INTRINSIC_W_CHAIN:
1074 IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1076 case ISD::INTRINSIC_WO_CHAIN:
1077 IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1080 llvm_unreachable("Invalid intrinsic");
1086 return SDValue(); // Don't custom lower most intrinsics.
1088 case Intrinsic::wasm_lsda: {
1089 EVT VT = Op.getValueType();
1090 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1091 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
1092 auto &Context = MF.getMMI().getContext();
1093 MCSymbol *S = Context.getOrCreateSymbol(Twine("GCC_except_table") +
1094 Twine(MF.getFunctionNumber()));
1095 return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1096 DAG.getMCSymbol(S, PtrVT));
1099 case Intrinsic::wasm_throw: {
1100 // We only support C++ exceptions for now
1101 int Tag = cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue();
1102 if (Tag != CPP_EXCEPTION)
1103 llvm_unreachable("Invalid tag!");
1104 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1105 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
1106 const char *SymName = MF.createExternalSymbolName("__cpp_exception");
1108 DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
1109 DAG.getTargetExternalSymbol(
1110 SymName, PtrVT, WebAssemblyII::MO_SYMBOL_EVENT));
1111 return DAG.getNode(WebAssemblyISD::THROW, DL,
1112 MVT::Other, // outchain type
1114 Op.getOperand(0), // inchain
1115 SymNode, // exception symbol
1116 Op.getOperand(3) // thrown value
1123 WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1124 SelectionDAG &DAG) const {
1125 // If sign extension operations are disabled, allow sext_inreg only if operand
1126 // is a vector extract. SIMD does not depend on sign extension operations, but
1127 // allowing sext_inreg in this context lets us have simple patterns to select
1128 // extract_lane_s instructions. Expanding sext_inreg everywhere would be
1129 // simpler in this file, but would necessitate large and brittle patterns to
1130 // undo the expansion and select extract_lane_s instructions.
1131 assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128());
1132 if (Op.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT)
1138 SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,
1139 SelectionDAG &DAG) const {
1141 const EVT VecT = Op.getValueType();
1142 const EVT LaneT = Op.getOperand(0).getValueType();
1143 const size_t Lanes = Op.getNumOperands();
1144 auto IsConstant = [](const SDValue &V) {
1145 return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP;
1148 // Find the most common operand, which is approximately the best to splat
1149 using Entry = std::pair<SDValue, size_t>;
1150 SmallVector<Entry, 16> ValueCounts;
1151 size_t NumConst = 0, NumDynamic = 0;
1152 for (const SDValue &Lane : Op->op_values()) {
1153 if (Lane.isUndef()) {
1155 } else if (IsConstant(Lane)) {
1160 auto CountIt = std::find_if(ValueCounts.begin(), ValueCounts.end(),
1161 [&Lane](Entry A) { return A.first == Lane; });
1162 if (CountIt == ValueCounts.end()) {
1163 ValueCounts.emplace_back(Lane, 1);
1169 std::max_element(ValueCounts.begin(), ValueCounts.end(),
1170 [](Entry A, Entry B) { return A.second < B.second; });
1171 assert(CommonIt != ValueCounts.end() && "Unexpected all-undef build_vector");
1172 SDValue SplatValue = CommonIt->first;
1173 size_t NumCommon = CommonIt->second;
1175 // If v128.const is available, consider using it instead of a splat
1176 if (Subtarget->hasUnimplementedSIMD128()) {
1177 // {i32,i64,f32,f64}.const opcode, and value
1178 const size_t ConstBytes = 1 + std::max(size_t(4), 16 / Lanes);
1179 // SIMD prefix and opcode
1180 const size_t SplatBytes = 2;
1181 const size_t SplatConstBytes = SplatBytes + ConstBytes;
1182 // SIMD prefix, opcode, and lane index
1183 const size_t ReplaceBytes = 3;
1184 const size_t ReplaceConstBytes = ReplaceBytes + ConstBytes;
1185 // SIMD prefix, v128.const opcode, and 128-bit value
1186 const size_t VecConstBytes = 18;
1187 // Initial v128.const and a replace_lane for each non-const operand
1188 const size_t ConstInitBytes = VecConstBytes + NumDynamic * ReplaceBytes;
1189 // Initial splat and all necessary replace_lanes
1190 const size_t SplatInitBytes =
1191 IsConstant(SplatValue)
1192 // Initial constant splat
1193 ? (SplatConstBytes +
1194 // Constant replace_lanes
1195 (NumConst - NumCommon) * ReplaceConstBytes +
1196 // Dynamic replace_lanes
1197 (NumDynamic * ReplaceBytes))
1198 // Initial dynamic splat
1200 // Constant replace_lanes
1201 (NumConst * ReplaceConstBytes) +
1202 // Dynamic replace_lanes
1203 (NumDynamic - NumCommon) * ReplaceBytes);
1204 if (ConstInitBytes < SplatInitBytes) {
1205 // Create build_vector that will lower to initial v128.const
1206 SmallVector<SDValue, 16> ConstLanes;
1207 for (const SDValue &Lane : Op->op_values()) {
1208 if (IsConstant(Lane)) {
1209 ConstLanes.push_back(Lane);
1210 } else if (LaneT.isFloatingPoint()) {
1211 ConstLanes.push_back(DAG.getConstantFP(0, DL, LaneT));
1213 ConstLanes.push_back(DAG.getConstant(0, DL, LaneT));
1216 SDValue Result = DAG.getBuildVector(VecT, DL, ConstLanes);
1217 // Add replace_lane instructions for non-const lanes
1218 for (size_t I = 0; I < Lanes; ++I) {
1219 const SDValue &Lane = Op->getOperand(I);
1220 if (!Lane.isUndef() && !IsConstant(Lane))
1221 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane,
1222 DAG.getConstant(I, DL, MVT::i32));
1227 // Use a splat for the initial vector
1228 SDValue Result = DAG.getSplatBuildVector(VecT, DL, SplatValue);
1229 // Add replace_lane instructions for other values
1230 for (size_t I = 0; I < Lanes; ++I) {
1231 const SDValue &Lane = Op->getOperand(I);
1232 if (Lane != SplatValue)
1233 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane,
1234 DAG.getConstant(I, DL, MVT::i32));
1240 WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
1241 SelectionDAG &DAG) const {
1243 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op.getNode())->getMask();
1244 MVT VecType = Op.getOperand(0).getSimpleValueType();
1245 assert(VecType.is128BitVector() && "Unexpected shuffle vector type");
1246 size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8;
1248 // Space for two vector args and sixteen mask indices
1251 Ops[OpIdx++] = Op.getOperand(0);
1252 Ops[OpIdx++] = Op.getOperand(1);
1254 // Expand mask indices to byte indices and materialize them as operands
1255 for (int M : Mask) {
1256 for (size_t J = 0; J < LaneBytes; ++J) {
1257 // Lower undefs (represented by -1 in mask) to zero
1258 uint64_t ByteIndex = M == -1 ? 0 : (uint64_t)M * LaneBytes + J;
1259 Ops[OpIdx++] = DAG.getConstant(ByteIndex, DL, MVT::i32);
1263 return DAG.getNode(WebAssemblyISD::SHUFFLE, DL, Op.getValueType(), Ops);
1267 WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op,
1268 SelectionDAG &DAG) const {
1269 // Allow constant lane indices, expand variable lane indices
1270 SDNode *IdxNode = Op.getOperand(Op.getNumOperands() - 1).getNode();
1271 if (isa<ConstantSDNode>(IdxNode) || IdxNode->isUndef())
1274 // Perform default expansion
1278 static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) {
1279 EVT LaneT = Op.getSimpleValueType().getVectorElementType();
1280 // 32-bit and 64-bit unrolled shifts will have proper semantics
1281 if (LaneT.bitsGE(MVT::i32))
1282 return DAG.UnrollVectorOp(Op.getNode());
1283 // Otherwise mask the shift value to get proper semantics from 32-bit shift
1285 SDValue ShiftVal = Op.getOperand(1);
1286 uint64_t MaskVal = LaneT.getSizeInBits() - 1;
1287 SDValue MaskedShiftVal = DAG.getNode(
1288 ISD::AND, // mask opcode
1289 DL, ShiftVal.getValueType(), // masked value type
1290 ShiftVal, // original shift value operand
1291 DAG.getConstant(MaskVal, DL, ShiftVal.getValueType()) // mask operand
1294 return DAG.UnrollVectorOp(
1295 DAG.getNode(Op.getOpcode(), // original shift opcode
1296 DL, Op.getValueType(), // original return type
1297 Op.getOperand(0), // original vector operand,
1298 MaskedShiftVal // new masked shift value operand
1303 SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op,
1304 SelectionDAG &DAG) const {
1307 // Only manually lower vector shifts
1308 assert(Op.getSimpleValueType().isVector());
1310 // Expand all vector shifts until V8 fixes its implementation
1311 // TODO: remove this once V8 is fixed
1312 if (!Subtarget->hasUnimplementedSIMD128())
1313 return unrollVectorShift(Op, DAG);
1315 // Unroll non-splat vector shifts
1316 BuildVectorSDNode *ShiftVec;
1318 if (!(ShiftVec = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode())) ||
1319 !(SplatVal = ShiftVec->getSplatValue()))
1320 return unrollVectorShift(Op, DAG);
1322 // All splats except i64x2 const splats are handled by patterns
1323 auto *SplatConst = dyn_cast<ConstantSDNode>(SplatVal);
1324 if (!SplatConst || Op.getSimpleValueType() != MVT::v2i64)
1327 // i64x2 const splats are custom lowered to avoid unnecessary wraps
1329 switch (Op.getOpcode()) {
1331 Opcode = WebAssemblyISD::VEC_SHL;
1334 Opcode = WebAssemblyISD::VEC_SHR_S;
1337 Opcode = WebAssemblyISD::VEC_SHR_U;
1340 llvm_unreachable("unexpected opcode");
1342 APInt Shift = SplatConst->getAPIntValue().zextOrTrunc(32);
1343 return DAG.getNode(Opcode, DL, Op.getValueType(), Op.getOperand(0),
1344 DAG.getConstant(Shift, DL, MVT::i32));
1347 //===----------------------------------------------------------------------===//
1348 // WebAssembly Optimization Hooks
1349 //===----------------------------------------------------------------------===//