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 // We don't know the microarchitecture here, so just reduce register pressure.
49 setSchedulingPreference(Sched::RegPressure);
50 // Tell ISel that we have a stack pointer.
51 setStackPointerRegisterToSaveRestore(
52 Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32);
53 // Set up the register classes.
54 addRegisterClass(MVT::i32, &WebAssembly::I32RegClass);
55 addRegisterClass(MVT::i64, &WebAssembly::I64RegClass);
56 addRegisterClass(MVT::f32, &WebAssembly::F32RegClass);
57 addRegisterClass(MVT::f64, &WebAssembly::F64RegClass);
58 if (Subtarget->hasSIMD128()) {
59 addRegisterClass(MVT::v16i8, &WebAssembly::V128RegClass);
60 addRegisterClass(MVT::v8i16, &WebAssembly::V128RegClass);
61 addRegisterClass(MVT::v4i32, &WebAssembly::V128RegClass);
62 addRegisterClass(MVT::v4f32, &WebAssembly::V128RegClass);
64 if (Subtarget->hasUnimplementedSIMD128()) {
65 addRegisterClass(MVT::v2i64, &WebAssembly::V128RegClass);
66 addRegisterClass(MVT::v2f64, &WebAssembly::V128RegClass);
68 // Compute derived properties from the register classes.
69 computeRegisterProperties(Subtarget->getRegisterInfo());
71 setOperationAction(ISD::GlobalAddress, MVTPtr, Custom);
72 setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom);
73 setOperationAction(ISD::JumpTable, MVTPtr, Custom);
74 setOperationAction(ISD::BlockAddress, MVTPtr, Custom);
75 setOperationAction(ISD::BRIND, MVT::Other, Custom);
77 // Take the default expansion for va_arg, va_copy, and va_end. There is no
78 // default action for va_start, so we do that custom.
79 setOperationAction(ISD::VASTART, MVT::Other, Custom);
80 setOperationAction(ISD::VAARG, MVT::Other, Expand);
81 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
82 setOperationAction(ISD::VAEND, MVT::Other, Expand);
84 for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
85 // Don't expand the floating-point types to constant pools.
86 setOperationAction(ISD::ConstantFP, T, Legal);
87 // Expand floating-point comparisons.
88 for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
89 ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
90 setCondCodeAction(CC, T, Expand);
91 // Expand floating-point library function operators.
93 {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM, ISD::FMA})
94 setOperationAction(Op, T, Expand);
95 // Note supported floating-point library function operators that otherwise
98 {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT, ISD::FRINT})
99 setOperationAction(Op, T, Legal);
100 // Support minimum and maximum, which otherwise default to expand.
101 setOperationAction(ISD::FMINIMUM, T, Legal);
102 setOperationAction(ISD::FMAXIMUM, T, Legal);
103 // WebAssembly currently has no builtin f16 support.
104 setOperationAction(ISD::FP16_TO_FP, T, Expand);
105 setOperationAction(ISD::FP_TO_FP16, T, Expand);
106 setLoadExtAction(ISD::EXTLOAD, T, MVT::f16, Expand);
107 setTruncStoreAction(T, MVT::f16, Expand);
110 // Expand unavailable integer operations.
112 {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU,
113 ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS,
114 ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) {
115 for (auto T : {MVT::i32, MVT::i64})
116 setOperationAction(Op, T, Expand);
117 if (Subtarget->hasSIMD128())
118 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
119 setOperationAction(Op, T, Expand);
120 if (Subtarget->hasUnimplementedSIMD128())
121 setOperationAction(Op, MVT::v2i64, Expand);
124 // SIMD-specific configuration
125 if (Subtarget->hasSIMD128()) {
126 // Support saturating add for i8x16 and i16x8
127 for (auto Op : {ISD::SADDSAT, ISD::UADDSAT})
128 for (auto T : {MVT::v16i8, MVT::v8i16})
129 setOperationAction(Op, T, Legal);
131 // Custom lower BUILD_VECTORs to minimize number of replace_lanes
132 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
133 setOperationAction(ISD::BUILD_VECTOR, T, Custom);
134 if (Subtarget->hasUnimplementedSIMD128())
135 for (auto T : {MVT::v2i64, MVT::v2f64})
136 setOperationAction(ISD::BUILD_VECTOR, T, Custom);
138 // We have custom shuffle lowering to expose the shuffle mask
139 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
140 setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom);
141 if (Subtarget->hasUnimplementedSIMD128())
142 for (auto T: {MVT::v2i64, MVT::v2f64})
143 setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom);
145 // Custom lowering since wasm shifts must have a scalar shift amount
146 for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL}) {
147 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
148 setOperationAction(Op, T, Custom);
149 if (Subtarget->hasUnimplementedSIMD128())
150 setOperationAction(Op, MVT::v2i64, Custom);
153 // Custom lower lane accesses to expand out variable indices
154 for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT}) {
155 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
156 setOperationAction(Op, T, Custom);
157 if (Subtarget->hasUnimplementedSIMD128())
158 for (auto T : {MVT::v2i64, MVT::v2f64})
159 setOperationAction(Op, T, Custom);
162 // There is no i64x2.mul instruction
163 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
165 // There are no vector select instructions
166 for (auto Op : {ISD::VSELECT, ISD::SELECT_CC, ISD::SELECT}) {
167 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32})
168 setOperationAction(Op, T, Expand);
169 if (Subtarget->hasUnimplementedSIMD128())
170 for (auto T : {MVT::v2i64, MVT::v2f64})
171 setOperationAction(Op, T, Expand);
174 // Expand integer operations supported for scalars but not SIMD
175 for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP, ISD::SDIV, ISD::UDIV,
176 ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR}) {
177 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
178 setOperationAction(Op, T, Expand);
179 if (Subtarget->hasUnimplementedSIMD128())
180 setOperationAction(Op, MVT::v2i64, Expand);
183 // Expand float operations supported for scalars but not SIMD
184 for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT,
185 ISD::FCOPYSIGN, ISD::FLOG, ISD::FLOG2, ISD::FLOG10,
186 ISD::FEXP, ISD::FEXP2, ISD::FRINT}) {
187 setOperationAction(Op, MVT::v4f32, Expand);
188 if (Subtarget->hasUnimplementedSIMD128())
189 setOperationAction(Op, MVT::v2f64, Expand);
192 // Expand additional SIMD ops that V8 hasn't implemented yet
193 if (!Subtarget->hasUnimplementedSIMD128()) {
194 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
195 setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
199 // As a special case, these operators use the type to mean the type to
201 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
202 if (!Subtarget->hasSignExt()) {
203 // Sign extends are legal only when extending a vector extract
204 auto Action = Subtarget->hasSIMD128() ? Custom : Expand;
205 for (auto T : {MVT::i8, MVT::i16, MVT::i32})
206 setOperationAction(ISD::SIGN_EXTEND_INREG, T, Action);
208 for (auto T : MVT::integer_vector_valuetypes())
209 setOperationAction(ISD::SIGN_EXTEND_INREG, T, Expand);
211 // Dynamic stack allocation: use the default expansion.
212 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
213 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
214 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVTPtr, Expand);
216 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
217 setOperationAction(ISD::CopyToReg, MVT::Other, Custom);
219 // Expand these forms; we pattern-match the forms that we can handle in isel.
220 for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64})
221 for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
222 setOperationAction(Op, T, Expand);
224 // We have custom switch handling.
225 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
227 // WebAssembly doesn't have:
228 // - Floating-point extending loads.
229 // - Floating-point truncating stores.
230 // - i1 extending loads.
231 // - extending/truncating SIMD loads/stores
232 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
233 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
234 for (auto T : MVT::integer_valuetypes())
235 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
236 setLoadExtAction(Ext, T, MVT::i1, Promote);
237 if (Subtarget->hasSIMD128()) {
238 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32,
240 for (auto MemT : MVT::vector_valuetypes()) {
241 if (MVT(T) != MemT) {
242 setTruncStoreAction(T, MemT, Expand);
243 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
244 setLoadExtAction(Ext, T, MemT, Expand);
250 // Don't do anything clever with build_pairs
251 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
253 // Trap lowers to wasm unreachable
254 setOperationAction(ISD::TRAP, MVT::Other, Legal);
256 // Exception handling intrinsics
257 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
258 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
260 setMaxAtomicSizeInBitsSupported(64);
262 if (Subtarget->hasBulkMemory()) {
263 // Use memory.copy and friends over multiple loads and stores
264 MaxStoresPerMemcpy = 1;
265 MaxStoresPerMemcpyOptSize = 1;
266 MaxStoresPerMemmove = 1;
267 MaxStoresPerMemmoveOptSize = 1;
268 MaxStoresPerMemset = 1;
269 MaxStoresPerMemsetOptSize = 1;
272 // Override the __gnu_f2h_ieee/__gnu_h2f_ieee names so that the f32 name is
273 // consistent with the f64 and f128 names.
274 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
275 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
277 // Define the emscripten name for return address helper.
278 // TODO: when implementing other WASM backends, make this generic or only do
279 // this on emscripten depending on what they end up doing.
280 setLibcallName(RTLIB::RETURN_ADDRESS, "emscripten_return_address");
282 // Always convert switches to br_tables unless there is only one case, which
283 // is equivalent to a simple branch. This reduces code size for wasm, and we
284 // defer possible jump table optimizations to the VM.
285 setMinimumJumpTableEntries(2);
288 TargetLowering::AtomicExpansionKind
289 WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
290 // We have wasm instructions for these
291 switch (AI->getOperation()) {
292 case AtomicRMWInst::Add:
293 case AtomicRMWInst::Sub:
294 case AtomicRMWInst::And:
295 case AtomicRMWInst::Or:
296 case AtomicRMWInst::Xor:
297 case AtomicRMWInst::Xchg:
298 return AtomicExpansionKind::None;
302 return AtomicExpansionKind::CmpXChg;
305 FastISel *WebAssemblyTargetLowering::createFastISel(
306 FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo) const {
307 return WebAssembly::createFastISel(FuncInfo, LibInfo);
310 MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
312 unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
313 if (BitWidth > 1 && BitWidth < 8)
317 // The shift will be lowered to a libcall, and compiler-rt libcalls expect
318 // the count to be an i32.
320 assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
321 "32-bit shift counts ought to be enough for anyone");
324 MVT Result = MVT::getIntegerVT(BitWidth);
325 assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
326 "Unable to represent scalar shift amount type");
330 // Lower an fp-to-int conversion operator from the LLVM opcode, which has an
331 // undefined result on invalid/overflow, to the WebAssembly opcode, which
332 // traps on invalid/overflow.
333 static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL,
334 MachineBasicBlock *BB,
335 const TargetInstrInfo &TII,
336 bool IsUnsigned, bool Int64,
337 bool Float64, unsigned LoweredOpcode) {
338 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
340 unsigned OutReg = MI.getOperand(0).getReg();
341 unsigned InReg = MI.getOperand(1).getReg();
343 unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32;
344 unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32;
345 unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32;
346 unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32;
347 unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32;
348 unsigned Eqz = WebAssembly::EQZ_I32;
349 unsigned And = WebAssembly::AND_I32;
350 int64_t Limit = Int64 ? INT64_MIN : INT32_MIN;
351 int64_t Substitute = IsUnsigned ? 0 : Limit;
352 double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit;
353 auto &Context = BB->getParent()->getFunction().getContext();
354 Type *Ty = Float64 ? Type::getDoubleTy(Context) : Type::getFloatTy(Context);
356 const BasicBlock *LLVMBB = BB->getBasicBlock();
357 MachineFunction *F = BB->getParent();
358 MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(LLVMBB);
359 MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(LLVMBB);
360 MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(LLVMBB);
362 MachineFunction::iterator It = ++BB->getIterator();
363 F->insert(It, FalseMBB);
364 F->insert(It, TrueMBB);
365 F->insert(It, DoneMBB);
367 // Transfer the remainder of BB and its successor edges to DoneMBB.
368 DoneMBB->splice(DoneMBB->begin(), BB, std::next(MI.getIterator()), BB->end());
369 DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
371 BB->addSuccessor(TrueMBB);
372 BB->addSuccessor(FalseMBB);
373 TrueMBB->addSuccessor(DoneMBB);
374 FalseMBB->addSuccessor(DoneMBB);
376 unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg;
377 Tmp0 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
378 Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
379 CmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
380 EqzReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
381 FalseReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
382 TrueReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
384 MI.eraseFromParent();
385 // For signed numbers, we can do a single comparison to determine whether
386 // fabs(x) is within range.
390 BuildMI(BB, DL, TII.get(Abs), Tmp0).addReg(InReg);
392 BuildMI(BB, DL, TII.get(FConst), Tmp1)
393 .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, CmpVal)));
394 BuildMI(BB, DL, TII.get(LT), CmpReg).addReg(Tmp0).addReg(Tmp1);
396 // For unsigned numbers, we have to do a separate comparison with zero.
398 Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
399 unsigned SecondCmpReg =
400 MRI.createVirtualRegister(&WebAssembly::I32RegClass);
401 unsigned AndReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
402 BuildMI(BB, DL, TII.get(FConst), Tmp1)
403 .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, 0.0)));
404 BuildMI(BB, DL, TII.get(GE), SecondCmpReg).addReg(Tmp0).addReg(Tmp1);
405 BuildMI(BB, DL, TII.get(And), AndReg).addReg(CmpReg).addReg(SecondCmpReg);
409 BuildMI(BB, DL, TII.get(Eqz), EqzReg).addReg(CmpReg);
411 // Create the CFG diamond to select between doing the conversion or using
412 // the substitute value.
413 BuildMI(BB, DL, TII.get(WebAssembly::BR_IF)).addMBB(TrueMBB).addReg(EqzReg);
414 BuildMI(FalseMBB, DL, TII.get(LoweredOpcode), FalseReg).addReg(InReg);
415 BuildMI(FalseMBB, DL, TII.get(WebAssembly::BR)).addMBB(DoneMBB);
416 BuildMI(TrueMBB, DL, TII.get(IConst), TrueReg).addImm(Substitute);
417 BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(TargetOpcode::PHI), OutReg)
426 MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter(
427 MachineInstr &MI, MachineBasicBlock *BB) const {
428 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
429 DebugLoc DL = MI.getDebugLoc();
431 switch (MI.getOpcode()) {
433 llvm_unreachable("Unexpected instr type to insert");
434 case WebAssembly::FP_TO_SINT_I32_F32:
435 return LowerFPToInt(MI, DL, BB, TII, false, false, false,
436 WebAssembly::I32_TRUNC_S_F32);
437 case WebAssembly::FP_TO_UINT_I32_F32:
438 return LowerFPToInt(MI, DL, BB, TII, true, false, false,
439 WebAssembly::I32_TRUNC_U_F32);
440 case WebAssembly::FP_TO_SINT_I64_F32:
441 return LowerFPToInt(MI, DL, BB, TII, false, true, false,
442 WebAssembly::I64_TRUNC_S_F32);
443 case WebAssembly::FP_TO_UINT_I64_F32:
444 return LowerFPToInt(MI, DL, BB, TII, true, true, false,
445 WebAssembly::I64_TRUNC_U_F32);
446 case WebAssembly::FP_TO_SINT_I32_F64:
447 return LowerFPToInt(MI, DL, BB, TII, false, false, true,
448 WebAssembly::I32_TRUNC_S_F64);
449 case WebAssembly::FP_TO_UINT_I32_F64:
450 return LowerFPToInt(MI, DL, BB, TII, true, false, true,
451 WebAssembly::I32_TRUNC_U_F64);
452 case WebAssembly::FP_TO_SINT_I64_F64:
453 return LowerFPToInt(MI, DL, BB, TII, false, true, true,
454 WebAssembly::I64_TRUNC_S_F64);
455 case WebAssembly::FP_TO_UINT_I64_F64:
456 return LowerFPToInt(MI, DL, BB, TII, true, true, true,
457 WebAssembly::I64_TRUNC_U_F64);
458 llvm_unreachable("Unexpected instruction to emit with custom inserter");
463 WebAssemblyTargetLowering::getTargetNodeName(unsigned Opcode) const {
464 switch (static_cast<WebAssemblyISD::NodeType>(Opcode)) {
465 case WebAssemblyISD::FIRST_NUMBER:
467 #define HANDLE_NODETYPE(NODE) \
468 case WebAssemblyISD::NODE: \
469 return "WebAssemblyISD::" #NODE;
470 #include "WebAssemblyISD.def"
471 #undef HANDLE_NODETYPE
476 std::pair<unsigned, const TargetRegisterClass *>
477 WebAssemblyTargetLowering::getRegForInlineAsmConstraint(
478 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
479 // First, see if this is a constraint that directly corresponds to a
480 // WebAssembly register class.
481 if (Constraint.size() == 1) {
482 switch (Constraint[0]) {
484 assert(VT != MVT::iPTR && "Pointer MVT not expected here");
485 if (Subtarget->hasSIMD128() && VT.isVector()) {
486 if (VT.getSizeInBits() == 128)
487 return std::make_pair(0U, &WebAssembly::V128RegClass);
489 if (VT.isInteger() && !VT.isVector()) {
490 if (VT.getSizeInBits() <= 32)
491 return std::make_pair(0U, &WebAssembly::I32RegClass);
492 if (VT.getSizeInBits() <= 64)
493 return std::make_pair(0U, &WebAssembly::I64RegClass);
501 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
504 bool WebAssemblyTargetLowering::isCheapToSpeculateCttz() const {
505 // Assume ctz is a relatively cheap operation.
509 bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz() const {
510 // Assume clz is a relatively cheap operation.
514 bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL,
516 Type *Ty, unsigned AS,
517 Instruction *I) const {
518 // WebAssembly offsets are added as unsigned without wrapping. The
519 // isLegalAddressingMode gives us no way to determine if wrapping could be
520 // happening, so we approximate this by accepting only non-negative offsets.
524 // WebAssembly has no scale register operands.
528 // Everything else is legal.
532 bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses(
533 EVT /*VT*/, unsigned /*AddrSpace*/, unsigned /*Align*/,
534 MachineMemOperand::Flags /*Flags*/, bool *Fast) const {
535 // WebAssembly supports unaligned accesses, though it should be declared
536 // with the p2align attribute on loads and stores which do so, and there
537 // may be a performance impact. We tell LLVM they're "fast" because
538 // for the kinds of things that LLVM uses this for (merging adjacent stores
539 // of constants, etc.), WebAssembly implementations will either want the
540 // unaligned access or they'll split anyway.
546 bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT,
547 AttributeList Attr) const {
548 // The current thinking is that wasm engines will perform this optimization,
549 // so we can save on code size.
553 EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
557 return VT.changeVectorElementTypeToInteger();
559 return TargetLowering::getSetCCResultType(DL, C, VT);
562 bool WebAssemblyTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
565 unsigned Intrinsic) const {
567 case Intrinsic::wasm_atomic_notify:
568 Info.opc = ISD::INTRINSIC_W_CHAIN;
569 Info.memVT = MVT::i32;
570 Info.ptrVal = I.getArgOperand(0);
573 // atomic.notify instruction does not really load the memory specified with
574 // this argument, but MachineMemOperand should either be load or store, so
575 // we set this to a load.
576 // FIXME Volatile isn't really correct, but currently all LLVM atomic
577 // instructions are treated as volatiles in the backend, so we should be
578 // consistent. The same applies for wasm_atomic_wait intrinsics too.
579 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
581 case Intrinsic::wasm_atomic_wait_i32:
582 Info.opc = ISD::INTRINSIC_W_CHAIN;
583 Info.memVT = MVT::i32;
584 Info.ptrVal = I.getArgOperand(0);
587 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
589 case Intrinsic::wasm_atomic_wait_i64:
590 Info.opc = ISD::INTRINSIC_W_CHAIN;
591 Info.memVT = MVT::i64;
592 Info.ptrVal = I.getArgOperand(0);
595 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
602 //===----------------------------------------------------------------------===//
603 // WebAssembly Lowering private implementation.
604 //===----------------------------------------------------------------------===//
606 //===----------------------------------------------------------------------===//
608 //===----------------------------------------------------------------------===//
610 static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) {
611 MachineFunction &MF = DAG.getMachineFunction();
612 DAG.getContext()->diagnose(
613 DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc()));
616 // Test whether the given calling convention is supported.
617 static bool callingConvSupported(CallingConv::ID CallConv) {
618 // We currently support the language-independent target-independent
619 // conventions. We don't yet have a way to annotate calls with properties like
620 // "cold", and we don't have any call-clobbered registers, so these are mostly
621 // all handled the same.
622 return CallConv == CallingConv::C || CallConv == CallingConv::Fast ||
623 CallConv == CallingConv::Cold ||
624 CallConv == CallingConv::PreserveMost ||
625 CallConv == CallingConv::PreserveAll ||
626 CallConv == CallingConv::CXX_FAST_TLS;
630 WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI,
631 SmallVectorImpl<SDValue> &InVals) const {
632 SelectionDAG &DAG = CLI.DAG;
634 SDValue Chain = CLI.Chain;
635 SDValue Callee = CLI.Callee;
636 MachineFunction &MF = DAG.getMachineFunction();
637 auto Layout = MF.getDataLayout();
639 CallingConv::ID CallConv = CLI.CallConv;
640 if (!callingConvSupported(CallConv))
642 "WebAssembly doesn't support language-specific or target-specific "
643 "calling conventions yet");
644 if (CLI.IsPatchPoint)
645 fail(DL, DAG, "WebAssembly doesn't support patch point yet");
647 // Fail if tail calls are required but not enabled
648 if (!Subtarget->hasTailCall()) {
649 if ((CallConv == CallingConv::Fast && CLI.IsTailCall &&
650 MF.getTarget().Options.GuaranteedTailCallOpt) ||
651 (CLI.CS && CLI.CS.isMustTailCall()))
652 fail(DL, DAG, "WebAssembly 'tail-call' feature not enabled");
653 CLI.IsTailCall = false;
656 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
658 fail(DL, DAG, "WebAssembly doesn't support more than 1 returned value yet");
660 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
661 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
662 unsigned NumFixedArgs = 0;
663 for (unsigned I = 0; I < Outs.size(); ++I) {
664 const ISD::OutputArg &Out = Outs[I];
665 SDValue &OutVal = OutVals[I];
666 if (Out.Flags.isNest())
667 fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
668 if (Out.Flags.isInAlloca())
669 fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
670 if (Out.Flags.isInConsecutiveRegs())
671 fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
672 if (Out.Flags.isInConsecutiveRegsLast())
673 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
674 if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) {
675 auto &MFI = MF.getFrameInfo();
676 int FI = MFI.CreateStackObject(Out.Flags.getByValSize(),
677 Out.Flags.getByValAlign(),
680 DAG.getConstant(Out.Flags.getByValSize(), DL, MVT::i32);
681 SDValue FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
682 Chain = DAG.getMemcpy(
683 Chain, DL, FINode, OutVal, SizeNode, Out.Flags.getByValAlign(),
684 /*isVolatile*/ false, /*AlwaysInline=*/false,
685 /*isTailCall*/ false, MachinePointerInfo(), MachinePointerInfo());
688 // Count the number of fixed args *after* legalization.
689 NumFixedArgs += Out.IsFixed;
692 bool IsVarArg = CLI.IsVarArg;
693 auto PtrVT = getPointerTy(Layout);
695 // Analyze operands of the call, assigning locations to each operand.
696 SmallVector<CCValAssign, 16> ArgLocs;
697 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
700 // Outgoing non-fixed arguments are placed in a buffer. First
701 // compute their offsets and the total amount of buffer space needed.
702 for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) {
703 const ISD::OutputArg &Out = Outs[I];
704 SDValue &Arg = OutVals[I];
705 EVT VT = Arg.getValueType();
706 assert(VT != MVT::iPTR && "Legalized args should be concrete");
707 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
708 unsigned Align = std::max(Out.Flags.getOrigAlign(),
709 Layout.getABITypeAlignment(Ty));
710 unsigned Offset = CCInfo.AllocateStack(Layout.getTypeAllocSize(Ty),
712 CCInfo.addLoc(CCValAssign::getMem(ArgLocs.size(), VT.getSimpleVT(),
713 Offset, VT.getSimpleVT(),
718 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
721 if (IsVarArg && NumBytes) {
722 // For non-fixed arguments, next emit stores to store the argument values
723 // to the stack buffer at the offsets computed above.
724 int FI = MF.getFrameInfo().CreateStackObject(NumBytes,
725 Layout.getStackAlignment(),
728 SmallVector<SDValue, 8> Chains;
730 make_range(OutVals.begin() + NumFixedArgs, OutVals.end())) {
731 assert(ArgLocs[ValNo].getValNo() == ValNo &&
732 "ArgLocs should remain in order and only hold varargs args");
733 unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
734 FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
735 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, FINode,
736 DAG.getConstant(Offset, DL, PtrVT));
738 DAG.getStore(Chain, DL, Arg, Add,
739 MachinePointerInfo::getFixedStack(MF, FI, Offset), 0));
742 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
743 } else if (IsVarArg) {
744 FINode = DAG.getIntPtrConstant(0, DL);
747 if (Callee->getOpcode() == ISD::GlobalAddress) {
748 // If the callee is a GlobalAddress node (quite common, every direct call
749 // is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress
750 // doesn't at MO_GOT which is not needed for direct calls.
751 GlobalAddressSDNode* GA = cast<GlobalAddressSDNode>(Callee);
752 Callee = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
753 getPointerTy(DAG.getDataLayout()),
755 Callee = DAG.getNode(WebAssemblyISD::Wrapper, DL,
756 getPointerTy(DAG.getDataLayout()), Callee);
759 // Compute the operands for the CALLn node.
760 SmallVector<SDValue, 16> Ops;
761 Ops.push_back(Chain);
762 Ops.push_back(Callee);
764 // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
766 Ops.append(OutVals.begin(),
767 IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
768 // Add a pointer to the vararg buffer.
770 Ops.push_back(FINode);
772 SmallVector<EVT, 8> InTys;
773 for (const auto &In : Ins) {
774 assert(!In.Flags.isByVal() && "byval is not valid for return values");
775 assert(!In.Flags.isNest() && "nest is not valid for return values");
776 if (In.Flags.isInAlloca())
777 fail(DL, DAG, "WebAssembly hasn't implemented inalloca return values");
778 if (In.Flags.isInConsecutiveRegs())
779 fail(DL, DAG, "WebAssembly hasn't implemented cons regs return values");
780 if (In.Flags.isInConsecutiveRegsLast())
782 "WebAssembly hasn't implemented cons regs last return values");
783 // Ignore In.getOrigAlign() because all our arguments are passed in
785 InTys.push_back(In.VT);
788 if (CLI.IsTailCall) {
789 // ret_calls do not return values to the current frame
790 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
791 return DAG.getNode(WebAssemblyISD::RET_CALL, DL, NodeTys, Ops);
794 InTys.push_back(MVT::Other);
795 SDVTList InTyList = DAG.getVTList(InTys);
797 DAG.getNode(Ins.empty() ? WebAssemblyISD::CALL0 : WebAssemblyISD::CALL1,
802 InVals.push_back(Res);
803 Chain = Res.getValue(1);
809 bool WebAssemblyTargetLowering::CanLowerReturn(
810 CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
811 const SmallVectorImpl<ISD::OutputArg> &Outs,
812 LLVMContext & /*Context*/) const {
813 // WebAssembly can't currently handle returning tuples.
814 return Outs.size() <= 1;
817 SDValue WebAssemblyTargetLowering::LowerReturn(
818 SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
819 const SmallVectorImpl<ISD::OutputArg> &Outs,
820 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
821 SelectionDAG &DAG) const {
822 assert(Outs.size() <= 1 && "WebAssembly can only return up to one value");
823 if (!callingConvSupported(CallConv))
824 fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
826 SmallVector<SDValue, 4> RetOps(1, Chain);
827 RetOps.append(OutVals.begin(), OutVals.end());
828 Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps);
830 // Record the number and types of the return values.
831 for (const ISD::OutputArg &Out : Outs) {
832 assert(!Out.Flags.isByVal() && "byval is not valid for return values");
833 assert(!Out.Flags.isNest() && "nest is not valid for return values");
834 assert(Out.IsFixed && "non-fixed return value is not valid");
835 if (Out.Flags.isInAlloca())
836 fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
837 if (Out.Flags.isInConsecutiveRegs())
838 fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
839 if (Out.Flags.isInConsecutiveRegsLast())
840 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
846 SDValue WebAssemblyTargetLowering::LowerFormalArguments(
847 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
848 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
849 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
850 if (!callingConvSupported(CallConv))
851 fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
853 MachineFunction &MF = DAG.getMachineFunction();
854 auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>();
856 // Set up the incoming ARGUMENTS value, which serves to represent the liveness
857 // of the incoming values before they're represented by virtual registers.
858 MF.getRegInfo().addLiveIn(WebAssembly::ARGUMENTS);
860 for (const ISD::InputArg &In : Ins) {
861 if (In.Flags.isInAlloca())
862 fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
863 if (In.Flags.isNest())
864 fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
865 if (In.Flags.isInConsecutiveRegs())
866 fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
867 if (In.Flags.isInConsecutiveRegsLast())
868 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
869 // Ignore In.getOrigAlign() because all our arguments are passed in
871 InVals.push_back(In.Used ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT,
872 DAG.getTargetConstant(InVals.size(),
874 : DAG.getUNDEF(In.VT));
876 // Record the number and types of arguments.
877 MFI->addParam(In.VT);
880 // Varargs are copied into a buffer allocated by the caller, and a pointer to
881 // the buffer is passed as an argument.
883 MVT PtrVT = getPointerTy(MF.getDataLayout());
884 unsigned VarargVreg =
885 MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrVT));
886 MFI->setVarargBufferVreg(VarargVreg);
887 Chain = DAG.getCopyToReg(
888 Chain, DL, VarargVreg,
889 DAG.getNode(WebAssemblyISD::ARGUMENT, DL, PtrVT,
890 DAG.getTargetConstant(Ins.size(), DL, MVT::i32)));
891 MFI->addParam(PtrVT);
894 // Record the number and types of arguments and results.
895 SmallVector<MVT, 4> Params;
896 SmallVector<MVT, 4> Results;
897 computeSignatureVTs(MF.getFunction().getFunctionType(), MF.getFunction(),
898 DAG.getTarget(), Params, Results);
899 for (MVT VT : Results)
901 // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify
902 // the param logic here with ComputeSignatureVTs
903 assert(MFI->getParams().size() == Params.size() &&
904 std::equal(MFI->getParams().begin(), MFI->getParams().end(),
910 void WebAssemblyTargetLowering::ReplaceNodeResults(
911 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
912 switch (N->getOpcode()) {
913 case ISD::SIGN_EXTEND_INREG:
914 // Do not add any results, signifying that N should not be custom lowered
915 // after all. This happens because simd128 turns on custom lowering for
916 // SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an
921 "ReplaceNodeResults not implemented for this op for WebAssembly!");
925 //===----------------------------------------------------------------------===//
926 // Custom lowering hooks.
927 //===----------------------------------------------------------------------===//
929 SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
930 SelectionDAG &DAG) const {
932 switch (Op.getOpcode()) {
934 llvm_unreachable("unimplemented operation lowering");
936 case ISD::FrameIndex:
937 return LowerFrameIndex(Op, DAG);
938 case ISD::GlobalAddress:
939 return LowerGlobalAddress(Op, DAG);
940 case ISD::ExternalSymbol:
941 return LowerExternalSymbol(Op, DAG);
943 return LowerJumpTable(Op, DAG);
945 return LowerBR_JT(Op, DAG);
947 return LowerVASTART(Op, DAG);
948 case ISD::BlockAddress:
950 fail(DL, DAG, "WebAssembly hasn't implemented computed gotos");
952 case ISD::RETURNADDR:
953 return LowerRETURNADDR(Op, DAG);
955 return LowerFRAMEADDR(Op, DAG);
957 return LowerCopyToReg(Op, DAG);
958 case ISD::EXTRACT_VECTOR_ELT:
959 case ISD::INSERT_VECTOR_ELT:
960 return LowerAccessVectorElement(Op, DAG);
961 case ISD::INTRINSIC_VOID:
962 case ISD::INTRINSIC_WO_CHAIN:
963 case ISD::INTRINSIC_W_CHAIN:
964 return LowerIntrinsic(Op, DAG);
965 case ISD::SIGN_EXTEND_INREG:
966 return LowerSIGN_EXTEND_INREG(Op, DAG);
967 case ISD::BUILD_VECTOR:
968 return LowerBUILD_VECTOR(Op, DAG);
969 case ISD::VECTOR_SHUFFLE:
970 return LowerVECTOR_SHUFFLE(Op, DAG);
974 return LowerShift(Op, DAG);
978 SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op,
979 SelectionDAG &DAG) const {
980 SDValue Src = Op.getOperand(2);
981 if (isa<FrameIndexSDNode>(Src.getNode())) {
982 // CopyToReg nodes don't support FrameIndex operands. Other targets select
983 // the FI to some LEA-like instruction, but since we don't have that, we
984 // need to insert some kind of instruction that can take an FI operand and
985 // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
986 // local.copy between Op and its FI operand.
987 SDValue Chain = Op.getOperand(0);
989 unsigned Reg = cast<RegisterSDNode>(Op.getOperand(1))->getReg();
990 EVT VT = Src.getValueType();
991 SDValue Copy(DAG.getMachineNode(VT == MVT::i32 ? WebAssembly::COPY_I32
992 : WebAssembly::COPY_I64,
995 return Op.getNode()->getNumValues() == 1
996 ? DAG.getCopyToReg(Chain, DL, Reg, Copy)
997 : DAG.getCopyToReg(Chain, DL, Reg, Copy,
998 Op.getNumOperands() == 4 ? Op.getOperand(3)
1004 SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op,
1005 SelectionDAG &DAG) const {
1006 int FI = cast<FrameIndexSDNode>(Op)->getIndex();
1007 return DAG.getTargetFrameIndex(FI, Op.getValueType());
1010 SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op,
1011 SelectionDAG &DAG) const {
1014 if (!Subtarget->getTargetTriple().isOSEmscripten()) {
1016 "Non-Emscripten WebAssembly hasn't implemented "
1017 "__builtin_return_address");
1021 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
1024 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1025 return makeLibCall(DAG, RTLIB::RETURN_ADDRESS, Op.getValueType(),
1026 {DAG.getConstant(Depth, DL, MVT::i32)}, false, DL)
1030 SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op,
1031 SelectionDAG &DAG) const {
1032 // Non-zero depths are not supported by WebAssembly currently. Use the
1033 // legalizer's default expansion, which is to return 0 (what this function is
1034 // documented to do).
1035 if (Op.getConstantOperandVal(0) > 0)
1038 DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true);
1039 EVT VT = Op.getValueType();
1041 Subtarget->getRegisterInfo()->getFrameRegister(DAG.getMachineFunction());
1042 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), FP, VT);
1045 SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
1046 SelectionDAG &DAG) const {
1048 const auto *GA = cast<GlobalAddressSDNode>(Op);
1049 EVT VT = Op.getValueType();
1050 assert(GA->getTargetFlags() == 0 &&
1051 "Unexpected target flags on generic GlobalAddressSDNode");
1052 if (GA->getAddressSpace() != 0)
1053 fail(DL, DAG, "WebAssembly only expects the 0 address space");
1055 unsigned OperandFlags = 0;
1056 if (isPositionIndependent()) {
1057 const GlobalValue *GV = GA->getGlobal();
1058 if (getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV)) {
1059 MachineFunction &MF = DAG.getMachineFunction();
1060 MVT PtrVT = getPointerTy(MF.getDataLayout());
1061 const char *BaseName;
1062 if (GV->getValueType()->isFunctionTy()) {
1063 BaseName = MF.createExternalSymbolName("__table_base");
1064 OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL;
1067 BaseName = MF.createExternalSymbolName("__memory_base");
1068 OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL;
1071 DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
1072 DAG.getTargetExternalSymbol(BaseName, PtrVT));
1074 SDValue SymAddr = DAG.getNode(
1075 WebAssemblyISD::WrapperPIC, DL, VT,
1076 DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT, GA->getOffset(),
1079 return DAG.getNode(ISD::ADD, DL, VT, BaseAddr, SymAddr);
1081 OperandFlags = WebAssemblyII::MO_GOT;
1085 return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1086 DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT,
1087 GA->getOffset(), OperandFlags));
1091 WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op,
1092 SelectionDAG &DAG) const {
1094 const auto *ES = cast<ExternalSymbolSDNode>(Op);
1095 EVT VT = Op.getValueType();
1096 assert(ES->getTargetFlags() == 0 &&
1097 "Unexpected target flags on generic ExternalSymbolSDNode");
1098 return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1099 DAG.getTargetExternalSymbol(ES->getSymbol(), VT));
1102 SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op,
1103 SelectionDAG &DAG) const {
1104 // There's no need for a Wrapper node because we always incorporate a jump
1105 // table operand into a BR_TABLE instruction, rather than ever
1106 // materializing it in a register.
1107 const JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1108 return DAG.getTargetJumpTable(JT->getIndex(), Op.getValueType(),
1109 JT->getTargetFlags());
1112 SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
1113 SelectionDAG &DAG) const {
1115 SDValue Chain = Op.getOperand(0);
1116 const auto *JT = cast<JumpTableSDNode>(Op.getOperand(1));
1117 SDValue Index = Op.getOperand(2);
1118 assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags");
1120 SmallVector<SDValue, 8> Ops;
1121 Ops.push_back(Chain);
1122 Ops.push_back(Index);
1124 MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
1125 const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs;
1127 // Add an operand for each case.
1128 for (auto MBB : MBBs)
1129 Ops.push_back(DAG.getBasicBlock(MBB));
1131 // TODO: For now, we just pick something arbitrary for a default case for now.
1132 // We really want to sniff out the guard and put in the real default case (and
1133 // delete the guard).
1134 Ops.push_back(DAG.getBasicBlock(MBBs[0]));
1136 return DAG.getNode(WebAssemblyISD::BR_TABLE, DL, MVT::Other, Ops);
1139 SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
1140 SelectionDAG &DAG) const {
1142 EVT PtrVT = getPointerTy(DAG.getMachineFunction().getDataLayout());
1144 auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>();
1145 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1147 SDValue ArgN = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1148 MFI->getVarargBufferVreg(), PtrVT);
1149 return DAG.getStore(Op.getOperand(0), DL, ArgN, Op.getOperand(1),
1150 MachinePointerInfo(SV), 0);
1153 SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op,
1154 SelectionDAG &DAG) const {
1155 MachineFunction &MF = DAG.getMachineFunction();
1157 switch (Op.getOpcode()) {
1158 case ISD::INTRINSIC_VOID:
1159 case ISD::INTRINSIC_W_CHAIN:
1160 IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1162 case ISD::INTRINSIC_WO_CHAIN:
1163 IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1166 llvm_unreachable("Invalid intrinsic");
1172 return SDValue(); // Don't custom lower most intrinsics.
1174 case Intrinsic::wasm_lsda: {
1175 EVT VT = Op.getValueType();
1176 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1177 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
1178 auto &Context = MF.getMMI().getContext();
1179 MCSymbol *S = Context.getOrCreateSymbol(Twine("GCC_except_table") +
1180 Twine(MF.getFunctionNumber()));
1181 return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1182 DAG.getMCSymbol(S, PtrVT));
1185 case Intrinsic::wasm_throw: {
1186 // We only support C++ exceptions for now
1187 int Tag = cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue();
1188 if (Tag != CPP_EXCEPTION)
1189 llvm_unreachable("Invalid tag!");
1190 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1191 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
1192 const char *SymName = MF.createExternalSymbolName("__cpp_exception");
1193 SDValue SymNode = DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
1194 DAG.getTargetExternalSymbol(SymName, PtrVT));
1195 return DAG.getNode(WebAssemblyISD::THROW, DL,
1196 MVT::Other, // outchain type
1198 Op.getOperand(0), // inchain
1199 SymNode, // exception symbol
1200 Op.getOperand(3) // thrown value
1207 WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1208 SelectionDAG &DAG) const {
1210 // If sign extension operations are disabled, allow sext_inreg only if operand
1211 // is a vector extract. SIMD does not depend on sign extension operations, but
1212 // allowing sext_inreg in this context lets us have simple patterns to select
1213 // extract_lane_s instructions. Expanding sext_inreg everywhere would be
1214 // simpler in this file, but would necessitate large and brittle patterns to
1215 // undo the expansion and select extract_lane_s instructions.
1216 assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128());
1217 if (Op.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1218 const SDValue &Extract = Op.getOperand(0);
1219 MVT VecT = Extract.getOperand(0).getSimpleValueType();
1220 MVT ExtractedLaneT = static_cast<VTSDNode *>(Op.getOperand(1).getNode())
1224 MVT::getVectorVT(ExtractedLaneT, 128 / ExtractedLaneT.getSizeInBits());
1225 if (ExtractedVecT == VecT)
1227 // Bitcast vector to appropriate type to ensure ISel pattern coverage
1228 const SDValue &Index = Extract.getOperand(1);
1230 static_cast<ConstantSDNode *>(Index.getNode())->getZExtValue();
1232 ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements();
1235 DAG.getConstant(IndexVal * Scale, DL, Index.getValueType());
1236 SDValue NewExtract = DAG.getNode(
1237 ISD::EXTRACT_VECTOR_ELT, DL, Extract.getValueType(),
1238 DAG.getBitcast(ExtractedVecT, Extract.getOperand(0)), NewIndex);
1239 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, Op.getValueType(),
1240 NewExtract, Op.getOperand(1));
1246 SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,
1247 SelectionDAG &DAG) const {
1249 const EVT VecT = Op.getValueType();
1250 const EVT LaneT = Op.getOperand(0).getValueType();
1251 const size_t Lanes = Op.getNumOperands();
1252 auto IsConstant = [](const SDValue &V) {
1253 return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP;
1256 // Find the most common operand, which is approximately the best to splat
1257 using Entry = std::pair<SDValue, size_t>;
1258 SmallVector<Entry, 16> ValueCounts;
1259 size_t NumConst = 0, NumDynamic = 0;
1260 for (const SDValue &Lane : Op->op_values()) {
1261 if (Lane.isUndef()) {
1263 } else if (IsConstant(Lane)) {
1268 auto CountIt = std::find_if(ValueCounts.begin(), ValueCounts.end(),
1269 [&Lane](Entry A) { return A.first == Lane; });
1270 if (CountIt == ValueCounts.end()) {
1271 ValueCounts.emplace_back(Lane, 1);
1277 std::max_element(ValueCounts.begin(), ValueCounts.end(),
1278 [](Entry A, Entry B) { return A.second < B.second; });
1279 assert(CommonIt != ValueCounts.end() && "Unexpected all-undef build_vector");
1280 SDValue SplatValue = CommonIt->first;
1281 size_t NumCommon = CommonIt->second;
1283 // If v128.const is available, consider using it instead of a splat
1284 if (Subtarget->hasUnimplementedSIMD128()) {
1285 // {i32,i64,f32,f64}.const opcode, and value
1286 const size_t ConstBytes = 1 + std::max(size_t(4), 16 / Lanes);
1287 // SIMD prefix and opcode
1288 const size_t SplatBytes = 2;
1289 const size_t SplatConstBytes = SplatBytes + ConstBytes;
1290 // SIMD prefix, opcode, and lane index
1291 const size_t ReplaceBytes = 3;
1292 const size_t ReplaceConstBytes = ReplaceBytes + ConstBytes;
1293 // SIMD prefix, v128.const opcode, and 128-bit value
1294 const size_t VecConstBytes = 18;
1295 // Initial v128.const and a replace_lane for each non-const operand
1296 const size_t ConstInitBytes = VecConstBytes + NumDynamic * ReplaceBytes;
1297 // Initial splat and all necessary replace_lanes
1298 const size_t SplatInitBytes =
1299 IsConstant(SplatValue)
1300 // Initial constant splat
1301 ? (SplatConstBytes +
1302 // Constant replace_lanes
1303 (NumConst - NumCommon) * ReplaceConstBytes +
1304 // Dynamic replace_lanes
1305 (NumDynamic * ReplaceBytes))
1306 // Initial dynamic splat
1308 // Constant replace_lanes
1309 (NumConst * ReplaceConstBytes) +
1310 // Dynamic replace_lanes
1311 (NumDynamic - NumCommon) * ReplaceBytes);
1312 if (ConstInitBytes < SplatInitBytes) {
1313 // Create build_vector that will lower to initial v128.const
1314 SmallVector<SDValue, 16> ConstLanes;
1315 for (const SDValue &Lane : Op->op_values()) {
1316 if (IsConstant(Lane)) {
1317 ConstLanes.push_back(Lane);
1318 } else if (LaneT.isFloatingPoint()) {
1319 ConstLanes.push_back(DAG.getConstantFP(0, DL, LaneT));
1321 ConstLanes.push_back(DAG.getConstant(0, DL, LaneT));
1324 SDValue Result = DAG.getBuildVector(VecT, DL, ConstLanes);
1325 // Add replace_lane instructions for non-const lanes
1326 for (size_t I = 0; I < Lanes; ++I) {
1327 const SDValue &Lane = Op->getOperand(I);
1328 if (!Lane.isUndef() && !IsConstant(Lane))
1329 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane,
1330 DAG.getConstant(I, DL, MVT::i32));
1335 // Use a splat for the initial vector
1336 SDValue Result = DAG.getSplatBuildVector(VecT, DL, SplatValue);
1337 // Add replace_lane instructions for other values
1338 for (size_t I = 0; I < Lanes; ++I) {
1339 const SDValue &Lane = Op->getOperand(I);
1340 if (Lane != SplatValue)
1341 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane,
1342 DAG.getConstant(I, DL, MVT::i32));
1348 WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
1349 SelectionDAG &DAG) const {
1351 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op.getNode())->getMask();
1352 MVT VecType = Op.getOperand(0).getSimpleValueType();
1353 assert(VecType.is128BitVector() && "Unexpected shuffle vector type");
1354 size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8;
1356 // Space for two vector args and sixteen mask indices
1359 Ops[OpIdx++] = Op.getOperand(0);
1360 Ops[OpIdx++] = Op.getOperand(1);
1362 // Expand mask indices to byte indices and materialize them as operands
1363 for (int M : Mask) {
1364 for (size_t J = 0; J < LaneBytes; ++J) {
1365 // Lower undefs (represented by -1 in mask) to zero
1366 uint64_t ByteIndex = M == -1 ? 0 : (uint64_t)M * LaneBytes + J;
1367 Ops[OpIdx++] = DAG.getConstant(ByteIndex, DL, MVT::i32);
1371 return DAG.getNode(WebAssemblyISD::SHUFFLE, DL, Op.getValueType(), Ops);
1375 WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op,
1376 SelectionDAG &DAG) const {
1377 // Allow constant lane indices, expand variable lane indices
1378 SDNode *IdxNode = Op.getOperand(Op.getNumOperands() - 1).getNode();
1379 if (isa<ConstantSDNode>(IdxNode) || IdxNode->isUndef())
1382 // Perform default expansion
1386 static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) {
1387 EVT LaneT = Op.getSimpleValueType().getVectorElementType();
1388 // 32-bit and 64-bit unrolled shifts will have proper semantics
1389 if (LaneT.bitsGE(MVT::i32))
1390 return DAG.UnrollVectorOp(Op.getNode());
1391 // Otherwise mask the shift value to get proper semantics from 32-bit shift
1393 SDValue ShiftVal = Op.getOperand(1);
1394 uint64_t MaskVal = LaneT.getSizeInBits() - 1;
1395 SDValue MaskedShiftVal = DAG.getNode(
1396 ISD::AND, // mask opcode
1397 DL, ShiftVal.getValueType(), // masked value type
1398 ShiftVal, // original shift value operand
1399 DAG.getConstant(MaskVal, DL, ShiftVal.getValueType()) // mask operand
1402 return DAG.UnrollVectorOp(
1403 DAG.getNode(Op.getOpcode(), // original shift opcode
1404 DL, Op.getValueType(), // original return type
1405 Op.getOperand(0), // original vector operand,
1406 MaskedShiftVal // new masked shift value operand
1411 SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op,
1412 SelectionDAG &DAG) const {
1415 // Only manually lower vector shifts
1416 assert(Op.getSimpleValueType().isVector());
1418 // Expand all vector shifts until V8 fixes its implementation
1419 // TODO: remove this once V8 is fixed
1420 if (!Subtarget->hasUnimplementedSIMD128())
1421 return unrollVectorShift(Op, DAG);
1423 // Unroll non-splat vector shifts
1424 BuildVectorSDNode *ShiftVec;
1426 if (!(ShiftVec = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode())) ||
1427 !(SplatVal = ShiftVec->getSplatValue()))
1428 return unrollVectorShift(Op, DAG);
1430 // All splats except i64x2 const splats are handled by patterns
1431 auto *SplatConst = dyn_cast<ConstantSDNode>(SplatVal);
1432 if (!SplatConst || Op.getSimpleValueType() != MVT::v2i64)
1435 // i64x2 const splats are custom lowered to avoid unnecessary wraps
1437 switch (Op.getOpcode()) {
1439 Opcode = WebAssemblyISD::VEC_SHL;
1442 Opcode = WebAssemblyISD::VEC_SHR_S;
1445 Opcode = WebAssemblyISD::VEC_SHR_U;
1448 llvm_unreachable("unexpected opcode");
1450 APInt Shift = SplatConst->getAPIntValue().zextOrTrunc(32);
1451 return DAG.getNode(Opcode, DL, Op.getValueType(), Op.getOperand(0),
1452 DAG.getConstant(Shift, DL, MVT::i32));
1455 //===----------------------------------------------------------------------===//
1456 // WebAssembly Optimization Hooks
1457 //===----------------------------------------------------------------------===//