1 //=- WebAssemblyISelLowering.cpp - WebAssembly DAG Lowering Implementation -==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file implements the WebAssemblyTargetLowering class.
13 //===----------------------------------------------------------------------===//
15 #include "WebAssemblyISelLowering.h"
16 #include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
17 #include "WebAssemblyMachineFunctionInfo.h"
18 #include "WebAssemblySubtarget.h"
19 #include "WebAssemblyTargetMachine.h"
20 #include "llvm/CodeGen/Analysis.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineJumpTableInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/IR/DiagnosticInfo.h"
27 #include "llvm/IR/DiagnosticPrinter.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Target/TargetOptions.h"
36 #define DEBUG_TYPE "wasm-lower"
38 WebAssemblyTargetLowering::WebAssemblyTargetLowering(
39 const TargetMachine &TM, const WebAssemblySubtarget &STI)
40 : TargetLowering(TM), Subtarget(&STI) {
41 auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32;
43 // Booleans always contain 0 or 1.
44 setBooleanContents(ZeroOrOneBooleanContent);
45 // WebAssembly does not produce floating-point exceptions on normal floating
47 setHasFloatingPointExceptions(false);
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 // Compute derived properties from the register classes.
65 computeRegisterProperties(Subtarget->getRegisterInfo());
67 setOperationAction(ISD::GlobalAddress, MVTPtr, Custom);
68 setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom);
69 setOperationAction(ISD::JumpTable, MVTPtr, Custom);
70 setOperationAction(ISD::BlockAddress, MVTPtr, Custom);
71 setOperationAction(ISD::BRIND, MVT::Other, Custom);
73 // Take the default expansion for va_arg, va_copy, and va_end. There is no
74 // default action for va_start, so we do that custom.
75 setOperationAction(ISD::VASTART, MVT::Other, Custom);
76 setOperationAction(ISD::VAARG, MVT::Other, Expand);
77 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
78 setOperationAction(ISD::VAEND, MVT::Other, Expand);
80 for (auto T : {MVT::f32, MVT::f64}) {
81 // Don't expand the floating-point types to constant pools.
82 setOperationAction(ISD::ConstantFP, T, Legal);
83 // Expand floating-point comparisons.
84 for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
85 ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
86 setCondCodeAction(CC, T, Expand);
87 // Expand floating-point library function operators.
88 for (auto Op : {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM,
90 setOperationAction(Op, T, Expand);
91 // Note supported floating-point library function operators that otherwise
94 {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT, ISD::FRINT})
95 setOperationAction(Op, T, Legal);
96 // Support minnan and maxnan, which otherwise default to expand.
97 setOperationAction(ISD::FMINNAN, T, Legal);
98 setOperationAction(ISD::FMAXNAN, T, Legal);
99 // WebAssembly currently has no builtin f16 support.
100 setOperationAction(ISD::FP16_TO_FP, T, Expand);
101 setOperationAction(ISD::FP_TO_FP16, T, Expand);
102 setLoadExtAction(ISD::EXTLOAD, T, MVT::f16, Expand);
103 setTruncStoreAction(T, MVT::f16, Expand);
106 for (auto T : {MVT::i32, MVT::i64}) {
107 // Expand unavailable integer operations.
109 {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI,
110 ISD::MULHS, ISD::MULHU, ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS,
111 ISD::SRA_PARTS, ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC,
113 setOperationAction(Op, T, Expand);
117 // As a special case, these operators use the type to mean the type to
119 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
120 if (!Subtarget->hasSignExt()) {
121 for (auto T : {MVT::i8, MVT::i16, MVT::i32})
122 setOperationAction(ISD::SIGN_EXTEND_INREG, T, Expand);
125 // Dynamic stack allocation: use the default expansion.
126 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
127 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
128 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVTPtr, Expand);
130 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
131 setOperationAction(ISD::CopyToReg, MVT::Other, Custom);
133 // Expand these forms; we pattern-match the forms that we can handle in isel.
134 for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64})
135 for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
136 setOperationAction(Op, T, Expand);
138 // We have custom switch handling.
139 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
141 // WebAssembly doesn't have:
142 // - Floating-point extending loads.
143 // - Floating-point truncating stores.
144 // - i1 extending loads.
145 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
146 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
147 for (auto T : MVT::integer_valuetypes())
148 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
149 setLoadExtAction(Ext, T, MVT::i1, Promote);
151 // Trap lowers to wasm unreachable
152 setOperationAction(ISD::TRAP, MVT::Other, Legal);
154 // Exception handling intrinsics
155 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
157 setMaxAtomicSizeInBitsSupported(64);
160 FastISel *WebAssemblyTargetLowering::createFastISel(
161 FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo) const {
162 return WebAssembly::createFastISel(FuncInfo, LibInfo);
165 bool WebAssemblyTargetLowering::isOffsetFoldingLegal(
166 const GlobalAddressSDNode * /*GA*/) const {
167 // All offsets can be folded.
171 MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
173 unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
174 if (BitWidth > 1 && BitWidth < 8) BitWidth = 8;
177 // The shift will be lowered to a libcall, and compiler-rt libcalls expect
178 // the count to be an i32.
180 assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
181 "32-bit shift counts ought to be enough for anyone");
184 MVT Result = MVT::getIntegerVT(BitWidth);
185 assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
186 "Unable to represent scalar shift amount type");
190 // Lower an fp-to-int conversion operator from the LLVM opcode, which has an
191 // undefined result on invalid/overflow, to the WebAssembly opcode, which
192 // traps on invalid/overflow.
193 static MachineBasicBlock *
197 MachineBasicBlock *BB,
198 const TargetInstrInfo &TII,
202 unsigned LoweredOpcode
204 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
206 unsigned OutReg = MI.getOperand(0).getReg();
207 unsigned InReg = MI.getOperand(1).getReg();
209 unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32;
210 unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32;
211 unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32;
212 unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32;
213 unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32;
214 unsigned Eqz = WebAssembly::EQZ_I32;
215 unsigned And = WebAssembly::AND_I32;
216 int64_t Limit = Int64 ? INT64_MIN : INT32_MIN;
217 int64_t Substitute = IsUnsigned ? 0 : Limit;
218 double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit;
219 auto &Context = BB->getParent()->getFunction().getContext();
220 Type *Ty = Float64 ? Type::getDoubleTy(Context) : Type::getFloatTy(Context);
222 const BasicBlock *LLVM_BB = BB->getBasicBlock();
223 MachineFunction *F = BB->getParent();
224 MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(LLVM_BB);
225 MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(LLVM_BB);
226 MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(LLVM_BB);
228 MachineFunction::iterator It = ++BB->getIterator();
229 F->insert(It, FalseMBB);
230 F->insert(It, TrueMBB);
231 F->insert(It, DoneMBB);
233 // Transfer the remainder of BB and its successor edges to DoneMBB.
234 DoneMBB->splice(DoneMBB->begin(), BB,
235 std::next(MachineBasicBlock::iterator(MI)),
237 DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
239 BB->addSuccessor(TrueMBB);
240 BB->addSuccessor(FalseMBB);
241 TrueMBB->addSuccessor(DoneMBB);
242 FalseMBB->addSuccessor(DoneMBB);
244 unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg;
245 Tmp0 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
246 Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
247 CmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
248 EqzReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
249 FalseReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
250 TrueReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
252 MI.eraseFromParent();
253 // For signed numbers, we can do a single comparison to determine whether
254 // fabs(x) is within range.
258 BuildMI(BB, DL, TII.get(Abs), Tmp0)
261 BuildMI(BB, DL, TII.get(FConst), Tmp1)
262 .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, CmpVal)));
263 BuildMI(BB, DL, TII.get(LT), CmpReg)
267 // For unsigned numbers, we have to do a separate comparison with zero.
269 Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
270 unsigned SecondCmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
271 unsigned AndReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
272 BuildMI(BB, DL, TII.get(FConst), Tmp1)
273 .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, 0.0)));
274 BuildMI(BB, DL, TII.get(GE), SecondCmpReg)
277 BuildMI(BB, DL, TII.get(And), AndReg)
279 .addReg(SecondCmpReg);
283 BuildMI(BB, DL, TII.get(Eqz), EqzReg)
286 // Create the CFG diamond to select between doing the conversion or using
287 // the substitute value.
288 BuildMI(BB, DL, TII.get(WebAssembly::BR_IF))
291 BuildMI(FalseMBB, DL, TII.get(LoweredOpcode), FalseReg)
293 BuildMI(FalseMBB, DL, TII.get(WebAssembly::BR))
295 BuildMI(TrueMBB, DL, TII.get(IConst), TrueReg)
297 BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(TargetOpcode::PHI), OutReg)
307 WebAssemblyTargetLowering::EmitInstrWithCustomInserter(
309 MachineBasicBlock *BB
311 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
312 DebugLoc DL = MI.getDebugLoc();
314 switch (MI.getOpcode()) {
315 default: llvm_unreachable("Unexpected instr type to insert");
316 case WebAssembly::FP_TO_SINT_I32_F32:
317 return LowerFPToInt(MI, DL, BB, TII, false, false, false,
318 WebAssembly::I32_TRUNC_S_F32);
319 case WebAssembly::FP_TO_UINT_I32_F32:
320 return LowerFPToInt(MI, DL, BB, TII, true, false, false,
321 WebAssembly::I32_TRUNC_U_F32);
322 case WebAssembly::FP_TO_SINT_I64_F32:
323 return LowerFPToInt(MI, DL, BB, TII, false, true, false,
324 WebAssembly::I64_TRUNC_S_F32);
325 case WebAssembly::FP_TO_UINT_I64_F32:
326 return LowerFPToInt(MI, DL, BB, TII, true, true, false,
327 WebAssembly::I64_TRUNC_U_F32);
328 case WebAssembly::FP_TO_SINT_I32_F64:
329 return LowerFPToInt(MI, DL, BB, TII, false, false, true,
330 WebAssembly::I32_TRUNC_S_F64);
331 case WebAssembly::FP_TO_UINT_I32_F64:
332 return LowerFPToInt(MI, DL, BB, TII, true, false, true,
333 WebAssembly::I32_TRUNC_U_F64);
334 case WebAssembly::FP_TO_SINT_I64_F64:
335 return LowerFPToInt(MI, DL, BB, TII, false, true, true,
336 WebAssembly::I64_TRUNC_S_F64);
337 case WebAssembly::FP_TO_UINT_I64_F64:
338 return LowerFPToInt(MI, DL, BB, TII, true, true, true,
339 WebAssembly::I64_TRUNC_U_F64);
340 llvm_unreachable("Unexpected instruction to emit with custom inserter");
344 const char *WebAssemblyTargetLowering::getTargetNodeName(
345 unsigned Opcode) const {
346 switch (static_cast<WebAssemblyISD::NodeType>(Opcode)) {
347 case WebAssemblyISD::FIRST_NUMBER:
349 #define HANDLE_NODETYPE(NODE) \
350 case WebAssemblyISD::NODE: \
351 return "WebAssemblyISD::" #NODE;
352 #include "WebAssemblyISD.def"
353 #undef HANDLE_NODETYPE
358 std::pair<unsigned, const TargetRegisterClass *>
359 WebAssemblyTargetLowering::getRegForInlineAsmConstraint(
360 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
361 // First, see if this is a constraint that directly corresponds to a
362 // WebAssembly register class.
363 if (Constraint.size() == 1) {
364 switch (Constraint[0]) {
366 assert(VT != MVT::iPTR && "Pointer MVT not expected here");
367 if (Subtarget->hasSIMD128() && VT.isVector()) {
368 if (VT.getSizeInBits() == 128)
369 return std::make_pair(0U, &WebAssembly::V128RegClass);
371 if (VT.isInteger() && !VT.isVector()) {
372 if (VT.getSizeInBits() <= 32)
373 return std::make_pair(0U, &WebAssembly::I32RegClass);
374 if (VT.getSizeInBits() <= 64)
375 return std::make_pair(0U, &WebAssembly::I64RegClass);
383 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
386 bool WebAssemblyTargetLowering::isCheapToSpeculateCttz() const {
387 // Assume ctz is a relatively cheap operation.
391 bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz() const {
392 // Assume clz is a relatively cheap operation.
396 bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL,
400 Instruction *I) const {
401 // WebAssembly offsets are added as unsigned without wrapping. The
402 // isLegalAddressingMode gives us no way to determine if wrapping could be
403 // happening, so we approximate this by accepting only non-negative offsets.
404 if (AM.BaseOffs < 0) return false;
406 // WebAssembly has no scale register operands.
407 if (AM.Scale != 0) return false;
409 // Everything else is legal.
413 bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses(
414 EVT /*VT*/, unsigned /*AddrSpace*/, unsigned /*Align*/, bool *Fast) const {
415 // WebAssembly supports unaligned accesses, though it should be declared
416 // with the p2align attribute on loads and stores which do so, and there
417 // may be a performance impact. We tell LLVM they're "fast" because
418 // for the kinds of things that LLVM uses this for (merging adjacent stores
419 // of constants, etc.), WebAssembly implementations will either want the
420 // unaligned access or they'll split anyway.
421 if (Fast) *Fast = true;
425 bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT,
426 AttributeList Attr) const {
427 // The current thinking is that wasm engines will perform this optimization,
428 // so we can save on code size.
432 EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
436 return VT.changeVectorElementTypeToInteger();
438 return TargetLowering::getSetCCResultType(DL, C, VT);
441 //===----------------------------------------------------------------------===//
442 // WebAssembly Lowering private implementation.
443 //===----------------------------------------------------------------------===//
445 //===----------------------------------------------------------------------===//
447 //===----------------------------------------------------------------------===//
449 static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *msg) {
450 MachineFunction &MF = DAG.getMachineFunction();
451 DAG.getContext()->diagnose(
452 DiagnosticInfoUnsupported(MF.getFunction(), msg, DL.getDebugLoc()));
455 // Test whether the given calling convention is supported.
456 static bool CallingConvSupported(CallingConv::ID CallConv) {
457 // We currently support the language-independent target-independent
458 // conventions. We don't yet have a way to annotate calls with properties like
459 // "cold", and we don't have any call-clobbered registers, so these are mostly
460 // all handled the same.
461 return CallConv == CallingConv::C || CallConv == CallingConv::Fast ||
462 CallConv == CallingConv::Cold ||
463 CallConv == CallingConv::PreserveMost ||
464 CallConv == CallingConv::PreserveAll ||
465 CallConv == CallingConv::CXX_FAST_TLS;
468 SDValue WebAssemblyTargetLowering::LowerCall(
469 CallLoweringInfo &CLI, SmallVectorImpl<SDValue> &InVals) const {
470 SelectionDAG &DAG = CLI.DAG;
472 SDValue Chain = CLI.Chain;
473 SDValue Callee = CLI.Callee;
474 MachineFunction &MF = DAG.getMachineFunction();
475 auto Layout = MF.getDataLayout();
477 CallingConv::ID CallConv = CLI.CallConv;
478 if (!CallingConvSupported(CallConv))
480 "WebAssembly doesn't support language-specific or target-specific "
481 "calling conventions yet");
482 if (CLI.IsPatchPoint)
483 fail(DL, DAG, "WebAssembly doesn't support patch point yet");
485 // WebAssembly doesn't currently support explicit tail calls. If they are
486 // required, fail. Otherwise, just disable them.
487 if ((CallConv == CallingConv::Fast && CLI.IsTailCall &&
488 MF.getTarget().Options.GuaranteedTailCallOpt) ||
489 (CLI.CS && CLI.CS.isMustTailCall()))
490 fail(DL, DAG, "WebAssembly doesn't support tail call yet");
491 CLI.IsTailCall = false;
493 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
495 fail(DL, DAG, "WebAssembly doesn't support more than 1 returned value yet");
497 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
498 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
499 unsigned NumFixedArgs = 0;
500 for (unsigned i = 0; i < Outs.size(); ++i) {
501 const ISD::OutputArg &Out = Outs[i];
502 SDValue &OutVal = OutVals[i];
503 if (Out.Flags.isNest())
504 fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
505 if (Out.Flags.isInAlloca())
506 fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
507 if (Out.Flags.isInConsecutiveRegs())
508 fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
509 if (Out.Flags.isInConsecutiveRegsLast())
510 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
511 if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) {
512 auto &MFI = MF.getFrameInfo();
513 int FI = MFI.CreateStackObject(Out.Flags.getByValSize(),
514 Out.Flags.getByValAlign(),
517 DAG.getConstant(Out.Flags.getByValSize(), DL, MVT::i32);
518 SDValue FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
519 Chain = DAG.getMemcpy(
520 Chain, DL, FINode, OutVal, SizeNode, Out.Flags.getByValAlign(),
521 /*isVolatile*/ false, /*AlwaysInline=*/false,
522 /*isTailCall*/ false, MachinePointerInfo(), MachinePointerInfo());
525 // Count the number of fixed args *after* legalization.
526 NumFixedArgs += Out.IsFixed;
529 bool IsVarArg = CLI.IsVarArg;
530 auto PtrVT = getPointerTy(Layout);
532 // Analyze operands of the call, assigning locations to each operand.
533 SmallVector<CCValAssign, 16> ArgLocs;
534 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
537 // Outgoing non-fixed arguments are placed in a buffer. First
538 // compute their offsets and the total amount of buffer space needed.
540 make_range(OutVals.begin() + NumFixedArgs, OutVals.end())) {
541 EVT VT = Arg.getValueType();
542 assert(VT != MVT::iPTR && "Legalized args should be concrete");
543 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
544 unsigned Offset = CCInfo.AllocateStack(Layout.getTypeAllocSize(Ty),
545 Layout.getABITypeAlignment(Ty));
546 CCInfo.addLoc(CCValAssign::getMem(ArgLocs.size(), VT.getSimpleVT(),
547 Offset, VT.getSimpleVT(),
552 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
555 if (IsVarArg && NumBytes) {
556 // For non-fixed arguments, next emit stores to store the argument values
557 // to the stack buffer at the offsets computed above.
558 int FI = MF.getFrameInfo().CreateStackObject(NumBytes,
559 Layout.getStackAlignment(),
562 SmallVector<SDValue, 8> Chains;
564 make_range(OutVals.begin() + NumFixedArgs, OutVals.end())) {
565 assert(ArgLocs[ValNo].getValNo() == ValNo &&
566 "ArgLocs should remain in order and only hold varargs args");
567 unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
568 FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
569 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, FINode,
570 DAG.getConstant(Offset, DL, PtrVT));
571 Chains.push_back(DAG.getStore(
573 MachinePointerInfo::getFixedStack(MF, FI, Offset), 0));
576 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
577 } else if (IsVarArg) {
578 FINode = DAG.getIntPtrConstant(0, DL);
581 // Compute the operands for the CALLn node.
582 SmallVector<SDValue, 16> Ops;
583 Ops.push_back(Chain);
584 Ops.push_back(Callee);
586 // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
588 Ops.append(OutVals.begin(),
589 IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
590 // Add a pointer to the vararg buffer.
591 if (IsVarArg) Ops.push_back(FINode);
593 SmallVector<EVT, 8> InTys;
594 for (const auto &In : Ins) {
595 assert(!In.Flags.isByVal() && "byval is not valid for return values");
596 assert(!In.Flags.isNest() && "nest is not valid for return values");
597 if (In.Flags.isInAlloca())
598 fail(DL, DAG, "WebAssembly hasn't implemented inalloca return values");
599 if (In.Flags.isInConsecutiveRegs())
600 fail(DL, DAG, "WebAssembly hasn't implemented cons regs return values");
601 if (In.Flags.isInConsecutiveRegsLast())
603 "WebAssembly hasn't implemented cons regs last return values");
604 // Ignore In.getOrigAlign() because all our arguments are passed in
606 InTys.push_back(In.VT);
608 InTys.push_back(MVT::Other);
609 SDVTList InTyList = DAG.getVTList(InTys);
611 DAG.getNode(Ins.empty() ? WebAssemblyISD::CALL0 : WebAssemblyISD::CALL1,
616 InVals.push_back(Res);
617 Chain = Res.getValue(1);
623 bool WebAssemblyTargetLowering::CanLowerReturn(
624 CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
625 const SmallVectorImpl<ISD::OutputArg> &Outs,
626 LLVMContext & /*Context*/) const {
627 // WebAssembly can't currently handle returning tuples.
628 return Outs.size() <= 1;
631 SDValue WebAssemblyTargetLowering::LowerReturn(
632 SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
633 const SmallVectorImpl<ISD::OutputArg> &Outs,
634 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
635 SelectionDAG &DAG) const {
636 assert(Outs.size() <= 1 && "WebAssembly can only return up to one value");
637 if (!CallingConvSupported(CallConv))
638 fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
640 SmallVector<SDValue, 4> RetOps(1, Chain);
641 RetOps.append(OutVals.begin(), OutVals.end());
642 Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps);
644 // Record the number and types of the return values.
645 for (const ISD::OutputArg &Out : Outs) {
646 assert(!Out.Flags.isByVal() && "byval is not valid for return values");
647 assert(!Out.Flags.isNest() && "nest is not valid for return values");
648 assert(Out.IsFixed && "non-fixed return value is not valid");
649 if (Out.Flags.isInAlloca())
650 fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
651 if (Out.Flags.isInConsecutiveRegs())
652 fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
653 if (Out.Flags.isInConsecutiveRegsLast())
654 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
660 SDValue WebAssemblyTargetLowering::LowerFormalArguments(
661 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
662 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
663 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
664 if (!CallingConvSupported(CallConv))
665 fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
667 MachineFunction &MF = DAG.getMachineFunction();
668 auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>();
670 // Set up the incoming ARGUMENTS value, which serves to represent the liveness
671 // of the incoming values before they're represented by virtual registers.
672 MF.getRegInfo().addLiveIn(WebAssembly::ARGUMENTS);
674 for (const ISD::InputArg &In : Ins) {
675 if (In.Flags.isInAlloca())
676 fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
677 if (In.Flags.isNest())
678 fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
679 if (In.Flags.isInConsecutiveRegs())
680 fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
681 if (In.Flags.isInConsecutiveRegsLast())
682 fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
683 // Ignore In.getOrigAlign() because all our arguments are passed in
687 ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT,
688 DAG.getTargetConstant(InVals.size(), DL, MVT::i32))
689 : DAG.getUNDEF(In.VT));
691 // Record the number and types of arguments.
692 MFI->addParam(In.VT);
695 // Varargs are copied into a buffer allocated by the caller, and a pointer to
696 // the buffer is passed as an argument.
698 MVT PtrVT = getPointerTy(MF.getDataLayout());
699 unsigned VarargVreg =
700 MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrVT));
701 MFI->setVarargBufferVreg(VarargVreg);
702 Chain = DAG.getCopyToReg(
703 Chain, DL, VarargVreg,
704 DAG.getNode(WebAssemblyISD::ARGUMENT, DL, PtrVT,
705 DAG.getTargetConstant(Ins.size(), DL, MVT::i32)));
706 MFI->addParam(PtrVT);
709 // Record the number and types of results.
710 SmallVector<MVT, 4> Params;
711 SmallVector<MVT, 4> Results;
712 ComputeSignatureVTs(MF.getFunction(), DAG.getTarget(), Params, Results);
713 for (MVT VT : Results)
719 //===----------------------------------------------------------------------===//
720 // Custom lowering hooks.
721 //===----------------------------------------------------------------------===//
723 SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
724 SelectionDAG &DAG) const {
726 switch (Op.getOpcode()) {
728 llvm_unreachable("unimplemented operation lowering");
730 case ISD::FrameIndex:
731 return LowerFrameIndex(Op, DAG);
732 case ISD::GlobalAddress:
733 return LowerGlobalAddress(Op, DAG);
734 case ISD::ExternalSymbol:
735 return LowerExternalSymbol(Op, DAG);
737 return LowerJumpTable(Op, DAG);
739 return LowerBR_JT(Op, DAG);
741 return LowerVASTART(Op, DAG);
742 case ISD::BlockAddress:
744 fail(DL, DAG, "WebAssembly hasn't implemented computed gotos");
746 case ISD::RETURNADDR: // Probably nothing meaningful can be returned here.
747 fail(DL, DAG, "WebAssembly hasn't implemented __builtin_return_address");
750 return LowerFRAMEADDR(Op, DAG);
752 return LowerCopyToReg(Op, DAG);
753 case ISD::INTRINSIC_WO_CHAIN:
754 return LowerINTRINSIC_WO_CHAIN(Op, DAG);
758 SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op,
759 SelectionDAG &DAG) const {
760 SDValue Src = Op.getOperand(2);
761 if (isa<FrameIndexSDNode>(Src.getNode())) {
762 // CopyToReg nodes don't support FrameIndex operands. Other targets select
763 // the FI to some LEA-like instruction, but since we don't have that, we
764 // need to insert some kind of instruction that can take an FI operand and
765 // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
766 // copy_local between Op and its FI operand.
767 SDValue Chain = Op.getOperand(0);
769 unsigned Reg = cast<RegisterSDNode>(Op.getOperand(1))->getReg();
770 EVT VT = Src.getValueType();
772 DAG.getMachineNode(VT == MVT::i32 ? WebAssembly::COPY_I32
773 : WebAssembly::COPY_I64,
776 return Op.getNode()->getNumValues() == 1
777 ? DAG.getCopyToReg(Chain, DL, Reg, Copy)
778 : DAG.getCopyToReg(Chain, DL, Reg, Copy, Op.getNumOperands() == 4
785 SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op,
786 SelectionDAG &DAG) const {
787 int FI = cast<FrameIndexSDNode>(Op)->getIndex();
788 return DAG.getTargetFrameIndex(FI, Op.getValueType());
791 SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op,
792 SelectionDAG &DAG) const {
793 // Non-zero depths are not supported by WebAssembly currently. Use the
794 // legalizer's default expansion, which is to return 0 (what this function is
795 // documented to do).
796 if (Op.getConstantOperandVal(0) > 0)
799 DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true);
800 EVT VT = Op.getValueType();
802 Subtarget->getRegisterInfo()->getFrameRegister(DAG.getMachineFunction());
803 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), FP, VT);
806 SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
807 SelectionDAG &DAG) const {
809 const auto *GA = cast<GlobalAddressSDNode>(Op);
810 EVT VT = Op.getValueType();
811 assert(GA->getTargetFlags() == 0 &&
812 "Unexpected target flags on generic GlobalAddressSDNode");
813 if (GA->getAddressSpace() != 0)
814 fail(DL, DAG, "WebAssembly only expects the 0 address space");
816 WebAssemblyISD::Wrapper, DL, VT,
817 DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT, GA->getOffset()));
820 SDValue WebAssemblyTargetLowering::LowerExternalSymbol(
821 SDValue Op, SelectionDAG &DAG) const {
823 const auto *ES = cast<ExternalSymbolSDNode>(Op);
824 EVT VT = Op.getValueType();
825 assert(ES->getTargetFlags() == 0 &&
826 "Unexpected target flags on generic ExternalSymbolSDNode");
827 // Set the TargetFlags to 0x1 which indicates that this is a "function"
828 // symbol rather than a data symbol. We do this unconditionally even though
829 // we don't know anything about the symbol other than its name, because all
830 // external symbols used in target-independent SelectionDAG code are for
832 return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
833 DAG.getTargetExternalSymbol(ES->getSymbol(), VT,
834 /*TargetFlags=*/0x1));
837 SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op,
838 SelectionDAG &DAG) const {
839 // There's no need for a Wrapper node because we always incorporate a jump
840 // table operand into a BR_TABLE instruction, rather than ever
841 // materializing it in a register.
842 const JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
843 return DAG.getTargetJumpTable(JT->getIndex(), Op.getValueType(),
844 JT->getTargetFlags());
847 SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
848 SelectionDAG &DAG) const {
850 SDValue Chain = Op.getOperand(0);
851 const auto *JT = cast<JumpTableSDNode>(Op.getOperand(1));
852 SDValue Index = Op.getOperand(2);
853 assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags");
855 SmallVector<SDValue, 8> Ops;
856 Ops.push_back(Chain);
857 Ops.push_back(Index);
859 MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
860 const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs;
862 // Add an operand for each case.
863 for (auto MBB : MBBs) Ops.push_back(DAG.getBasicBlock(MBB));
865 // TODO: For now, we just pick something arbitrary for a default case for now.
866 // We really want to sniff out the guard and put in the real default case (and
867 // delete the guard).
868 Ops.push_back(DAG.getBasicBlock(MBBs[0]));
870 return DAG.getNode(WebAssemblyISD::BR_TABLE, DL, MVT::Other, Ops);
873 SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
874 SelectionDAG &DAG) const {
876 EVT PtrVT = getPointerTy(DAG.getMachineFunction().getDataLayout());
878 auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>();
879 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
881 SDValue ArgN = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
882 MFI->getVarargBufferVreg(), PtrVT);
883 return DAG.getStore(Op.getOperand(0), DL, ArgN, Op.getOperand(1),
884 MachinePointerInfo(SV), 0);
888 WebAssemblyTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
889 SelectionDAG &DAG) const {
890 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
894 return {}; // Don't custom lower most intrinsics.
896 case Intrinsic::wasm_lsda:
897 // TODO For now, just return 0 not to crash
898 return DAG.getConstant(0, DL, Op.getValueType());
902 //===----------------------------------------------------------------------===//
903 // WebAssembly Optimization Hooks
904 //===----------------------------------------------------------------------===//