blob: 0df8f3e0e09c065b6ead980ab4850a8c7d25f3be [file] [log] [blame]
//=- WebAssemblyISelLowering.cpp - WebAssembly DAG Lowering Implementation -==//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
/// \file
/// This file implements the WebAssemblyTargetLowering class.
#include "WebAssemblyISelLowering.h"
#include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
#include "Utils/WebAssemblyTypeUtilities.h"
#include "Utils/WebAssemblyUtilities.h"
#include "WebAssemblyMachineFunctionInfo.h"
#include "WebAssemblySubtarget.h"
#include "WebAssemblyTargetMachine.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/DiagnosticPrinter.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsWebAssembly.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
#define DEBUG_TYPE "wasm-lower"
const TargetMachine &TM, const WebAssemblySubtarget &STI)
: TargetLowering(TM), Subtarget(&STI) {
auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32;
// Booleans always contain 0 or 1.
// Except in SIMD vectors
// We don't know the microarchitecture here, so just reduce register pressure.
// Tell ISel that we have a stack pointer.
Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32);
// Set up the register classes.
addRegisterClass(MVT::i32, &WebAssembly::I32RegClass);
addRegisterClass(MVT::i64, &WebAssembly::I64RegClass);
addRegisterClass(MVT::f32, &WebAssembly::F32RegClass);
addRegisterClass(MVT::f64, &WebAssembly::F64RegClass);
if (Subtarget->hasSIMD128()) {
addRegisterClass(MVT::v16i8, &WebAssembly::V128RegClass);
addRegisterClass(MVT::v8i16, &WebAssembly::V128RegClass);
addRegisterClass(MVT::v4i32, &WebAssembly::V128RegClass);
addRegisterClass(MVT::v4f32, &WebAssembly::V128RegClass);
addRegisterClass(MVT::v2i64, &WebAssembly::V128RegClass);
addRegisterClass(MVT::v2f64, &WebAssembly::V128RegClass);
if (Subtarget->hasReferenceTypes()) {
addRegisterClass(MVT::externref, &WebAssembly::EXTERNREFRegClass);
addRegisterClass(MVT::funcref, &WebAssembly::FUNCREFRegClass);
// Compute derived properties from the register classes.
// Transform loads and stores to pointers in address space 1 to loads and
// stores to WebAssembly global variables, outside linear memory.
for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) {
setOperationAction(ISD::LOAD, T, Custom);
setOperationAction(ISD::STORE, T, Custom);
if (Subtarget->hasSIMD128()) {
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
MVT::v2f64}) {
setOperationAction(ISD::LOAD, T, Custom);
setOperationAction(ISD::STORE, T, Custom);
if (Subtarget->hasReferenceTypes()) {
// We need custom load and store lowering for both externref, funcref and
// Other. The MVT::Other here represents tables of reference types.
for (auto T : {MVT::externref, MVT::funcref, MVT::Other}) {
setOperationAction(ISD::LOAD, T, Custom);
setOperationAction(ISD::STORE, T, Custom);
setOperationAction(ISD::GlobalAddress, MVTPtr, Custom);
setOperationAction(ISD::GlobalTLSAddress, MVTPtr, Custom);
setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom);
setOperationAction(ISD::JumpTable, MVTPtr, Custom);
setOperationAction(ISD::BlockAddress, MVTPtr, Custom);
setOperationAction(ISD::BRIND, MVT::Other, Custom);
// Take the default expansion for va_arg, va_copy, and va_end. There is no
// default action for va_start, so we do that custom.
setOperationAction(ISD::VASTART, MVT::Other, Custom);
setOperationAction(ISD::VAARG, MVT::Other, Expand);
setOperationAction(ISD::VACOPY, MVT::Other, Expand);
setOperationAction(ISD::VAEND, MVT::Other, Expand);
for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
// Don't expand the floating-point types to constant pools.
setOperationAction(ISD::ConstantFP, T, Legal);
// Expand floating-point comparisons.
setCondCodeAction(CC, T, Expand);
// Expand floating-point library function operators.
for (auto Op :
setOperationAction(Op, T, Expand);
// Note supported floating-point library function operators that otherwise
// default to expand.
for (auto Op :
setOperationAction(Op, T, Legal);
// Support minimum and maximum, which otherwise default to expand.
setOperationAction(ISD::FMINIMUM, T, Legal);
setOperationAction(ISD::FMAXIMUM, T, Legal);
// WebAssembly currently has no builtin f16 support.
setOperationAction(ISD::FP16_TO_FP, T, Expand);
setOperationAction(ISD::FP_TO_FP16, T, Expand);
setLoadExtAction(ISD::EXTLOAD, T, MVT::f16, Expand);
setTruncStoreAction(T, MVT::f16, Expand);
// Expand unavailable integer operations.
for (auto Op :
for (auto T : {MVT::i32, MVT::i64})
setOperationAction(Op, T, Expand);
if (Subtarget->hasSIMD128())
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
setOperationAction(Op, T, Expand);
if (Subtarget->hasNontrappingFPToInt())
for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT})
for (auto T : {MVT::i32, MVT::i64})
setOperationAction(Op, T, Custom);
// SIMD-specific configuration
if (Subtarget->hasSIMD128()) {
// Hoist bitcasts out of shuffles
// Combine extends of extract_subvectors into widening ops
// Combine int_to_fp or fp_extend of extract_vectors and vice versa into
// conversions ops
// Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa
// into conversion ops
// Support saturating add for i8x16 and i16x8
for (auto Op : {ISD::SADDSAT, ISD::UADDSAT})
for (auto T : {MVT::v16i8, MVT::v8i16})
setOperationAction(Op, T, Legal);
// Support integer abs
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
setOperationAction(ISD::ABS, T, Legal);
// Custom lower BUILD_VECTORs to minimize number of replace_lanes
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
setOperationAction(ISD::BUILD_VECTOR, T, Custom);
// We have custom shuffle lowering to expose the shuffle mask
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom);
// Custom lowering since wasm shifts must have a scalar shift amount
for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL})
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
setOperationAction(Op, T, Custom);
// Custom lower lane accesses to expand out variable indices
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
setOperationAction(Op, T, Custom);
// There is no i8x16.mul instruction
setOperationAction(ISD::MUL, MVT::v16i8, Expand);
// There is no vector conditional select instruction
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
setOperationAction(ISD::SELECT_CC, T, Expand);
// Expand integer operations supported for scalars but not SIMD
for (auto Op :
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
setOperationAction(Op, T, Expand);
// But we do have integer min and max operations
for (auto Op : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
setOperationAction(Op, T, Legal);
// And we have popcnt for i8x16. It can be used to expand ctlz/cttz.
setOperationAction(ISD::CTPOP, MVT::v16i8, Legal);
setOperationAction(ISD::CTLZ, MVT::v16i8, Expand);
setOperationAction(ISD::CTTZ, MVT::v16i8, Expand);
// Custom lower bit counting operations for other types to scalarize them.
for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP})
for (auto T : {MVT::v8i16, MVT::v4i32, MVT::v2i64})
setOperationAction(Op, T, Custom);
// Expand float operations supported for scalars but not SIMD
for (auto T : {MVT::v4f32, MVT::v2f64})
setOperationAction(Op, T, Expand);
// Unsigned comparison operations are unavailable for i64x2 vectors.
setCondCodeAction(CC, MVT::v2i64, Custom);
// 64x2 conversions are not in the spec
for (auto Op :
for (auto T : {MVT::v2i64, MVT::v2f64})
setOperationAction(Op, T, Expand);
// But saturating fp_to_int converstions are
for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT})
setOperationAction(Op, MVT::v4i32, Custom);
// As a special case, these operators use the type to mean the type to
// sign-extend from.
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
if (!Subtarget->hasSignExt()) {
// Sign extends are legal only when extending a vector extract
auto Action = Subtarget->hasSIMD128() ? Custom : Expand;
for (auto T : {MVT::i8, MVT::i16, MVT::i32})
setOperationAction(ISD::SIGN_EXTEND_INREG, T, Action);
for (auto T : MVT::integer_fixedlen_vector_valuetypes())
setOperationAction(ISD::SIGN_EXTEND_INREG, T, Expand);
// Dynamic stack allocation: use the default expansion.
setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVTPtr, Expand);
setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
setOperationAction(ISD::FrameIndex, MVT::i64, Custom);
setOperationAction(ISD::CopyToReg, MVT::Other, Custom);
// Expand these forms; we pattern-match the forms that we can handle in isel.
for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64})
for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
setOperationAction(Op, T, Expand);
// We have custom switch handling.
setOperationAction(ISD::BR_JT, MVT::Other, Custom);
// WebAssembly doesn't have:
// - Floating-point extending loads.
// - Floating-point truncating stores.
// - i1 extending loads.
// - truncating SIMD stores and most extending loads
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
for (auto T : MVT::integer_valuetypes())
setLoadExtAction(Ext, T, MVT::i1, Promote);
if (Subtarget->hasSIMD128()) {
for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32,
MVT::v2f64}) {
for (auto MemT : MVT::fixedlen_vector_valuetypes()) {
if (MVT(T) != MemT) {
setTruncStoreAction(T, MemT, Expand);
setLoadExtAction(Ext, T, MemT, Expand);
// But some vector extending loads are legal
setLoadExtAction(Ext, MVT::v8i16, MVT::v8i8, Legal);
setLoadExtAction(Ext, MVT::v4i32, MVT::v4i16, Legal);
setLoadExtAction(Ext, MVT::v2i64, MVT::v2i32, Legal);
setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Legal);
// Don't do anything clever with build_pairs
setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
// Trap lowers to wasm unreachable
setOperationAction(ISD::TRAP, MVT::Other, Legal);
setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
// Exception handling intrinsics
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
// Override the __gnu_f2h_ieee/__gnu_h2f_ieee names so that the f32 name is
// consistent with the f64 and f128 names.
setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
// Define the emscripten name for return address helper.
// TODO: when implementing other Wasm backends, make this generic or only do
// this on emscripten depending on what they end up doing.
setLibcallName(RTLIB::RETURN_ADDRESS, "emscripten_return_address");
// Always convert switches to br_tables unless there is only one case, which
// is equivalent to a simple branch. This reduces code size for wasm, and we
// defer possible jump table optimizations to the VM.
MVT WebAssemblyTargetLowering::getPointerTy(const DataLayout &DL,
uint32_t AS) const {
if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF)
return MVT::externref;
if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF)
return MVT::funcref;
return TargetLowering::getPointerTy(DL, AS);
MVT WebAssemblyTargetLowering::getPointerMemTy(const DataLayout &DL,
uint32_t AS) const {
if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF)
return MVT::externref;
if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF)
return MVT::funcref;
return TargetLowering::getPointerMemTy(DL, AS);
WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
// We have wasm instructions for these
switch (AI->getOperation()) {
case AtomicRMWInst::Add:
case AtomicRMWInst::Sub:
case AtomicRMWInst::And:
case AtomicRMWInst::Or:
case AtomicRMWInst::Xor:
case AtomicRMWInst::Xchg:
return AtomicExpansionKind::None;
return AtomicExpansionKind::CmpXChg;
bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
// Implementation copied from X86TargetLowering.
unsigned Opc = VecOp.getOpcode();
// Assume target opcodes can't be scalarized.
// TODO - do we have any exceptions?
return false;
// If the vector op is not supported, try to convert to scalar.
EVT VecVT = VecOp.getValueType();
if (!isOperationLegalOrCustomOrPromote(Opc, VecVT))
return true;
// If the vector op is supported, but the scalar op is not, the transform may
// not be worthwhile.
EVT ScalarVT = VecVT.getScalarType();
return isOperationLegalOrCustomOrPromote(Opc, ScalarVT);
FastISel *WebAssemblyTargetLowering::createFastISel(
FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo) const {
return WebAssembly::createFastISel(FuncInfo, LibInfo);
MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
EVT VT) const {
unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
if (BitWidth > 1 && BitWidth < 8)
BitWidth = 8;
if (BitWidth > 64) {
// The shift will be lowered to a libcall, and compiler-rt libcalls expect
// the count to be an i32.
BitWidth = 32;
assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
"32-bit shift counts ought to be enough for anyone");
MVT Result = MVT::getIntegerVT(BitWidth);
"Unable to represent scalar shift amount type");
return Result;
// Lower an fp-to-int conversion operator from the LLVM opcode, which has an
// undefined result on invalid/overflow, to the WebAssembly opcode, which
// traps on invalid/overflow.
static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL,
MachineBasicBlock *BB,
const TargetInstrInfo &TII,
bool IsUnsigned, bool Int64,
bool Float64, unsigned LoweredOpcode) {
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
Register OutReg = MI.getOperand(0).getReg();
Register InReg = MI.getOperand(1).getReg();
unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32;
unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32;
unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32;
unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32;
unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32;
unsigned Eqz = WebAssembly::EQZ_I32;
unsigned And = WebAssembly::AND_I32;
int64_t Limit = Int64 ? INT64_MIN : INT32_MIN;
int64_t Substitute = IsUnsigned ? 0 : Limit;
double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit;
auto &Context = BB->getParent()->getFunction().getContext();
Type *Ty = Float64 ? Type::getDoubleTy(Context) : Type::getFloatTy(Context);
const BasicBlock *LLVMBB = BB->getBasicBlock();
MachineFunction *F = BB->getParent();
MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(LLVMBB);
MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(LLVMBB);
MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(LLVMBB);
MachineFunction::iterator It = ++BB->getIterator();
F->insert(It, FalseMBB);
F->insert(It, TrueMBB);
F->insert(It, DoneMBB);
// Transfer the remainder of BB and its successor edges to DoneMBB.
DoneMBB->splice(DoneMBB->begin(), BB, std::next(MI.getIterator()), BB->end());
unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg;
Tmp0 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
CmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
EqzReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
FalseReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
TrueReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
// For signed numbers, we can do a single comparison to determine whether
// fabs(x) is within range.
if (IsUnsigned) {
Tmp0 = InReg;
} else {
BuildMI(BB, DL, TII.get(Abs), Tmp0).addReg(InReg);
BuildMI(BB, DL, TII.get(FConst), Tmp1)
.addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, CmpVal)));
BuildMI(BB, DL, TII.get(LT), CmpReg).addReg(Tmp0).addReg(Tmp1);
// For unsigned numbers, we have to do a separate comparison with zero.
if (IsUnsigned) {
Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
Register SecondCmpReg =
Register AndReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
BuildMI(BB, DL, TII.get(FConst), Tmp1)
.addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, 0.0)));
BuildMI(BB, DL, TII.get(GE), SecondCmpReg).addReg(Tmp0).addReg(Tmp1);
BuildMI(BB, DL, TII.get(And), AndReg).addReg(CmpReg).addReg(SecondCmpReg);
CmpReg = AndReg;
BuildMI(BB, DL, TII.get(Eqz), EqzReg).addReg(CmpReg);
// Create the CFG diamond to select between doing the conversion or using
// the substitute value.
BuildMI(BB, DL, TII.get(WebAssembly::BR_IF)).addMBB(TrueMBB).addReg(EqzReg);
BuildMI(FalseMBB, DL, TII.get(LoweredOpcode), FalseReg).addReg(InReg);
BuildMI(FalseMBB, DL, TII.get(WebAssembly::BR)).addMBB(DoneMBB);
BuildMI(TrueMBB, DL, TII.get(IConst), TrueReg).addImm(Substitute);
BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(TargetOpcode::PHI), OutReg)
return DoneMBB;
static MachineBasicBlock *
LowerCallResults(MachineInstr &CallResults, DebugLoc DL, MachineBasicBlock *BB,
const WebAssemblySubtarget *Subtarget,
const TargetInstrInfo &TII) {
MachineInstr &CallParams = *CallResults.getPrevNode();
assert(CallParams.getOpcode() == WebAssembly::CALL_PARAMS);
assert(CallResults.getOpcode() == WebAssembly::CALL_RESULTS ||
CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS);
bool IsIndirect = CallParams.getOperand(0).isReg();
bool IsRetCall = CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS;
bool IsFuncrefCall = false;
if (IsIndirect) {
Register Reg = CallParams.getOperand(0).getReg();
const MachineFunction *MF = BB->getParent();
const MachineRegisterInfo &MRI = MF->getRegInfo();
const TargetRegisterClass *TRC = MRI.getRegClass(Reg);
IsFuncrefCall = (TRC == &WebAssembly::FUNCREFRegClass);
assert(!IsFuncrefCall || Subtarget->hasReferenceTypes());
unsigned CallOp;
if (IsIndirect && IsRetCall) {
CallOp = WebAssembly::RET_CALL_INDIRECT;
} else if (IsIndirect) {
CallOp = WebAssembly::CALL_INDIRECT;
} else if (IsRetCall) {
CallOp = WebAssembly::RET_CALL;
} else {
CallOp = WebAssembly::CALL;
MachineFunction &MF = *BB->getParent();
const MCInstrDesc &MCID = TII.get(CallOp);
MachineInstrBuilder MIB(MF, MF.CreateMachineInstr(MCID, DL));
// See if we must truncate the function pointer.
// CALL_INDIRECT takes an i32, but in wasm64 we represent function pointers
// as 64-bit for uniformity with other pointer types.
// See also: WebAssemblyFastISel::selectCall
if (IsIndirect && MF.getSubtarget<WebAssemblySubtarget>().hasAddr64()) {
Register Reg32 =
auto &FnPtr = CallParams.getOperand(0);
BuildMI(*BB, CallResults.getIterator(), DL,
TII.get(WebAssembly::I32_WRAP_I64), Reg32)
// Move the function pointer to the end of the arguments for indirect calls
if (IsIndirect) {
auto FnPtr = CallParams.getOperand(0);
// For funcrefs, call_indirect is done through __funcref_call_table and the
// funcref is always installed in slot 0 of the table, therefore instead of having
// the function pointer added at the end of the params list, a zero (the index in
// __funcref_call_table is added).
if (IsFuncrefCall) {
Register RegZero =
MachineInstrBuilder MIBC0 =
BuildMI(MF, DL, TII.get(WebAssembly::CONST_I32), RegZero).addImm(0);
BB->insert(CallResults.getIterator(), MIBC0);
MachineInstrBuilder(MF, CallParams).addReg(RegZero);
} else
for (auto Def : CallResults.defs())
if (IsIndirect) {
// Placeholder for the type index.
// The table into which this call_indirect indexes.
MCSymbolWasm *Table = IsFuncrefCall
? WebAssembly::getOrCreateFuncrefCallTableSymbol(
MF.getContext(), Subtarget)
: WebAssembly::getOrCreateFunctionTableSymbol(
MF.getContext(), Subtarget);
if (Subtarget->hasReferenceTypes()) {
} else {
// For the MVP there is at most one table whose number is 0, but we can't
// write a table symbol or issue relocations. Instead we just ensure the
// table is live and write a zero.
for (auto Use : CallParams.uses())
BB->insert(CallResults.getIterator(), MIB);
// If this is a funcref call, to avoid hidden GC roots, we need to clear the
// table slot with ref.null upon call_indirect return.
// This generates the following code, which comes right after a call_indirect
// of a funcref:
// i32.const 0
// ref.null func
// table.set __funcref_call_table
if (IsIndirect && IsFuncrefCall) {
MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol(
MF.getContext(), Subtarget);
Register RegZero =
MachineInstr *Const0 =
BuildMI(MF, DL, TII.get(WebAssembly::CONST_I32), RegZero).addImm(0);
BB->insertAfter(MIB.getInstr()->getIterator(), Const0);
Register RegFuncref =
MachineInstr *RefNull =
BuildMI(MF, DL, TII.get(WebAssembly::REF_NULL_FUNCREF), RegFuncref)
BB->insertAfter(Const0->getIterator(), RefNull);
MachineInstr *TableSet =
BuildMI(MF, DL, TII.get(WebAssembly::TABLE_SET_FUNCREF))
BB->insertAfter(RefNull->getIterator(), TableSet);
return BB;
MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter(
MachineInstr &MI, MachineBasicBlock *BB) const {
const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
DebugLoc DL = MI.getDebugLoc();
switch (MI.getOpcode()) {
llvm_unreachable("Unexpected instr type to insert");
case WebAssembly::FP_TO_SINT_I32_F32:
return LowerFPToInt(MI, DL, BB, TII, false, false, false,
case WebAssembly::FP_TO_UINT_I32_F32:
return LowerFPToInt(MI, DL, BB, TII, true, false, false,
case WebAssembly::FP_TO_SINT_I64_F32:
return LowerFPToInt(MI, DL, BB, TII, false, true, false,
case WebAssembly::FP_TO_UINT_I64_F32:
return LowerFPToInt(MI, DL, BB, TII, true, true, false,
case WebAssembly::FP_TO_SINT_I32_F64:
return LowerFPToInt(MI, DL, BB, TII, false, false, true,
case WebAssembly::FP_TO_UINT_I32_F64:
return LowerFPToInt(MI, DL, BB, TII, true, false, true,
case WebAssembly::FP_TO_SINT_I64_F64:
return LowerFPToInt(MI, DL, BB, TII, false, true, true,
case WebAssembly::FP_TO_UINT_I64_F64:
return LowerFPToInt(MI, DL, BB, TII, true, true, true,
case WebAssembly::CALL_RESULTS:
case WebAssembly::RET_CALL_RESULTS:
return LowerCallResults(MI, DL, BB, Subtarget, TII);
const char *
WebAssemblyTargetLowering::getTargetNodeName(unsigned Opcode) const {
switch (static_cast<WebAssemblyISD::NodeType>(Opcode)) {
case WebAssemblyISD::FIRST_NUMBER:
case WebAssemblyISD::FIRST_MEM_OPCODE:
case WebAssemblyISD::NODE: \
return "WebAssemblyISD::" #NODE;
#include "WebAssemblyISD.def"
return nullptr;
std::pair<unsigned, const TargetRegisterClass *>
const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
// First, see if this is a constraint that directly corresponds to a
// WebAssembly register class.
if (Constraint.size() == 1) {
switch (Constraint[0]) {
case 'r':
assert(VT != MVT::iPTR && "Pointer MVT not expected here");
if (Subtarget->hasSIMD128() && VT.isVector()) {
if (VT.getSizeInBits() == 128)
return std::make_pair(0U, &WebAssembly::V128RegClass);
if (VT.isInteger() && !VT.isVector()) {
if (VT.getSizeInBits() <= 32)
return std::make_pair(0U, &WebAssembly::I32RegClass);
if (VT.getSizeInBits() <= 64)
return std::make_pair(0U, &WebAssembly::I64RegClass);
if (VT.isFloatingPoint() && !VT.isVector()) {
switch (VT.getSizeInBits()) {
case 32:
return std::make_pair(0U, &WebAssembly::F32RegClass);
case 64:
return std::make_pair(0U, &WebAssembly::F64RegClass);
return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
bool WebAssemblyTargetLowering::isCheapToSpeculateCttz() const {
// Assume ctz is a relatively cheap operation.
return true;
bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz() const {
// Assume clz is a relatively cheap operation.
return true;
bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL,
const AddrMode &AM,
Type *Ty, unsigned AS,
Instruction *I) const {
// WebAssembly offsets are added as unsigned without wrapping. The
// isLegalAddressingMode gives us no way to determine if wrapping could be
// happening, so we approximate this by accepting only non-negative offsets.
if (AM.BaseOffs < 0)
return false;
// WebAssembly has no scale register operands.
if (AM.Scale != 0)
return false;
// Everything else is legal.
return true;
bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses(
EVT /*VT*/, unsigned /*AddrSpace*/, Align /*Align*/,
MachineMemOperand::Flags /*Flags*/, bool *Fast) const {
// WebAssembly supports unaligned accesses, though it should be declared
// with the p2align attribute on loads and stores which do so, and there
// may be a performance impact. We tell LLVM they're "fast" because
// for the kinds of things that LLVM uses this for (merging adjacent stores
// of constants, etc.), WebAssembly implementations will either want the
// unaligned access or they'll split anyway.
if (Fast)
*Fast = true;
return true;
bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT,
AttributeList Attr) const {
// The current thinking is that wasm engines will perform this optimization,
// so we can save on code size.
return true;
bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
EVT ExtT = ExtVal.getValueType();
EVT MemT = cast<LoadSDNode>(ExtVal->getOperand(0))->getValueType(0);
return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) ||
(ExtT == MVT::v4i32 && MemT == MVT::v4i16) ||
(ExtT == MVT::v2i64 && MemT == MVT::v2i32);
bool WebAssemblyTargetLowering::isOffsetFoldingLegal(
const GlobalAddressSDNode *GA) const {
// Wasm doesn't support function addresses with offsets
const GlobalValue *GV = GA->getGlobal();
return isa<Function>(GV) ? false : TargetLowering::isOffsetFoldingLegal(GA);
EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
LLVMContext &C,
EVT VT) const {
if (VT.isVector())
return VT.changeVectorElementTypeToInteger();
// So far, all branch instructions in Wasm take an I32 condition.
// The default TargetLowering::getSetCCResultType returns the pointer size,
// which would be useful to reduce instruction counts when testing
// against 64-bit pointers/values if at some point Wasm supports that.
return EVT::getIntegerVT(C, 32);
bool WebAssemblyTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
const CallInst &I,
MachineFunction &MF,
unsigned Intrinsic) const {
switch (Intrinsic) {
case Intrinsic::wasm_memory_atomic_notify:
Info.memVT = MVT::i32;
Info.ptrVal = I.getArgOperand(0);
Info.offset = 0;
Info.align = Align(4);
// atomic.notify instruction does not really load the memory specified with
// this argument, but MachineMemOperand should either be load or store, so
// we set this to a load.
// FIXME Volatile isn't really correct, but currently all LLVM atomic
// instructions are treated as volatiles in the backend, so we should be
// consistent. The same applies for wasm_atomic_wait intrinsics too.
Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
return true;
case Intrinsic::wasm_memory_atomic_wait32:
Info.memVT = MVT::i32;
Info.ptrVal = I.getArgOperand(0);
Info.offset = 0;
Info.align = Align(4);
Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
return true;
case Intrinsic::wasm_memory_atomic_wait64:
Info.memVT = MVT::i64;
Info.ptrVal = I.getArgOperand(0);
Info.offset = 0;
Info.align = Align(8);
Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
return true;
return false;
void WebAssemblyTargetLowering::computeKnownBitsForTargetNode(
const SDValue Op, KnownBits &Known, const APInt &DemandedElts,
const SelectionDAG &DAG, unsigned Depth) const {
switch (Op.getOpcode()) {
unsigned IntNo = Op.getConstantOperandVal(0);
switch (IntNo) {
case Intrinsic::wasm_bitmask: {
unsigned BitWidth = Known.getBitWidth();
EVT VT = Op.getOperand(1).getSimpleValueType();
unsigned PossibleBits = VT.getVectorNumElements();
APInt ZeroMask = APInt::getHighBitsSet(BitWidth, BitWidth - PossibleBits);
Known.Zero |= ZeroMask;
WebAssemblyTargetLowering::getPreferredVectorAction(MVT VT) const {
if (VT.isFixedLengthVector()) {
MVT EltVT = VT.getVectorElementType();
// We have legal vector types with these lane types, so widening the
// vector would let us use some of the lanes directly without having to
// extend or truncate values.
if (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 ||
EltVT == MVT::i64 || EltVT == MVT::f32 || EltVT == MVT::f64)
return TypeWidenVector;
return TargetLoweringBase::getPreferredVectorAction(VT);
// WebAssembly Lowering private implementation.
// Lowering Code
static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) {
MachineFunction &MF = DAG.getMachineFunction();
DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc()));
// Test whether the given calling convention is supported.
static bool callingConvSupported(CallingConv::ID CallConv) {
// We currently support the language-independent target-independent
// conventions. We don't yet have a way to annotate calls with properties like
// "cold", and we don't have any call-clobbered registers, so these are mostly
// all handled the same.
return CallConv == CallingConv::C || CallConv == CallingConv::Fast ||
CallConv == CallingConv::Cold ||
CallConv == CallingConv::PreserveMost ||
CallConv == CallingConv::PreserveAll ||
CallConv == CallingConv::CXX_FAST_TLS ||
CallConv == CallingConv::WASM_EmscriptenInvoke ||
CallConv == CallingConv::Swift;
WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI,
SmallVectorImpl<SDValue> &InVals) const {
SelectionDAG &DAG = CLI.DAG;
SDValue Chain = CLI.Chain;
SDValue Callee = CLI.Callee;
MachineFunction &MF = DAG.getMachineFunction();
auto Layout = MF.getDataLayout();
CallingConv::ID CallConv = CLI.CallConv;
if (!callingConvSupported(CallConv))
fail(DL, DAG,
"WebAssembly doesn't support language-specific or target-specific "
"calling conventions yet");
if (CLI.IsPatchPoint)
fail(DL, DAG, "WebAssembly doesn't support patch point yet");
if (CLI.IsTailCall) {
auto NoTail = [&](const char *Msg) {
if (CLI.CB && CLI.CB->isMustTailCall())
fail(DL, DAG, Msg);
CLI.IsTailCall = false;
if (!Subtarget->hasTailCall())
NoTail("WebAssembly 'tail-call' feature not enabled");
// Varargs calls cannot be tail calls because the buffer is on the stack
if (CLI.IsVarArg)
NoTail("WebAssembly does not support varargs tail calls");
// Do not tail call unless caller and callee return types match
const Function &F = MF.getFunction();
const TargetMachine &TM = getTargetMachine();
Type *RetTy = F.getReturnType();
SmallVector<MVT, 4> CallerRetTys;
SmallVector<MVT, 4> CalleeRetTys;
computeLegalValueVTs(F, TM, RetTy, CallerRetTys);
computeLegalValueVTs(F, TM, CLI.RetTy, CalleeRetTys);
bool TypesMatch = CallerRetTys.size() == CalleeRetTys.size() &&
std::equal(CallerRetTys.begin(), CallerRetTys.end(),
if (!TypesMatch)
NoTail("WebAssembly tail call requires caller and callee return types to "
// If pointers to local stack values are passed, we cannot tail call
if (CLI.CB) {
for (auto &Arg : CLI.CB->args()) {
Value *Val = Arg.get();
// Trace the value back through pointer operations
while (true) {
Value *Src = Val->stripPointerCastsAndAliases();
if (auto *GEP = dyn_cast<GetElementPtrInst>(Src))
Src = GEP->getPointerOperand();
if (Val == Src)
Val = Src;
if (isa<AllocaInst>(Val)) {
"WebAssembly does not support tail calling with stack arguments");
SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
// The generic code may have added an sret argument. If we're lowering an
// invoke function, the ABI requires that the function pointer be the first
// argument, so we may have to swap the arguments.
if (CallConv == CallingConv::WASM_EmscriptenInvoke && Outs.size() >= 2 &&
Outs[0].Flags.isSRet()) {
std::swap(Outs[0], Outs[1]);
std::swap(OutVals[0], OutVals[1]);
bool HasSwiftSelfArg = false;
bool HasSwiftErrorArg = false;
unsigned NumFixedArgs = 0;
for (unsigned I = 0; I < Outs.size(); ++I) {
const ISD::OutputArg &Out = Outs[I];
SDValue &OutVal = OutVals[I];
HasSwiftSelfArg |= Out.Flags.isSwiftSelf();
HasSwiftErrorArg |= Out.Flags.isSwiftError();
if (Out.Flags.isNest())
fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
if (Out.Flags.isInAlloca())
fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
if (Out.Flags.isInConsecutiveRegs())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
if (Out.Flags.isInConsecutiveRegsLast())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) {
auto &MFI = MF.getFrameInfo();
int FI = MFI.CreateStackObject(Out.Flags.getByValSize(),
SDValue SizeNode =
DAG.getConstant(Out.Flags.getByValSize(), DL, MVT::i32);
SDValue FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
Chain = DAG.getMemcpy(
Chain, DL, FINode, OutVal, SizeNode, Out.Flags.getNonZeroByValAlign(),
/*isVolatile*/ false, /*AlwaysInline=*/false,
/*isTailCall*/ false, MachinePointerInfo(), MachinePointerInfo());
OutVal = FINode;
// Count the number of fixed args *after* legalization.
NumFixedArgs += Out.IsFixed;
bool IsVarArg = CLI.IsVarArg;
auto PtrVT = getPointerTy(Layout);
// For swiftcc, emit additional swiftself and swifterror arguments
// if there aren't. These additional arguments are also added for callee
// signature They are necessary to match callee and caller signature for
// indirect call.
if (CallConv == CallingConv::Swift) {
if (!HasSwiftSelfArg) {
ISD::OutputArg Arg;
SDValue ArgVal = DAG.getUNDEF(PtrVT);
if (!HasSwiftErrorArg) {
ISD::OutputArg Arg;
SDValue ArgVal = DAG.getUNDEF(PtrVT);
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
if (IsVarArg) {
// Outgoing non-fixed arguments are placed in a buffer. First
// compute their offsets and the total amount of buffer space needed.
for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) {
const ISD::OutputArg &Out = Outs[I];
SDValue &Arg = OutVals[I];
EVT VT = Arg.getValueType();
assert(VT != MVT::iPTR && "Legalized args should be concrete");
Type *Ty = VT.getTypeForEVT(*DAG.getContext());
Align Alignment =
std::max(Out.Flags.getNonZeroOrigAlign(), Layout.getABITypeAlign(Ty));
unsigned Offset =
CCInfo.AllocateStack(Layout.getTypeAllocSize(Ty), Alignment);
CCInfo.addLoc(CCValAssign::getMem(ArgLocs.size(), VT.getSimpleVT(),
Offset, VT.getSimpleVT(),
unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
SDValue FINode;
if (IsVarArg && NumBytes) {
// For non-fixed arguments, next emit stores to store the argument values
// to the stack buffer at the offsets computed above.
int FI = MF.getFrameInfo().CreateStackObject(NumBytes,
unsigned ValNo = 0;
SmallVector<SDValue, 8> Chains;
for (SDValue Arg : drop_begin(OutVals, NumFixedArgs)) {
assert(ArgLocs[ValNo].getValNo() == ValNo &&
"ArgLocs should remain in order and only hold varargs args");
unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, FINode,
DAG.getConstant(Offset, DL, PtrVT));
DAG.getStore(Chain, DL, Arg, Add,
MachinePointerInfo::getFixedStack(MF, FI, Offset)));
if (!Chains.empty())
Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
} else if (IsVarArg) {
FINode = DAG.getIntPtrConstant(0, DL);
if (Callee->getOpcode() == ISD::GlobalAddress) {
// If the callee is a GlobalAddress node (quite common, every direct call
// is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress
// doesn't at MO_GOT which is not needed for direct calls.
GlobalAddressSDNode* GA = cast<GlobalAddressSDNode>(Callee);
Callee = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
Callee = DAG.getNode(WebAssemblyISD::Wrapper, DL,
getPointerTy(DAG.getDataLayout()), Callee);
// Compute the operands for the CALLn node.
SmallVector<SDValue, 16> Ops;
// Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
// isn't reliable.
IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
// Add a pointer to the vararg buffer.
if (IsVarArg)
SmallVector<EVT, 8> InTys;
for (const auto &In : Ins) {
assert(!In.Flags.isByVal() && "byval is not valid for return values");
assert(!In.Flags.isNest() && "nest is not valid for return values");
if (In.Flags.isInAlloca())
fail(DL, DAG, "WebAssembly hasn't implemented inalloca return values");
if (In.Flags.isInConsecutiveRegs())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs return values");
if (In.Flags.isInConsecutiveRegsLast())
fail(DL, DAG,
"WebAssembly hasn't implemented cons regs last return values");
// Ignore In.getNonZeroOrigAlign() because all our arguments are passed in
// registers.
// Lastly, if this is a call to a funcref we need to add an instruction
// table.set to the chain and transform the call.
if (CLI.CB &&
WebAssembly::isFuncrefType(CLI.CB->getCalledOperand()->getType())) {
// In the absence of function references proposal where a funcref call is
// lowered to call_ref, using reference types we generate a table.set to set
// the funcref to a special table used solely for this purpose, followed by
// a call_indirect. Here we just generate the table set, and return the
// SDValue of the table.set so that LowerCall can finalize the lowering by
// generating the call_indirect.
SDValue Chain = Ops[0];
MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol(
MF.getContext(), Subtarget);
SDValue Sym = DAG.getMCSymbol(Table, PtrVT);
SDValue TableSlot = DAG.getConstant(0, DL, MVT::i32);
SDValue TableSetOps[] = {Chain, Sym, TableSlot, Callee};
SDValue TableSet = DAG.getMemIntrinsicNode(
WebAssemblyISD::TABLE_SET, DL, DAG.getVTList(MVT::Other), TableSetOps,
// Machine Mem Operand args
Ops[0] = TableSet; // The new chain is the TableSet itself
if (CLI.IsTailCall) {
// ret_calls do not return values to the current frame
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
return DAG.getNode(WebAssemblyISD::RET_CALL, DL, NodeTys, Ops);
SDVTList InTyList = DAG.getVTList(InTys);
SDValue Res = DAG.getNode(WebAssemblyISD::CALL, DL, InTyList, Ops);
for (size_t I = 0; I < Ins.size(); ++I)
// Return the chain
return Res.getValue(Ins.size());
bool WebAssemblyTargetLowering::CanLowerReturn(
CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
const SmallVectorImpl<ISD::OutputArg> &Outs,
LLVMContext & /*Context*/) const {
// WebAssembly can only handle returning tuples with multivalue enabled
return Subtarget->hasMultivalue() || Outs.size() <= 1;
SDValue WebAssemblyTargetLowering::LowerReturn(
SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
SelectionDAG &DAG) const {
assert((Subtarget->hasMultivalue() || Outs.size() <= 1) &&
"MVP WebAssembly can only return up to one value");
if (!callingConvSupported(CallConv))
fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
SmallVector<SDValue, 4> RetOps(1, Chain);
RetOps.append(OutVals.begin(), OutVals.end());
Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps);
// Record the number and types of the return values.
for (const ISD::OutputArg &Out : Outs) {
assert(!Out.Flags.isByVal() && "byval is not valid for return values");
assert(!Out.Flags.isNest() && "nest is not valid for return values");
assert(Out.IsFixed && "non-fixed return value is not valid");
if (Out.Flags.isInAlloca())
fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
if (Out.Flags.isInConsecutiveRegs())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
if (Out.Flags.isInConsecutiveRegsLast())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
return Chain;
SDValue WebAssemblyTargetLowering::LowerFormalArguments(
SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
if (!callingConvSupported(CallConv))
fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
MachineFunction &MF = DAG.getMachineFunction();
auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>();
// Set up the incoming ARGUMENTS value, which serves to represent the liveness
// of the incoming values before they're represented by virtual registers.
bool HasSwiftErrorArg = false;
bool HasSwiftSelfArg = false;
for (const ISD::InputArg &In : Ins) {
HasSwiftSelfArg |= In.Flags.isSwiftSelf();
HasSwiftErrorArg |= In.Flags.isSwiftError();
if (In.Flags.isInAlloca())
fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
if (In.Flags.isNest())
fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
if (In.Flags.isInConsecutiveRegs())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
if (In.Flags.isInConsecutiveRegsLast())
fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
// Ignore In.getNonZeroOrigAlign() because all our arguments are passed in
// registers.
InVals.push_back(In.Used ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT,
DL, MVT::i32))
: DAG.getUNDEF(In.VT));
// Record the number and types of arguments.
// For swiftcc, emit additional swiftself and swifterror arguments
// if there aren't. These additional arguments are also added for callee
// signature They are necessary to match callee and caller signature for
// indirect call.
auto PtrVT = getPointerTy(MF.getDataLayout());
if (CallConv == CallingConv::Swift) {
if (!HasSwiftSelfArg) {
if (!HasSwiftErrorArg) {
// Varargs are copied into a buffer allocated by the caller, and a pointer to
// the buffer is passed as an argument.
if (IsVarArg) {
MVT PtrVT = getPointerTy(MF.getDataLayout());
Register VarargVreg =
Chain = DAG.getCopyToReg(
Chain, DL, VarargVreg,
DAG.getNode(WebAssemblyISD::ARGUMENT, DL, PtrVT,
DAG.getTargetConstant(Ins.size(), DL, MVT::i32)));
// Record the number and types of arguments and results.
SmallVector<MVT, 4> Params;
SmallVector<MVT, 4> Results;
computeSignatureVTs(MF.getFunction().getFunctionType(), &MF.getFunction(),
MF.getFunction(), DAG.getTarget(), Params, Results);
for (MVT VT : Results)
// TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify
// the param logic here with ComputeSignatureVTs
assert(MFI->getParams().size() == Params.size() &&
std::equal(MFI->getParams().begin(), MFI->getParams().end(),
return Chain;
void WebAssemblyTargetLowering::ReplaceNodeResults(
SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
switch (N->getOpcode()) {
// Do not add any results, signifying that N should not be custom lowered
// after all. This happens because simd128 turns on custom lowering for
// SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an
// illegal type.
"ReplaceNodeResults not implemented for this op for WebAssembly!");
// Custom lowering hooks.
SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
switch (Op.getOpcode()) {
llvm_unreachable("unimplemented operation lowering");
return SDValue();
case ISD::FrameIndex:
return LowerFrameIndex(Op, DAG);
case ISD::GlobalAddress:
return LowerGlobalAddress(Op, DAG);
case ISD::GlobalTLSAddress:
return LowerGlobalTLSAddress(Op, DAG);
case ISD::ExternalSymbol:
return LowerExternalSymbol(Op, DAG);
case ISD::JumpTable:
return LowerJumpTable(Op, DAG);
case ISD::BR_JT:
return LowerBR_JT(Op, DAG);
return LowerVASTART(Op, DAG);
case ISD::BlockAddress:
case ISD::BRIND:
fail(DL, DAG, "WebAssembly hasn't implemented computed gotos");
return SDValue();
return LowerRETURNADDR(Op, DAG);
return LowerFRAMEADDR(Op, DAG);
case ISD::CopyToReg:
return LowerCopyToReg(Op, DAG);
return LowerAccessVectorElement(Op, DAG);
return LowerIntrinsic(Op, DAG);
return LowerBUILD_VECTOR(Op, DAG);
return LowerVECTOR_SHUFFLE(Op, DAG);
case ISD::SETCC:
return LowerSETCC(Op, DAG);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL:
return LowerShift(Op, DAG);
return LowerFP_TO_INT_SAT(Op, DAG);
case ISD::LOAD:
return LowerLoad(Op, DAG);
case ISD::STORE:
return LowerStore(Op, DAG);
case ISD::CTPOP:
case ISD::CTLZ:
case ISD::CTTZ:
return DAG.UnrollVectorOp(Op.getNode());
static bool IsWebAssemblyGlobal(SDValue Op) {
if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op))
return WebAssembly::isWasmVarAddressSpace(GA->getAddressSpace());
return false;
static Optional<unsigned> IsWebAssemblyLocal(SDValue Op, SelectionDAG &DAG) {
const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Op);
if (!FI)
return None;
auto &MF = DAG.getMachineFunction();
return WebAssemblyFrameLowering::getLocalForStackObject(MF, FI->getIndex());
static bool IsWebAssemblyTable(SDValue Op) {
const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
if (GA && WebAssembly::isWasmVarAddressSpace(GA->getAddressSpace())) {
const GlobalValue *Value = GA->getGlobal();
const Type *Ty = Value->getValueType();
if (Ty->isArrayTy() && WebAssembly::isRefType(Ty->getArrayElementType()))
return true;
return false;
// This function will accept as Op any access to a table, so Op can
// be the actual table or an offset into the table.
static bool IsWebAssemblyTableWithOffset(SDValue Op) {
if (Op->getOpcode() == ISD::ADD && Op->getNumOperands() == 2)
return (Op->getOperand(1).getSimpleValueType() == MVT::i32 &&
IsWebAssemblyTableWithOffset(Op->getOperand(0))) ||
(Op->getOperand(0).getSimpleValueType() == MVT::i32 &&
return IsWebAssemblyTable(Op);
// Helper for table pattern matching used in LowerStore and LowerLoad
bool WebAssemblyTargetLowering::MatchTableForLowering(SelectionDAG &DAG,
const SDLoc &DL,
const SDValue &Base,
GlobalAddressSDNode *&GA,
SDValue &Idx) const {
// We expect the following graph for a load of the form:
// table[<var> + <constant offset>]
// Case 1:
// externref = load t1
// t1: i32 = add t2, i32:<constant offset>
// t2: i32 = add tX, table
// This is in some cases simplified to just:
// Case 2:
// externref = load t1
// t1: i32 = add t2, i32:tX
// So, unfortunately we need to check for both cases and if we are in the
// first case extract the table GlobalAddressNode and build a new node tY
// that's tY: i32 = add i32:<constant offset>, i32:tX
if (IsWebAssemblyTable(Base)) {
GA = cast<GlobalAddressSDNode>(Base);
Idx = DAG.getConstant(0, DL, MVT::i32);
} else {
GA = dyn_cast<GlobalAddressSDNode>(Base->getOperand(0));
if (GA) {
// We are in Case 2 above.
Idx = Base->getOperand(1);
if (!Idx || GA->getNumValues() != 1 || Idx->getNumValues() != 1)
return false;
} else {
// This might be Case 1 above (or an error)
SDValue V = Base->getOperand(0);
GA = dyn_cast<GlobalAddressSDNode>(V->getOperand(1));
if (V->getOpcode() != ISD::ADD || V->getNumOperands() != 2 || !GA)
return false;
SDValue IdxV = DAG.getNode(ISD::ADD, DL, MVT::i32, Base->getOperand(1),
Idx = IdxV;
return true;
SDValue WebAssemblyTargetLowering::LowerStore(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
StoreSDNode *SN = cast<StoreSDNode>(Op.getNode());
const SDValue &Value = SN->getValue();
const SDValue &Base = SN->getBasePtr();
const SDValue &Offset = SN->getOffset();
if (IsWebAssemblyTableWithOffset(Base)) {
if (!Offset->isUndef())
"unexpected offset when loading from webassembly table", false);
SDValue Idx;
GlobalAddressSDNode *GA;
if (!MatchTableForLowering(DAG, DL, Base, GA, Idx))
report_fatal_error("failed pattern matching for lowering table store",
SDVTList Tys = DAG.getVTList(MVT::Other);
SDValue TableSetOps[] = {SN->getChain(), SDValue(GA, 0), Idx, Value};
SDValue TableSet =
DAG.getMemIntrinsicNode(WebAssemblyISD::TABLE_SET, DL, Tys, TableSetOps,
SN->getMemoryVT(), SN->getMemOperand());
return TableSet;
if (IsWebAssemblyGlobal(Base)) {
if (!Offset->isUndef())
report_fatal_error("unexpected offset when storing to webassembly global",
SDVTList Tys = DAG.getVTList(MVT::Other);
SDValue Ops[] = {SN->getChain(), Value, Base};
return DAG.getMemIntrinsicNode(WebAssemblyISD::GLOBAL_SET, DL, Tys, Ops,
SN->getMemoryVT(), SN->getMemOperand());
if (Optional<unsigned> Local = IsWebAssemblyLocal(Base, DAG)) {
if (!Offset->isUndef())
report_fatal_error("unexpected offset when storing to webassembly local",
SDValue Idx = DAG.getTargetConstant(*Local, Base, MVT::i32);
SDVTList Tys = DAG.getVTList(MVT::Other); // The chain.
SDValue Ops[] = {SN->getChain(), Idx, Value};
return DAG.getNode(WebAssemblyISD::LOCAL_SET, DL, Tys, Ops);
return Op;
SDValue WebAssemblyTargetLowering::LowerLoad(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
LoadSDNode *LN = cast<LoadSDNode>(Op.getNode());
const SDValue &Base = LN->getBasePtr();
const SDValue &Offset = LN->getOffset();
if (IsWebAssemblyTableWithOffset(Base)) {
if (!Offset->isUndef())
"unexpected offset when loading from webassembly table", false);
GlobalAddressSDNode *GA;
SDValue Idx;
if (!MatchTableForLowering(DAG, DL, Base, GA, Idx))
report_fatal_error("failed pattern matching for lowering table load",
SDVTList Tys = DAG.getVTList(LN->getValueType(0), MVT::Other);
SDValue TableGetOps[] = {LN->getChain(), SDValue(GA, 0), Idx};
SDValue TableGet =
DAG.getMemIntrinsicNode(WebAssemblyISD::TABLE_GET, DL, Tys, TableGetOps,
LN->getMemoryVT(), LN->getMemOperand());
return TableGet;
if (IsWebAssemblyGlobal(Base)) {
if (!Offset->isUndef())
"unexpected offset when loading from webassembly global", false);
SDVTList Tys = DAG.getVTList(LN->getValueType(0), MVT::Other);
SDValue Ops[] = {LN->getChain(), Base};
return DAG.getMemIntrinsicNode(WebAssemblyISD::GLOBAL_GET, DL, Tys, Ops,
LN->getMemoryVT(), LN->getMemOperand());
if (Optional<unsigned> Local = IsWebAssemblyLocal(Base, DAG)) {
if (!Offset->isUndef())
"unexpected offset when loading from webassembly local", false);
SDValue Idx = DAG.getTargetConstant(*Local, Base, MVT::i32);
EVT LocalVT = LN->getValueType(0);
SDValue LocalGet = DAG.getNode(WebAssemblyISD::LOCAL_GET, DL, LocalVT,
{LN->getChain(), Idx});
SDValue Result = DAG.getMergeValues({LocalGet, LN->getChain()}, DL);
assert(Result->getNumValues() == 2 && "Loads must carry a chain!");
return Result;
return Op;
SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op,
SelectionDAG &DAG) const {
SDValue Src = Op.getOperand(2);
if (isa<FrameIndexSDNode>(Src.getNode())) {
// CopyToReg nodes don't support FrameIndex operands. Other targets select
// the FI to some LEA-like instruction, but since we don't have that, we
// need to insert some kind of instruction that can take an FI operand and
// produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
// local.copy between Op and its FI operand.
SDValue Chain = Op.getOperand(0);
SDLoc DL(Op);
unsigned Reg = cast<RegisterSDNode>(Op.getOperand(1))->getReg();
EVT VT = Src.getValueType();
SDValue Copy(DAG.getMachineNode(VT == MVT::i32 ? WebAssembly::COPY_I32
: WebAssembly::COPY_I64,
DL, VT, Src),
return Op.getNode()->getNumValues() == 1
? DAG.getCopyToReg(Chain, DL, Reg, Copy)
: DAG.getCopyToReg(Chain, DL, Reg, Copy,
Op.getNumOperands() == 4 ? Op.getOperand(3)
: SDValue());
return SDValue();
SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op,
SelectionDAG &DAG) const {
int FI = cast<FrameIndexSDNode>(Op)->getIndex();
return DAG.getTargetFrameIndex(FI, Op.getValueType());
SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
if (!Subtarget->getTargetTriple().isOSEmscripten()) {
fail(DL, DAG,
"Non-Emscripten WebAssembly hasn't implemented "
return SDValue();
if (verifyReturnAddressArgumentIsConstant(Op, DAG))
return SDValue();
unsigned Depth = Op.getConstantOperandVal(0);
MakeLibCallOptions CallOptions;
return makeLibCall(DAG, RTLIB::RETURN_ADDRESS, Op.getValueType(),
{DAG.getConstant(Depth, DL, MVT::i32)}, CallOptions, DL)
SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op,
SelectionDAG &DAG) const {
// Non-zero depths are not supported by WebAssembly currently. Use the
// legalizer's default expansion, which is to return 0 (what this function is
// documented to do).
if (Op.getConstantOperandVal(0) > 0)
return SDValue();
EVT VT = Op.getValueType();
Register FP =
return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), FP, VT);
WebAssemblyTargetLowering::LowerGlobalTLSAddress(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
const auto *GA = cast<GlobalAddressSDNode>(Op);
MachineFunction &MF = DAG.getMachineFunction();
if (!MF.getSubtarget<WebAssemblySubtarget>().hasBulkMemory())
report_fatal_error("cannot use thread-local storage without bulk memory",
const GlobalValue *GV = GA->getGlobal();
// Currently Emscripten does not support dynamic linking with threads.
// Therefore, if we have thread-local storage, only the local-exec model
// is possible.
// TODO: remove this and implement proper TLS models once Emscripten
// supports dynamic linking with threads.
if (GV->getThreadLocalMode() != GlobalValue::LocalExecTLSModel &&
!Subtarget->getTargetTriple().isOSEmscripten()) {
report_fatal_error("only -ftls-model=local-exec is supported for now on "
"non-Emscripten OSes: variable " +
auto model = GV->getThreadLocalMode();
// Unsupported TLS modes
assert(model != GlobalValue::NotThreadLocal);
assert(model != GlobalValue::InitialExecTLSModel);
if (model == GlobalValue::LocalExecTLSModel ||
model == GlobalValue::LocalDynamicTLSModel ||
(model == GlobalValue::GeneralDynamicTLSModel &&
getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV))) {
// For DSO-local TLS variables we use offset from __tls_base
MVT PtrVT = getPointerTy(DAG.getDataLayout());
auto GlobalGet = PtrVT == MVT::i64 ? WebAssembly::GLOBAL_GET_I64
: WebAssembly::GLOBAL_GET_I32;
const char *BaseName = MF.createExternalSymbolName("__tls_base");
SDValue BaseAddr(
DAG.getMachineNode(GlobalGet, DL, PtrVT,
DAG.getTargetExternalSymbol(BaseName, PtrVT)),
SDValue TLSOffset = DAG.getTargetGlobalAddress(
GV, DL, PtrVT, GA->getOffset(), WebAssemblyII::MO_TLS_BASE_REL);
SDValue SymOffset =
DAG.getNode(WebAssemblyISD::WrapperREL, DL, PtrVT, TLSOffset);
return DAG.getNode(ISD::ADD, DL, PtrVT, BaseAddr, SymOffset);
assert(model == GlobalValue::GeneralDynamicTLSModel);
EVT VT = Op.getValueType();
return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT,
SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
const auto *GA = cast<GlobalAddressSDNode>(Op);
EVT VT = Op.getValueType();
assert(GA->getTargetFlags() == 0 &&
"Unexpected target flags on generic GlobalAddressSDNode");
if (!WebAssembly::isValidAddressSpace(GA->getAddressSpace()))
fail(DL, DAG, "Invalid address space for WebAssembly target");
unsigned OperandFlags = 0;
if (isPositionIndependent()) {
const GlobalValue *GV = GA->getGlobal();
if (getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV)) {
MachineFunction &MF = DAG.getMachineFunction();
MVT PtrVT = getPointerTy(MF.getDataLayout());
const char *BaseName;
if (GV->getValueType()->isFunctionTy()) {
BaseName = MF.createExternalSymbolName("__table_base");
OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL;
else {
BaseName = MF.createExternalSymbolName("__memory_base");
OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL;
SDValue BaseAddr =
DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
DAG.getTargetExternalSymbol(BaseName, PtrVT));
SDValue SymAddr = DAG.getNode(
WebAssemblyISD::WrapperREL, DL, VT,
DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT, GA->getOffset(),
return DAG.getNode(ISD::ADD, DL, VT, BaseAddr, SymAddr);
OperandFlags = WebAssemblyII::MO_GOT;
return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT,
GA->getOffset(), OperandFlags));
WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
const auto *ES = cast<ExternalSymbolSDNode>(Op);
EVT VT = Op.getValueType();
assert(ES->getTargetFlags() == 0 &&
"Unexpected target flags on generic ExternalSymbolSDNode");
return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
DAG.getTargetExternalSymbol(ES->getSymbol(), VT));
SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op,
SelectionDAG &DAG) const {
// There's no need for a Wrapper node because we always incorporate a jump
// table operand into a BR_TABLE instruction, rather than ever
// materializing it in a register.
const JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
return DAG.getTargetJumpTable(JT->getIndex(), Op.getValueType(),
SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
SDValue Chain = Op.getOperand(0);
const auto *JT = cast<JumpTableSDNode>(Op.getOperand(1));
SDValue Index = Op.getOperand(2);
assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags");
SmallVector<SDValue, 8> Ops;
MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs;
// Add an operand for each case.
for (auto MBB : MBBs)
// Add the first MBB as a dummy default target for now. This will be replaced
// with the proper default target (and the preceding range check eliminated)
// if possible by WebAssemblyFixBrTableDefaults.
return DAG.getNode(WebAssemblyISD::BR_TABLE, DL, MVT::Other, Ops);
SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
EVT PtrVT = getPointerTy(DAG.getMachineFunction().getDataLayout());
auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>();
const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
SDValue ArgN = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
MFI->getVarargBufferVreg(), PtrVT);
return DAG.getStore(Op.getOperand(0), DL, ArgN, Op.getOperand(1),
SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op,
SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
unsigned IntNo;
switch (Op.getOpcode()) {
IntNo = Op.getConstantOperandVal(1);
IntNo = Op.getConstantOperandVal(0);
llvm_unreachable("Invalid intrinsic");
SDLoc DL(Op);
switch (IntNo) {
return SDValue(); // Don't custom lower most intrinsics.
case Intrinsic::wasm_lsda: {
auto PtrVT = getPointerTy(MF.getDataLayout());
const char *SymName = MF.createExternalSymbolName(
"GCC_except_table" + std::to_string(MF.getFunctionNumber()));
if (isPositionIndependent()) {
SDValue Node = DAG.getTargetExternalSymbol(
SymName, PtrVT, WebAssemblyII::MO_MEMORY_BASE_REL);
const char *BaseName = MF.createExternalSymbolName("__memory_base");
SDValue BaseAddr =
DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
DAG.getTargetExternalSymbol(BaseName, PtrVT));
SDValue SymAddr =
DAG.getNode(WebAssemblyISD::WrapperREL, DL, PtrVT, Node);
return DAG.getNode(ISD::ADD, DL, PtrVT, BaseAddr, SymAddr);
SDValue Node = DAG.getTargetExternalSymbol(SymName, PtrVT);
return DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT, Node);
case Intrinsic::wasm_shuffle: {
// Drop in-chain and replace undefs, but otherwise pass through unchanged
SDValue Ops[18];
size_t OpIdx = 0;
Ops[OpIdx++] = Op.getOperand(1);
Ops[OpIdx++] = Op.getOperand(2);
while (OpIdx < 18) {
const SDValue &MaskIdx = Op.getOperand(OpIdx + 1);
if (MaskIdx.isUndef() ||
cast<ConstantSDNode>(MaskIdx.getNode())->getZExtValue() >= 32) {
Ops[OpIdx++] = DAG.getConstant(0, DL, MVT::i32);
} else {
Ops[OpIdx++] = MaskIdx;
return DAG.getNode(WebAssemblyISD::SHUFFLE, DL, Op.getValueType(), Ops);
WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
// If sign extension operations are disabled, allow sext_inreg only if operand
// is a vector extract of an i8 or i16 lane. SIMD does not depend on sign
// extension operations, but allowing sext_inreg in this context lets us have
// simple patterns to select extract_lane_s instructions. Expanding sext_inreg
// everywhere would be simpler in this file, but would necessitate large and
// brittle patterns to undo the expansion and select extract_lane_s
// instructions.
assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128());
if (Op.getOperand(0).getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return SDValue();
const SDValue &Extract = Op.getOperand(0);
MVT VecT = Extract.getOperand(0).getSimpleValueType();
if (VecT.getVectorElementType().getSizeInBits() > 32)
return SDValue();
MVT ExtractedLaneT =
MVT ExtractedVecT =
MVT::getVectorVT(ExtractedLaneT, 128 / ExtractedLaneT.getSizeInBits());
if (ExtractedVecT == VecT)
return Op;
// Bitcast vector to appropriate type to ensure ISel pattern coverage
const SDNode *Index = Extract.getOperand(1).getNode();
if (!isa<ConstantSDNode>(Index))
return SDValue();
unsigned IndexVal = cast<ConstantSDNode>(Index)->getZExtValue();
unsigned Scale =
ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements();
assert(Scale > 1);
SDValue NewIndex =
DAG.getConstant(IndexVal * Scale, DL, Index->getValueType(0));
SDValue NewExtract = DAG.getNode(
ISD::EXTRACT_VECTOR_ELT, DL, Extract.getValueType(),
DAG.getBitcast(ExtractedVecT, Extract.getOperand(0)), NewIndex);
return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, Op.getValueType(), NewExtract,
static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
SDLoc DL(Op);
if (Op.getValueType() != MVT::v2f64)
return SDValue();
auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec,
unsigned &Index) -> bool {
switch (Op.getOpcode()) {
Opcode = WebAssemblyISD::CONVERT_LOW_S;
Opcode = WebAssemblyISD::CONVERT_LOW_U;
Opcode = WebAssemblyISD::PROMOTE_LOW;
return false;
auto ExtractVector = Op.getOperand(0);
if (ExtractVector.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return false;
if (!isa<ConstantSDNode>(ExtractVector.getOperand(1).getNode()))
return false;
SrcVec = ExtractVector.getOperand(0);
Index = ExtractVector.getConstantOperandVal(1);
return true;
unsigned LHSOpcode, RHSOpcode, LHSIndex, RHSIndex;
SDValue LHSSrcVec, RHSSrcVec;
if (!GetConvertedLane(Op.getOperand(0), LHSOpcode, LHSSrcVec, LHSIndex) ||
!GetConvertedLane(Op.getOperand(1), RHSOpcode, RHSSrcVec, RHSIndex))
return SDValue();
if (LHSOpcode != RHSOpcode)
return SDValue();
MVT ExpectedSrcVT;
switch (LHSOpcode) {
case WebAssemblyISD::CONVERT_LOW_S:
case WebAssemblyISD::CONVERT_LOW_U:
ExpectedSrcVT = MVT::v4i32;
case WebAssemblyISD::PROMOTE_LOW:
ExpectedSrcVT = MVT::v4f32;
if (LHSSrcVec.getValueType() != ExpectedSrcVT)
return SDValue();
auto Src = LHSSrcVec;
if (LHSIndex != 0 || RHSIndex != 1 || LHSSrcVec != RHSSrcVec) {
// Shuffle the source vector so that the converted lanes are the low lanes.
Src = DAG.getVectorShuffle(
ExpectedSrcVT, DL, LHSSrcVec, RHSSrcVec,
{static_cast<int>(LHSIndex), static_cast<int>(RHSIndex) + 4, -1, -1});
return DAG.getNode(LHSOpcode, DL, MVT::v2f64, Src);
SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,
SelectionDAG &DAG) const {
if (auto ConvertLow = LowerConvertLow(Op, DAG))
return ConvertLow;
SDLoc DL(Op);
const EVT VecT = Op.getValueType();
const EVT LaneT = Op.getOperand(0).getValueType();
const size_t Lanes = Op.getNumOperands();
bool CanSwizzle = VecT == MVT::v16i8;
// BUILD_VECTORs are lowered to the instruction that initializes the highest
// possible number of lanes at once followed by a sequence of replace_lane
// instructions to individually initialize any remaining lanes.
// TODO: Tune this. For example, lanewise swizzling is very expensive, so
// swizzled lanes should be given greater weight.
// TODO: Investigate looping rather than always extracting/replacing specific
// lanes to fill gaps.
auto IsConstant = [](const SDValue &V) {
return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP;
// Returns the source vector and index vector pair if they exist. Checks for:
// (extract_vector_elt
// $src,
// (sign_extend_inreg (extract_vector_elt $indices, $i))
// )
auto GetSwizzleSrcs = [](size_t I, const SDValue &Lane) {
auto Bail = std::make_pair(SDValue(), SDValue());
if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return Bail;
const SDValue &SwizzleSrc = Lane->getOperand(0);
const SDValue &IndexExt = Lane->getOperand(1);
if (IndexExt->getOpcode() != ISD::SIGN_EXTEND_INREG)
return Bail;
const SDValue &Index = IndexExt->getOperand(0);
if (Index->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return Bail;
const SDValue &SwizzleIndices = Index->getOperand(0);
if (SwizzleSrc.getValueType() != MVT::v16i8 ||
SwizzleIndices.getValueType() != MVT::v16i8 ||
Index->getOperand(1)->getOpcode() != ISD::Constant ||
Index->getConstantOperandVal(1) != I)
return Bail;
return std::make_pair(SwizzleSrc, SwizzleIndices);
// If the lane is extracted from another vector at a constant index, return
// that vector. The source vector must not have more lanes than the dest
// because the shufflevector indices are in terms of the destination lanes and
// would not be able to address the smaller individual source lanes.
auto GetShuffleSrc = [&](const SDValue &Lane) {
if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return SDValue();
if (!isa<ConstantSDNode>(Lane->getOperand(1).getNode()))
return SDValue();
if (Lane->getOperand(0).getValueType().getVectorNumElements() >
return SDValue();
return Lane->getOperand(0);
using ValueEntry = std::pair<SDValue, size_t>;
SmallVector<ValueEntry, 16> SplatValueCounts;
using SwizzleEntry = std::pair<std::pair<SDValue, SDValue>, size_t>;
SmallVector<SwizzleEntry, 16> SwizzleCounts;
using ShuffleEntry = std::pair<SDValue, size_t>;
SmallVector<ShuffleEntry, 16> ShuffleCounts;
auto AddCount = [](auto &Counts, const auto &Val) {
auto CountIt =
llvm::find_if(Counts, [&Val](auto E) { return E.first == Val; });
if (CountIt == Counts.end()) {
Counts.emplace_back(Val, 1);
} else {
auto GetMostCommon = [](auto &Counts) {
auto CommonIt =
std::max_element(Counts.begin(), Counts.end(),
[](auto A, auto B) { return A.second < B.second; });
assert(CommonIt != Counts.end() && "Unexpected all-undef build_vector");
return *CommonIt;
size_t NumConstantLanes = 0;
// Count eligible lanes for each type of vector creation op
for (size_t I = 0; I < Lanes; ++I) {
const SDValue &Lane = Op->getOperand(I);
if (Lane.isUndef())
AddCount(SplatValueCounts, Lane);
if (IsConstant(Lane))
if (auto ShuffleSrc = GetShuffleSrc(Lane))
AddCount(ShuffleCounts, ShuffleSrc);
if (CanSwizzle) {
auto SwizzleSrcs = GetSwizzleSrcs(I, Lane);
if (SwizzleSrcs.first)
AddCount(SwizzleCounts, SwizzleSrcs);
SDValue SplatValue;
size_t NumSplatLanes;
std::tie(SplatValue, NumSplatLanes) = GetMostCommon(SplatValueCounts);
SDValue SwizzleSrc;
SDValue SwizzleIndices;
size_t NumSwizzleLanes = 0;
if (SwizzleCounts.size())
std::forward_as_tuple(std::tie(SwizzleSrc, SwizzleIndices),
NumSwizzleLanes) = GetMostCommon(SwizzleCounts);
// Shuffles can draw from up to two vectors, so find the two most common
// sources.
SDValue ShuffleSrc1, ShuffleSrc2;
size_t NumShuffleLanes = 0;
if (ShuffleCounts.size()) {
std::tie(ShuffleSrc1, NumShuffleLanes) = GetMostCommon(ShuffleCounts);
[&](const auto &Pair) { return Pair.first == ShuffleSrc1; });
if (ShuffleCounts.size()) {
size_t AdditionalShuffleLanes;
std::tie(ShuffleSrc2, AdditionalShuffleLanes) =
NumShuffleLanes += AdditionalShuffleLanes;
// Predicate returning true if the lane is properly initialized by the
// original instruction
std::function<bool(size_t, const SDValue &)> IsLaneConstructed;
SDValue Result;
// Prefer swizzles over shuffles over vector consts over splats
if (NumSwizzleLanes >= NumShuffleLanes &&
NumSwizzleLanes >= NumConstantLanes && NumSwizzleLanes >= NumSplatLanes) {
Result = DAG.getNode(WebAssemblyISD::SWIZZLE, DL, VecT, SwizzleSrc,
auto Swizzled = std::make_pair(SwizzleSrc, SwizzleIndices);
IsLaneConstructed = [&, Swizzled](size_t I, const SDValue &Lane) {
return Swizzled == GetSwizzleSrcs(I, Lane);
} else if (NumShuffleLanes >= NumConstantLanes &&
NumShuffleLanes >= NumSplatLanes) {
size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits() / 8;
size_t DestLaneCount = VecT.getVectorNumElements();
size_t Scale1 = 1;
size_t Scale2 = 1;
SDValue Src1 = ShuffleSrc1;
SDValue Src2 = ShuffleSrc2 ? ShuffleSrc2 : DAG.getUNDEF(VecT);
if (Src1.getValueType() != VecT) {
size_t LaneSize =
Src1.getValueType().getVectorElementType().getFixedSizeInBits() / 8;
assert(LaneSize > DestLaneSize);
Scale1 = LaneSize / DestLaneSize;
Src1 = DAG.getBitcast(VecT, Src1);
if (Src2.getValueType() != VecT) {
size_t LaneSize =
Src2.getValueType().getVectorElementType().getFixedSizeInBits() / 8;
assert(LaneSize > DestLaneSize);
Scale2 = LaneSize / DestLaneSize;
Src2 = DAG.getBitcast(VecT, Src2);
int Mask[16];
assert(DestLaneCount <= 16);
for (size_t I = 0; I < DestLaneCount; ++I) {
const SDValue &Lane = Op->getOperand(I);
SDValue Src = GetShuffleSrc(Lane);
if (Src == ShuffleSrc1) {
Mask[I] = Lane->getConstantOperandVal(1) * Scale1;
} else if (Src && Src == ShuffleSrc2) {
Mask[I] = DestLaneCount + Lane->getConstantOperandVal(1) * Scale2;
} else {
Mask[I] = -1;
ArrayRef<int> MaskRef(Mask, DestLaneCount);
Result = DAG.getVectorShuffle(VecT, DL, Src1, Src2, MaskRef);
IsLaneConstructed = [&](size_t, const SDValue &Lane) {
auto Src = GetShuffleSrc(Lane);
return Src == ShuffleSrc1 || (Src && Src == ShuffleSrc2);
} else if (NumConstantLanes >= NumSplatLanes) {
SmallVector<SDValue, 16> ConstLanes;
for (const SDValue &Lane : Op->op_values()) {
if (IsConstant(Lane)) {
// Values may need to be fixed so that they will sign extend to be
// within the expected range during ISel. Check whether the value is in
// bounds based on the lane bit width and if it is out of bounds, lop
// off the extra bits and subtract 2^n to reflect giving the high bit
// value -2^(n-1) rather than +2^(n-1). Skip the i64 case because it
// cannot possibly be out of range.
auto *Const = dyn_cast<ConstantSDNode>(Lane.getNode());
int64_t Val = Const ? Const->getSExtValue() : 0;
uint64_t LaneBits = 128 / Lanes;
assert((LaneBits == 64 || Val >= -(1ll << (LaneBits - 1))) &&
"Unexpected out of bounds negative value");
if (Const && LaneBits != 64 && Val > (1ll << (LaneBits - 1)) - 1) {
auto NewVal = ((uint64_t)Val % (1ll << LaneBits)) - (1ll << LaneBits);
ConstLanes.push_back(DAG.getConstant(NewVal, SDLoc(Lane), LaneT));
} else {
} else if (LaneT.isFloatingPoint()) {
ConstLanes.push_back(DAG.getConstantFP(0, DL, LaneT));
} else {
ConstLanes.push_back(DAG.getConstant(0, DL, LaneT));
Result = DAG.getBuildVector(VecT, DL, ConstLanes);
IsLaneConstructed = [&IsConstant](size_t _, const SDValue &Lane) {
return IsConstant(Lane);
} else {
// Use a splat, but possibly a load_splat
LoadSDNode *SplattedLoad;
if ((SplattedLoad = dyn_cast<LoadSDNode>(SplatValue)) &&
SplattedLoad->getMemoryVT() == VecT.getVectorElementType()) {
Result = DAG.getMemIntrinsicNode(
WebAssemblyISD::LOAD_SPLAT, DL, DAG.getVTList(VecT),
{SplattedLoad->getChain(), SplattedLoad->getBasePtr(),
SplattedLoad->getMemoryVT(), SplattedLoad->getMemOperand());
} else {
Result = DAG.getSplatBuildVector(VecT, DL, SplatValue);
IsLaneConstructed = [&SplatValue](size_t _, const SDValue &Lane) {
return Lane == SplatValue;
// Add replace_lane instructions for any unhandled values
for (size_t I = 0; I < Lanes; ++I) {
const SDValue &Lane = Op->getOperand(I);
if (!Lane.isUndef() && !IsLaneConstructed(I, Lane))
Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane,