Google Git
Sign in
cobalt / cobalt / 0e7d696c3ce6a2ef566d0fabc20213b6a651f942 / . / src / third_party / v8 / src / wasm / baseline / liftoff-compiler.cc
blob: 5066468c5ef969e8c0723e44390a749df6034d36 [file] [log] [blame]
// Copyright 2017 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "src/wasm/baseline/liftoff-compiler.h"
#include "src/base/optional.h"
#include "src/codegen/assembler-inl.h"
// TODO(clemensb): Remove dependences on compiler stuff.
#include "src/codegen/external-reference.h"
#include "src/codegen/interface-descriptors.h"
#include "src/codegen/machine-type.h"
#include "src/codegen/macro-assembler-inl.h"
#include "src/compiler/linkage.h"
#include "src/compiler/wasm-compiler.h"
#include "src/logging/counters.h"
#include "src/logging/log.h"
#include "src/objects/smi.h"
#include "src/tracing/trace-event.h"
#include "src/utils/ostreams.h"
#include "src/utils/utils.h"
#include "src/wasm/baseline/liftoff-assembler.h"
#include "src/wasm/function-body-decoder-impl.h"
#include "src/wasm/function-compiler.h"
#include "src/wasm/memory-tracing.h"
#include "src/wasm/object-access.h"
#include "src/wasm/simd-shuffle.h"
#include "src/wasm/wasm-debug.h"
#include "src/wasm/wasm-engine.h"
#include "src/wasm/wasm-linkage.h"
#include "src/wasm/wasm-objects.h"
#include "src/wasm/wasm-opcodes-inl.h"
namespace v8 {
namespace internal {
namespace wasm {
constexpr auto kRegister = LiftoffAssembler::VarState::kRegister;
constexpr auto kIntConst = LiftoffAssembler::VarState::kIntConst;
constexpr auto kStack = LiftoffAssembler::VarState::kStack;
namespace {
#define __ asm_.
#define TRACE(...) \
do { \
if (FLAG_trace_liftoff) PrintF("[liftoff] " __VA_ARGS__); \
} while (false)
#define WASM_INSTANCE_OBJECT_FIELD_OFFSET(name) \
ObjectAccess::ToTagged(WasmInstanceObject::k##name##Offset)
template <int expected_size, int actual_size>
struct assert_field_size {
static_assert(expected_size == actual_size,
"field in WasmInstance does not have the expected size");
static constexpr int size = actual_size;
};
#define WASM_INSTANCE_OBJECT_FIELD_SIZE(name) \
FIELD_SIZE(WasmInstanceObject::k##name##Offset)
#define LOAD_INSTANCE_FIELD(dst, name, load_size) \
__ LoadFromInstance(dst, WASM_INSTANCE_OBJECT_FIELD_OFFSET(name), \
assert_field_size<WASM_INSTANCE_OBJECT_FIELD_SIZE(name), \
load_size>::size);
#define LOAD_TAGGED_PTR_INSTANCE_FIELD(dst, name) \
static_assert(WASM_INSTANCE_OBJECT_FIELD_SIZE(name) == kTaggedSize, \
"field in WasmInstance does not have the expected size"); \
__ LoadTaggedPointerFromInstance(dst, \
WASM_INSTANCE_OBJECT_FIELD_OFFSET(name));
#ifdef DEBUG
#define DEBUG_CODE_COMMENT(str) \
do { \
__ RecordComment(str); \
} while (false)
#else
#define DEBUG_CODE_COMMENT(str) ((void)0)
#endif
constexpr LoadType::LoadTypeValue kPointerLoadType =
kSystemPointerSize == 8 ? LoadType::kI64Load : LoadType::kI32Load;
constexpr ValueType kPointerValueType =
kSystemPointerSize == 8 ? kWasmI64 : kWasmI32;
#if V8_TARGET_ARCH_ARM64
// On ARM64, the Assembler keeps track of pointers to Labels to resolve
// branches to distant targets. Moving labels would confuse the Assembler,
// thus store the label on the heap and keep a unique_ptr.
class MovableLabel {
public:
MOVE_ONLY_NO_DEFAULT_CONSTRUCTOR(MovableLabel);
MovableLabel() : label_(new Label()) {}
Label* get() { return label_.get(); }
private:
std::unique_ptr<Label> label_;
};
#else
// On all other platforms, just store the Label directly.
class MovableLabel {
public:
MOVE_ONLY_WITH_DEFAULT_CONSTRUCTORS(MovableLabel);
Label* get() { return &label_; }
private:
Label label_;
};
#endif
compiler::CallDescriptor* GetLoweredCallDescriptor(
Zone* zone, compiler::CallDescriptor* call_desc) {
return kSystemPointerSize == 4
? compiler::GetI32WasmCallDescriptor(zone, call_desc)
: call_desc;
}
constexpr ValueType kSupportedTypesArr[] = {
kWasmI32, kWasmI64, kWasmF32, kWasmF64,
kWasmS128, kWasmExternRef, kWasmFuncRef};
constexpr Vector<const ValueType> kSupportedTypes =
ArrayVector(kSupportedTypesArr);
constexpr ValueType kSupportedTypesWithoutRefsArr[] = {
kWasmI32, kWasmI64, kWasmF32, kWasmF64, kWasmS128};
constexpr Vector<const ValueType> kSupportedTypesWithoutRefs =
ArrayVector(kSupportedTypesWithoutRefsArr);
constexpr Condition GetCompareCondition(WasmOpcode opcode) {
switch (opcode) {
case kExprI32Eq:
return kEqual;
case kExprI32Ne:
return kUnequal;
case kExprI32LtS:
return kSignedLessThan;
case kExprI32LtU:
return kUnsignedLessThan;
case kExprI32GtS:
return kSignedGreaterThan;
case kExprI32GtU:
return kUnsignedGreaterThan;
case kExprI32LeS:
return kSignedLessEqual;
case kExprI32LeU:
return kUnsignedLessEqual;
case kExprI32GeS:
return kSignedGreaterEqual;
case kExprI32GeU:
return kUnsignedGreaterEqual;
default:
#if V8_HAS_CXX14_CONSTEXPR
UNREACHABLE();
#else
// We need to return something for old compilers here.
return kEqual;
#endif
}
}
// Builds a {DebugSideTable}.
class DebugSideTableBuilder {
public:
enum AssumeSpilling {
// All register values will be spilled before the pc covered by the debug
// side table entry. Register slots will be marked as stack slots in the
// generated debug side table entry.
kAssumeSpilling,
// Register slots will be written out as they are.
kAllowRegisters,
// Register slots cannot appear since we already spilled.
kDidSpill
};
class EntryBuilder {
public:
explicit EntryBuilder(int pc_offset,
std::vector<DebugSideTable::Entry::Value> values)
: pc_offset_(pc_offset), values_(std::move(values)) {}
DebugSideTable::Entry ToTableEntry() {
return DebugSideTable::Entry{pc_offset_, std::move(values_)};
}
int pc_offset() const { return pc_offset_; }
void set_pc_offset(int new_pc_offset) { pc_offset_ = new_pc_offset; }
private:
int pc_offset_;
std::vector<DebugSideTable::Entry::Value> values_;
};
// Adds a new entry, and returns a pointer to a builder for modifying that
// entry ({stack_height} includes {num_locals}).
EntryBuilder* NewEntry(int pc_offset, int num_locals, int stack_height,
LiftoffAssembler::VarState* stack_state,
AssumeSpilling assume_spilling) {
DCHECK_LE(num_locals, stack_height);
// Record stack types.
std::vector<DebugSideTable::Entry::Value> values(stack_height);
for (int i = 0; i < stack_height; ++i) {
const auto& slot = stack_state[i];
values[i].type = slot.type();
values[i].stack_offset = slot.offset();
switch (slot.loc()) {
case kIntConst:
values[i].kind = DebugSideTable::Entry::kConstant;
values[i].i32_const = slot.i32_const();
break;
case kRegister:
DCHECK_NE(kDidSpill, assume_spilling);
if (assume_spilling == kAllowRegisters) {
values[i].kind = DebugSideTable::Entry::kRegister;
values[i].reg_code = slot.reg().liftoff_code();
break;
}
DCHECK_EQ(kAssumeSpilling, assume_spilling);
V8_FALLTHROUGH;
case kStack:
values[i].kind = DebugSideTable::Entry::kStack;
values[i].stack_offset = slot.offset();
break;
}
}
entries_.emplace_back(pc_offset, std::move(values));
return &entries_.back();
}
void SetNumLocals(int num_locals) {
DCHECK_EQ(-1, num_locals_);
DCHECK_LE(0, num_locals);
num_locals_ = num_locals;
}
std::unique_ptr<DebugSideTable> GenerateDebugSideTable() {
DCHECK_LE(0, num_locals_);
std::vector<DebugSideTable::Entry> entries;
entries.reserve(entries_.size());
for (auto& entry : entries_) entries.push_back(entry.ToTableEntry());
std::sort(entries.begin(), entries.end(),
[](DebugSideTable::Entry& a, DebugSideTable::Entry& b) {
return a.pc_offset() < b.pc_offset();
});
return std::make_unique<DebugSideTable>(num_locals_, std::move(entries));
}
private:
int num_locals_ = -1;
std::list<EntryBuilder> entries_;
};
class LiftoffCompiler {
public:
// TODO(clemensb): Make this a template parameter.
static constexpr Decoder::ValidateFlag validate = Decoder::kBooleanValidation;
using Value = ValueBase<validate>;
static constexpr auto kI32 = ValueType::kI32;
static constexpr auto kI64 = ValueType::kI64;
static constexpr auto kF32 = ValueType::kF32;
static constexpr auto kF64 = ValueType::kF64;
static constexpr auto kS128 = ValueType::kS128;
struct ElseState {
MovableLabel label;
LiftoffAssembler::CacheState state;
};
struct Control : public ControlBase<Value, validate> {
std::unique_ptr<ElseState> else_state;
LiftoffAssembler::CacheState label_state;
MovableLabel label;
MOVE_ONLY_NO_DEFAULT_CONSTRUCTOR(Control);
template <typename... Args>
explicit Control(Args&&... args) V8_NOEXCEPT
: ControlBase(std::forward<Args>(args)...) {}
};
using FullDecoder = WasmFullDecoder<validate, LiftoffCompiler>;
// For debugging, we need to spill registers before a trap or a stack check to
// be able to inspect them.
struct SpilledRegistersForInspection : public ZoneObject {
struct Entry {
int offset;
LiftoffRegister reg;
ValueType type;
};
ZoneVector<Entry> entries;
explicit SpilledRegistersForInspection(Zone* zone) : entries(zone) {}
};
struct OutOfLineCode {
MovableLabel label;
MovableLabel continuation;
WasmCode::RuntimeStubId stub;
WasmCodePosition position;
LiftoffRegList regs_to_save;
uint32_t pc; // for trap handler.
// These two pointers will only be used for debug code:
DebugSideTableBuilder::EntryBuilder* debug_sidetable_entry_builder;
SpilledRegistersForInspection* spilled_registers;
// Named constructors:
static OutOfLineCode Trap(
WasmCode::RuntimeStubId s, WasmCodePosition pos, uint32_t pc,
DebugSideTableBuilder::EntryBuilder* debug_sidetable_entry_builder,
SpilledRegistersForInspection* spilled_registers) {
DCHECK_LT(0, pos);
return {{},
{},
s,
pos,
{},
pc,
debug_sidetable_entry_builder,
spilled_registers};
}
static OutOfLineCode StackCheck(
WasmCodePosition pos, LiftoffRegList regs_to_save,
SpilledRegistersForInspection* spilled_regs,
DebugSideTableBuilder::EntryBuilder* debug_sidetable_entry_builder) {
return {{}, {}, WasmCode::kWasmStackGuard, pos,
regs_to_save, 0, debug_sidetable_entry_builder, spilled_regs};
}
};
LiftoffCompiler(compiler::CallDescriptor* call_descriptor,
CompilationEnv* env, Zone* compilation_zone,
std::unique_ptr<AssemblerBuffer> buffer,
DebugSideTableBuilder* debug_sidetable_builder,
ForDebugging for_debugging, int func_index,
Vector<int> breakpoints = {}, int dead_breakpoint = 0)
: asm_(std::move(buffer)),
descriptor_(
GetLoweredCallDescriptor(compilation_zone, call_descriptor)),
env_(env),
debug_sidetable_builder_(debug_sidetable_builder),
for_debugging_(for_debugging),
func_index_(func_index),
out_of_line_code_(compilation_zone),
source_position_table_builder_(compilation_zone),
protected_instructions_(compilation_zone),
compilation_zone_(compilation_zone),
safepoint_table_builder_(compilation_zone_),
next_breakpoint_ptr_(breakpoints.begin()),
next_breakpoint_end_(breakpoints.end()),
dead_breakpoint_(dead_breakpoint) {
if (breakpoints.empty()) {
next_breakpoint_ptr_ = next_breakpoint_end_ = nullptr;
}
}
bool did_bailout() const { return bailout_reason_ != kSuccess; }
LiftoffBailoutReason bailout_reason() const { return bailout_reason_; }
void GetCode(CodeDesc* desc) {
asm_.GetCode(nullptr, desc, &safepoint_table_builder_,
Assembler::kNoHandlerTable);
}
OwnedVector<uint8_t> GetSourcePositionTable() {
return source_position_table_builder_.ToSourcePositionTableVector();
}
OwnedVector<uint8_t> GetProtectedInstructionsData() const {
return OwnedVector<uint8_t>::Of(
Vector<const uint8_t>::cast(VectorOf(protected_instructions_)));
}
uint32_t GetTotalFrameSlotCountForGC() const {
return __ GetTotalFrameSlotCountForGC();
}
void unsupported(FullDecoder* decoder, LiftoffBailoutReason reason,
const char* detail) {
DCHECK_NE(kSuccess, reason);
if (did_bailout()) return;
bailout_reason_ = reason;
TRACE("unsupported: %s\n", detail);
decoder->errorf(decoder->pc_offset(), "unsupported liftoff operation: %s",
detail);
UnuseLabels(decoder);
}
bool DidAssemblerBailout(FullDecoder* decoder) {
if (decoder->failed() || !__ did_bailout()) return false;
unsupported(decoder, __ bailout_reason(), __ bailout_detail());
return true;
}
LiftoffBailoutReason BailoutReasonForType(ValueType type) {
switch (type.kind()) {
case ValueType::kS128:
return kSimd;
case ValueType::kOptRef:
case ValueType::kRef:
if (type.is_reference_to(HeapType::kExn)) {
return kExceptionHandling;
} else {
return kRefTypes;
}
case ValueType::kBottom:
return kMultiValue;
default:
return kOtherReason;
}
}
bool CheckSupportedType(FullDecoder* decoder,
Vector<const ValueType> supported_types,
ValueType type, const char* context) {
// Special case for kWasm128 which requires specific hardware support.
if (type == kWasmS128 && (!CpuFeatures::SupportsWasmSimd128())) {
unsupported(decoder, kSimd, "simd");
return false;
}
// Check supported types.
for (ValueType supported : supported_types) {
if (type == supported) return true;
}
LiftoffBailoutReason bailout_reason = BailoutReasonForType(type);
EmbeddedVector<char, 128> buffer;
SNPrintF(buffer, "%s %s", type.name().c_str(), context);
unsupported(decoder, bailout_reason, buffer.begin());
return false;
}
int GetSafepointTableOffset() const {
return safepoint_table_builder_.GetCodeOffset();
}
void UnuseLabels(FullDecoder* decoder) {
#ifdef DEBUG
auto Unuse = [](Label* label) {
label->Unuse();
label->UnuseNear();
};
// Unuse all labels now, otherwise their destructor will fire a DCHECK error
// if they where referenced before.
uint32_t control_depth = decoder ? decoder->control_depth() : 0;
for (uint32_t i = 0; i < control_depth; ++i) {
Control* c = decoder->control_at(i);
Unuse(c->label.get());
if (c->else_state) Unuse(c->else_state->label.get());
}
for (auto& ool : out_of_line_code_) Unuse(ool.label.get());
#endif
}
void StartFunction(FullDecoder* decoder) {
if (FLAG_trace_liftoff && !FLAG_trace_wasm_decoder) {
StdoutStream{} << "hint: add --trace-wasm-decoder to also see the wasm "
"instructions being decoded\n";
}
int num_locals = decoder->num_locals();
__ set_num_locals(num_locals);
for (int i = 0; i < num_locals; ++i) {
ValueType type = decoder->local_type(i);
__ set_local_type(i, type);
}
}
// Returns the number of inputs processed (1 or 2).
uint32_t ProcessParameter(ValueType type, uint32_t input_idx) {
const bool needs_pair = needs_gp_reg_pair(type);
const ValueType reg_type = needs_pair ? kWasmI32 : type;
const RegClass rc = reg_class_for(reg_type);
auto LoadToReg = [this, reg_type, rc](compiler::LinkageLocation location,
LiftoffRegList pinned) {
if (location.IsRegister()) {
DCHECK(!location.IsAnyRegister());
return LiftoffRegister::from_external_code(rc, reg_type,
location.AsRegister());
}
DCHECK(location.IsCallerFrameSlot());
LiftoffRegister reg = __ GetUnusedRegister(rc, pinned);
__ LoadCallerFrameSlot(reg, -location.AsCallerFrameSlot(), reg_type);
return reg;
};
LiftoffRegister reg =
LoadToReg(descriptor_->GetInputLocation(input_idx), {});
if (needs_pair) {
LiftoffRegister reg2 =
LoadToReg(descriptor_->GetInputLocation(input_idx + 1),
LiftoffRegList::ForRegs(reg));
reg = LiftoffRegister::ForPair(reg.gp(), reg2.gp());
}
__ PushRegister(type, reg);
return needs_pair ? 2 : 1;
}
void StackCheck(WasmCodePosition position) {
DEBUG_CODE_COMMENT("stack check");
if (!FLAG_wasm_stack_checks || !env_->runtime_exception_support) return;
LiftoffRegList regs_to_save = __ cache_state()->used_registers;
SpilledRegistersForInspection* spilled_regs = nullptr;
Register limit_address = __ GetUnusedRegister(kGpReg, {}).gp();
if (V8_UNLIKELY(for_debugging_)) {
regs_to_save = {};
spilled_regs = GetSpilledRegistersForInspection();
}
out_of_line_code_.push_back(OutOfLineCode::StackCheck(
position, regs_to_save, spilled_regs,
RegisterDebugSideTableEntry(DebugSideTableBuilder::kAssumeSpilling)));
OutOfLineCode& ool = out_of_line_code_.back();
LOAD_INSTANCE_FIELD(limit_address, StackLimitAddress, kSystemPointerSize);
__ StackCheck(ool.label.get(), limit_address);
__ bind(ool.continuation.get());
}
bool SpillLocalsInitially(FullDecoder* decoder, uint32_t num_params) {
int actual_locals = __ num_locals() - num_params;
DCHECK_LE(0, actual_locals);
constexpr int kNumCacheRegisters = NumRegs(kLiftoffAssemblerGpCacheRegs);
// If we have many locals, we put them on the stack initially. This avoids
// having to spill them on merge points. Use of these initial values should
// be rare anyway.
if (actual_locals > kNumCacheRegisters / 2) return true;
// If there are locals which are not i32 or i64, we also spill all locals,
// because other types cannot be initialized to constants.
for (uint32_t param_idx = num_params; param_idx < __ num_locals();
++param_idx) {
ValueType type = decoder->local_type(param_idx);
if (type != kWasmI32 && type != kWasmI64) return true;
}
return false;
}
void TierUpFunction(FullDecoder* decoder) {
__ CallRuntimeStub(WasmCode::kWasmTriggerTierUp);
DefineSafepoint();
}
void TraceFunctionEntry(FullDecoder* decoder) {
DEBUG_CODE_COMMENT("trace function entry");
__ SpillAllRegisters();
source_position_table_builder_.AddPosition(
__ pc_offset(), SourcePosition(decoder->position()), false);
__ CallRuntimeStub(WasmCode::kWasmTraceEnter);
DefineSafepoint();
}
void StartFunctionBody(FullDecoder* decoder, Control* block) {
for (uint32_t i = 0; i < __ num_locals(); ++i) {
if (!CheckSupportedType(decoder,
FLAG_experimental_liftoff_extern_ref
? kSupportedTypes
: kSupportedTypesWithoutRefs,
__ local_type(i), "param"))
return;
}
// Input 0 is the call target, the instance is at 1.
constexpr int kInstanceParameterIndex = 1;
// Store the instance parameter to a special stack slot.
compiler::LinkageLocation instance_loc =
descriptor_->GetInputLocation(kInstanceParameterIndex);
DCHECK(instance_loc.IsRegister());
DCHECK(!instance_loc.IsAnyRegister());
Register instance_reg = Register::from_code(instance_loc.AsRegister());
DCHECK_EQ(kWasmInstanceRegister, instance_reg);
// Parameter 0 is the instance parameter.
uint32_t num_params =
static_cast<uint32_t>(decoder->sig_->parameter_count());
__ CodeEntry();
DEBUG_CODE_COMMENT("enter frame");
__ EnterFrame(StackFrame::WASM);
__ set_has_frame(true);
pc_offset_stack_frame_construction_ = __ PrepareStackFrame();
// {PrepareStackFrame} is the first platform-specific assembler method.
// If this failed, we can bail out immediately, avoiding runtime overhead
// and potential failures because of other unimplemented methods.
// A platform implementing {PrepareStackFrame} must ensure that we can
// finish compilation without errors even if we hit unimplemented
// LiftoffAssembler methods.
if (DidAssemblerBailout(decoder)) return;
// Process parameters.
if (num_params) DEBUG_CODE_COMMENT("process parameters");
__ SpillInstance(instance_reg);
// Input 0 is the code target, 1 is the instance. First parameter at 2.
uint32_t input_idx = kInstanceParameterIndex + 1;
for (uint32_t param_idx = 0; param_idx < num_params; ++param_idx) {
input_idx += ProcessParameter(__ local_type(param_idx), input_idx);
}
int params_size = __ TopSpillOffset();
DCHECK_EQ(input_idx, descriptor_->InputCount());
// Initialize locals beyond parameters.
if (num_params < __ num_locals()) DEBUG_CODE_COMMENT("init locals");
if (SpillLocalsInitially(decoder, num_params)) {
for (uint32_t param_idx = num_params; param_idx < __ num_locals();
++param_idx) {
ValueType type = decoder->local_type(param_idx);
__ PushStack(type);
}
int spill_size = __ TopSpillOffset() - params_size;
__ FillStackSlotsWithZero(params_size, spill_size);
} else {
for (uint32_t param_idx = num_params; param_idx < __ num_locals();
++param_idx) {
ValueType type = decoder->local_type(param_idx);
__ PushConstant(type, int32_t{0});
}
}
if (FLAG_experimental_liftoff_extern_ref) {
// Initialize all reference type locals with ref.null.
for (uint32_t param_idx = num_params; param_idx < __ num_locals();
++param_idx) {
ValueType type = decoder->local_type(param_idx);
if (type.is_reference_type()) {
Register isolate_root = __ GetUnusedRegister(kGpReg, {}).gp();
// We can re-use the isolate_root register as result register.
Register result = isolate_root;
LOAD_INSTANCE_FIELD(isolate_root, IsolateRoot, kSystemPointerSize);
__ LoadTaggedPointer(
result, isolate_root, no_reg,
IsolateData::root_slot_offset(RootIndex::kNullValue), {});
__ Spill(__ cache_state()->stack_state.back().offset(),
LiftoffRegister(result), type);
}
}
}
DCHECK_EQ(__ num_locals(), __ cache_state()->stack_height());
if (V8_UNLIKELY(debug_sidetable_builder_)) {
debug_sidetable_builder_->SetNumLocals(__ num_locals());
}
// The function-prologue stack check is associated with position 0, which
// is never a position of any instruction in the function.
StackCheck(0);
if (FLAG_wasm_dynamic_tiering) {
// TODO(arobin): Avoid spilling registers unconditionally.
__ SpillAllRegisters();
DEBUG_CODE_COMMENT("dynamic tiering");
LiftoffRegList pinned;
// Get the number of calls array address.
LiftoffRegister array_address =
pinned.set(__ GetUnusedRegister(kGpReg, pinned));
LOAD_INSTANCE_FIELD(array_address.gp(), NumLiftoffFunctionCallsArray,
kSystemPointerSize);
// Compute the correct offset in the array.
uint32_t offset =
kInt32Size * declared_function_index(env_->module, func_index_);
// Get the number of calls and update it.
LiftoffRegister old_number_of_calls =
pinned.set(__ GetUnusedRegister(kGpReg, pinned));
LiftoffRegister new_number_of_calls =
pinned.set(__ GetUnusedRegister(kGpReg, pinned));
__ Load(old_number_of_calls, array_address.gp(), no_reg, offset,
LoadType::kI32Load, pinned);
__ emit_i32_addi(new_number_of_calls.gp(), old_number_of_calls.gp(), 1);
__ Store(array_address.gp(), no_reg, offset, new_number_of_calls,
StoreType::kI32Store, pinned);
// Emit the runtime call if necessary.
Label no_tierup;
// Check if the number of calls is a power of 2.
__ emit_i32_and(old_number_of_calls.gp(), old_number_of_calls.gp(),
new_number_of_calls.gp());
// Unary "unequal" means "different from zero".
__ emit_cond_jump(kUnequal, &no_tierup, kWasmI32,
old_number_of_calls.gp());
TierUpFunction(decoder);
__ bind(&no_tierup);
}
if (FLAG_trace_wasm) TraceFunctionEntry(decoder);
}
void GenerateOutOfLineCode(OutOfLineCode* ool) {
DEBUG_CODE_COMMENT(
(std::string("out of line: ") + GetRuntimeStubName(ool->stub)).c_str());
__ bind(ool->label.get());
const bool is_stack_check = ool->stub == WasmCode::kWasmStackGuard;
const bool is_mem_out_of_bounds =
ool->stub == WasmCode::kThrowWasmTrapMemOutOfBounds;
if (is_mem_out_of_bounds && env_->use_trap_handler) {
uint32_t pc = static_cast<uint32_t>(__ pc_offset());
DCHECK_EQ(pc, __ pc_offset());
protected_instructions_.emplace_back(
trap_handler::ProtectedInstructionData{ool->pc, pc});
}
if (!env_->runtime_exception_support) {
// We cannot test calls to the runtime in cctest/test-run-wasm.
// Therefore we emit a call to C here instead of a call to the runtime.
// In this mode, we never generate stack checks.
DCHECK(!is_stack_check);
__ CallTrapCallbackForTesting();
DEBUG_CODE_COMMENT("leave frame");
__ LeaveFrame(StackFrame::WASM);
__ DropStackSlotsAndRet(
static_cast<uint32_t>(descriptor_->StackParameterCount()));
return;
}
// We cannot both push and spill registers.
DCHECK(ool->regs_to_save.is_empty() || ool->spilled_registers == nullptr);
if (!ool->regs_to_save.is_empty()) {
__ PushRegisters(ool->regs_to_save);
} else if (V8_UNLIKELY(ool->spilled_registers != nullptr)) {
for (auto& entry : ool->spilled_registers->entries) {
__ Spill(entry.offset, entry.reg, entry.type);
}
}
source_position_table_builder_.AddPosition(
__ pc_offset(), SourcePosition(ool->position), true);
__ CallRuntimeStub(ool->stub);
// TODO(ahaas): Define a proper safepoint here.
safepoint_table_builder_.DefineSafepoint(&asm_, Safepoint::kNoLazyDeopt);
DCHECK_EQ(!debug_sidetable_builder_, !ool->debug_sidetable_entry_builder);
if (V8_UNLIKELY(ool->debug_sidetable_entry_builder)) {
ool->debug_sidetable_entry_builder->set_pc_offset(__ pc_offset());
}
DCHECK_EQ(ool->continuation.get()->is_bound(), is_stack_check);
if (!ool->regs_to_save.is_empty()) __ PopRegisters(ool->regs_to_save);
if (is_stack_check) {
if (V8_UNLIKELY(ool->spilled_registers != nullptr)) {
DCHECK(for_debugging_);
for (auto& entry : ool->spilled_registers->entries) {
__ Fill(entry.reg, entry.offset, entry.type);
}
}
__ emit_jump(ool->continuation.get());
} else {
__ AssertUnreachable(AbortReason::kUnexpectedReturnFromWasmTrap);
}
}
void FinishFunction(FullDecoder* decoder) {
if (DidAssemblerBailout(decoder)) return;
for (OutOfLineCode& ool : out_of_line_code_) {
GenerateOutOfLineCode(&ool);
}
__ PatchPrepareStackFrame(pc_offset_stack_frame_construction_,
__ GetTotalFrameSize());
__ FinishCode();
safepoint_table_builder_.Emit(&asm_, __ GetTotalFrameSlotCountForGC());
__ MaybeEmitOutOfLineConstantPool();
// The previous calls may have also generated a bailout.
DidAssemblerBailout(decoder);
}
void OnFirstError(FullDecoder* decoder) {
if (!did_bailout()) bailout_reason_ = kDecodeError;
UnuseLabels(decoder);
asm_.AbortCompilation();
}
V8_NOINLINE void EmitDebuggingInfo(FullDecoder* decoder, WasmOpcode opcode) {
DCHECK(for_debugging_);
if (!WasmOpcodes::IsBreakable(opcode)) return;
bool has_breakpoint = false;
if (next_breakpoint_ptr_) {
if (*next_breakpoint_ptr_ == 0) {
// A single breakpoint at offset 0 indicates stepping.
DCHECK_EQ(next_breakpoint_ptr_ + 1, next_breakpoint_end_);
has_breakpoint = true;
} else {
while (next_breakpoint_ptr_ != next_breakpoint_end_ &&
*next_breakpoint_ptr_ < decoder->position()) {
// Skip unreachable breakpoints.
++next_breakpoint_ptr_;
}
if (next_breakpoint_ptr_ == next_breakpoint_end_) {
next_breakpoint_ptr_ = next_breakpoint_end_ = nullptr;
} else if (*next_breakpoint_ptr_ == decoder->position()) {
has_breakpoint = true;
}
}
}
if (has_breakpoint) {
EmitBreakpoint(decoder);
// Once we emitted a breakpoint, we don't need to check the "hook on
// function call" any more.
checked_hook_on_function_call_ = true;
} else if (!checked_hook_on_function_call_) {
checked_hook_on_function_call_ = true;
// Check the "hook on function call" flag. If set, trigger a break.
DEBUG_CODE_COMMENT("check hook on function call");
Register flag = __ GetUnusedRegister(kGpReg, {}).gp();
LOAD_INSTANCE_FIELD(flag, HookOnFunctionCallAddress, kSystemPointerSize);
Label no_break;
__ Load(LiftoffRegister{flag}, flag, no_reg, 0, LoadType::kI32Load8U, {});
// Unary "equal" means "equals zero".
__ emit_cond_jump(kEqual, &no_break, kWasmI32, flag);
EmitBreakpoint(decoder);
__ bind(&no_break);
} else if (dead_breakpoint_ == decoder->position()) {
DCHECK(!next_breakpoint_ptr_ ||
*next_breakpoint_ptr_ != dead_breakpoint_);
// The top frame is paused at this position, but the breakpoint was
// removed. Adding a dead breakpoint here ensures that the source
// position exists, and that the offset to the return address is the
// same as in the old code.
Label cont;
__ emit_jump(&cont);
EmitBreakpoint(decoder);
__ bind(&cont);
}
}
void NextInstruction(FullDecoder* decoder, WasmOpcode opcode) {
// Add a single check, so that the fast path can be inlined while
// {EmitDebuggingInfo} stays outlined.
if (V8_UNLIKELY(for_debugging_)) EmitDebuggingInfo(decoder, opcode);
TraceCacheState(decoder);
#ifdef DEBUG
SLOW_DCHECK(__ ValidateCacheState());
if (WasmOpcodes::IsPrefixOpcode(opcode)) {
opcode = decoder->read_prefixed_opcode<Decoder::kFullValidation>(
decoder->pc());
}
DEBUG_CODE_COMMENT(WasmOpcodes::OpcodeName(opcode));
#endif
}
void EmitBreakpoint(FullDecoder* decoder) {
DEBUG_CODE_COMMENT("breakpoint");
DCHECK(for_debugging_);
source_position_table_builder_.AddPosition(
__ pc_offset(), SourcePosition(decoder->position()), true);
__ CallRuntimeStub(WasmCode::kWasmDebugBreak);
// TODO(ahaas): Define a proper safepoint here.
safepoint_table_builder_.DefineSafepoint(&asm_, Safepoint::kNoLazyDeopt);
RegisterDebugSideTableEntry(DebugSideTableBuilder::kAllowRegisters);
}
void Block(FullDecoder* decoder, Control* block) {}
void Loop(FullDecoder* decoder, Control* loop) {
// Before entering a loop, spill all locals to the stack, in order to free
// the cache registers, and to avoid unnecessarily reloading stack values
// into registers at branches.
// TODO(clemensb): Come up with a better strategy here, involving
// pre-analysis of the function.
__ SpillLocals();
__ PrepareLoopArgs(loop->start_merge.arity);
// Loop labels bind at the beginning of the block.
__ bind(loop->label.get());
// Save the current cache state for the merge when jumping to this loop.
loop->label_state.Split(*__ cache_state());
// Execute a stack check in the loop header.
StackCheck(decoder->position());
}
void Try(FullDecoder* decoder, Control* block) {
unsupported(decoder, kExceptionHandling, "try");
}
void Catch(FullDecoder* decoder, Control* block, Value* exception) {
unsupported(decoder, kExceptionHandling, "catch");
}
void If(FullDecoder* decoder, const Value& cond, Control* if_block) {
DCHECK_EQ(if_block, decoder->control_at(0));
DCHECK(if_block->is_if());
// Allocate the else state.
if_block->else_state = std::make_unique<ElseState>();
// Test the condition, jump to else if zero.
Register value = __ PopToRegister().gp();
__ emit_cond_jump(kEqual, if_block->else_state->label.get(), kWasmI32,
value);
// Store the state (after popping the value) for executing the else branch.
if_block->else_state->state.Split(*__ cache_state());
}
void FallThruTo(FullDecoder* decoder, Control* c) {
if (c->end_merge.reached) {
__ MergeFullStackWith(c->label_state, *__ cache_state());
} else {
c->label_state.Split(*__ cache_state());
}
TraceCacheState(decoder);
}
void FinishOneArmedIf(FullDecoder* decoder, Control* c) {
DCHECK(c->is_onearmed_if());
if (c->end_merge.reached) {
// Someone already merged to the end of the if. Merge both arms into that.
if (c->reachable()) {
// Merge the if state into the end state.
__ MergeFullStackWith(c->label_state, *__ cache_state());
__ emit_jump(c->label.get());
}
// Merge the else state into the end state.
__ bind(c->else_state->label.get());
__ MergeFullStackWith(c->label_state, c->else_state->state);
__ cache_state()->Steal(c->label_state);
} else if (c->reachable()) {
// No merge yet at the end of the if, but we need to create a merge for
// the both arms of this if. Thus init the merge point from the else
// state, then merge the if state into that.
DCHECK_EQ(c->start_merge.arity, c->end_merge.arity);
c->label_state.InitMerge(c->else_state->state, __ num_locals(),
c->start_merge.arity, c->stack_depth);
__ MergeFullStackWith(c->label_state, *__ cache_state());
__ emit_jump(c->label.get());
// Merge the else state into the end state.
__ bind(c->else_state->label.get());
__ MergeFullStackWith(c->label_state, c->else_state->state);
__ cache_state()->Steal(c->label_state);
} else {
// No merge needed, just continue with the else state.
__ bind(c->else_state->label.get());
__ cache_state()->Steal(c->else_state->state);
}
}
void PopControl(FullDecoder* decoder, Control* c) {
if (c->is_loop()) return; // A loop just falls through.
if (c->is_onearmed_if()) {
// Special handling for one-armed ifs.
FinishOneArmedIf(decoder, c);
} else if (c->end_merge.reached) {
// There is a merge already. Merge our state into that, then continue with
// that state.
if (c->reachable()) {
__ MergeFullStackWith(c->label_state, *__ cache_state());
}
__ cache_state()->Steal(c->label_state);
} else {
// No merge, just continue with our current state.
}
if (!c->label.get()->is_bound()) __ bind(c->label.get());
}
void EndControl(FullDecoder* decoder, Control* c) {}
void GenerateCCall(const LiftoffRegister* result_regs, const FunctionSig* sig,
ValueType out_argument_type,
const LiftoffRegister* arg_regs,
ExternalReference ext_ref) {
// Before making a call, spill all cache registers.
__ SpillAllRegisters();
// Store arguments on our stack, then align the stack for calling to C.
int param_bytes = 0;
for (ValueType param_type : sig->parameters()) {
param_bytes += param_type.element_size_bytes();
}
int out_arg_bytes = out_argument_type == kWasmStmt
? 0
: out_argument_type.element_size_bytes();
int stack_bytes = std::max(param_bytes, out_arg_bytes);
__ CallC(sig, arg_regs, result_regs, out_argument_type, stack_bytes,
ext_ref);
}
template <typename EmitFn, typename... Args>
typename std::enable_if<!std::is_member_function_pointer<EmitFn>::value>::type
CallEmitFn(EmitFn fn, Args... args) {
fn(args...);
}
template <typename EmitFn, typename... Args>
typename std::enable_if<std::is_member_function_pointer<EmitFn>::value>::type
CallEmitFn(EmitFn fn, Args... args) {
(asm_.*fn)(ConvertAssemblerArg(args)...);
}
// Wrap a {LiftoffRegister} with implicit conversions to {Register} and
// {DoubleRegister}.
struct AssemblerRegisterConverter {
LiftoffRegister reg;
operator LiftoffRegister() { return reg; }
operator Register() { return reg.gp(); }
operator DoubleRegister() { return reg.fp(); }
};
// Convert {LiftoffRegister} to {AssemblerRegisterConverter}, other types stay
// unchanged.
template <typename T>
typename std::conditional<std::is_same<LiftoffRegister, T>::value,
AssemblerRegisterConverter, T>::type
ConvertAssemblerArg(T t) {
return {t};
}
template <typename EmitFn, typename ArgType>
struct EmitFnWithFirstArg {
EmitFn fn;
ArgType first_arg;
};
template <typename EmitFn, typename ArgType>
EmitFnWithFirstArg<EmitFn, ArgType> BindFirst(EmitFn fn, ArgType arg) {
return {fn, arg};
}
template <typename EmitFn, typename T, typename... Args>
void CallEmitFn(EmitFnWithFirstArg<EmitFn, T> bound_fn, Args... args) {
CallEmitFn(bound_fn.fn, bound_fn.first_arg, ConvertAssemblerArg(args)...);
}
template <ValueType::Kind src_type, ValueType::Kind result_type, class EmitFn>
void EmitUnOp(EmitFn fn) {
constexpr RegClass src_rc = reg_class_for(src_type);
constexpr RegClass result_rc = reg_class_for(result_type);
LiftoffRegister src = __ PopToRegister();
LiftoffRegister dst = src_rc == result_rc
? __ GetUnusedRegister(result_rc, {src}, {})
: __ GetUnusedRegister(result_rc, {});
CallEmitFn(fn, dst, src);
__ PushRegister(ValueType::Primitive(result_type), dst);
}
template <ValueType::Kind type>
void EmitFloatUnOpWithCFallback(
bool (LiftoffAssembler::*emit_fn)(DoubleRegister, DoubleRegister),
ExternalReference (*fallback_fn)()) {
auto emit_with_c_fallback = [=](LiftoffRegister dst, LiftoffRegister src) {
if ((asm_.*emit_fn)(dst.fp(), src.fp())) return;
ExternalReference ext_ref = fallback_fn();
ValueType sig_reps[] = {ValueType::Primitive(type)};
FunctionSig sig(0, 1, sig_reps);
GenerateCCall(&dst, &sig, ValueType::Primitive(type), &src, ext_ref);
};
EmitUnOp<type, type>(emit_with_c_fallback);
}
enum TypeConversionTrapping : bool { kCanTrap = true, kNoTrap = false };
template <ValueType::Kind dst_type, ValueType::Kind src_type,
TypeConversionTrapping can_trap>
void EmitTypeConversion(WasmOpcode opcode, ExternalReference (*fallback_fn)(),
WasmCodePosition trap_position) {
static constexpr RegClass src_rc = reg_class_for(src_type);
static constexpr RegClass dst_rc = reg_class_for(dst_type);
LiftoffRegister src = __ PopToRegister();
LiftoffRegister dst = src_rc == dst_rc
? __ GetUnusedRegister(dst_rc, {src}, {})
: __ GetUnusedRegister(dst_rc, {});
DCHECK_EQ(!!can_trap, trap_position > 0);
Label* trap = can_trap ? AddOutOfLineTrap(
trap_position,
WasmCode::kThrowWasmTrapFloatUnrepresentable)
: nullptr;
if (!__ emit_type_conversion(opcode, dst, src, trap)) {
DCHECK_NOT_NULL(fallback_fn);
ExternalReference ext_ref = fallback_fn();
if (can_trap) {
// External references for potentially trapping conversions return int.
ValueType sig_reps[] = {kWasmI32, ValueType::Primitive(src_type)};
FunctionSig sig(1, 1, sig_reps);
LiftoffRegister ret_reg =
__ GetUnusedRegister(kGpReg, LiftoffRegList::ForRegs(dst));
LiftoffRegister dst_regs[] = {ret_reg, dst};
GenerateCCall(dst_regs, &sig, ValueType::Primitive(dst_type), &src,
ext_ref);
__ emit_cond_jump(kEqual, trap, kWasmI32, ret_reg.gp());
} else {
ValueType sig_reps[] = {ValueType::Primitive(src_type)};
FunctionSig sig(0, 1, sig_reps);
GenerateCCall(&dst, &sig, ValueType::Primitive(dst_type), &src,
ext_ref);
}
}
__ PushRegister(ValueType::Primitive(dst_type), dst);
}
void UnOp(FullDecoder* decoder, WasmOpcode opcode, const Value& value,
Value* result) {
#define CASE_I32_UNOP(opcode, fn) \
case kExpr##opcode: \
return EmitUnOp<kI32, kI32>(&LiftoffAssembler::emit_##fn);
#define CASE_I64_UNOP(opcode, fn) \
case kExpr##opcode: \
return EmitUnOp<kI64, kI64>(&LiftoffAssembler::emit_##fn);
#define CASE_FLOAT_UNOP(opcode, type, fn) \
case kExpr##opcode: \
return EmitUnOp<k##type, k##type>(&LiftoffAssembler::emit_##fn);
#define CASE_FLOAT_UNOP_WITH_CFALLBACK(opcode, type, fn) \
case kExpr##opcode: \
return EmitFloatUnOpWithCFallback<k##type>(&LiftoffAssembler::emit_##fn, \
&ExternalReference::wasm_##fn);
#define CASE_TYPE_CONVERSION(opcode, dst_type, src_type, ext_ref, can_trap) \
case kExpr##opcode: \
return EmitTypeConversion<k##dst_type, k##src_type, can_trap>( \
kExpr##opcode, ext_ref, can_trap ? decoder->position() : 0);
switch (opcode) {
CASE_I32_UNOP(I32Clz, i32_clz)
CASE_I32_UNOP(I32Ctz, i32_ctz)
CASE_FLOAT_UNOP(F32Abs, F32, f32_abs)
CASE_FLOAT_UNOP(F32Neg, F32, f32_neg)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F32Ceil, F32, f32_ceil)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F32Floor, F32, f32_floor)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F32Trunc, F32, f32_trunc)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F32NearestInt, F32, f32_nearest_int)
CASE_FLOAT_UNOP(F32Sqrt, F32, f32_sqrt)
CASE_FLOAT_UNOP(F64Abs, F64, f64_abs)
CASE_FLOAT_UNOP(F64Neg, F64, f64_neg)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F64Ceil, F64, f64_ceil)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F64Floor, F64, f64_floor)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F64Trunc, F64, f64_trunc)
CASE_FLOAT_UNOP_WITH_CFALLBACK(F64NearestInt, F64, f64_nearest_int)
CASE_FLOAT_UNOP(F64Sqrt, F64, f64_sqrt)
CASE_TYPE_CONVERSION(I32ConvertI64, I32, I64, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I32SConvertF32, I32, F32, nullptr, kCanTrap)
CASE_TYPE_CONVERSION(I32UConvertF32, I32, F32, nullptr, kCanTrap)
CASE_TYPE_CONVERSION(I32SConvertF64, I32, F64, nullptr, kCanTrap)
CASE_TYPE_CONVERSION(I32UConvertF64, I32, F64, nullptr, kCanTrap)
CASE_TYPE_CONVERSION(I32ReinterpretF32, I32, F32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I64SConvertI32, I64, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I64UConvertI32, I64, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I64SConvertF32, I64, F32,
&ExternalReference::wasm_float32_to_int64, kCanTrap)
CASE_TYPE_CONVERSION(I64UConvertF32, I64, F32,
&ExternalReference::wasm_float32_to_uint64, kCanTrap)
CASE_TYPE_CONVERSION(I64SConvertF64, I64, F64,
&ExternalReference::wasm_float64_to_int64, kCanTrap)
CASE_TYPE_CONVERSION(I64UConvertF64, I64, F64,
&ExternalReference::wasm_float64_to_uint64, kCanTrap)
CASE_TYPE_CONVERSION(I64ReinterpretF64, I64, F64, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F32SConvertI32, F32, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F32UConvertI32, F32, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F32SConvertI64, F32, I64,
&ExternalReference::wasm_int64_to_float32, kNoTrap)
CASE_TYPE_CONVERSION(F32UConvertI64, F32, I64,
&ExternalReference::wasm_uint64_to_float32, kNoTrap)
CASE_TYPE_CONVERSION(F32ConvertF64, F32, F64, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F32ReinterpretI32, F32, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F64SConvertI32, F64, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F64UConvertI32, F64, I32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F64SConvertI64, F64, I64,
&ExternalReference::wasm_int64_to_float64, kNoTrap)
CASE_TYPE_CONVERSION(F64UConvertI64, F64, I64,
&ExternalReference::wasm_uint64_to_float64, kNoTrap)
CASE_TYPE_CONVERSION(F64ConvertF32, F64, F32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(F64ReinterpretI64, F64, I64, nullptr, kNoTrap)
CASE_I32_UNOP(I32SExtendI8, i32_signextend_i8)
CASE_I32_UNOP(I32SExtendI16, i32_signextend_i16)
CASE_I64_UNOP(I64SExtendI8, i64_signextend_i8)
CASE_I64_UNOP(I64SExtendI16, i64_signextend_i16)
CASE_I64_UNOP(I64SExtendI32, i64_signextend_i32)
CASE_I64_UNOP(I64Clz, i64_clz)
CASE_I64_UNOP(I64Ctz, i64_ctz)
CASE_TYPE_CONVERSION(I32SConvertSatF32, I32, F32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I32UConvertSatF32, I32, F32, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I32SConvertSatF64, I32, F64, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I32UConvertSatF64, I32, F64, nullptr, kNoTrap)
CASE_TYPE_CONVERSION(I64SConvertSatF32, I64, F32,
&ExternalReference::wasm_float32_to_int64_sat,
kNoTrap)
CASE_TYPE_CONVERSION(I64UConvertSatF32, I64, F32,
&ExternalReference::wasm_float32_to_uint64_sat,
kNoTrap)
CASE_TYPE_CONVERSION(I64SConvertSatF64, I64, F64,
&ExternalReference::wasm_float64_to_int64_sat,
kNoTrap)
CASE_TYPE_CONVERSION(I64UConvertSatF64, I64, F64,
&ExternalReference::wasm_float64_to_uint64_sat,
kNoTrap)
case kExprI32Eqz:
DCHECK(decoder->lookahead(0, kExprI32Eqz));
if (decoder->lookahead(1, kExprBrIf) && !for_debugging_) {
DCHECK(!has_outstanding_op());
outstanding_op_ = kExprI32Eqz;
break;
}
return EmitUnOp<kI32, kI32>(&LiftoffAssembler::emit_i32_eqz);
case kExprI64Eqz:
return EmitUnOp<kI64, kI32>(&LiftoffAssembler::emit_i64_eqz);
case kExprI32Popcnt:
return EmitUnOp<kI32, kI32>(
[=](LiftoffRegister dst, LiftoffRegister src) {
if (__ emit_i32_popcnt(dst.gp(), src.gp())) return;
ValueType sig_i_i_reps[] = {kWasmI32, kWasmI32};
FunctionSig sig_i_i(1, 1, sig_i_i_reps);
GenerateCCall(&dst, &sig_i_i, kWasmStmt, &src,
ExternalReference::wasm_word32_popcnt());
});
case kExprI64Popcnt:
return EmitUnOp<kI64, kI64>(
[=](LiftoffRegister dst, LiftoffRegister src) {
if (__ emit_i64_popcnt(dst, src)) return;
// The c function returns i32. We will zero-extend later.
ValueType sig_i_l_reps[] = {kWasmI32, kWasmI64};
FunctionSig sig_i_l(1, 1, sig_i_l_reps);
LiftoffRegister c_call_dst = kNeedI64RegPair ? dst.low() : dst;
GenerateCCall(&c_call_dst, &sig_i_l, kWasmStmt, &src,
ExternalReference::wasm_word64_popcnt());
// Now zero-extend the result to i64.
__ emit_type_conversion(kExprI64UConvertI32, dst, c_call_dst,
nullptr);
});
case kExprRefIsNull:
unsupported(decoder, kRefTypes, "ref_is_null");
return;
default:
UNREACHABLE();
}
#undef CASE_I32_UNOP
#undef CASE_I64_UNOP
#undef CASE_FLOAT_UNOP
#undef CASE_FLOAT_UNOP_WITH_CFALLBACK
#undef CASE_TYPE_CONVERSION
}
template <ValueType::Kind src_type, ValueType::Kind result_type,
typename EmitFn, typename EmitFnImm>
void EmitBinOpImm(EmitFn fn, EmitFnImm fnImm) {
static constexpr RegClass src_rc = reg_class_for(src_type);
static constexpr RegClass result_rc = reg_class_for(result_type);
LiftoffAssembler::VarState rhs_slot = __ cache_state()->stack_state.back();
// Check if the RHS is an immediate.
if (rhs_slot.is_const()) {
__ cache_state()->stack_state.pop_back();
int32_t imm = rhs_slot.i32_const();
LiftoffRegister lhs = __ PopToRegister();
// Either reuse {lhs} for {dst}, or choose a register (pair) which does
// not overlap, for easier code generation.
LiftoffRegList pinned = LiftoffRegList::ForRegs(lhs);
LiftoffRegister dst = src_rc == result_rc
? __ GetUnusedRegister(result_rc, {lhs}, pinned)
: __ GetUnusedRegister(result_rc, pinned);
CallEmitFn(fnImm, dst, lhs, imm);
__ PushRegister(ValueType::Primitive(result_type), dst);
} else {
// The RHS was not an immediate.
EmitBinOp<src_type, result_type>(fn);
}
}
template <ValueType::Kind src_type, ValueType::Kind result_type,
bool swap_lhs_rhs = false, typename EmitFn>
void EmitBinOp(EmitFn fn) {
static constexpr RegClass src_rc = reg_class_for(src_type);
static constexpr RegClass result_rc = reg_class_for(result_type);
LiftoffRegister rhs = __ PopToRegister();
LiftoffRegister lhs = __ PopToRegister(LiftoffRegList::ForRegs(rhs));
LiftoffRegister dst = src_rc == result_rc
? __ GetUnusedRegister(result_rc, {lhs, rhs}, {})
: __ GetUnusedRegister(result_rc, {});
if (swap_lhs_rhs) std::swap(lhs, rhs);
CallEmitFn(fn, dst, lhs, rhs);
__ PushRegister(ValueType::Primitive(result_type), dst);
}
void EmitDivOrRem64CCall(LiftoffRegister dst, LiftoffRegister lhs,
LiftoffRegister rhs, ExternalReference ext_ref,
Label* trap_by_zero,
Label* trap_unrepresentable = nullptr) {
// Cannot emit native instructions, build C call.
LiftoffRegister ret =
__ GetUnusedRegister(kGpReg, LiftoffRegList::ForRegs(dst));
LiftoffRegister tmp =
__ GetUnusedRegister(kGpReg, LiftoffRegList::ForRegs(dst, ret));
LiftoffRegister arg_regs[] = {lhs, rhs};
LiftoffRegister result_regs[] = {ret, dst};
ValueType sig_types[] = {kWasmI32, kWasmI64, kWasmI64};
// <i64, i64> -> i32 (with i64 output argument)
FunctionSig sig(1, 2, sig_types);
GenerateCCall(result_regs, &sig, kWasmI64, arg_regs, ext_ref);
__ LoadConstant(tmp, WasmValue(int32_t{0}));
__ emit_cond_jump(kEqual, trap_by_zero, kWasmI32, ret.gp(), tmp.gp());
if (trap_unrepresentable) {
__ LoadConstant(tmp, WasmValue(int32_t{-1}));
__ emit_cond_jump(kEqual, trap_unrepresentable, kWasmI32, ret.gp(),
tmp.gp());
}
}
template <WasmOpcode opcode>
void EmitI32CmpOp(FullDecoder* decoder) {
DCHECK(decoder->lookahead(0, opcode));
if (decoder->lookahead(1, kExprBrIf) && !for_debugging_) {
DCHECK(!has_outstanding_op());
outstanding_op_ = opcode;
return;
}
return EmitBinOp<kI32, kI32>(BindFirst(&LiftoffAssembler::emit_i32_set_cond,
GetCompareCondition(opcode)));
}
void BinOp(FullDecoder* decoder, WasmOpcode opcode, const Value& lhs,
const Value& rhs, Value* result) {
#define CASE_I64_SHIFTOP(opcode, fn) \
case kExpr##opcode: \
return EmitBinOpImm<kI64, kI64>( \
[=](LiftoffRegister dst, LiftoffRegister src, \
LiftoffRegister amount) { \
__ emit_##fn(dst, src, \
amount.is_gp_pair() ? amount.low_gp() : amount.gp()); \
}, \
&LiftoffAssembler::emit_##fn##i);
#define CASE_CCALL_BINOP(opcode, type, ext_ref_fn) \
case kExpr##opcode: \
return EmitBinOp<k##type, k##type>( \
[=](LiftoffRegister dst, LiftoffRegister lhs, LiftoffRegister rhs) { \
LiftoffRegister args[] = {lhs, rhs}; \
auto ext_ref = ExternalReference::ext_ref_fn(); \
ValueType sig_reps[] = {kWasm##type, kWasm##type, kWasm##type}; \
const bool out_via_stack = kWasm##type == kWasmI64; \
FunctionSig sig(out_via_stack ? 0 : 1, 2, sig_reps); \
ValueType out_arg_type = out_via_stack ? kWasmI64 : kWasmStmt; \
GenerateCCall(&dst, &sig, out_arg_type, args, ext_ref); \
});
switch (opcode) {
case kExprI32Add:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_add,
&LiftoffAssembler::emit_i32_addi);
case kExprI32Sub:
return EmitBinOp<kI32, kI32>(&LiftoffAssembler::emit_i32_sub);
case kExprI32Mul:
return EmitBinOp<kI32, kI32>(&LiftoffAssembler::emit_i32_mul);
case kExprI32And:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_and,
&LiftoffAssembler::emit_i32_andi);
case kExprI32Ior:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_or,
&LiftoffAssembler::emit_i32_ori);
case kExprI32Xor:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_xor,
&LiftoffAssembler::emit_i32_xori);
case kExprI32Eq:
return EmitI32CmpOp<kExprI32Eq>(decoder);
case kExprI32Ne:
return EmitI32CmpOp<kExprI32Ne>(decoder);
case kExprI32LtS:
return EmitI32CmpOp<kExprI32LtS>(decoder);
case kExprI32LtU:
return EmitI32CmpOp<kExprI32LtU>(decoder);
case kExprI32GtS:
return EmitI32CmpOp<kExprI32GtS>(decoder);
case kExprI32GtU:
return EmitI32CmpOp<kExprI32GtU>(decoder);
case kExprI32LeS:
return EmitI32CmpOp<kExprI32LeS>(decoder);
case kExprI32LeU:
return EmitI32CmpOp<kExprI32LeU>(decoder);
case kExprI32GeS:
return EmitI32CmpOp<kExprI32GeS>(decoder);
case kExprI32GeU:
return EmitI32CmpOp<kExprI32GeU>(decoder);
case kExprI64Add:
return EmitBinOpImm<kI64, kI64>(&LiftoffAssembler::emit_i64_add,
&LiftoffAssembler::emit_i64_addi);
case kExprI64Sub:
return EmitBinOp<kI64, kI64>(&LiftoffAssembler::emit_i64_sub);
case kExprI64Mul:
return EmitBinOp<kI64, kI64>(&LiftoffAssembler::emit_i64_mul);
case kExprI64And:
return EmitBinOpImm<kI64, kI64>(&LiftoffAssembler::emit_i64_and,
&LiftoffAssembler::emit_i64_andi);
case kExprI64Ior:
return EmitBinOpImm<kI64, kI64>(&LiftoffAssembler::emit_i64_or,
&LiftoffAssembler::emit_i64_ori);
case kExprI64Xor:
return EmitBinOpImm<kI64, kI64>(&LiftoffAssembler::emit_i64_xor,
&LiftoffAssembler::emit_i64_xori);
case kExprI64Eq:
return EmitBinOp<kI64, kI32>(
BindFirst(&LiftoffAssembler::emit_i64_set_cond, kEqual));
case kExprI64Ne:
return EmitBinOp<kI64, kI32>(
BindFirst(&LiftoffAssembler::emit_i64_set_cond, kUnequal));
case kExprI64LtS:
return EmitBinOp<kI64, kI32>(
BindFirst(&LiftoffAssembler::emit_i64_set_cond, kSignedLessThan));
case kExprI64LtU:
return EmitBinOp<kI64, kI32>(
BindFirst(&LiftoffAssembler::emit_i64_set_cond, kUnsignedLessThan));
case kExprI64GtS:
return EmitBinOp<kI64, kI32>(BindFirst(
&LiftoffAssembler::emit_i64_set_cond, kSignedGreaterThan));
case kExprI64GtU:
return EmitBinOp<kI64, kI32>(BindFirst(
&LiftoffAssembler::emit_i64_set_cond, kUnsignedGreaterThan));
case kExprI64LeS:
return EmitBinOp<kI64, kI32>(
BindFirst(&LiftoffAssembler::emit_i64_set_cond, kSignedLessEqual));
case kExprI64LeU:
return EmitBinOp<kI64, kI32>(BindFirst(
&LiftoffAssembler::emit_i64_set_cond, kUnsignedLessEqual));
case kExprI64GeS:
return EmitBinOp<kI64, kI32>(BindFirst(
&LiftoffAssembler::emit_i64_set_cond, kSignedGreaterEqual));
case kExprI64GeU:
return EmitBinOp<kI64, kI32>(BindFirst(
&LiftoffAssembler::emit_i64_set_cond, kUnsignedGreaterEqual));
case kExprF32Eq:
return EmitBinOp<kF32, kI32>(
BindFirst(&LiftoffAssembler::emit_f32_set_cond, kEqual));
case kExprF32Ne:
return EmitBinOp<kF32, kI32>(
BindFirst(&LiftoffAssembler::emit_f32_set_cond, kUnequal));
case kExprF32Lt:
return EmitBinOp<kF32, kI32>(
BindFirst(&LiftoffAssembler::emit_f32_set_cond, kUnsignedLessThan));
case kExprF32Gt:
return EmitBinOp<kF32, kI32>(BindFirst(
&LiftoffAssembler::emit_f32_set_cond, kUnsignedGreaterThan));
case kExprF32Le:
return EmitBinOp<kF32, kI32>(BindFirst(
&LiftoffAssembler::emit_f32_set_cond, kUnsignedLessEqual));
case kExprF32Ge:
return EmitBinOp<kF32, kI32>(BindFirst(
&LiftoffAssembler::emit_f32_set_cond, kUnsignedGreaterEqual));
case kExprF64Eq:
return EmitBinOp<kF64, kI32>(
BindFirst(&LiftoffAssembler::emit_f64_set_cond, kEqual));
case kExprF64Ne:
return EmitBinOp<kF64, kI32>(
BindFirst(&LiftoffAssembler::emit_f64_set_cond, kUnequal));
case kExprF64Lt:
return EmitBinOp<kF64, kI32>(
BindFirst(&LiftoffAssembler::emit_f64_set_cond, kUnsignedLessThan));
case kExprF64Gt:
return EmitBinOp<kF64, kI32>(BindFirst(
&LiftoffAssembler::emit_f64_set_cond, kUnsignedGreaterThan));
case kExprF64Le:
return EmitBinOp<kF64, kI32>(BindFirst(
&LiftoffAssembler::emit_f64_set_cond, kUnsignedLessEqual));
case kExprF64Ge:
return EmitBinOp<kF64, kI32>(BindFirst(
&LiftoffAssembler::emit_f64_set_cond, kUnsignedGreaterEqual));
case kExprI32Shl:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_shl,
&LiftoffAssembler::emit_i32_shli);
case kExprI32ShrS:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_sar,
&LiftoffAssembler::emit_i32_sari);
case kExprI32ShrU:
return EmitBinOpImm<kI32, kI32>(&LiftoffAssembler::emit_i32_shr,
&LiftoffAssembler::emit_i32_shri);
CASE_CCALL_BINOP(I32Rol, I32, wasm_word32_rol)
CASE_CCALL_BINOP(I32Ror, I32, wasm_word32_ror)
CASE_I64_SHIFTOP(I64Shl, i64_shl)
CASE_I64_SHIFTOP(I64ShrS, i64_sar)
CASE_I64_SHIFTOP(I64ShrU, i64_shr)
CASE_CCALL_BINOP(I64Rol, I64, wasm_word64_rol)
CASE_CCALL_BINOP(I64Ror, I64, wasm_word64_ror)
case kExprF32Add:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_add);
case kExprF32Sub:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_sub);
case kExprF32Mul:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_mul);
case kExprF32Div:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_div);
case kExprF32Min:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_min);
case kExprF32Max:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_max);
case kExprF32CopySign:
return EmitBinOp<kF32, kF32>(&LiftoffAssembler::emit_f32_copysign);
case kExprF64Add:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_add);
case kExprF64Sub:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_sub);
case kExprF64Mul:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_mul);
case kExprF64Div:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_div);
case kExprF64Min:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_min);
case kExprF64Max:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_max);
case kExprF64CopySign:
return EmitBinOp<kF64, kF64>(&LiftoffAssembler::emit_f64_copysign);
case kExprI32DivS:
return EmitBinOp<kI32, kI32>([this, decoder](LiftoffRegister dst,
LiftoffRegister lhs,
LiftoffRegister rhs) {
WasmCodePosition position = decoder->position();
AddOutOfLineTrap(position, WasmCode::kThrowWasmTrapDivByZero);
// Adding the second trap might invalidate the pointer returned for
// the first one, thus get both pointers afterwards.
AddOutOfLineTrap(position,
WasmCode::kThrowWasmTrapDivUnrepresentable);
Label* div_by_zero = out_of_line_code_.end()[-2].label.get();
Label* div_unrepresentable = out_of_line_code_.end()[-1].label.get();
__ emit_i32_divs(dst.gp(), lhs.gp(), rhs.gp(), div_by_zero,
div_unrepresentable);
});
case kExprI32DivU:
return EmitBinOp<kI32, kI32>(
[this, decoder](LiftoffRegister dst, LiftoffRegister lhs,
LiftoffRegister rhs) {
Label* div_by_zero = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapDivByZero);
__ emit_i32_divu(dst.gp(), lhs.gp(), rhs.gp(), div_by_zero);
});
case kExprI32RemS:
return EmitBinOp<kI32, kI32>(
[this, decoder](LiftoffRegister dst, LiftoffRegister lhs,
LiftoffRegister rhs) {
Label* rem_by_zero = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapRemByZero);
__ emit_i32_rems(dst.gp(), lhs.gp(), rhs.gp(), rem_by_zero);
});
case kExprI32RemU:
return EmitBinOp<kI32, kI32>(
[this, decoder](LiftoffRegister dst, LiftoffRegister lhs,
LiftoffRegister rhs) {
Label* rem_by_zero = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapRemByZero);
__ emit_i32_remu(dst.gp(), lhs.gp(), rhs.gp(), rem_by_zero);
});
case kExprI64DivS:
return EmitBinOp<kI64, kI64>([this, decoder](LiftoffRegister dst,
LiftoffRegister lhs,
LiftoffRegister rhs) {
WasmCodePosition position = decoder->position();
AddOutOfLineTrap(position, WasmCode::kThrowWasmTrapDivByZero);
// Adding the second trap might invalidate the pointer returned for
// the first one, thus get both pointers afterwards.
AddOutOfLineTrap(position,
WasmCode::kThrowWasmTrapDivUnrepresentable);
Label* div_by_zero = out_of_line_code_.end()[-2].label.get();
Label* div_unrepresentable = out_of_line_code_.end()[-1].label.get();
if (!__ emit_i64_divs(dst, lhs, rhs, div_by_zero,
div_unrepresentable)) {
ExternalReference ext_ref = ExternalReference::wasm_int64_div();
EmitDivOrRem64CCall(dst, lhs, rhs, ext_ref, div_by_zero,
div_unrepresentable);
}
});
case kExprI64DivU:
return EmitBinOp<kI64, kI64>([this, decoder](LiftoffRegister dst,
LiftoffRegister lhs,
LiftoffRegister rhs) {
Label* div_by_zero = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapDivByZero);
if (!__ emit_i64_divu(dst, lhs, rhs, div_by_zero)) {
ExternalReference ext_ref = ExternalReference::wasm_uint64_div();
EmitDivOrRem64CCall(dst, lhs, rhs, ext_ref, div_by_zero);
}
});
case kExprI64RemS:
return EmitBinOp<kI64, kI64>(
[this, decoder](LiftoffRegister dst, LiftoffRegister lhs,
LiftoffRegister rhs) {
Label* rem_by_zero = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapRemByZero);
if (!__ emit_i64_rems(dst, lhs, rhs, rem_by_zero)) {
ExternalReference ext_ref = ExternalReference::wasm_int64_mod();
EmitDivOrRem64CCall(dst, lhs, rhs, ext_ref, rem_by_zero);
}
});
case kExprI64RemU:
return EmitBinOp<kI64, kI64>([this, decoder](LiftoffRegister dst,
LiftoffRegister lhs,
LiftoffRegister rhs) {
Label* rem_by_zero = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapRemByZero);
if (!__ emit_i64_remu(dst, lhs, rhs, rem_by_zero)) {
ExternalReference ext_ref = ExternalReference::wasm_uint64_mod();
EmitDivOrRem64CCall(dst, lhs, rhs, ext_ref, rem_by_zero);
}
});
default:
UNREACHABLE();
}
#undef CASE_I64_SHIFTOP
#undef CASE_CCALL_BINOP
}
void I32Const(FullDecoder* decoder, Value* result, int32_t value) {
__ PushConstant(kWasmI32, value);
}
void I64Const(FullDecoder* decoder, Value* result, int64_t value) {
// The {VarState} stores constant values as int32_t, thus we only store
// 64-bit constants in this field if it fits in an int32_t. Larger values
// cannot be used as immediate value anyway, so we can also just put them in
// a register immediately.
int32_t value_i32 = static_cast<int32_t>(value);
if (value_i32 == value) {
__ PushConstant(kWasmI64, value_i32);
} else {
LiftoffRegister reg = __ GetUnusedRegister(reg_class_for(kWasmI64), {});
__ LoadConstant(reg, WasmValue(value));
__ PushRegister(kWasmI64, reg);
}
}
void F32Const(FullDecoder* decoder, Value* result, float value) {
LiftoffRegister reg = __ GetUnusedRegister(kFpReg, {});
__ LoadConstant(reg, WasmValue(value));
__ PushRegister(kWasmF32, reg);
}
void F64Const(FullDecoder* decoder, Value* result, double value) {
LiftoffRegister reg = __ GetUnusedRegister(kFpReg, {});
__ LoadConstant(reg, WasmValue(value));
__ PushRegister(kWasmF64, reg);
}
void RefNull(FullDecoder* decoder, ValueType type, Value*) {
if (!FLAG_experimental_liftoff_extern_ref) {
unsupported(decoder, kRefTypes, "ref_null");
return;
}
Register isolate_root = __ GetUnusedRegister(kGpReg, {}).gp();
// We can re-use the isolate_root register as result register.
Register result = isolate_root;
LOAD_INSTANCE_FIELD(isolate_root, IsolateRoot, kSystemPointerSize);
__ LoadTaggedPointer(result, isolate_root, no_reg,
IsolateData::root_slot_offset(RootIndex::kNullValue),
{});
__ PushRegister(type, LiftoffRegister(result));
}
void RefFunc(FullDecoder* decoder, uint32_t function_index, Value* result) {
unsupported(decoder, kRefTypes, "func");
}
void RefAsNonNull(FullDecoder* decoder, const Value& arg, Value* result) {
unsupported(decoder, kRefTypes, "ref.as_non_null");
}
void Drop(FullDecoder* decoder, const Value& value) {
auto& slot = __ cache_state()->stack_state.back();
// If the dropped slot contains a register, decrement it's use count.
if (slot.is_reg()) __ cache_state()->dec_used(slot.reg());
__ cache_state()->stack_state.pop_back();
}
void TraceFunctionExit(FullDecoder* decoder) {
DEBUG_CODE_COMMENT("trace function exit");
// Before making the runtime call, spill all cache registers.
__ SpillAllRegisters();
LiftoffRegList pinned;
// Get a register to hold the stack slot for the return value.
LiftoffRegister info = pinned.set(__ GetUnusedRegister(kGpReg, pinned));
__ AllocateStackSlot(info.gp(), sizeof(int64_t));
// Store the return value if there is exactly one. Multiple return values
// are not handled yet.
size_t num_returns = decoder->sig_->return_count();
if (num_returns == 1) {
ValueType return_type = decoder->sig_->GetReturn(0);
LiftoffRegister return_reg =
__ LoadToRegister(__ cache_state()->stack_state.back(), pinned);
__ Store(info.gp(), no_reg, 0, return_reg,
StoreType::ForValueType(return_type), pinned);
}
// Put the parameter in its place.
WasmTraceExitDescriptor descriptor;
DCHECK_EQ(0, descriptor.GetStackParameterCount());
DCHECK_EQ(1, descriptor.GetRegisterParameterCount());
Register param_reg = descriptor.GetRegisterParameter(0);
if (info.gp() != param_reg) {
__ Move(param_reg, info.gp(), LiftoffAssembler::kWasmIntPtr);
}
source_position_table_builder_.AddPosition(
__ pc_offset(), SourcePosition(decoder->position()), false);
__ CallRuntimeStub(WasmCode::kWasmTraceExit);
DefineSafepoint();
__ DeallocateStackSlot(sizeof(int64_t));
}
void ReturnImpl(FullDecoder* decoder) {
if (FLAG_trace_wasm) TraceFunctionExit(decoder);
size_t num_returns = decoder->sig_->return_count();
if (num_returns > 0) __ MoveToReturnLocations(decoder->sig_, descriptor_);
DEBUG_CODE_COMMENT("leave frame");
__ LeaveFrame(StackFrame::WASM);
__ DropStackSlotsAndRet(
static_cast<uint32_t>(descriptor_->StackParameterCount()));
}
void DoReturn(FullDecoder* decoder, Vector<Value> /*values*/) {
ReturnImpl(decoder);
}
void LocalGet(FullDecoder* decoder, Value* result,
const LocalIndexImmediate<validate>& imm) {
auto local_slot = __ cache_state()->stack_state[imm.index];
__ cache_state()->stack_state.emplace_back(
local_slot.type(), __ NextSpillOffset(local_slot.type()));
auto* slot = &__ cache_state()->stack_state.back();
if (local_slot.is_reg()) {
__ cache_state()->inc_used(local_slot.reg());
slot->MakeRegister(local_slot.reg());
} else if (local_slot.is_const()) {
slot->MakeConstant(local_slot.i32_const());
} else {
DCHECK(local_slot.is_stack());
auto rc = reg_class_for(local_slot.type());
LiftoffRegister reg = __ GetUnusedRegister(rc, {});
__ cache_state()->inc_used(reg);
slot->MakeRegister(reg);
__ Fill(reg, local_slot.offset(), local_slot.type());
}
}
void LocalSetFromStackSlot(LiftoffAssembler::VarState* dst_slot,
uint32_t local_index) {
auto& state = *__ cache_state();
auto& src_slot = state.stack_state.back();
ValueType type = dst_slot->type();
if (dst_slot->is_reg()) {
LiftoffRegister slot_reg = dst_slot->reg();
if (state.get_use_count(slot_reg) == 1) {
__ Fill(dst_slot->reg(), src_slot.offset(), type);
return;
}
state.dec_used(slot_reg);
dst_slot->MakeStack();
}
DCHECK_EQ(type, __ local_type(local_index));
RegClass rc = reg_class_for(type);
LiftoffRegister dst_reg = __ GetUnusedRegister(rc, {});
__ Fill(dst_reg, src_slot.offset(), type);
*dst_slot = LiftoffAssembler::VarState(type, dst_reg, dst_slot->offset());
__ cache_state()->inc_used(dst_reg);
}
void LocalSet(uint32_t local_index, bool is_tee) {
auto& state = *__ cache_state();
auto& source_slot = state.stack_state.back();
auto& target_slot = state.stack_state[local_index];
switch (source_slot.loc()) {
case kRegister:
if (target_slot.is_reg()) state.dec_used(target_slot.reg());
target_slot.Copy(source_slot);
if (is_tee) state.inc_used(target_slot.reg());
break;
case kIntConst:
if (target_slot.is_reg()) state.dec_used(target_slot.reg());
target_slot.Copy(source_slot);
break;
case kStack:
LocalSetFromStackSlot(&target_slot, local_index);
break;
}
if (!is_tee) __ cache_state()->stack_state.pop_back();
}
void LocalSet(FullDecoder* decoder, const Value& value,
const LocalIndexImmediate<validate>& imm) {
LocalSet(imm.index, false);
}
void LocalTee(FullDecoder* decoder, const Value& value, Value* result,
const LocalIndexImmediate<validate>& imm) {
LocalSet(imm.index, true);
}
void AllocateLocals(FullDecoder* decoder, Vector<Value> local_values) {
// TODO(7748): Introduce typed functions bailout reason
unsupported(decoder, kGC, "let");
}
void DeallocateLocals(FullDecoder* decoder, uint32_t count) {
// TODO(7748): Introduce typed functions bailout reason
unsupported(decoder, kGC, "let");
}
Register GetGlobalBaseAndOffset(const WasmGlobal* global,
LiftoffRegList* pinned, uint32_t* offset) {
Register addr = pinned->set(__ GetUnusedRegister(kGpReg, {})).gp();
if (global->mutability && global->imported) {
LOAD_INSTANCE_FIELD(addr, ImportedMutableGlobals, kSystemPointerSize);
__ Load(LiftoffRegister(addr), addr, no_reg,
global->index * sizeof(Address), kPointerLoadType, *pinned);
*offset = 0;
} else {
LOAD_INSTANCE_FIELD(addr, GlobalsStart, kSystemPointerSize);
*offset = global->offset;
}
return addr;
}
void GlobalGet(FullDecoder* decoder, Value* result,
const GlobalIndexImmediate<validate>& imm) {
const auto* global = &env_->module->globals[imm.index];
if (!CheckSupportedType(decoder,
FLAG_experimental_liftoff_extern_ref
? kSupportedTypes
: kSupportedTypesWithoutRefs,
global->type, "global")) {
return;
}
if (global->type.is_reference_type()) {
if (global->mutability && global->imported) {
unsupported(decoder, kRefTypes, "imported mutable globals");
return;
}
LiftoffRegList pinned;
Register globals_buffer =
pinned.set(__ GetUnusedRegister(kGpReg, pinned)).gp();
LOAD_TAGGED_PTR_INSTANCE_FIELD(globals_buffer, TaggedGlobalsBuffer);
Register value = pinned.set(__ GetUnusedRegister(kGpReg, pinned)).gp();
__ LoadTaggedPointer(value, globals_buffer, no_reg,
wasm::ObjectAccess::ElementOffsetInTaggedFixedArray(
imm.global->offset),
pinned);
__ PushRegister(global->type, LiftoffRegister(value));
return;
}
LiftoffRegList pinned;
uint32_t offset = 0;
Register addr = GetGlobalBaseAndOffset(global, &pinned, &offset);
LiftoffRegister value =
pinned.set(__ GetUnusedRegister(reg_class_for(global->type), pinned));
LoadType type = LoadType::ForValueType(global->type);
__ Load(value, addr, no_reg, offset, type, pinned, nullptr, true);
__ PushRegister(global->type, value);
}
void GlobalSet(FullDecoder* decoder, const Value& value,
const GlobalIndexImmediate<validate>& imm) {
auto* global = &env_->module->globals[imm.index];
if (!CheckSupportedType(decoder,
FLAG_experimental_liftoff_extern_ref
? kSupportedTypes
: kSupportedTypesWithoutRefs,
global->type, "global")) {
return;
}
if (global->type.is_reference_type()) {
if (global->mutability && global->imported) {
unsupported(decoder, kRefTypes, "imported mutable globals");
return;
}
LiftoffRegList pinned;
Register globals_buffer =
pinned.set(__ GetUnusedRegister(kGpReg, pinned)).gp();
LOAD_TAGGED_PTR_INSTANCE_FIELD(globals_buffer, TaggedGlobalsBuffer);
LiftoffRegister value = pinned.set(__ PopToRegister(pinned));
__ StoreTaggedPointer(globals_buffer,
wasm::ObjectAccess::ElementOffsetInTaggedFixedArray(
imm.global->offset),
value, pinned);
return;
}
LiftoffRegList pinned;
uint32_t offset = 0;
Register addr = GetGlobalBaseAndOffset(global, &pinned, &offset);
LiftoffRegister reg = pinned.set(__ PopToRegister(pinned));
StoreType type = StoreType::ForValueType(global->type);
__ Store(addr, no_reg, offset, reg, type, {}, nullptr, true);
}
void TableGet(FullDecoder* decoder, const Value& index, Value* result,
const TableIndexImmediate<validate>& imm) {
unsupported(decoder, kRefTypes, "table_get");
}
void TableSet(FullDecoder* decoder, const Value& index, const Value& value,
const TableIndexImmediate<validate>& imm) {
unsupported(decoder, kRefTypes, "table_set");
}
void Unreachable(FullDecoder* decoder) {
Label* unreachable_label = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapUnreachable);
__ emit_jump(unreachable_label);
__ AssertUnreachable(AbortReason::kUnexpectedReturnFromWasmTrap);
}
void Select(FullDecoder* decoder, const Value& cond, const Value& fval,
const Value& tval, Value* result) {
LiftoffRegList pinned;
Register condition = pinned.set(__ PopToRegister()).gp();
ValueType type = __ cache_state()->stack_state.end()[-1].type();
DCHECK_EQ(type, __ cache_state()->stack_state.end()[-2].type());
LiftoffRegister false_value = pinned.set(__ PopToRegister(pinned));
LiftoffRegister true_value = __ PopToRegister(pinned);
LiftoffRegister dst = __ GetUnusedRegister(true_value.reg_class(),
{true_value, false_value}, {});
if (!__ emit_select(dst, condition, true_value, false_value, type)) {
// Emit generic code (using branches) instead.
Label cont;
Label case_false;
__ emit_cond_jump(kEqual, &case_false, kWasmI32, condition);
if (dst != true_value) __ Move(dst, true_value, type);
__ emit_jump(&cont);
__ bind(&case_false);
if (dst != false_value) __ Move(dst, false_value, type);
__ bind(&cont);
}
__ PushRegister(type, dst);
}
void BrImpl(Control* target) {
if (!target->br_merge()->reached) {
target->label_state.InitMerge(*__ cache_state(), __ num_locals(),
target->br_merge()->arity,
target->stack_depth);
}
__ MergeStackWith(target->label_state, target->br_merge()->arity);
__ jmp(target->label.get());
}
void Br(FullDecoder* decoder, Control* target) { BrImpl(target); }
void BrOrRet(FullDecoder* decoder, uint32_t depth) {
if (depth == decoder->control_depth() - 1) {
ReturnImpl(decoder);
} else {
BrImpl(decoder->control_at(depth));
}
}
void BrIf(FullDecoder* decoder, const Value& /* cond */, uint32_t depth) {
// Before branching, materialize all constants. This avoids repeatedly
// materializing them for each conditional branch.
// TODO(clemensb): Do the same for br_table.
if (depth != decoder->control_depth() - 1) {
__ MaterializeMergedConstants(
decoder->control_at(depth)->br_merge()->arity);
}
Label cont_false;
Register value = __ PopToRegister().gp();
if (!has_outstanding_op()) {
// Unary "equal" means "equals zero".
__ emit_cond_jump(kEqual, &cont_false, kWasmI32, value);
} else if (outstanding_op_ == kExprI32Eqz) {
// Unary "unequal" means "not equals zero".
__ emit_cond_jump(kUnequal, &cont_false, kWasmI32, value);
outstanding_op_ = kNoOutstandingOp;
} else {
// Otherwise, it's an i32 compare opcode.
Condition cond = NegateCondition(GetCompareCondition(outstanding_op_));
Register rhs = value;
Register lhs = __ PopToRegister(LiftoffRegList::ForRegs(rhs)).gp();
__ emit_cond_jump(cond, &cont_false, kWasmI32, lhs, rhs);
outstanding_op_ = kNoOutstandingOp;
}
BrOrRet(decoder, depth);
__ bind(&cont_false);
}
// Generate a branch table case, potentially reusing previously generated
// stack transfer code.
void GenerateBrCase(FullDecoder* decoder, uint32_t br_depth,
std::map<uint32_t, MovableLabel>* br_targets) {
MovableLabel& label = (*br_targets)[br_depth];
if (label.get()->is_bound()) {
__ jmp(label.get());
} else {
__ bind(label.get());
BrOrRet(decoder, br_depth);
}
}
// Generate a branch table for input in [min, max).
// TODO(wasm): Generate a real branch table (like TF TableSwitch).
void GenerateBrTable(FullDecoder* decoder, LiftoffRegister tmp,
LiftoffRegister value, uint32_t min, uint32_t max,
BranchTableIterator<validate>* table_iterator,
std::map<uint32_t, MovableLabel>* br_targets) {
DCHECK_LT(min, max);
// Check base case.
if (max == min + 1) {
DCHECK_EQ(min, table_iterator->cur_index());
GenerateBrCase(decoder, table_iterator->next(), br_targets);
return;
}
uint32_t split = min + (max - min) / 2;
Label upper_half;
__ LoadConstant(tmp, WasmValue(split));
__ emit_cond_jump(kUnsignedGreaterEqual, &upper_half, kWasmI32, value.gp(),
tmp.gp());
// Emit br table for lower half:
GenerateBrTable(decoder, tmp, value, min, split, table_iterator,
br_targets);
__ bind(&upper_half);
// table_iterator will trigger a DCHECK if we don't stop decoding now.
if (did_bailout()) return;
// Emit br table for upper half:
GenerateBrTable(decoder, tmp, value, split, max, table_iterator,
br_targets);
}
void BrTable(FullDecoder* decoder, const BranchTableImmediate<validate>& imm,
const Value& key) {
LiftoffRegList pinned;
LiftoffRegister value = pinned.set(__ PopToRegister());
BranchTableIterator<validate> table_iterator(decoder, imm);
std::map<uint32_t, MovableLabel> br_targets;
if (imm.table_count > 0) {
LiftoffRegister tmp = __ GetUnusedRegister(kGpReg, pinned);
__ LoadConstant(tmp, WasmValue(uint32_t{imm.table_count}));
Label case_default;
__ emit_cond_jump(kUnsignedGreaterEqual, &case_default, kWasmI32,
value.gp(), tmp.gp());
GenerateBrTable(decoder, tmp, value, 0, imm.table_count, &table_iterator,
&br_targets);
__ bind(&case_default);
// table_iterator will trigger a DCHECK if we don't stop decoding now.
if (did_bailout()) return;
}
// Generate the default case.
GenerateBrCase(decoder, table_iterator.next(), &br_targets);
DCHECK(!table_iterator.has_next());
}
void Else(FullDecoder* decoder, Control* c) {
if (c->reachable()) {
if (!c->end_merge.reached) {
c->label_state.InitMerge(*__ cache_state(), __ num_locals(),
c->end_merge.arity, c->stack_depth);
}
__ MergeFullStackWith(c->label_state, *__ cache_state());
__ emit_jump(c->label.get());
}
__ bind(c->else_state->label.get());
__ cache_state()->Steal(c->else_state->state);
}
SpilledRegistersForInspection* GetSpilledRegistersForInspection() {
DCHECK(for_debugging_);
// If we are generating debugging code, we really need to spill all
// registers to make them inspectable when stopping at the trap.
auto* spilled = compilation_zone_->New<SpilledRegistersForInspection>(
compilation_zone_);
for (uint32_t i = 0, e = __ cache_state()->stack_height(); i < e; ++i) {
auto& slot = __ cache_state()->stack_state[i];
if (!slot.is_reg()) continue;
spilled->entries.push_back(SpilledRegistersForInspection::Entry{
slot.offset(), slot.reg(), slot.type()});
}
return spilled;
}
Label* AddOutOfLineTrap(WasmCodePosition position,
WasmCode::RuntimeStubId stub, uint32_t pc = 0) {
DCHECK(FLAG_wasm_bounds_checks);
out_of_line_code_.push_back(OutOfLineCode::Trap(
stub, position, pc,
RegisterDebugSideTableEntry(DebugSideTableBuilder::kAssumeSpilling),
V8_UNLIKELY(for_debugging_) ? GetSpilledRegistersForInspection()
: nullptr));
return out_of_line_code_.back().label.get();
}
enum ForceCheck : bool { kDoForceCheck = true, kDontForceCheck = false };
// Returns true if the memory access is statically known to be out of bounds
// (a jump to the trap was generated then); return false otherwise.
bool BoundsCheckMem(FullDecoder* decoder, uint32_t access_size,
uint64_t offset, Register index, LiftoffRegList pinned,
ForceCheck force_check) {
// If the offset does not fit in a uintptr_t, this can never succeed on this
// machine.
const bool statically_oob =
offset > std::numeric_limits<uintptr_t>::max() ||
!base::IsInBounds<uintptr_t>(offset, access_size,
env_->max_memory_size);
if (!force_check && !statically_oob &&
(!FLAG_wasm_bounds_checks || env_->use_trap_handler)) {
return false;
}
// TODO(wasm): This adds protected instruction information for the jump
// instruction we are about to generate. It would be better to just not add
// protected instruction info when the pc is 0.
Label* trap_label = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapMemOutOfBounds,
env_->use_trap_handler ? __ pc_offset() : 0);
if (statically_oob) {
__ emit_jump(trap_label);
decoder->SetSucceedingCodeDynamicallyUnreachable();
return true;
}
uintptr_t end_offset = offset + access_size - 1u;
// If the end offset is larger than the smallest memory, dynamically check
// the end offset against the actual memory size, which is not known at
// compile time. Otherwise, only one check is required (see below).
LiftoffRegister end_offset_reg =
pinned.set(__ GetUnusedRegister(kGpReg, pinned));
Register mem_size = __ GetUnusedRegister(kGpReg, pinned).gp();
LOAD_INSTANCE_FIELD(mem_size, MemorySize, kSystemPointerSize);
__ LoadConstant(end_offset_reg, WasmValue::ForUintPtr(end_offset));
if (end_offset >= env_->min_memory_size) {
__ emit_cond_jump(kUnsignedGreaterEqual, trap_label,
LiftoffAssembler::kWasmIntPtr, end_offset_reg.gp(),
mem_size);
}
// Just reuse the end_offset register for computing the effective size.
LiftoffRegister effective_size_reg = end_offset_reg;
__ emit_ptrsize_sub(effective_size_reg.gp(), mem_size, end_offset_reg.gp());
__ emit_u32_to_intptr(index, index);
__ emit_cond_jump(kUnsignedGreaterEqual, trap_label,
LiftoffAssembler::kWasmIntPtr, index,
effective_size_reg.gp());
return false;
}
void AlignmentCheckMem(FullDecoder* decoder, uint32_t access_size,
uint32_t offset, Register index,
LiftoffRegList pinned) {
Label* trap_label = AddOutOfLineTrap(
decoder->position(), WasmCode::kThrowWasmTrapUnalignedAccess, 0);
Register address = __ GetUnusedRegister(kGpReg, pinned).gp();
const uint32_t align_mask = access_size - 1;
if ((offset & align_mask) == 0) {
// If {offset} is aligned, we can produce faster code.
// TODO(ahaas): On Intel, the "test" instruction implicitly computes the
// AND of two operands. We could introduce a new variant of
// {emit_cond_jump} to use the "test" instruction without the "and" here.
// Then we can also avoid using the temp register here.
__ emit_i32_andi(address, index, align_mask);
__ emit_cond_jump(kUnequal, trap_label, kWasmI32, address);
return;
}
__ emit_i32_addi(address, index, offset);
__ emit_i32_andi(address, address, align_mask);
__ emit_cond_jump(kUnequal, trap_label, kWasmI32, address);
}
void TraceMemoryOperation(bool is_store, MachineRepresentation rep,
Register index, uint32_t offset,
WasmCodePosition position) {
// Before making the runtime call, spill all cache registers.
__ SpillAllRegisters();
LiftoffRegList pinned = LiftoffRegList::ForRegs(index);
// Get one register for computing the effective offset (offset + index).
LiftoffRegister effective_offset =
pinned.set(__ GetUnusedRegister(kGpReg, pinned));
__ LoadConstant(effective_offset, WasmValue(offset));
__ emit_i32_add(effective_offset.gp(), effective_offset.gp(), index);
// Get a register to hold the stack slot for MemoryTracingInfo.
LiftoffRegister info = pinned.set(__ GetUnusedRegister(kGpReg, pinned));
// Allocate stack slot for MemoryTracingInfo.
__ AllocateStackSlot(info.gp(), sizeof(MemoryTracingInfo));
// Reuse the {effective_offset} register for all information to be stored in
// the MemoryTracingInfo struct.
LiftoffRegister data = effective_offset;
// Now store all information into the MemoryTracingInfo struct.
if (kSystemPointerSize == 8) {
// Zero-extend the effective offset to u64.
CHECK(__ emit_type_conversion(kExprI64UConvertI32, data, effective_offset,
nullptr));
}
__ Store(
info.gp(), no_reg, offsetof(MemoryTracingInfo, offset), data,
kSystemPointerSize == 8 ? StoreType::kI64Store : StoreType::kI32Store,
pinned);
__ LoadConstant(data, WasmValue(is_store ? 1 : 0));
__ Store(info.gp(), no_reg, offsetof(MemoryTracingInfo, is_store), data,
StoreType::kI32Store8, pinned);
__ LoadConstant(data, WasmValue(static_cast<int>(rep)));
__ Store(info.gp(), no_reg, offsetof(MemoryTracingInfo, mem_rep), data,
StoreType::kI32Store8, pinned);
WasmTraceMemoryDescriptor descriptor;
DCHECK_EQ(0, descriptor.GetStackParameterCount());
DCHECK_EQ(1, descriptor.GetRegisterParameterCount());
Register param_reg = descriptor.GetRegisterParameter(0);
if (info.gp() != param_reg) {
__ Move(param_reg, info.gp(), LiftoffAssembler::kWasmIntPtr);
}
source_position_table_builder_.AddPosition(__ pc_offset(),
SourcePosition(position), false);
__ CallRuntimeStub(WasmCode::kWasmTraceMemory);
DefineSafepoint();
__ DeallocateStackSlot(sizeof(MemoryTracingInfo));
}
Register AddMemoryMasking(Register index, uint32_t* offset,
LiftoffRegList* pinned) {
if (!FLAG_untrusted_code_mitigations || env_->use_trap_handler) {
return index;
}
DEBUG_CODE_COMMENT("mask memory index");
// Make sure that we can overwrite {index}.
if (__ cache_state()->is_used(LiftoffRegister(index))) {
Register old_index = index;
pinned->clear(LiftoffRegister(old_index));
index = pinned->set(__ GetUnusedRegister(kGpReg, *pinned)).gp();
if (index != old_index) __ Move(index, old_index, kWasmI32);
}
Register tmp = __ GetUnusedRegister(kGpReg, *pinned).gp();
__ emit_ptrsize_addi(index, index, *offset);
LOAD_INSTANCE_FIELD(tmp, MemoryMask, kSystemPointerSize);
__ emit_ptrsize_and(index, index, tmp);
*offset = 0;
return index;
}
void LoadMem(FullDecoder* decoder, LoadType type,
const MemoryAccessImmediate<validate>& imm,
const Value& index_val, Value* result) {
ValueType value_type = type.value_type();
if (!CheckSupportedType(decoder, kSupportedTypes, value_type, "load"))
return;
LiftoffRegList pinned;
Register index = pinned.set(__ PopToRegister()).gp();
if (BoundsCheckMem(decoder, type.size(), imm.offset, index, pinned,
kDontForceCheck)) {
return;
}
uint32_t offset = imm.offset;
index = AddMemoryMasking(index, &offset, &pinned);
DEBUG_CODE_COMMENT("load from memory");
Register addr = pinned.set(__ GetUnusedRegister(kGpReg, pinned)).gp();
LOAD_INSTANCE_FIELD(addr, MemoryStart, kSystemPointerSize);
RegClass rc = reg_class_for(value_type);
LiftoffRegister value = pinned.set(__ GetUnusedRegister(rc, pinned));
uint32_t protected_load_pc = 0;
__ Load(value, addr, index, offset, type, pinned, &protected_load_pc, true);
if (env_->use_trap_handler) {
AddOutOfLineTrap(decoder->position(),
WasmCode::kThrowWasmTrapMemOutOfBounds,
protected_load_pc);
}
__ PushRegister(value_type, value);
if (FLAG_trace_wasm_memory) {
TraceMemoryOperation(false, type.mem_type().representation(), index,
offset, decoder->position());
}
}
void LoadTransform(FullDecoder* decoder, LoadType type,
LoadTransformationKind transform,
const MemoryAccessImmediate<validate>& imm,
const Value& index_val, Value* result) {
// LoadTransform requires SIMD support, so check for it here. If
// unsupported, bailout and let TurboFan lower the code.
if (!CheckSupportedType(decoder, kSupportedTypes, kWasmS128,
"LoadTransform")) {
return;
}
LiftoffRegList pinned;
Register index = pinned.set(__ PopToRegister()).gp();
// For load splats and load zero, LoadType is the size of the load, and for
// load extends, LoadType is the size of the lane, and it always loads 8
// bytes.
uint32_t access_size =
transform == LoadTransformationKind::kExtend ? 8 : type.size();
if (BoundsCheckMem(decoder, access_size, imm.offset, index, pinned,
kDontForceCheck)) {
return;
}
uint32_t offset = imm.offset;
index = AddMemoryMasking(index, &offset, &pinned);
DEBUG_CODE_COMMENT("load with transformation");
Register addr = __ GetUnusedRegister(kGpReg, pinned).gp();
LOAD_INSTANCE_FIELD(addr, MemoryStart, kSystemPointerSize);
LiftoffRegister value = __ GetUnusedRegister(reg_class_for(kS128), {});
uint32_t protected_load_pc = 0;
__ LoadTransform(value, addr, index, offset, type, transform,
&protected_load_pc);
if (env_->use_trap_handler) {
AddOutOfLineTrap(decoder->position(),
WasmCode::kThrowWasmTrapMemOutOfBounds,
protected_load_pc);
}
__ PushRegister(ValueType::Primitive(kS128), value);
if (FLAG_trace_wasm_memory) {
// Again load extend is different.
MachineRepresentation mem_rep =
transform == LoadTransformationKind::kExtend
? MachineRepresentation::kWord64
: type.mem_type().representation();
TraceMemoryOperation(false, mem_rep, index, offset, decoder->position());
}
}
void LoadLane(FullDecoder* decoder, LoadType type, const Value& value,
const Value& index, const MemoryAccessImmediate<validate>& imm,
const uint8_t laneidx, Value* result) {
unsupported(decoder, kSimd, "simd load lane");
}
void StoreMem(FullDecoder* decoder, StoreType type,
const MemoryAccessImmediate<validate>& imm,
const Value& index_val, const Value& value_val) {
ValueType value_type = type.value_type();
if (!CheckSupportedType(decoder, kSupportedTypes, value_type, "store"))
return;
LiftoffRegList pinned;
LiftoffRegister value = pinned.set(__ PopToRegister());
Register index = pinned.set(__ PopToRegister(pinned)).gp();
if (BoundsCheckMem(decoder, type.size(), imm.offset, index, pinned,
kDontForceCheck)) {
return;
}
uint32_t offset = imm.offset;
index = AddMemoryMasking(index, &offset, &pinned);
DEBUG_CODE_COMMENT("store to memory");
Register addr = pinned.set(__ GetUnusedRegister(kGpReg, pinned)).gp();
LOAD_INSTANCE_FIELD(addr, MemoryStart, kSystemPointerSize);
uint32_t protected_store_pc = 0;
LiftoffRegList outer_pinned;
if (FLAG_trace_wasm_memory) outer_pinned.set(index);
__ Store(addr, index, offset, value, type, outer_pinned,
&protected_store_pc, true);
if (env_->use_trap_handler) {
AddOutOfLineTrap(decoder->position(),
WasmCode::kThrowWasmTrapMemOutOfBounds,
protected_store_pc);
}
if (FLAG_trace_wasm_memory) {
TraceMemoryOperation(true, type.mem_rep(), index, offset,
decoder->position());
}
}
void StoreLane(FullDecoder* decoder, StoreType type,
const MemoryAccessImmediate<validate>& imm, const Value& index,
const Value& value, const uint8_t laneidx) {
unsupported(decoder, kSimd, "simd load lane");
}
void CurrentMemoryPages(FullDecoder* decoder, Value* result) {
Register mem_size = __ GetUnusedRegister(kGpReg, {}).gp();
LOAD_INSTANCE_FIELD(mem_size, MemorySize, kSystemPointerSize);
__ emit_ptrsize_shri(mem_size, mem_size, kWasmPageSizeLog2);
__ PushRegister(kWasmI32, LiftoffRegister(mem_size));
}
void MemoryGrow(FullDecoder* decoder, const Value& value, Value* result_val) {
// Pop the input, then spill all cache registers to make the runtime call.
LiftoffRegList pinned;
LiftoffRegister input = pinned.set(__ PopToRegister());
__ SpillAllRegisters();
constexpr Register kGpReturnReg = kGpReturnRegisters[0];
static_assert(kLiftoffAssemblerGpCacheRegs & kGpReturnReg.bit(),
"first return register is a cache register (needs more "
"complex code here otherwise)");
LiftoffRegister result = pinned.set(LiftoffRegister(kGpReturnReg));
WasmMemoryGrowDescriptor descriptor;
DCHECK_EQ(0, descriptor.GetStackParameterCount());
DCHECK_EQ(1, descriptor.GetRegisterParameterCount());
DCHECK_EQ(kWasmI32.machine_type(), descriptor.GetParameterType(0));
Register param_reg = descriptor.GetRegisterParameter(0);
if (input.gp() != param_reg) __ Move(param_reg, input.gp(), kWasmI32);
__ CallRuntimeStub(WasmCode::kWasmMemoryGrow);
DefineSafepoint();
RegisterDebugSideTableEntry(DebugSideTableBuilder::kDidSpill);
if (kReturnRegister0 != result.gp()) {
__ Move(result.gp(), kReturnRegister0, kWasmI32);
}
__ PushRegister(kWasmI32, result);
}
DebugSideTableBuilder::EntryBuilder* RegisterDebugSideTableEntry(
DebugSideTableBuilder::AssumeSpilling assume_spilling) {
if (V8_LIKELY(!debug_sidetable_builder_)) return nullptr;
int stack_height = static_cast<int>(__ cache_state()->stack_height());
return debug_sidetable_builder_->NewEntry(
__ pc_offset(), __ num_locals(), stack_height,
__ cache_state()->stack_state.begin(), assume_spilling);
}
enum CallKind : bool { kReturnCall = true, kNoReturnCall = false };
void CallDirect(FullDecoder* decoder,
const CallFunctionImmediate<validate>& imm,
const Value args[], Value[]) {
CallDirect(decoder, imm, args, nullptr, kNoReturnCall);
}
void CallIndirect(FullDecoder* decoder, const Value& index_val,
const CallIndirectImmediate<validate>& imm,
const Value args[], Value returns[]) {
CallIndirect(decoder, index_val, imm, kNoReturnCall);
}
void CallRef(FullDecoder* decoder, const Value& func_ref,
const FunctionSig* sig, uint32_t sig_index, const Value args[],
Value returns[]) {
unsupported(decoder, kRefTypes, "call_ref");
}
void ReturnCall(FullDecoder* decoder,
const CallFunctionImmediate<validate>& imm,
const Value args[]) {
CallDirect(decoder, imm, args, nullptr, kReturnCall);
}
void ReturnCallIndirect(FullDecoder* decoder, const Value& index_val,
const CallIndirectImmediate<validate>& imm,
const Value args[]) {
CallIndirect(decoder, index_val, imm, kReturnCall);
}
void ReturnCallRef(FullDecoder* decoder, const Value& func_ref,
const FunctionSig* sig, uint32_t sig_index,
const Value args[]) {
unsupported(decoder, kRefTypes, "call_ref");
}
void BrOnNull(FullDecoder* decoder, const Value& ref_object, uint32_t depth) {
unsupported(decoder, kRefTypes, "br_on_null");
}
template <ValueType::Kind src_type, ValueType::Kind result_type,
typename EmitFn>
void EmitTerOp(EmitFn fn) {
static constexpr RegClass src_rc = reg_class_for(src_type);
static constexpr RegClass result_rc = reg_class_for(result_type);
LiftoffRegister src3 = __ PopToRegister();
LiftoffRegister src2 = __ PopToRegister(LiftoffRegList::ForRegs(src3));
LiftoffRegister src1 =
__ PopToRegister(LiftoffRegList::ForRegs(src3, src2));
// Reusing src1 and src2 will complicate codegen for select for some
// backend, so we allow only reusing src3 (the mask), and pin src1 and src2.
LiftoffRegister dst =
src_rc == result_rc
? __ GetUnusedRegister(result_rc, {src3},
LiftoffRegList::ForRegs(src1, src2))
: __ GetUnusedRegister(result_rc, {});
CallEmitFn(fn, dst, src1, src2, src3);
__ PushRegister(ValueType::Primitive(result_type), dst);
}
template <typename EmitFn, typename EmitFnImm>
void EmitSimdShiftOp(EmitFn fn, EmitFnImm fnImm) {
static constexpr RegClass result_rc = reg_class_for(ValueType::kS128);
LiftoffAssembler::VarState rhs_slot = __ cache_state()->stack_state.back();
// Check if the RHS is an immediate.
if (rhs_slot.is_const()) {
__ cache_state()->stack_state.pop_back();
int32_t imm = rhs_slot.i32_const();
LiftoffRegister operand = __ PopToRegister();
LiftoffRegister dst = __ GetUnusedRegister(result_rc, {operand}, {});
CallEmitFn(fnImm, dst, operand, imm);
__ PushRegister(kWasmS128, dst);
} else {
LiftoffRegister count = __ PopToRegister();
LiftoffRegister operand = __ PopToRegister();
LiftoffRegister dst = __ GetUnusedRegister(result_rc, {operand}, {});
CallEmitFn(fn, dst, operand, count);
__ PushRegister(kWasmS128, dst);
}
}
void EmitSimdFloatRoundingOpWithCFallback(
bool (LiftoffAssembler::*emit_fn)(LiftoffRegister, LiftoffRegister),
ExternalReference (*ext_ref)()) {
static constexpr RegClass rc = reg_class_for(kWasmS128);
LiftoffRegister src = __ PopToRegister();
LiftoffRegister dst = __ GetUnusedRegister(rc, {src}, {});
if (!(asm_.*emit_fn)(dst, src)) {
// Return v128 via stack for ARM.
ValueType sig_v_s_reps[] = {kWasmS128};
FunctionSig sig_v_s(0, 1, sig_v_s_reps);
GenerateCCall(&dst, &sig_v_s, kWasmS128, &src, ext_ref());
}
__ PushRegister(kWasmS128, dst);
}
void SimdOp(FullDecoder* decoder, WasmOpcode opcode, Vector<Value> args,
Value* result) {
if (!CpuFeatures::SupportsWasmSimd128()) {
return unsupported(decoder, kSimd, "simd");
}
switch (opcode) {
case wasm::kExprI8x16Swizzle:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_swizzle);
case wasm::kExprI8x16Splat:
return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i8x16_splat);
case wasm::kExprI16x8Splat:
return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i16x8_splat);
case wasm::kExprI32x4Splat:
return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i32x4_splat);
case wasm::kExprI64x2Splat:
return EmitUnOp<kI64, kS128>(&LiftoffAssembler::emit_i64x2_splat);
case wasm::kExprF32x4Splat:
return EmitUnOp<kF32, kS128>(&LiftoffAssembler::emit_f32x4_splat);
case wasm::kExprF64x2Splat:
return EmitUnOp<kF64, kS128>(&LiftoffAssembler::emit_f64x2_splat);
case wasm::kExprI8x16Eq:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_eq);
case wasm::kExprI8x16Ne:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_ne);
case wasm::kExprI8x16LtS:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i8x16_gt_s);
case wasm::kExprI8x16LtU:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i8x16_gt_u);
case wasm::kExprI8x16GtS:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_gt_s);
case wasm::kExprI8x16GtU:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_gt_u);
case wasm::kExprI8x16LeS:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i8x16_ge_s);
case wasm::kExprI8x16LeU:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i8x16_ge_u);
case wasm::kExprI8x16GeS:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_ge_s);
case wasm::kExprI8x16GeU:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_ge_u);
case wasm::kExprI16x8Eq:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_eq);
case wasm::kExprI16x8Ne:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_ne);
case wasm::kExprI16x8LtS:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i16x8_gt_s);
case wasm::kExprI16x8LtU:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i16x8_gt_u);
case wasm::kExprI16x8GtS:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_gt_s);
case wasm::kExprI16x8GtU:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_gt_u);
case wasm::kExprI16x8LeS:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i16x8_ge_s);
case wasm::kExprI16x8LeU:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::emit_i16x8_ge_u);
case wasm::kExprI16x8GeS:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_ge_s);
case wasm::kExprI16x8GeU:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_ge_u);
case wasm::kExprI32x4Eq:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_eq);
case wasm::kExprI32x4Ne:
return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_ne);
case wasm::kExprI32x4LtS:
return EmitBinOp<kS128, kS128, true>(
&LiftoffAssembler::