blob: 8ef70780996f9b9e6ef5eeeadb756ca45d1cd101 [file] [log] [blame]
/*
* Copyright (C) 2016-2019 Apple Inc. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY APPLE INC. ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "config.h"
#include "WasmB3IRGenerator.h"
#if ENABLE(WEBASSEMBLY)
#include "AllowMacroScratchRegisterUsageIf.h"
#include "B3BasicBlockInlines.h"
#include "B3CCallValue.h"
#include "B3Compile.h"
#include "B3ConstPtrValue.h"
#include "B3FixSSA.h"
#include "B3Generate.h"
#include "B3InsertionSet.h"
#include "B3SlotBaseValue.h"
#include "B3StackmapGenerationParams.h"
#include "B3SwitchValue.h"
#include "B3UpsilonValue.h"
#include "B3Validate.h"
#include "B3ValueInlines.h"
#include "B3ValueKey.h"
#include "B3Variable.h"
#include "B3VariableValue.h"
#include "B3WasmAddressValue.h"
#include "B3WasmBoundsCheckValue.h"
#include "DisallowMacroScratchRegisterUsage.h"
#include "JSCInlines.h"
#include "JSWebAssemblyInstance.h"
#include "ScratchRegisterAllocator.h"
#include "VirtualRegister.h"
#include "WasmCallingConvention.h"
#include "WasmContextInlines.h"
#include "WasmExceptionType.h"
#include "WasmFunctionParser.h"
#include "WasmInstance.h"
#include "WasmMemory.h"
#include "WasmOMGPlan.h"
#include "WasmOSREntryData.h"
#include "WasmOpcodeOrigin.h"
#include "WasmOperations.h"
#include "WasmSignatureInlines.h"
#include "WasmThunks.h"
#include <limits>
#include <wtf/Optional.h>
#include <wtf/StdLibExtras.h>
void dumpProcedure(void* ptr)
{
JSC::B3::Procedure* proc = static_cast<JSC::B3::Procedure*>(ptr);
proc->dump(WTF::dataFile());
}
namespace JSC { namespace Wasm {
using namespace B3;
namespace {
namespace WasmB3IRGeneratorInternal {
static constexpr bool verbose = false;
}
}
class B3IRGenerator {
public:
using ResultList = Vector<Value*, 1>; // Value must be a Phi
struct ControlData {
ControlData(Procedure& proc, Origin origin, BlockSignature signature, BlockType type, BasicBlock* continuation, BasicBlock* special = nullptr)
: controlBlockType(type)
, signature(signature)
, continuation(continuation)
, special(special)
{
if (type == BlockType::Loop) {
for (unsigned i = 0; i < signature->argumentCount(); ++i)
phis.append(proc.add<Value>(Phi, toB3Type(signature->returnType(i)), origin));
} else {
for (unsigned i = 0; i < signature->returnCount(); ++i)
phis.append(proc.add<Value>(Phi, toB3Type(signature->returnType(i)), origin));
}
}
ControlData()
{
}
void dump(PrintStream& out) const
{
switch (blockType()) {
case BlockType::If:
out.print("If: ");
break;
case BlockType::Block:
out.print("Block: ");
break;
case BlockType::Loop:
out.print("Loop: ");
break;
case BlockType::TopLevel:
out.print("TopLevel: ");
break;
}
out.print("Continuation: ", *continuation, ", Special: ");
if (special)
out.print(*special);
else
out.print("None");
}
BlockType blockType() const { return controlBlockType; }
bool hasNonVoidresult() const { return signature->returnsVoid(); }
BasicBlock* targetBlockForBranch()
{
if (blockType() == BlockType::Loop)
return special;
return continuation;
}
void convertIfToBlock()
{
ASSERT(blockType() == BlockType::If);
controlBlockType = BlockType::Block;
special = nullptr;
}
private:
friend class B3IRGenerator;
BlockType controlBlockType;
BlockSignature signature;
BasicBlock* continuation;
BasicBlock* special;
ResultList phis;
};
using ExpressionType = Value*;
using ExpressionList = Vector<ExpressionType, 1>;
using Stack = ExpressionList;
using ControlType = ControlData;
using ControlEntry = FunctionParser<B3IRGenerator>::ControlEntry;
static constexpr ExpressionType emptyExpression() { return nullptr; }
typedef String ErrorType;
typedef Unexpected<ErrorType> UnexpectedResult;
typedef Expected<std::unique_ptr<InternalFunction>, ErrorType> Result;
typedef Expected<void, ErrorType> PartialResult;
template <typename ...Args>
NEVER_INLINE UnexpectedResult WARN_UNUSED_RETURN fail(Args... args) const
{
using namespace FailureHelper; // See ADL comment in WasmParser.h.
return UnexpectedResult(makeString("WebAssembly.Module failed compiling: "_s, makeString(args)...));
}
#define WASM_COMPILE_FAIL_IF(condition, ...) do { \
if (UNLIKELY(condition)) \
return fail(__VA_ARGS__); \
} while (0)
B3IRGenerator(const ModuleInformation&, Procedure&, InternalFunction*, Vector<UnlinkedWasmToWasmCall>&, unsigned& osrEntryScratchBufferSize, MemoryMode, CompilationMode, unsigned functionIndex, unsigned loopIndexForOSREntry, TierUpCount*, ThrowWasmException);
PartialResult WARN_UNUSED_RETURN addArguments(const Signature&);
PartialResult WARN_UNUSED_RETURN addLocal(Type, uint32_t);
ExpressionType addConstant(Type, uint64_t);
// References
PartialResult WARN_UNUSED_RETURN addRefIsNull(ExpressionType& value, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN addRefFunc(uint32_t index, ExpressionType& result);
// Tables
PartialResult WARN_UNUSED_RETURN addTableGet(unsigned, ExpressionType& index, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN addTableSet(unsigned, ExpressionType& index, ExpressionType& value);
PartialResult WARN_UNUSED_RETURN addTableSize(unsigned, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN addTableGrow(unsigned, ExpressionType& fill, ExpressionType& delta, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN addTableFill(unsigned, ExpressionType& offset, ExpressionType& fill, ExpressionType& count);
// Locals
PartialResult WARN_UNUSED_RETURN getLocal(uint32_t index, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN setLocal(uint32_t index, ExpressionType value);
// Globals
PartialResult WARN_UNUSED_RETURN getGlobal(uint32_t index, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN setGlobal(uint32_t index, ExpressionType value);
// Memory
PartialResult WARN_UNUSED_RETURN load(LoadOpType, ExpressionType pointer, ExpressionType& result, uint32_t offset);
PartialResult WARN_UNUSED_RETURN store(StoreOpType, ExpressionType pointer, ExpressionType value, uint32_t offset);
PartialResult WARN_UNUSED_RETURN addGrowMemory(ExpressionType delta, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN addCurrentMemory(ExpressionType& result);
// Basic operators
template<OpType>
PartialResult WARN_UNUSED_RETURN addOp(ExpressionType arg, ExpressionType& result);
template<OpType>
PartialResult WARN_UNUSED_RETURN addOp(ExpressionType left, ExpressionType right, ExpressionType& result);
PartialResult WARN_UNUSED_RETURN addSelect(ExpressionType condition, ExpressionType nonZero, ExpressionType zero, ExpressionType& result);
// Control flow
ControlData WARN_UNUSED_RETURN addTopLevel(BlockSignature);
PartialResult WARN_UNUSED_RETURN addBlock(BlockSignature, Stack& enclosingStack, ControlType& newBlock, Stack& newStack);
PartialResult WARN_UNUSED_RETURN addLoop(BlockSignature, Stack& enclosingStack, ControlType& block, Stack& newStack, uint32_t loopIndex);
PartialResult WARN_UNUSED_RETURN addIf(ExpressionType condition, BlockSignature, Stack& enclosingStack, ControlType& result, Stack& newStack);
PartialResult WARN_UNUSED_RETURN addElse(ControlData&, const Stack&);
PartialResult WARN_UNUSED_RETURN addElseToUnreachable(ControlData&);
PartialResult WARN_UNUSED_RETURN addReturn(const ControlData&, const ExpressionList& returnValues);
PartialResult WARN_UNUSED_RETURN addBranch(ControlData&, ExpressionType condition, const Stack& returnValues);
PartialResult WARN_UNUSED_RETURN addSwitch(ExpressionType condition, const Vector<ControlData*>& targets, ControlData& defaultTargets, const Stack& expressionStack);
PartialResult WARN_UNUSED_RETURN endBlock(ControlEntry&, Stack& expressionStack);
PartialResult WARN_UNUSED_RETURN addEndToUnreachable(ControlEntry&, const Stack& = { });
// Calls
PartialResult WARN_UNUSED_RETURN addCall(uint32_t calleeIndex, const Signature&, Vector<ExpressionType>& args, ResultList& results);
PartialResult WARN_UNUSED_RETURN addCallIndirect(unsigned tableIndex, const Signature&, Vector<ExpressionType>& args, ResultList& results);
PartialResult WARN_UNUSED_RETURN addUnreachable();
B3::Value* createCallPatchpoint(BasicBlock*, Origin, const Signature&, Vector<ExpressionType>& args, const ScopedLambda<void(PatchpointValue*)>& patchpointFunctor);
void dump(const Vector<ControlEntry>& controlStack, const Stack* expressionStack);
void setParser(FunctionParser<B3IRGenerator>* parser) { m_parser = parser; };
Value* constant(B3::Type, uint64_t bits, Optional<Origin> = WTF::nullopt);
Value* framePointer();
void insertConstants();
B3::Type toB3ResultType(BlockSignature);
private:
void emitExceptionCheck(CCallHelpers&, ExceptionType);
void emitEntryTierUpCheck();
void emitLoopTierUpCheck(uint32_t loopIndex);
void emitWriteBarrierForJSWrapper();
ExpressionType emitCheckAndPreparePointer(ExpressionType pointer, uint32_t offset, uint32_t sizeOfOp);
B3::Kind memoryKind(B3::Opcode memoryOp);
ExpressionType emitLoadOp(LoadOpType, ExpressionType pointer, uint32_t offset);
void emitStoreOp(StoreOpType, ExpressionType pointer, ExpressionType value, uint32_t offset);
void unify(const ExpressionType phi, const ExpressionType source);
void unifyValuesWithBlock(const Stack& resultStack, const ResultList& stack);
void emitChecksForModOrDiv(B3::Opcode, ExpressionType left, ExpressionType right);
int32_t WARN_UNUSED_RETURN fixupPointerPlusOffset(ExpressionType&, uint32_t);
void restoreWasmContextInstance(Procedure&, BasicBlock*, Value*);
enum class RestoreCachedStackLimit { No, Yes };
void restoreWebAssemblyGlobalState(RestoreCachedStackLimit, const MemoryInformation&, Value* instance, Procedure&, BasicBlock*);
Origin origin();
uint32_t outerLoopIndex() const
{
if (m_outerLoops.isEmpty())
return UINT32_MAX;
return m_outerLoops.last();
}
FunctionParser<B3IRGenerator>* m_parser { nullptr };
const ModuleInformation& m_info;
const MemoryMode m_mode { MemoryMode::BoundsChecking };
const CompilationMode m_compilationMode { CompilationMode::BBQMode };
const unsigned m_functionIndex { UINT_MAX };
const unsigned m_loopIndexForOSREntry { UINT_MAX };
TierUpCount* m_tierUp { nullptr };
Procedure& m_proc;
BasicBlock* m_rootBlock { nullptr };
BasicBlock* m_currentBlock { nullptr };
Vector<uint32_t> m_outerLoops;
Vector<Variable*> m_locals;
Vector<UnlinkedWasmToWasmCall>& m_unlinkedWasmToWasmCalls; // List each call site and the function index whose address it should be patched with.
unsigned& m_osrEntryScratchBufferSize;
HashMap<ValueKey, Value*> m_constantPool;
HashMap<BlockSignature, B3::Type> m_tupleMap;
InsertionSet m_constantInsertionValues;
Value* m_framePointer { nullptr };
GPRReg m_memoryBaseGPR { InvalidGPRReg };
GPRReg m_memorySizeGPR { InvalidGPRReg };
GPRReg m_wasmContextInstanceGPR { InvalidGPRReg };
bool m_makesCalls { false };
Value* m_instanceValue { nullptr }; // Always use the accessor below to ensure the instance value is materialized when used.
bool m_usesInstanceValue { false };
Value* instanceValue()
{
m_usesInstanceValue = true;
return m_instanceValue;
}
uint32_t m_maxNumJSCallArguments { 0 };
unsigned m_numImportFunctions;
};
// Memory accesses in WebAssembly have unsigned 32-bit offsets, whereas they have signed 32-bit offsets in B3.
int32_t B3IRGenerator::fixupPointerPlusOffset(ExpressionType& ptr, uint32_t offset)
{
if (static_cast<uint64_t>(offset) > static_cast<uint64_t>(std::numeric_limits<int32_t>::max())) {
ptr = m_currentBlock->appendNew<Value>(m_proc, Add, origin(), ptr, m_currentBlock->appendNew<Const64Value>(m_proc, origin(), offset));
return 0;
}
return offset;
}
void B3IRGenerator::restoreWasmContextInstance(Procedure& proc, BasicBlock* block, Value* arg)
{
if (Context::useFastTLS()) {
PatchpointValue* patchpoint = block->appendNew<PatchpointValue>(proc, B3::Void, Origin());
if (CCallHelpers::storeWasmContextInstanceNeedsMacroScratchRegister())
patchpoint->clobber(RegisterSet::macroScratchRegisters());
patchpoint->append(ConstrainedValue(arg, ValueRep::SomeRegister));
patchpoint->setGenerator(
[=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsageIf allowScratch(jit, CCallHelpers::storeWasmContextInstanceNeedsMacroScratchRegister());
jit.storeWasmContextInstance(params[0].gpr());
});
return;
}
// FIXME: Because WasmToWasm call clobbers wasmContextInstance register and does not restore it, we need to restore it in the caller side.
// This prevents us from using ArgumentReg to this (logically) immutable pinned register.
PatchpointValue* patchpoint = block->appendNew<PatchpointValue>(proc, B3::Void, Origin());
Effects effects = Effects::none();
effects.writesPinned = true;
effects.reads = B3::HeapRange::top();
patchpoint->effects = effects;
patchpoint->clobberLate(RegisterSet(m_wasmContextInstanceGPR));
patchpoint->append(arg, ValueRep::SomeRegister);
GPRReg wasmContextInstanceGPR = m_wasmContextInstanceGPR;
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& param) {
jit.move(param[0].gpr(), wasmContextInstanceGPR);
});
}
B3IRGenerator::B3IRGenerator(const ModuleInformation& info, Procedure& procedure, InternalFunction* compilation, Vector<UnlinkedWasmToWasmCall>& unlinkedWasmToWasmCalls, unsigned& osrEntryScratchBufferSize, MemoryMode mode, CompilationMode compilationMode, unsigned functionIndex, unsigned loopIndexForOSREntry, TierUpCount* tierUp, ThrowWasmException throwWasmException)
: m_info(info)
, m_mode(mode)
, m_compilationMode(compilationMode)
, m_functionIndex(functionIndex)
, m_loopIndexForOSREntry(loopIndexForOSREntry)
, m_tierUp(tierUp)
, m_proc(procedure)
, m_unlinkedWasmToWasmCalls(unlinkedWasmToWasmCalls)
, m_osrEntryScratchBufferSize(osrEntryScratchBufferSize)
, m_constantInsertionValues(m_proc)
, m_numImportFunctions(info.importFunctionCount())
{
m_rootBlock = m_proc.addBlock();
m_currentBlock = m_rootBlock;
// FIXME we don't really need to pin registers here if there's no memory. It makes wasm -> wasm thunks simpler for now. https://bugs.webkit.org/show_bug.cgi?id=166623
const PinnedRegisterInfo& pinnedRegs = PinnedRegisterInfo::get();
m_memoryBaseGPR = pinnedRegs.baseMemoryPointer;
m_proc.pinRegister(m_memoryBaseGPR);
m_wasmContextInstanceGPR = pinnedRegs.wasmContextInstancePointer;
if (!Context::useFastTLS())
m_proc.pinRegister(m_wasmContextInstanceGPR);
if (mode != MemoryMode::Signaling) {
m_memorySizeGPR = pinnedRegs.sizeRegister;
m_proc.pinRegister(m_memorySizeGPR);
}
if (throwWasmException)
Thunks::singleton().setThrowWasmException(throwWasmException);
if (info.memory) {
m_proc.setWasmBoundsCheckGenerator([=] (CCallHelpers& jit, GPRReg pinnedGPR) {
AllowMacroScratchRegisterUsage allowScratch(jit);
switch (m_mode) {
case MemoryMode::BoundsChecking:
ASSERT_UNUSED(pinnedGPR, m_memorySizeGPR == pinnedGPR);
break;
case MemoryMode::Signaling:
ASSERT_UNUSED(pinnedGPR, InvalidGPRReg == pinnedGPR);
break;
}
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsMemoryAccess);
});
switch (m_mode) {
case MemoryMode::BoundsChecking:
break;
case MemoryMode::Signaling:
// Most memory accesses in signaling mode don't do an explicit
// exception check because they can rely on fault handling to detect
// out-of-bounds accesses. FaultSignalHandler nonetheless needs the
// thunk to exist so that it can jump to that thunk.
if (UNLIKELY(!Thunks::singleton().stub(throwExceptionFromWasmThunkGenerator)))
CRASH();
break;
}
}
{
auto* calleeMoveLocation = &compilation->calleeMoveLocation;
static_assert(CallFrameSlot::codeBlock * sizeof(Register) < WasmCallingConvention::headerSizeInBytes, "We rely on this here for now.");
static_assert(CallFrameSlot::callee * sizeof(Register) < WasmCallingConvention::headerSizeInBytes, "We rely on this here for now.");
B3::PatchpointValue* getCalleePatchpoint = m_currentBlock->appendNew<B3::PatchpointValue>(m_proc, B3::Int64, Origin());
getCalleePatchpoint->resultConstraints = { B3::ValueRep::SomeRegister };
getCalleePatchpoint->effects = B3::Effects::none();
getCalleePatchpoint->setGenerator(
[=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
GPRReg result = params[0].gpr();
MacroAssembler::DataLabelPtr moveLocation = jit.moveWithPatch(MacroAssembler::TrustedImmPtr(nullptr), result);
jit.addLinkTask([calleeMoveLocation, moveLocation] (LinkBuffer& linkBuffer) {
*calleeMoveLocation = linkBuffer.locationOf<WasmEntryPtrTag>(moveLocation);
});
});
B3::Value* offsetOfCallee = m_currentBlock->appendNew<B3::Const64Value>(m_proc, Origin(), CallFrameSlot::callee * sizeof(Register));
m_currentBlock->appendNew<B3::MemoryValue>(m_proc, B3::Store, Origin(),
getCalleePatchpoint,
m_currentBlock->appendNew<B3::Value>(m_proc, B3::Add, Origin(), framePointer(), offsetOfCallee));
// FIXME: We shouldn't have to store zero into the CodeBlock* spot in the call frame,
// but there are places that interpret non-null CodeBlock slot to mean a valid CodeBlock.
// When doing unwinding, we'll need to verify that the entire runtime is OK with a non-null
// CodeBlock not implying that the CodeBlock is valid.
// https://bugs.webkit.org/show_bug.cgi?id=165321
B3::Value* offsetOfCodeBlock = m_currentBlock->appendNew<B3::Const64Value>(m_proc, Origin(), CallFrameSlot::codeBlock * sizeof(Register));
m_currentBlock->appendNew<B3::MemoryValue>(m_proc, B3::Store, Origin(),
m_currentBlock->appendNew<B3::Const64Value>(m_proc, Origin(), 0),
m_currentBlock->appendNew<B3::Value>(m_proc, B3::Add, Origin(), framePointer(), offsetOfCodeBlock));
}
{
B3::PatchpointValue* stackOverflowCheck = m_currentBlock->appendNew<B3::PatchpointValue>(m_proc, pointerType(), Origin());
m_instanceValue = stackOverflowCheck;
stackOverflowCheck->appendSomeRegister(framePointer());
stackOverflowCheck->clobber(RegisterSet::macroScratchRegisters());
if (!Context::useFastTLS()) {
// FIXME: Because WasmToWasm call clobbers wasmContextInstance register and does not restore it, we need to restore it in the caller side.
// This prevents us from using ArgumentReg to this (logically) immutable pinned register.
stackOverflowCheck->effects.writesPinned = false;
stackOverflowCheck->effects.readsPinned = true;
stackOverflowCheck->resultConstraints = { ValueRep::reg(m_wasmContextInstanceGPR) };
}
stackOverflowCheck->numGPScratchRegisters = 2;
stackOverflowCheck->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
const Checked<int32_t> wasmFrameSize = params.proc().frameSize();
const unsigned minimumParentCheckSize = WTF::roundUpToMultipleOf(stackAlignmentBytes(), 1024);
const unsigned extraFrameSize = WTF::roundUpToMultipleOf(stackAlignmentBytes(), std::max<uint32_t>(
// This allows us to elide stack checks for functions that are terminal nodes in the call
// tree, (e.g they don't make any calls) and have a small enough frame size. This works by
// having any such terminal node have its parent caller include some extra size in its
// own check for it. The goal here is twofold:
// 1. Emit less code.
// 2. Try to speed things up by skipping stack checks.
minimumParentCheckSize,
// This allows us to elide stack checks in the Wasm -> Embedder call IC stub. Since these will
// spill all arguments to the stack, we ensure that a stack check here covers the
// stack that such a stub would use.
(Checked<uint32_t>(m_maxNumJSCallArguments) * sizeof(Register) + JSCallingConvention::headerSizeInBytes).unsafeGet()
));
const int32_t checkSize = m_makesCalls ? (wasmFrameSize + extraFrameSize).unsafeGet() : wasmFrameSize.unsafeGet();
bool needUnderflowCheck = static_cast<unsigned>(checkSize) > Options::reservedZoneSize();
bool needsOverflowCheck = m_makesCalls || wasmFrameSize >= minimumParentCheckSize || needUnderflowCheck;
GPRReg contextInstance = Context::useFastTLS() ? params[0].gpr() : m_wasmContextInstanceGPR;
// This allows leaf functions to not do stack checks if their frame size is within
// certain limits since their caller would have already done the check.
if (needsOverflowCheck) {
AllowMacroScratchRegisterUsage allowScratch(jit);
GPRReg fp = params[1].gpr();
GPRReg scratch1 = params.gpScratch(0);
GPRReg scratch2 = params.gpScratch(1);
if (Context::useFastTLS())
jit.loadWasmContextInstance(contextInstance);
jit.loadPtr(CCallHelpers::Address(contextInstance, Instance::offsetOfCachedStackLimit()), scratch2);
jit.addPtr(CCallHelpers::TrustedImm32(-checkSize), fp, scratch1);
MacroAssembler::JumpList overflow;
if (UNLIKELY(needUnderflowCheck))
overflow.append(jit.branchPtr(CCallHelpers::Above, scratch1, fp));
overflow.append(jit.branchPtr(CCallHelpers::Below, scratch1, scratch2));
jit.addLinkTask([overflow] (LinkBuffer& linkBuffer) {
linkBuffer.link(overflow, CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(throwStackOverflowFromWasmThunkGenerator).code()));
});
} else if (m_usesInstanceValue && Context::useFastTLS()) {
// No overflow check is needed, but the instance values still needs to be correct.
AllowMacroScratchRegisterUsageIf allowScratch(jit, CCallHelpers::loadWasmContextInstanceNeedsMacroScratchRegister());
jit.loadWasmContextInstance(contextInstance);
} else {
// We said we'd return a pointer. We don't actually need to because it isn't used, but the patchpoint conservatively said it had effects (potential stack check) which prevent it from getting removed.
}
});
}
emitEntryTierUpCheck();
if (m_compilationMode == CompilationMode::OMGForOSREntryMode)
m_currentBlock = m_proc.addBlock();
}
void B3IRGenerator::restoreWebAssemblyGlobalState(RestoreCachedStackLimit restoreCachedStackLimit, const MemoryInformation& memory, Value* instance, Procedure& proc, BasicBlock* block)
{
restoreWasmContextInstance(proc, block, instance);
if (restoreCachedStackLimit == RestoreCachedStackLimit::Yes) {
// The Instance caches the stack limit, but also knows where its canonical location is.
Value* pointerToActualStackLimit = block->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfPointerToActualStackLimit()));
Value* actualStackLimit = block->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), pointerToActualStackLimit);
block->appendNew<MemoryValue>(m_proc, Store, origin(), actualStackLimit, instanceValue(), safeCast<int32_t>(Instance::offsetOfCachedStackLimit()));
}
if (!!memory) {
const PinnedRegisterInfo* pinnedRegs = &PinnedRegisterInfo::get();
RegisterSet clobbers;
clobbers.set(pinnedRegs->baseMemoryPointer);
clobbers.set(pinnedRegs->sizeRegister);
if (!isARM64())
clobbers.set(RegisterSet::macroScratchRegisters());
B3::PatchpointValue* patchpoint = block->appendNew<B3::PatchpointValue>(proc, B3::Void, origin());
Effects effects = Effects::none();
effects.writesPinned = true;
effects.reads = B3::HeapRange::top();
patchpoint->effects = effects;
patchpoint->clobber(clobbers);
patchpoint->numGPScratchRegisters = Gigacage::isEnabled(Gigacage::Primitive) ? 1 : 0;
patchpoint->append(instance, ValueRep::SomeRegister);
patchpoint->setGenerator([pinnedRegs] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
GPRReg baseMemory = pinnedRegs->baseMemoryPointer;
GPRReg scratchOrSize = Gigacage::isEnabled(Gigacage::Primitive) ? params.gpScratch(0) : pinnedRegs->sizeRegister;
jit.loadPtr(CCallHelpers::Address(params[0].gpr(), Instance::offsetOfCachedMemorySize()), pinnedRegs->sizeRegister);
jit.loadPtr(CCallHelpers::Address(params[0].gpr(), Instance::offsetOfCachedMemory()), baseMemory);
jit.cageConditionally(Gigacage::Primitive, baseMemory, pinnedRegs->sizeRegister, scratchOrSize);
});
}
}
void B3IRGenerator::emitExceptionCheck(CCallHelpers& jit, ExceptionType type)
{
jit.move(CCallHelpers::TrustedImm32(static_cast<uint32_t>(type)), GPRInfo::argumentGPR1);
auto jumpToExceptionStub = jit.jump();
jit.addLinkTask([jumpToExceptionStub] (LinkBuffer& linkBuffer) {
linkBuffer.link(jumpToExceptionStub, CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(throwExceptionFromWasmThunkGenerator).code()));
});
}
Value* B3IRGenerator::constant(B3::Type type, uint64_t bits, Optional<Origin> maybeOrigin)
{
auto result = m_constantPool.ensure(ValueKey(opcodeForConstant(type), type, static_cast<int64_t>(bits)), [&] {
Value* result = m_proc.addConstant(maybeOrigin ? *maybeOrigin : origin(), type, bits);
m_constantInsertionValues.insertValue(0, result);
return result;
});
return result.iterator->value;
}
Value* B3IRGenerator::framePointer()
{
if (!m_framePointer) {
m_framePointer = m_proc.add<B3::Value>(B3::FramePointer, Origin());
ASSERT(m_framePointer);
m_constantInsertionValues.insertValue(0, m_framePointer);
}
return m_framePointer;
}
void B3IRGenerator::insertConstants()
{
m_constantInsertionValues.execute(m_proc.at(0));
}
B3::Type B3IRGenerator::toB3ResultType(BlockSignature returnType)
{
if (returnType->returnsVoid())
return B3::Void;
if (returnType->returnCount() == 1)
return toB3Type(returnType->returnType(0));
auto result = m_tupleMap.ensure(returnType, [&] {
Vector<B3::Type> result;
for (unsigned i = 0; i < returnType->returnCount(); ++i)
result.append(toB3Type(returnType->returnType(i)));
return m_proc.addTuple(WTFMove(result));
});
return result.iterator->value;
}
auto B3IRGenerator::addLocal(Type type, uint32_t count) -> PartialResult
{
size_t newSize = m_locals.size() + count;
ASSERT(!(CheckedUint32(count) + m_locals.size()).hasOverflowed());
ASSERT(newSize <= maxFunctionLocals);
WASM_COMPILE_FAIL_IF(!m_locals.tryReserveCapacity(newSize), "can't allocate memory for ", newSize, " locals");
for (uint32_t i = 0; i < count; ++i) {
Variable* local = m_proc.addVariable(toB3Type(type));
m_locals.uncheckedAppend(local);
auto val = isSubtype(type, Anyref) ? JSValue::encode(jsNull()) : 0;
m_currentBlock->appendNew<VariableValue>(m_proc, Set, Origin(), local, constant(toB3Type(type), val, Origin()));
}
return { };
}
auto B3IRGenerator::addArguments(const Signature& signature) -> PartialResult
{
ASSERT(!m_locals.size());
WASM_COMPILE_FAIL_IF(!m_locals.tryReserveCapacity(signature.argumentCount()), "can't allocate memory for ", signature.argumentCount(), " arguments");
m_locals.grow(signature.argumentCount());
CallInformation wasmCallInfo = wasmCallingConvention().callInformationFor(signature, CallRole::Callee);
for (size_t i = 0; i < signature.argumentCount(); ++i) {
B3::Type type = toB3Type(signature.argument(i));
B3::Value* argument;
auto rep = wasmCallInfo.params[i];
if (rep.isReg()) {
argument = m_currentBlock->appendNew<B3::ArgumentRegValue>(m_proc, Origin(), rep.reg());
if (type == B3::Int32 || type == B3::Float)
argument = m_currentBlock->appendNew<B3::Value>(m_proc, B3::Trunc, Origin(), argument);
} else {
ASSERT(rep.isStack());
B3::Value* address = m_currentBlock->appendNew<B3::Value>(m_proc, B3::Add, Origin(), framePointer(),
m_currentBlock->appendNew<B3::Const64Value>(m_proc, Origin(), rep.offsetFromFP()));
argument = m_currentBlock->appendNew<B3::MemoryValue>(m_proc, B3::Load, type, Origin(), address);
}
Variable* argumentVariable = m_proc.addVariable(argument->type());
m_locals[i] = argumentVariable;
m_currentBlock->appendNew<VariableValue>(m_proc, Set, Origin(), argumentVariable, argument);
}
return { };
}
auto B3IRGenerator::addRefIsNull(ExpressionType& value, ExpressionType& result) -> PartialResult
{
result = m_currentBlock->appendNew<Value>(m_proc, B3::Equal, origin(), value, m_currentBlock->appendNew<Const64Value>(m_proc, origin(), JSValue::encode(jsNull())));
return { };
}
auto B3IRGenerator::addTableGet(unsigned tableIndex, ExpressionType& index, ExpressionType& result) -> PartialResult
{
// FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>.
result = m_currentBlock->appendNew<CCallValue>(m_proc, toB3Type(Anyref), origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(&getWasmTableElement, B3CCallPtrTag)),
instanceValue(), m_currentBlock->appendNew<Const32Value>(m_proc, origin(), tableIndex), index);
{
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), result, m_currentBlock->appendNew<Const64Value>(m_proc, origin(), 0)));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTableAccess);
});
}
return { };
}
auto B3IRGenerator::addTableSet(unsigned tableIndex, ExpressionType& index, ExpressionType& value) -> PartialResult
{
// FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>.
auto shouldThrow = m_currentBlock->appendNew<CCallValue>(m_proc, B3::Int32, origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(&setWasmTableElement, B3CCallPtrTag)),
instanceValue(), m_currentBlock->appendNew<Const32Value>(m_proc, origin(), tableIndex), index, value);
{
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), shouldThrow, m_currentBlock->appendNew<Const32Value>(m_proc, origin(), 0)));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTableAccess);
});
}
return { };
}
auto B3IRGenerator::addRefFunc(uint32_t index, ExpressionType& result) -> PartialResult
{
// FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>.
result = m_currentBlock->appendNew<CCallValue>(m_proc, B3::Int64, origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(&doWasmRefFunc, B3CCallPtrTag)),
instanceValue(), addConstant(Type::I32, index));
return { };
}
auto B3IRGenerator::addTableSize(unsigned tableIndex, ExpressionType& result) -> PartialResult
{
// FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>.
uint32_t (*doSize)(Instance*, unsigned) = [] (Instance* instance, unsigned tableIndex) -> uint32_t {
return instance->table(tableIndex)->length();
};
result = m_currentBlock->appendNew<CCallValue>(m_proc, toB3Type(I32), origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(doSize, B3CCallPtrTag)),
instanceValue(), m_currentBlock->appendNew<Const32Value>(m_proc, origin(), tableIndex));
return { };
}
auto B3IRGenerator::addTableGrow(unsigned tableIndex, ExpressionType& fill, ExpressionType& delta, ExpressionType& result) -> PartialResult
{
result = m_currentBlock->appendNew<CCallValue>(m_proc, toB3Type(I32), origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(&doWasmTableGrow, B3CCallPtrTag)),
instanceValue(), m_currentBlock->appendNew<Const32Value>(m_proc, origin(), tableIndex), fill, delta);
return { };
}
auto B3IRGenerator::addTableFill(unsigned tableIndex, ExpressionType& offset, ExpressionType& fill, ExpressionType& count) -> PartialResult
{
auto result = m_currentBlock->appendNew<CCallValue>(m_proc, toB3Type(I32), origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(&doWasmTableFill, B3CCallPtrTag)),
instanceValue(), m_currentBlock->appendNew<Const32Value>(m_proc, origin(), tableIndex), offset, fill, count);
{
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), result, m_currentBlock->appendNew<Const32Value>(m_proc, origin(), 0)));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTableAccess);
});
}
return { };
}
auto B3IRGenerator::getLocal(uint32_t index, ExpressionType& result) -> PartialResult
{
ASSERT(m_locals[index]);
result = m_currentBlock->appendNew<VariableValue>(m_proc, B3::Get, origin(), m_locals[index]);
return { };
}
auto B3IRGenerator::addUnreachable() -> PartialResult
{
B3::PatchpointValue* unreachable = m_currentBlock->appendNew<B3::PatchpointValue>(m_proc, B3::Void, origin());
unreachable->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::Unreachable);
});
unreachable->effects.terminal = true;
return { };
}
auto B3IRGenerator::addGrowMemory(ExpressionType delta, ExpressionType& result) -> PartialResult
{
int32_t (*growMemory)(void*, Instance*, int32_t) = [] (void* callFrame, Instance* instance, int32_t delta) -> int32_t {
instance->storeTopCallFrame(callFrame);
if (delta < 0)
return -1;
auto grown = instance->memory()->grow(PageCount(delta));
if (!grown) {
switch (grown.error()) {
case Memory::GrowFailReason::InvalidDelta:
case Memory::GrowFailReason::InvalidGrowSize:
case Memory::GrowFailReason::WouldExceedMaximum:
case Memory::GrowFailReason::OutOfMemory:
return -1;
}
RELEASE_ASSERT_NOT_REACHED();
}
return grown.value().pageCount();
};
result = m_currentBlock->appendNew<CCallValue>(m_proc, Int32, origin(),
m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(growMemory, B3CCallPtrTag)),
framePointer(), instanceValue(), delta);
restoreWebAssemblyGlobalState(RestoreCachedStackLimit::No, m_info.memory, instanceValue(), m_proc, m_currentBlock);
return { };
}
auto B3IRGenerator::addCurrentMemory(ExpressionType& result) -> PartialResult
{
static_assert(sizeof(decltype(static_cast<Memory*>(nullptr)->size())) == sizeof(uint64_t), "codegen relies on this size");
Value* size = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, Int64, origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfCachedMemorySize()));
constexpr uint32_t shiftValue = 16;
static_assert(PageCount::pageSize == 1ull << shiftValue, "This must hold for the code below to be correct.");
Value* numPages = m_currentBlock->appendNew<Value>(m_proc, ZShr, origin(),
size, m_currentBlock->appendNew<Const32Value>(m_proc, origin(), shiftValue));
result = m_currentBlock->appendNew<Value>(m_proc, Trunc, origin(), numPages);
return { };
}
auto B3IRGenerator::setLocal(uint32_t index, ExpressionType value) -> PartialResult
{
ASSERT(m_locals[index]);
m_currentBlock->appendNew<VariableValue>(m_proc, B3::Set, origin(), m_locals[index], value);
return { };
}
auto B3IRGenerator::getGlobal(uint32_t index, ExpressionType& result) -> PartialResult
{
Value* globalsArray = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfGlobals()));
result = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, toB3Type(m_info.globals[index].type), origin(), globalsArray, safeCast<int32_t>(index * sizeof(Register)));
return { };
}
auto B3IRGenerator::setGlobal(uint32_t index, ExpressionType value) -> PartialResult
{
ASSERT(toB3Type(m_info.globals[index].type) == value->type());
Value* globalsArray = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfGlobals()));
m_currentBlock->appendNew<MemoryValue>(m_proc, Store, origin(), value, globalsArray, safeCast<int32_t>(index * sizeof(Register)));
if (isSubtype(m_info.globals[index].type, Anyref))
emitWriteBarrierForJSWrapper();
return { };
}
inline void B3IRGenerator::emitWriteBarrierForJSWrapper()
{
Value* cell = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfOwner()));
Value* cellState = m_currentBlock->appendNew<MemoryValue>(m_proc, Load8Z, Int32, origin(), cell, safeCast<int32_t>(JSCell::cellStateOffset()));
Value* vm = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), cell, safeCast<int32_t>(JSWebAssemblyInstance::offsetOfVM()));
Value* threshold = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, Int32, origin(), vm, safeCast<int32_t>(VM::offsetOfHeapBarrierThreshold()));
BasicBlock* fenceCheckPath = m_proc.addBlock();
BasicBlock* fencePath = m_proc.addBlock();
BasicBlock* doSlowPath = m_proc.addBlock();
BasicBlock* continuation = m_proc.addBlock();
m_currentBlock->appendNewControlValue(m_proc, B3::Branch, origin(),
m_currentBlock->appendNew<Value>(m_proc, Above, origin(), cellState, threshold),
FrequentedBlock(continuation), FrequentedBlock(fenceCheckPath, FrequencyClass::Rare));
fenceCheckPath->addPredecessor(m_currentBlock);
continuation->addPredecessor(m_currentBlock);
m_currentBlock = fenceCheckPath;
Value* shouldFence = m_currentBlock->appendNew<MemoryValue>(m_proc, Load8Z, Int32, origin(), vm, safeCast<int32_t>(VM::offsetOfHeapMutatorShouldBeFenced()));
m_currentBlock->appendNewControlValue(m_proc, B3::Branch, origin(),
shouldFence,
FrequentedBlock(fencePath), FrequentedBlock(doSlowPath));
fencePath->addPredecessor(m_currentBlock);
doSlowPath->addPredecessor(m_currentBlock);
m_currentBlock = fencePath;
B3::PatchpointValue* doFence = m_currentBlock->appendNew<B3::PatchpointValue>(m_proc, B3::Void, origin());
doFence->setGenerator([] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
jit.memoryFence();
});
Value* cellStateLoadAfterFence = m_currentBlock->appendNew<MemoryValue>(m_proc, Load8Z, Int32, origin(), cell, safeCast<int32_t>(JSCell::cellStateOffset()));
m_currentBlock->appendNewControlValue(m_proc, B3::Branch, origin(),
m_currentBlock->appendNew<Value>(m_proc, Above, origin(), cellStateLoadAfterFence, m_currentBlock->appendNew<Const32Value>(m_proc, origin(), blackThreshold)),
FrequentedBlock(continuation), FrequentedBlock(doSlowPath, FrequencyClass::Rare));
doSlowPath->addPredecessor(m_currentBlock);
continuation->addPredecessor(m_currentBlock);
m_currentBlock = doSlowPath;
void (*writeBarrier)(JSWebAssemblyInstance*, VM*) = [] (JSWebAssemblyInstance* cell, VM* vm) -> void {
vm->heap.writeBarrierSlowPath(cell);
};
Value* writeBarrierAddress = m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(writeBarrier, B3CCallPtrTag));
m_currentBlock->appendNew<CCallValue>(m_proc, B3::Void, origin(), writeBarrierAddress, cell, vm);
m_currentBlock->appendNewControlValue(m_proc, Jump, origin(), continuation);
continuation->addPredecessor(m_currentBlock);
m_currentBlock = continuation;
}
inline Value* B3IRGenerator::emitCheckAndPreparePointer(ExpressionType pointer, uint32_t offset, uint32_t sizeOfOperation)
{
ASSERT(m_memoryBaseGPR);
switch (m_mode) {
case MemoryMode::BoundsChecking: {
// We're not using signal handling at all, we must therefore check that no memory access exceeds the current memory size.
ASSERT(m_memorySizeGPR);
ASSERT(sizeOfOperation + offset > offset);
m_currentBlock->appendNew<WasmBoundsCheckValue>(m_proc, origin(), m_memorySizeGPR, pointer, sizeOfOperation + offset - 1);
break;
}
case MemoryMode::Signaling: {
// We've virtually mapped 4GiB+redzone for this memory. Only the user-allocated pages are addressable, contiguously in range [0, current],
// and everything above is mapped PROT_NONE. We don't need to perform any explicit bounds check in the 4GiB range because WebAssembly register
// memory accesses are 32-bit. However WebAssembly register + offset accesses perform the addition in 64-bit which can push an access above
// the 32-bit limit (the offset is unsigned 32-bit). The redzone will catch most small offsets, and we'll explicitly bounds check any
// register + large offset access. We don't think this will be generated frequently.
//
// We could check that register + large offset doesn't exceed 4GiB+redzone since that's technically the limit we need to avoid overflowing the
// PROT_NONE region, but it's better if we use a smaller immediate because it can codegens better. We know that anything equal to or greater
// than the declared 'maximum' will trap, so we can compare against that number. If there was no declared 'maximum' then we still know that
// any access equal to or greater than 4GiB will trap, no need to add the redzone.
if (offset >= Memory::fastMappedRedzoneBytes()) {
size_t maximum = m_info.memory.maximum() ? m_info.memory.maximum().bytes() : std::numeric_limits<uint32_t>::max();
m_currentBlock->appendNew<WasmBoundsCheckValue>(m_proc, origin(), pointer, sizeOfOperation + offset - 1, maximum);
}
break;
}
}
pointer = m_currentBlock->appendNew<Value>(m_proc, ZExt32, origin(), pointer);
return m_currentBlock->appendNew<WasmAddressValue>(m_proc, origin(), pointer, m_memoryBaseGPR);
}
inline uint32_t sizeOfLoadOp(LoadOpType op)
{
switch (op) {
case LoadOpType::I32Load8S:
case LoadOpType::I32Load8U:
case LoadOpType::I64Load8S:
case LoadOpType::I64Load8U:
return 1;
case LoadOpType::I32Load16S:
case LoadOpType::I64Load16S:
case LoadOpType::I32Load16U:
case LoadOpType::I64Load16U:
return 2;
case LoadOpType::I32Load:
case LoadOpType::I64Load32S:
case LoadOpType::I64Load32U:
case LoadOpType::F32Load:
return 4;
case LoadOpType::I64Load:
case LoadOpType::F64Load:
return 8;
}
RELEASE_ASSERT_NOT_REACHED();
}
inline B3::Kind B3IRGenerator::memoryKind(B3::Opcode memoryOp)
{
if (m_mode == MemoryMode::Signaling)
return trapping(memoryOp);
return memoryOp;
}
inline Value* B3IRGenerator::emitLoadOp(LoadOpType op, ExpressionType pointer, uint32_t uoffset)
{
int32_t offset = fixupPointerPlusOffset(pointer, uoffset);
switch (op) {
case LoadOpType::I32Load8S: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load8S), origin(), pointer, offset);
}
case LoadOpType::I64Load8S: {
Value* value = m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load8S), origin(), pointer, offset);
return m_currentBlock->appendNew<Value>(m_proc, SExt32, origin(), value);
}
case LoadOpType::I32Load8U: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load8Z), origin(), pointer, offset);
}
case LoadOpType::I64Load8U: {
Value* value = m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load8Z), origin(), pointer, offset);
return m_currentBlock->appendNew<Value>(m_proc, ZExt32, origin(), value);
}
case LoadOpType::I32Load16S: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load16S), origin(), pointer, offset);
}
case LoadOpType::I64Load16S: {
Value* value = m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load16S), origin(), pointer, offset);
return m_currentBlock->appendNew<Value>(m_proc, SExt32, origin(), value);
}
case LoadOpType::I32Load16U: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load16Z), origin(), pointer, offset);
}
case LoadOpType::I64Load16U: {
Value* value = m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load16Z), origin(), pointer, offset);
return m_currentBlock->appendNew<Value>(m_proc, ZExt32, origin(), value);
}
case LoadOpType::I32Load: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load), Int32, origin(), pointer, offset);
}
case LoadOpType::I64Load32U: {
Value* value = m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load), Int32, origin(), pointer, offset);
return m_currentBlock->appendNew<Value>(m_proc, ZExt32, origin(), value);
}
case LoadOpType::I64Load32S: {
Value* value = m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load), Int32, origin(), pointer, offset);
return m_currentBlock->appendNew<Value>(m_proc, SExt32, origin(), value);
}
case LoadOpType::I64Load: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load), Int64, origin(), pointer, offset);
}
case LoadOpType::F32Load: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load), Float, origin(), pointer, offset);
}
case LoadOpType::F64Load: {
return m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Load), Double, origin(), pointer, offset);
}
}
RELEASE_ASSERT_NOT_REACHED();
}
auto B3IRGenerator::load(LoadOpType op, ExpressionType pointer, ExpressionType& result, uint32_t offset) -> PartialResult
{
ASSERT(pointer->type() == Int32);
if (UNLIKELY(sumOverflows<uint32_t>(offset, sizeOfLoadOp(op)))) {
// FIXME: Even though this is provably out of bounds, it's not a validation error, so we have to handle it
// as a runtime exception. However, this may change: https://bugs.webkit.org/show_bug.cgi?id=166435
B3::PatchpointValue* throwException = m_currentBlock->appendNew<B3::PatchpointValue>(m_proc, B3::Void, origin());
throwException->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsMemoryAccess);
});
switch (op) {
case LoadOpType::I32Load8S:
case LoadOpType::I32Load16S:
case LoadOpType::I32Load:
case LoadOpType::I32Load16U:
case LoadOpType::I32Load8U:
result = constant(Int32, 0);
break;
case LoadOpType::I64Load8S:
case LoadOpType::I64Load8U:
case LoadOpType::I64Load16S:
case LoadOpType::I64Load32U:
case LoadOpType::I64Load32S:
case LoadOpType::I64Load:
case LoadOpType::I64Load16U:
result = constant(Int64, 0);
break;
case LoadOpType::F32Load:
result = constant(Float, 0);
break;
case LoadOpType::F64Load:
result = constant(Double, 0);
break;
}
} else
result = emitLoadOp(op, emitCheckAndPreparePointer(pointer, offset, sizeOfLoadOp(op)), offset);
return { };
}
inline uint32_t sizeOfStoreOp(StoreOpType op)
{
switch (op) {
case StoreOpType::I32Store8:
case StoreOpType::I64Store8:
return 1;
case StoreOpType::I32Store16:
case StoreOpType::I64Store16:
return 2;
case StoreOpType::I32Store:
case StoreOpType::I64Store32:
case StoreOpType::F32Store:
return 4;
case StoreOpType::I64Store:
case StoreOpType::F64Store:
return 8;
}
RELEASE_ASSERT_NOT_REACHED();
}
inline void B3IRGenerator::emitStoreOp(StoreOpType op, ExpressionType pointer, ExpressionType value, uint32_t uoffset)
{
int32_t offset = fixupPointerPlusOffset(pointer, uoffset);
switch (op) {
case StoreOpType::I64Store8:
value = m_currentBlock->appendNew<Value>(m_proc, Trunc, origin(), value);
FALLTHROUGH;
case StoreOpType::I32Store8:
m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Store8), origin(), value, pointer, offset);
return;
case StoreOpType::I64Store16:
value = m_currentBlock->appendNew<Value>(m_proc, Trunc, origin(), value);
FALLTHROUGH;
case StoreOpType::I32Store16:
m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Store16), origin(), value, pointer, offset);
return;
case StoreOpType::I64Store32:
value = m_currentBlock->appendNew<Value>(m_proc, Trunc, origin(), value);
FALLTHROUGH;
case StoreOpType::I64Store:
case StoreOpType::I32Store:
case StoreOpType::F32Store:
case StoreOpType::F64Store:
m_currentBlock->appendNew<MemoryValue>(m_proc, memoryKind(Store), origin(), value, pointer, offset);
return;
}
RELEASE_ASSERT_NOT_REACHED();
}
auto B3IRGenerator::store(StoreOpType op, ExpressionType pointer, ExpressionType value, uint32_t offset) -> PartialResult
{
ASSERT(pointer->type() == Int32);
if (UNLIKELY(sumOverflows<uint32_t>(offset, sizeOfStoreOp(op)))) {
// FIXME: Even though this is provably out of bounds, it's not a validation error, so we have to handle it
// as a runtime exception. However, this may change: https://bugs.webkit.org/show_bug.cgi?id=166435
B3::PatchpointValue* throwException = m_currentBlock->appendNew<B3::PatchpointValue>(m_proc, B3::Void, origin());
throwException->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsMemoryAccess);
});
} else
emitStoreOp(op, emitCheckAndPreparePointer(pointer, offset, sizeOfStoreOp(op)), value, offset);
return { };
}
auto B3IRGenerator::addSelect(ExpressionType condition, ExpressionType nonZero, ExpressionType zero, ExpressionType& result) -> PartialResult
{
result = m_currentBlock->appendNew<Value>(m_proc, B3::Select, origin(), condition, nonZero, zero);
return { };
}
B3IRGenerator::ExpressionType B3IRGenerator::addConstant(Type type, uint64_t value)
{
return constant(toB3Type(type), value);
}
void B3IRGenerator::emitEntryTierUpCheck()
{
if (!m_tierUp)
return;
ASSERT(m_tierUp);
Value* countDownLocation = constant(pointerType(), reinterpret_cast<uint64_t>(&m_tierUp->m_counter), Origin());
PatchpointValue* patch = m_currentBlock->appendNew<PatchpointValue>(m_proc, B3::Void, Origin());
Effects effects = Effects::none();
// FIXME: we should have a more precise heap range for the tier up count.
effects.reads = B3::HeapRange::top();
effects.writes = B3::HeapRange::top();
patch->effects = effects;
patch->clobber(RegisterSet::macroScratchRegisters());
patch->append(countDownLocation, ValueRep::SomeRegister);
patch->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
CCallHelpers::Jump tierUp = jit.branchAdd32(CCallHelpers::PositiveOrZero, CCallHelpers::TrustedImm32(TierUpCount::functionEntryIncrement()), CCallHelpers::Address(params[0].gpr()));
CCallHelpers::Label tierUpResume = jit.label();
params.addLatePath([=] (CCallHelpers& jit) {
tierUp.link(&jit);
const unsigned extraPaddingBytes = 0;
RegisterSet registersToSpill = { };
registersToSpill.add(GPRInfo::argumentGPR1);
unsigned numberOfStackBytesUsedForRegisterPreservation = ScratchRegisterAllocator::preserveRegistersToStackForCall(jit, registersToSpill, extraPaddingBytes);
jit.move(MacroAssembler::TrustedImm32(m_functionIndex), GPRInfo::argumentGPR1);
MacroAssembler::Call call = jit.nearCall();
ScratchRegisterAllocator::restoreRegistersFromStackForCall(jit, registersToSpill, RegisterSet(), numberOfStackBytesUsedForRegisterPreservation, extraPaddingBytes);
jit.jump(tierUpResume);
jit.addLinkTask([=] (LinkBuffer& linkBuffer) {
MacroAssembler::repatchNearCall(linkBuffer.locationOfNearCall<NoPtrTag>(call), CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(triggerOMGEntryTierUpThunkGenerator).code()));
});
});
});
}
void B3IRGenerator::emitLoopTierUpCheck(uint32_t loopIndex)
{
uint32_t outerLoopIndex = this->outerLoopIndex();
m_outerLoops.append(loopIndex);
if (!m_tierUp)
return;
Origin origin = this->origin();
ASSERT(m_tierUp->osrEntryTriggers().size() == loopIndex);
m_tierUp->osrEntryTriggers().append(TierUpCount::TriggerReason::DontTrigger);
m_tierUp->outerLoops().append(outerLoopIndex);
Value* countDownLocation = constant(pointerType(), reinterpret_cast<uint64_t>(&m_tierUp->m_counter), origin);
Vector<ExpressionType> stackmap;
for (auto& local : m_locals) {
ExpressionType result = m_currentBlock->appendNew<VariableValue>(m_proc, B3::Get, origin, local);
stackmap.append(result);
}
for (unsigned controlIndex = 0; controlIndex < m_parser->controlStack().size(); ++controlIndex) {
auto& expressionStack = m_parser->controlStack()[controlIndex].enclosedExpressionStack;
for (Value* value : expressionStack)
stackmap.append(value);
}
PatchpointValue* patch = m_currentBlock->appendNew<PatchpointValue>(m_proc, B3::Void, origin);
Effects effects = Effects::none();
// FIXME: we should have a more precise heap range for the tier up count.
effects.reads = B3::HeapRange::top();
effects.writes = B3::HeapRange::top();
effects.exitsSideways = true;
patch->effects = effects;
patch->clobber(RegisterSet::macroScratchRegisters());
RegisterSet clobberLate;
clobberLate.add(GPRInfo::argumentGPR0);
patch->clobberLate(clobberLate);
patch->append(countDownLocation, ValueRep::SomeRegister);
patch->appendVectorWithRep(stackmap, ValueRep::ColdAny);
TierUpCount::TriggerReason* forceEntryTrigger = &(m_tierUp->osrEntryTriggers().last());
static_assert(!static_cast<uint8_t>(TierUpCount::TriggerReason::DontTrigger), "the JIT code assumes non-zero means 'enter'");
static_assert(sizeof(TierUpCount::TriggerReason) == 1, "branchTest8 assumes this size");
patch->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
CCallHelpers::Jump forceOSREntry = jit.branchTest8(CCallHelpers::NonZero, CCallHelpers::AbsoluteAddress(forceEntryTrigger));
CCallHelpers::Jump tierUp = jit.branchAdd32(CCallHelpers::PositiveOrZero, CCallHelpers::TrustedImm32(TierUpCount::loopIncrement()), CCallHelpers::Address(params[0].gpr()));
MacroAssembler::Label tierUpResume = jit.label();
OSREntryData& osrEntryData = m_tierUp->addOSREntryData(m_functionIndex, loopIndex);
// First argument is the countdown location.
for (unsigned i = 1; i < params.value()->numChildren(); ++i)
osrEntryData.values().constructAndAppend(params[i], params.value()->child(i)->type());
OSREntryData* osrEntryDataPtr = &osrEntryData;
params.addLatePath([=] (CCallHelpers& jit) {
AllowMacroScratchRegisterUsage allowScratch(jit);
forceOSREntry.link(&jit);
tierUp.link(&jit);
jit.probe(triggerOSREntryNow, osrEntryDataPtr);
jit.branchTestPtr(CCallHelpers::Zero, GPRInfo::argumentGPR0).linkTo(tierUpResume, &jit);
jit.farJump(GPRInfo::argumentGPR1, WasmEntryPtrTag);
});
});
}
auto B3IRGenerator::addLoop(BlockSignature signature, Stack& enclosingStack, ControlType& block, Stack& newStack, uint32_t loopIndex) -> PartialResult
{
BasicBlock* body = m_proc.addBlock();
BasicBlock* continuation = m_proc.addBlock();
block = ControlData(m_proc, origin(), signature, BlockType::Loop, continuation, body);
ExpressionList args;
{
unsigned offset = enclosingStack.size() - signature->argumentCount();
for (unsigned i = 0; i < signature->argumentCount(); ++i) {
Value* value = enclosingStack.at(offset + i);
auto* upsilon = m_currentBlock->appendNew<UpsilonValue>(m_proc, origin(), value);
Value* phi = block.phis[i];
body->append(phi);
upsilon->setPhi(phi);
newStack.append(phi);
}
enclosingStack.shrink(offset);
}
m_currentBlock->appendNewControlValue(m_proc, Jump, origin(), body);
if (loopIndex == m_loopIndexForOSREntry) {
dataLogLnIf(WasmB3IRGeneratorInternal::verbose, "Setting up for OSR entry");
m_currentBlock = m_rootBlock;
Value* pointer = m_rootBlock->appendNew<ArgumentRegValue>(m_proc, Origin(), GPRInfo::argumentGPR0);
unsigned indexInBuffer = 0;
auto loadFromScratchBuffer = [&] (B3::Type type) {
size_t offset = sizeof(uint64_t) * indexInBuffer++;
RELEASE_ASSERT(type.isNumeric());
return m_currentBlock->appendNew<MemoryValue>(m_proc, Load, type, origin(), pointer, offset);
};
for (auto& local : m_locals)
m_currentBlock->appendNew<VariableValue>(m_proc, Set, Origin(), local, loadFromScratchBuffer(local->type()));
for (unsigned controlIndex = 0; controlIndex < m_parser->controlStack().size(); ++controlIndex) {
const auto& data = m_parser->controlStack()[controlIndex].controlData;
auto& expressionStack = m_parser->controlStack()[controlIndex].enclosedExpressionStack;
// For each stack entry enclosed by this loop we need to replace the value with a phi so we can fill it on OSR entry.
BasicBlock* sourceBlock = nullptr;
unsigned blockIndex = 0;
B3::InsertionSet insertionSet(m_proc);
for (unsigned i = 0; i < expressionStack.size(); i++) {
auto* value = expressionStack[i];
if (value->isConstant())
continue;
if (value->owner != sourceBlock) {
insertionSet.execute(sourceBlock);
ASSERT(insertionSet.isEmpty());
dataLogLnIf(WasmB3IRGeneratorInternal::verbose && sourceBlock, "Executed insertion set into: ", *sourceBlock);
blockIndex = 0;
sourceBlock = value->owner;
}
while (sourceBlock->at(blockIndex++) != value)
ASSERT(blockIndex < sourceBlock->size());
ASSERT(sourceBlock->at(blockIndex - 1) == value);
auto* phi = data.continuation->appendNew<Value>(m_proc, Phi, value->type(), value->origin());
expressionStack[i] = phi;
m_currentBlock->appendNew<UpsilonValue>(m_proc, value->origin(), loadFromScratchBuffer(value->type()), phi);
auto* sourceUpsilon = m_proc.add<UpsilonValue>(value->origin(), value, phi);
insertionSet.insertValue(blockIndex, sourceUpsilon);
}
insertionSet.execute(sourceBlock);
}
m_osrEntryScratchBufferSize = indexInBuffer;
m_currentBlock->appendNewControlValue(m_proc, Jump, origin(), body);
body->addPredecessor(m_currentBlock);
}
m_currentBlock = body;
emitLoopTierUpCheck(loopIndex);
return { };
}
B3IRGenerator::ControlData B3IRGenerator::addTopLevel(BlockSignature signature)
{
return ControlData(m_proc, Origin(), signature, BlockType::TopLevel, m_proc.addBlock());
}
auto B3IRGenerator::addBlock(BlockSignature signature, Stack& enclosingStack, ControlType& newBlock, Stack& newStack) -> PartialResult
{
BasicBlock* continuation = m_proc.addBlock();
newStack = splitStack(signature, enclosingStack);
newBlock = ControlData(m_proc, origin(), signature, BlockType::Block, continuation);
return { };
}
auto B3IRGenerator::addIf(ExpressionType condition, BlockSignature signature, Stack& enclosingStack, ControlType& result, Stack& newStack) -> PartialResult
{
// FIXME: This needs to do some kind of stack passing.
BasicBlock* taken = m_proc.addBlock();
BasicBlock* notTaken = m_proc.addBlock();
BasicBlock* continuation = m_proc.addBlock();
m_currentBlock->appendNew<Value>(m_proc, B3::Branch, origin(), condition);
m_currentBlock->setSuccessors(FrequentedBlock(taken), FrequentedBlock(notTaken));
taken->addPredecessor(m_currentBlock);
notTaken->addPredecessor(m_currentBlock);
m_currentBlock = taken;
newStack = splitStack(signature, enclosingStack);
result = ControlData(m_proc, origin(), signature, BlockType::If, continuation, notTaken);
return { };
}
auto B3IRGenerator::addElse(ControlData& data, const Stack& currentStack) -> PartialResult
{
unifyValuesWithBlock(currentStack, data.phis);
m_currentBlock->appendNewControlValue(m_proc, Jump, origin(), data.continuation);
return addElseToUnreachable(data);
}
auto B3IRGenerator::addElseToUnreachable(ControlData& data) -> PartialResult
{
ASSERT(data.blockType() == BlockType::If);
m_currentBlock = data.special;
data.convertIfToBlock();
return { };
}
auto B3IRGenerator::addReturn(const ControlData&, const ExpressionList& returnValues) -> PartialResult
{
CallInformation wasmCallInfo = wasmCallingConvention().callInformationFor(m_parser->signature(), CallRole::Callee);
PatchpointValue* patch = m_proc.add<PatchpointValue>(B3::Void, origin());
patch->setGenerator([] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
auto calleeSaves = params.code().calleeSaveRegisterAtOffsetList();
for (RegisterAtOffset calleeSave : calleeSaves)
jit.load64ToReg(CCallHelpers::Address(GPRInfo::callFrameRegister, calleeSave.offset()), calleeSave.reg());
jit.emitFunctionEpilogue();
jit.ret();
});
patch->effects.terminal = true;
RELEASE_ASSERT(returnValues.size() >= wasmCallInfo.results.size());
unsigned offset = returnValues.size() - wasmCallInfo.results.size();
for (unsigned i = 0; i < wasmCallInfo.results.size(); ++i) {
B3::ValueRep rep = wasmCallInfo.results[i];
if (rep.isStack()) {
B3::Value* address = m_currentBlock->appendNew<B3::Value>(m_proc, B3::Add, Origin(), framePointer(), constant(pointerType(), rep.offsetFromFP()));
m_currentBlock->appendNew<B3::MemoryValue>(m_proc, B3::Store, Origin(), returnValues[offset + i], address);
} else {
ASSERT(rep.isReg());
patch->append(returnValues[offset + i], rep);
}
}
m_currentBlock->append(patch);
return { };
}
auto B3IRGenerator::addBranch(ControlData& data, ExpressionType condition, const Stack& returnValues) -> PartialResult
{
unifyValuesWithBlock(returnValues, data.phis);
BasicBlock* target = data.targetBlockForBranch();
if (condition) {
BasicBlock* continuation = m_proc.addBlock();
m_currentBlock->appendNew<Value>(m_proc, B3::Branch, origin(), condition);
m_currentBlock->setSuccessors(FrequentedBlock(target), FrequentedBlock(continuation));
target->addPredecessor(m_currentBlock);
continuation->addPredecessor(m_currentBlock);
m_currentBlock = continuation;
} else {
m_currentBlock->appendNewControlValue(m_proc, Jump, origin(), FrequentedBlock(target));
target->addPredecessor(m_currentBlock);
}
return { };
}
auto B3IRGenerator::addSwitch(ExpressionType condition, const Vector<ControlData*>& targets, ControlData& defaultTarget, const Stack& expressionStack) -> PartialResult
{
for (size_t i = 0; i < targets.size(); ++i)
unifyValuesWithBlock(expressionStack, targets[i]->phis);
unifyValuesWithBlock(expressionStack, defaultTarget.phis);
SwitchValue* switchValue = m_currentBlock->appendNew<SwitchValue>(m_proc, origin(), condition);
switchValue->setFallThrough(FrequentedBlock(defaultTarget.targetBlockForBranch()));
for (size_t i = 0; i < targets.size(); ++i)
switchValue->appendCase(SwitchCase(i, FrequentedBlock(targets[i]->targetBlockForBranch())));
return { };
}
auto B3IRGenerator::endBlock(ControlEntry& entry, Stack& expressionStack) -> PartialResult
{
ControlData& data = entry.controlData;
ASSERT(expressionStack.size() == data.signature->returnCount());
if (data.blockType() != BlockType::Loop)
unifyValuesWithBlock(expressionStack, data.phis);
m_currentBlock->appendNewControlValue(m_proc, Jump, origin(), data.continuation);
data.continuation->addPredecessor(m_currentBlock);
return addEndToUnreachable(entry, expressionStack);
}
auto B3IRGenerator::addEndToUnreachable(ControlEntry& entry, const Stack& expressionStack) -> PartialResult
{
ControlData& data = entry.controlData;
m_currentBlock = data.continuation;
if (data.blockType() == BlockType::If) {
data.special->appendNewControlValue(m_proc, Jump, origin(), m_currentBlock);
m_currentBlock->addPredecessor(data.special);
}
if (data.blockType() != BlockType::Loop) {
for (Value* result : data.phis) {
m_currentBlock->append(result);
entry.enclosedExpressionStack.append(result);
}
} else {
m_outerLoops.removeLast();
for (unsigned i = 0; i < data.signature->returnCount(); ++i) {
if (i < expressionStack.size())
entry.enclosedExpressionStack.append(expressionStack[i]);
else
entry.enclosedExpressionStack.append(constant(toB3Type(data.signature->returnType(i)), 0xbbadbeef));
}
}
// TopLevel does not have any code after this so we need to make sure we emit a return here.
if (data.blockType() == BlockType::TopLevel)
return addReturn(entry.controlData, entry.enclosedExpressionStack);
return { };
}
B3::Value* B3IRGenerator::createCallPatchpoint(BasicBlock* block, Origin origin, const Signature& signature, Vector<ExpressionType>& args, const ScopedLambda<void(PatchpointValue*)>& patchpointFunctor)
{
Vector<B3::ConstrainedValue> constrainedArguments;
CallInformation wasmCallInfo = wasmCallingConvention().callInformationFor(signature);
for (unsigned i = 0; i < args.size(); ++i)
constrainedArguments.append(B3::ConstrainedValue(args[i], wasmCallInfo.params[i]));
m_proc.requestCallArgAreaSizeInBytes(WTF::roundUpToMultipleOf(stackAlignmentBytes(), wasmCallInfo.headerAndArgumentStackSizeInBytes));
B3::Type returnType = toB3ResultType(&signature);
B3::PatchpointValue* patchpoint = block->appendNew<B3::PatchpointValue>(m_proc, returnType, origin);
patchpoint->clobberEarly(RegisterSet::macroScratchRegisters());
patchpoint->clobberLate(RegisterSet::volatileRegistersForJSCall());
patchpointFunctor(patchpoint);
patchpoint->appendVector(constrainedArguments);
if (returnType != B3::Void)
patchpoint->resultConstraints = WTFMove(wasmCallInfo.results);
return patchpoint;
}
auto B3IRGenerator::addCall(uint32_t functionIndex, const Signature& signature, Vector<ExpressionType>& args, ResultList& results) -> PartialResult
{
ASSERT(signature.argumentCount() == args.size());
m_makesCalls = true;
B3::Type returnType = toB3ResultType(&signature);
auto fillResults = [&] (Value* callResult) {
ASSERT(returnType == callResult->type());
switch (returnType.kind()) {
case B3::Void: {
break;
}
case B3::Tuple: {
const Vector<B3::Type>& tuple = m_proc.tupleForType(returnType);
ASSERT(signature.returnCount() == tuple.size());
for (unsigned i = 0; i < signature.returnCount(); ++i)
results.append(m_currentBlock->appendNew<ExtractValue>(m_proc, origin(), tuple[i], callResult, i));
break;
}
default: {
results.append(callResult);
break;
}
}
};
Vector<UnlinkedWasmToWasmCall>* unlinkedWasmToWasmCalls = &m_unlinkedWasmToWasmCalls;
if (m_info.isImportedFunctionFromFunctionIndexSpace(functionIndex)) {
m_maxNumJSCallArguments = std::max(m_maxNumJSCallArguments, static_cast<uint32_t>(args.size()));
// FIXME: imports can be linked here, instead of generating a patchpoint, because all import stubs are generated before B3 compilation starts. https://bugs.webkit.org/show_bug.cgi?id=166462
Value* targetInstance = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfTargetInstance(functionIndex)));
// The target instance is 0 unless the call is wasm->wasm.
Value* isWasmCall = m_currentBlock->appendNew<Value>(m_proc, NotEqual, origin(), targetInstance, m_currentBlock->appendNew<Const64Value>(m_proc, origin(), 0));
BasicBlock* isWasmBlock = m_proc.addBlock();
BasicBlock* isEmbedderBlock = m_proc.addBlock();
BasicBlock* continuation = m_proc.addBlock();
m_currentBlock->appendNewControlValue(m_proc, B3::Branch, origin(), isWasmCall, FrequentedBlock(isWasmBlock), FrequentedBlock(isEmbedderBlock));
Value* wasmCallResult = createCallPatchpoint(isWasmBlock, origin(), signature, args,
scopedLambdaRef<void(PatchpointValue*)>([=] (PatchpointValue* patchpoint) -> void {
patchpoint->effects.writesPinned = true;
patchpoint->effects.readsPinned = true;
// We need to clobber all potential pinned registers since we might be leaving the instance.
// We pessimistically assume we could be calling to something that is bounds checking.
// FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181
patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking));
patchpoint->setGenerator([unlinkedWasmToWasmCalls, functionIndex] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
AllowMacroScratchRegisterUsage allowScratch(jit);
CCallHelpers::Call call = jit.threadSafePatchableNearCall();
jit.addLinkTask([unlinkedWasmToWasmCalls, call, functionIndex] (LinkBuffer& linkBuffer) {
unlinkedWasmToWasmCalls->append({ linkBuffer.locationOfNearCall<WasmEntryPtrTag>(call), functionIndex });
});
});
}));
UpsilonValue* wasmCallResultUpsilon = returnType == B3::Void ? nullptr : isWasmBlock->appendNew<UpsilonValue>(m_proc, origin(), wasmCallResult);
isWasmBlock->appendNewControlValue(m_proc, Jump, origin(), continuation);
// FIXME: Let's remove this indirection by creating a PIC friendly IC
// for calls out to the embedder. This shouldn't be that hard to do. We could probably
// implement the IC to be over Context*.
// https://bugs.webkit.org/show_bug.cgi?id=170375
Value* jumpDestination = isEmbedderBlock->appendNew<MemoryValue>(m_proc,
Load, pointerType(), origin(), instanceValue(), safeCast<int32_t>(Instance::offsetOfWasmToEmbedderStub(functionIndex)));
Value* embedderCallResult = createCallPatchpoint(isEmbedderBlock, origin(), signature, args,
scopedLambdaRef<void(PatchpointValue*)>([=] (PatchpointValue* patchpoint) -> void {
patchpoint->effects.writesPinned = true;
patchpoint->effects.readsPinned = true;
patchpoint->append(jumpDestination, ValueRep::SomeRegister);
// We need to clobber all potential pinned registers since we might be leaving the instance.
// We pessimistically assume we could be calling to something that is bounds checking.
// FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181
patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking));
patchpoint->setGenerator([returnType] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
jit.call(params[params.proc().resultCount(returnType)].gpr(), WasmEntryPtrTag);
});
}));
UpsilonValue* embedderCallResultUpsilon = returnType == B3::Void ? nullptr : isEmbedderBlock->appendNew<UpsilonValue>(m_proc, origin(), embedderCallResult);
isEmbedderBlock->appendNewControlValue(m_proc, Jump, origin(), continuation);
m_currentBlock = continuation;
if (returnType != B3::Void) {
Value* phi = continuation->appendNew<Value>(m_proc, Phi, returnType, origin());
wasmCallResultUpsilon->setPhi(phi);
embedderCallResultUpsilon->setPhi(phi);
fillResults(phi);
}
// The call could have been to another WebAssembly instance, and / or could have modified our Memory.
restoreWebAssemblyGlobalState(RestoreCachedStackLimit::Yes, m_info.memory, instanceValue(), m_proc, continuation);
} else {
Value* patch = createCallPatchpoint(m_currentBlock, origin(), signature, args,
scopedLambdaRef<void(PatchpointValue*)>([=] (PatchpointValue* patchpoint) -> void {
patchpoint->effects.writesPinned = true;
patchpoint->effects.readsPinned = true;
patchpoint->setGenerator([unlinkedWasmToWasmCalls, functionIndex] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
AllowMacroScratchRegisterUsage allowScratch(jit);
CCallHelpers::Call call = jit.threadSafePatchableNearCall();
jit.addLinkTask([unlinkedWasmToWasmCalls, call, functionIndex] (LinkBuffer& linkBuffer) {
unlinkedWasmToWasmCalls->append({ linkBuffer.locationOfNearCall<WasmEntryPtrTag>(call), functionIndex });
});
});
}));
fillResults(patch);
}
return { };
}
auto B3IRGenerator::addCallIndirect(unsigned tableIndex, const Signature& signature, Vector<ExpressionType>& args, ResultList& results) -> PartialResult
{
ExpressionType calleeIndex = args.takeLast();
ASSERT(signature.argumentCount() == args.size());
m_makesCalls = true;
// Note: call indirect can call either WebAssemblyFunction or WebAssemblyWrapperFunction. Because
// WebAssemblyWrapperFunction is like calling into the embedder, we conservatively assume all call indirects
// can be to the embedder for our stack check calculation.
m_maxNumJSCallArguments = std::max(m_maxNumJSCallArguments, static_cast<uint32_t>(args.size()));
ExpressionType callableFunctionBuffer;
ExpressionType instancesBuffer;
ExpressionType callableFunctionBufferLength;
ExpressionType mask;
{
ExpressionType table = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(),
instanceValue(), safeCast<int32_t>(Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex)));
callableFunctionBuffer = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(),
table, safeCast<int32_t>(FuncRefTable::offsetOfFunctions()));
instancesBuffer = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(),
table, safeCast<int32_t>(FuncRefTable::offsetOfInstances()));
callableFunctionBufferLength = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, Int32, origin(),
table, safeCast<int32_t>(Table::offsetOfLength()));
mask = m_currentBlock->appendNew<Value>(m_proc, ZExt32, origin(),
m_currentBlock->appendNew<MemoryValue>(m_proc, Load, Int32, origin(),
table, safeCast<int32_t>(Table::offsetOfMask())));
}
// Check the index we are looking for is valid.
{
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, AboveEqual, origin(), calleeIndex, callableFunctionBufferLength));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsCallIndirect);
});
}
calleeIndex = m_currentBlock->appendNew<Value>(m_proc, ZExt32, origin(), calleeIndex);
if (Options::enableSpectreMitigations())
calleeIndex = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(), mask, calleeIndex);
ExpressionType callableFunction;
{
// Compute the offset in the table index space we are looking for.
ExpressionType offset = m_currentBlock->appendNew<Value>(m_proc, Mul, origin(),
calleeIndex, constant(pointerType(), sizeof(WasmToWasmImportableFunction)));
callableFunction = m_currentBlock->appendNew<Value>(m_proc, Add, origin(), callableFunctionBuffer, offset);
// Check that the WasmToWasmImportableFunction is initialized. We trap if it isn't. An "invalid" SignatureIndex indicates it's not initialized.
// FIXME: when we have trap handlers, we can just let the call fail because Signature::invalidIndex is 0. https://bugs.webkit.org/show_bug.cgi?id=177210
static_assert(sizeof(WasmToWasmImportableFunction::signatureIndex) == sizeof(uint64_t), "Load codegen assumes i64");
ExpressionType calleeSignatureIndex = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, Int64, origin(), callableFunction, safeCast<int32_t>(WasmToWasmImportableFunction::offsetOfSignatureIndex()));
{
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(),
calleeSignatureIndex,
m_currentBlock->appendNew<Const64Value>(m_proc, origin(), Signature::invalidIndex)));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::NullTableEntry);
});
}
// Check the signature matches the value we expect.
{
ExpressionType expectedSignatureIndex = m_currentBlock->appendNew<Const64Value>(m_proc, origin(), SignatureInformation::get(signature));
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, NotEqual, origin(), calleeSignatureIndex, expectedSignatureIndex));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::BadSignature);
});
}
}
// Do a context switch if needed.
{
Value* offset = m_currentBlock->appendNew<Value>(m_proc, Mul, origin(),
calleeIndex, constant(pointerType(), sizeof(Instance*)));
Value* newContextInstance = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(),
m_currentBlock->appendNew<Value>(m_proc, Add, origin(), instancesBuffer, offset));
BasicBlock* continuation = m_proc.addBlock();
BasicBlock* doContextSwitch = m_proc.addBlock();
Value* isSameContextInstance = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(),
newContextInstance, instanceValue());
m_currentBlock->appendNewControlValue(m_proc, B3::Branch, origin(),
isSameContextInstance, FrequentedBlock(continuation), FrequentedBlock(doContextSwitch));
PatchpointValue* patchpoint = doContextSwitch->appendNew<PatchpointValue>(m_proc, B3::Void, origin());
patchpoint->effects.writesPinned = true;
// We pessimistically assume we're calling something with BoundsChecking memory.
// FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181
patchpoint->clobber(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking));
patchpoint->clobber(RegisterSet::macroScratchRegisters());
patchpoint->append(newContextInstance, ValueRep::SomeRegister);
patchpoint->append(instanceValue(), ValueRep::SomeRegister);
patchpoint->numGPScratchRegisters = Gigacage::isEnabled(Gigacage::Primitive) ? 1 : 0;
patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
GPRReg newContextInstance = params[0].gpr();
GPRReg oldContextInstance = params[1].gpr();
const PinnedRegisterInfo& pinnedRegs = PinnedRegisterInfo::get();
GPRReg baseMemory = pinnedRegs.baseMemoryPointer;
ASSERT(newContextInstance != baseMemory);
jit.loadPtr(CCallHelpers::Address(oldContextInstance, Instance::offsetOfCachedStackLimit()), baseMemory);
jit.storePtr(baseMemory, CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedStackLimit()));
jit.storeWasmContextInstance(newContextInstance);
ASSERT(pinnedRegs.sizeRegister != baseMemory);
// FIXME: We should support more than one memory size register
// see: https://bugs.webkit.org/show_bug.cgi?id=162952
ASSERT(pinnedRegs.sizeRegister != newContextInstance);
GPRReg scratchOrSize = Gigacage::isEnabled(Gigacage::Primitive) ? params.gpScratch(0) : pinnedRegs.sizeRegister;
jit.loadPtr(CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedMemorySize()), pinnedRegs.sizeRegister); // Memory size.
jit.loadPtr(CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedMemory()), baseMemory); // Memory::void*.
jit.cageConditionally(Gigacage::Primitive, baseMemory, pinnedRegs.sizeRegister, scratchOrSize);
});
doContextSwitch->appendNewControlValue(m_proc, Jump, origin(), continuation);
m_currentBlock = continuation;
}
ExpressionType calleeCode = m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(),
m_currentBlock->appendNew<MemoryValue>(m_proc, Load, pointerType(), origin(), callableFunction,
safeCast<int32_t>(WasmToWasmImportableFunction::offsetOfEntrypointLoadLocation())));
B3::Type returnType = toB3ResultType(&signature);
ExpressionType callResult = createCallPatchpoint(m_currentBlock, origin(), signature, args,
scopedLambdaRef<void(PatchpointValue*)>([=] (PatchpointValue* patchpoint) -> void {
patchpoint->effects.writesPinned = true;
patchpoint->effects.readsPinned = true;
// We need to clobber all potential pinned registers since we might be leaving the instance.
// We pessimistically assume we're always calling something that is bounds checking so
// because the wasm->wasm thunk unconditionally overrides the size registers.
// FIXME: We should not have to do this, but the wasm->wasm stub assumes it can
// use all the pinned registers as scratch: https://bugs.webkit.org/show_bug.cgi?id=172181
patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking));
patchpoint->append(calleeCode, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
jit.call(params[params.proc().resultCount(returnType)].gpr(), WasmEntryPtrTag);
});
}));
switch (returnType.kind()) {
case B3::Void: {
break;
}
case B3::Tuple: {
const Vector<B3::Type>& tuple = m_proc.tupleForType(returnType);
for (unsigned i = 0; i < signature.returnCount(); ++i)
results.append(m_currentBlock->appendNew<ExtractValue>(m_proc, origin(), tuple[i], callResult, i));
break;
}
default: {
results.append(callResult);
break;
}
}
// The call could have been to another WebAssembly instance, and / or could have modified our Memory.
restoreWebAssemblyGlobalState(RestoreCachedStackLimit::Yes, m_info.memory, instanceValue(), m_proc, m_currentBlock);
return { };
}
void B3IRGenerator::unify(const ExpressionType phi, const ExpressionType source)
{
m_currentBlock->appendNew<UpsilonValue>(m_proc, origin(), source, phi);
}
void B3IRGenerator::unifyValuesWithBlock(const Stack& resultStack, const ResultList& result)
{
ASSERT(result.size() <= resultStack.size());
for (size_t i = 0; i < result.size(); ++i)
unify(result[result.size() - 1 - i], resultStack.at(resultStack.size() - 1 - i));
}
static void dumpExpressionStack(const CommaPrinter& comma, const B3IRGenerator::ExpressionList& expressionStack)
{
dataLog(comma, "ExpressionStack:");
for (const auto& expression : expressionStack)
dataLog(comma, *expression);
}
void B3IRGenerator::dump(const Vector<ControlEntry>& controlStack, const Stack* expressionStack)
{
dataLogLn("Constants:");
for (const auto& constant : m_constantPool)
dataLogLn(deepDump(m_proc, constant.value));
dataLogLn("Processing Graph:");
dataLog(m_proc);
dataLogLn("With current block:", *m_currentBlock);
dataLogLn("Control stack:");
ASSERT(controlStack.size());
for (size_t i = controlStack.size(); i--;) {
dataLog(" ", controlStack[i].controlData, ": ");
CommaPrinter comma(", ", "");
dumpExpressionStack(comma, *expressionStack);
expressionStack = &controlStack[i].enclosedExpressionStack;
dataLogLn();
}
dataLogLn();
}
auto B3IRGenerator::origin() -> Origin
{
OpcodeOrigin origin(m_parser->currentOpcode(), m_parser->currentOpcodeStartingOffset());
ASSERT(isValidOpType(static_cast<uint8_t>(origin.opcode())));
return bitwise_cast<Origin>(origin);
}
Expected<std::unique_ptr<InternalFunction>, String> parseAndCompile(CompilationContext& compilationContext, const FunctionData& function, const Signature& signature, Vector<UnlinkedWasmToWasmCall>& unlinkedWasmToWasmCalls, unsigned& osrEntryScratchBufferSize, const ModuleInformation& info, MemoryMode mode, CompilationMode compilationMode, uint32_t functionIndex, uint32_t loopIndexForOSREntry, TierUpCount* tierUp, ThrowWasmException throwWasmException)
{
auto result = makeUnique<InternalFunction>();
compilationContext.embedderEntrypointJIT = makeUnique<CCallHelpers>();
compilationContext.wasmEntrypointJIT = makeUnique<CCallHelpers>();
Procedure procedure;
procedure.setOriginPrinter([] (PrintStream& out, Origin origin) {
if (origin.data())
out.print("Wasm: ", bitwise_cast<OpcodeOrigin>(origin));
});
// This means we cannot use either StackmapGenerationParams::usedRegisters() or
// StackmapGenerationParams::unavailableRegisters(). In exchange for this concession, we
// don't strictly need to run Air::reportUsedRegisters(), which saves a bit of CPU time at
// optLevel=1.
procedure.setNeedsUsedRegisters(false);
procedure.setOptLevel(compilationMode == CompilationMode::BBQMode
? Options::webAssemblyBBQB3OptimizationLevel()
: Options::webAssemblyOMGOptimizationLevel());
B3IRGenerator irGenerator(info, procedure, result.get(), unlinkedWasmToWasmCalls, osrEntryScratchBufferSize, mode, compilationMode, functionIndex, loopIndexForOSREntry, tierUp, throwWasmException);
FunctionParser<B3IRGenerator> parser(irGenerator, function.data.data(), function.data.size(), signature, info);
WASM_FAIL_IF_HELPER_FAILS(parser.parse());
irGenerator.insertConstants();
procedure.resetReachability();
if (!ASSERT_DISABLED)
validate(procedure, "After parsing:\n");
dataLogIf(WasmB3IRGeneratorInternal::verbose, "Pre SSA: ", procedure);
fixSSA(procedure);
dataLogIf(WasmB3IRGeneratorInternal::verbose, "Post SSA: ", procedure);
{
B3::prepareForGeneration(procedure);
B3::generate(procedure, *compilationContext.wasmEntrypointJIT);
compilationContext.wasmEntrypointByproducts = procedure.releaseByproducts();
result->entrypoint.calleeSaveRegisters = procedure.calleeSaveRegisterAtOffsetList();
}
return result;
}
// Custom wasm ops. These are the ones too messy to do in wasm.json.
void B3IRGenerator::emitChecksForModOrDiv(B3::Opcode operation, ExpressionType left, ExpressionType right)
{
ASSERT(operation == Div || operation == Mod || operation == UDiv || operation == UMod);
const B3::Type type = left->type();
{
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), right, constant(type, 0)));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::DivisionByZero);
});
}
if (operation == Div) {
int64_t min = type == Int32 ? std::numeric_limits<int32_t>::min() : std::numeric_limits<int64_t>::min();
CheckValue* check = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(),
m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), left, constant(type, min)),
m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), right, constant(type, -1))));
check->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::IntegerOverflow);
});
}
}
template<>
auto B3IRGenerator::addOp<OpType::I32DivS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = Div;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, op, origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32RemS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = Mod;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, chill(op), origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32DivU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = UDiv;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, op, origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32RemU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = UMod;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, op, origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64DivS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = Div;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, op, origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64RemS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = Mod;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, chill(op), origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64DivU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = UDiv;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, op, origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64RemU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult
{
const B3::Opcode op = UMod;
emitChecksForModOrDiv(op, left, right);
result = m_currentBlock->appendNew<Value>(m_proc, op, origin(), left, right);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32Ctz>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int32, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.countTrailingZeros32(params[1].gpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64Ctz>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int64, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.countTrailingZeros64(params[1].gpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32Popcnt>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
#if CPU(X86_64)
if (MacroAssembler::supportsCountPopulation()) {
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int32, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.countPopulation32(params[1].gpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
#endif
uint32_t (*popcount)(int32_t) = [] (int32_t value) -> uint32_t { return __builtin_popcount(value); };
Value* funcAddress = m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(popcount, B3CCallPtrTag));
result = m_currentBlock->appendNew<CCallValue>(m_proc, Int32, origin(), Effects::none(), funcAddress, arg);
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64Popcnt>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
#if CPU(X86_64)
if (MacroAssembler::supportsCountPopulation()) {
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int64, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.countPopulation64(params[1].gpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
#endif
uint64_t (*popcount)(int64_t) = [] (int64_t value) -> uint64_t { return __builtin_popcountll(value); };
Value* funcAddress = m_currentBlock->appendNew<ConstPtrValue>(m_proc, origin(), tagCFunctionPtr<void*>(popcount, B3CCallPtrTag));
result = m_currentBlock->appendNew<CCallValue>(m_proc, Int64, origin(), Effects::none(), funcAddress, arg);
return { };
}
template<>
auto B3IRGenerator::addOp<F64ConvertUI64>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Double, origin());
if (isX86())
patchpoint->numGPScratchRegisters = 1;
patchpoint->clobber(RegisterSet::macroScratchRegisters());
patchpoint->append(ConstrainedValue(arg, ValueRep::SomeRegister));
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
#if CPU(X86_64)
jit.convertUInt64ToDouble(params[1].gpr(), params[0].fpr(), params.gpScratch(0));
#else
jit.convertUInt64ToDouble(params[1].gpr(), params[0].fpr());
#endif
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::F32ConvertUI64>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Float, origin());
if (isX86())
patchpoint->numGPScratchRegisters = 1;
patchpoint->clobber(RegisterSet::macroScratchRegisters());
patchpoint->append(ConstrainedValue(arg, ValueRep::SomeRegister));
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
#if CPU(X86_64)
jit.convertUInt64ToFloat(params[1].gpr(), params[0].fpr(), params.gpScratch(0));
#else
jit.convertUInt64ToFloat(params[1].gpr(), params[0].fpr());
#endif
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::F64Nearest>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Double, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.roundTowardNearestIntDouble(params[1].fpr(), params[0].fpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::F32Nearest>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Float, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.roundTowardNearestIntFloat(params[1].fpr(), params[0].fpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::F64Trunc>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Double, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.roundTowardZeroDouble(params[1].fpr(), params[0].fpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::F32Trunc>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Float, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.roundTowardZeroFloat(params[1].fpr(), params[0].fpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32TruncSF64>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Double, bitwise_cast<uint64_t>(-static_cast<double>(std::numeric_limits<int32_t>::min())));
Value* min = constant(Double, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int32_t>::min())));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterEqual, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int32, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.truncateDoubleToInt32(params[1].fpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32TruncSF32>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Float, bitwise_cast<uint32_t>(-static_cast<float>(std::numeric_limits<int32_t>::min())));
Value* min = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int32_t>::min())));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterEqual, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int32, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.truncateFloatToInt32(params[1].fpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32TruncUF64>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Double, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int32_t>::min()) * -2.0));
Value* min = constant(Double, bitwise_cast<uint64_t>(-1.0));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterThan, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int32, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.truncateDoubleToUint32(params[1].fpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I32TruncUF32>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int32_t>::min()) * static_cast<float>(-2.0)));
Value* min = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(-1.0)));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterThan, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int32, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.truncateFloatToUint32(params[1].fpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64TruncSF64>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Double, bitwise_cast<uint64_t>(-static_cast<double>(std::numeric_limits<int64_t>::min())));
Value* min = constant(Double, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int64_t>::min())));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterEqual, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int64, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.truncateDoubleToInt64(params[1].fpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64TruncUF64>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Double, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int64_t>::min()) * -2.0));
Value* min = constant(Double, bitwise_cast<uint64_t>(-1.0));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterThan, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
Value* signBitConstant;
if (isX86()) {
// Since x86 doesn't have an instruction to convert floating points to unsigned integers, we at least try to do the smart thing if
// the numbers are would be positive anyway as a signed integer. Since we cannot materialize constants into fprs we have b3 do it
// so we can pool them if needed.
signBitConstant = constant(Double, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<uint64_t>::max() - std::numeric_limits<int64_t>::max())));
}
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int64, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
if (isX86()) {
patchpoint->append(signBitConstant, ValueRep::SomeRegister);
patchpoint->numFPScratchRegisters = 1;
}
patchpoint->clobber(RegisterSet::macroScratchRegisters());
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
FPRReg scratch = InvalidFPRReg;
FPRReg constant = InvalidFPRReg;
if (isX86()) {
scratch = params.fpScratch(0);
constant = params[2].fpr();
}
jit.truncateDoubleToUint64(params[1].fpr(), params[0].gpr(), scratch, constant);
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64TruncSF32>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Float, bitwise_cast<uint32_t>(-static_cast<float>(std::numeric_limits<int64_t>::min())));
Value* min = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int64_t>::min())));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterEqual, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int64, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
jit.truncateFloatToInt64(params[1].fpr(), params[0].gpr());
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
template<>
auto B3IRGenerator::addOp<OpType::I64TruncUF32>(ExpressionType arg, ExpressionType& result) -> PartialResult
{
Value* max = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int64_t>::min()) * static_cast<float>(-2.0)));
Value* min = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(-1.0)));
Value* outOfBounds = m_currentBlock->appendNew<Value>(m_proc, BitAnd, origin(),
m_currentBlock->appendNew<Value>(m_proc, LessThan, origin(), arg, max),
m_currentBlock->appendNew<Value>(m_proc, GreaterThan, origin(), arg, min));
outOfBounds = m_currentBlock->appendNew<Value>(m_proc, Equal, origin(), outOfBounds, constant(Int32, 0));
CheckValue* trap = m_currentBlock->appendNew<CheckValue>(m_proc, Check, origin(), outOfBounds);
trap->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams&) {
this->emitExceptionCheck(jit, ExceptionType::OutOfBoundsTrunc);
});
Value* signBitConstant;
if (isX86()) {
// Since x86 doesn't have an instruction to convert floating points to unsigned integers, we at least try to do the smart thing if
// the numbers would be positive anyway as a signed integer. Since we cannot materialize constants into fprs we have b3 do it
// so we can pool them if needed.
signBitConstant = constant(Float, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<uint64_t>::max() - std::numeric_limits<int64_t>::max())));
}
PatchpointValue* patchpoint = m_currentBlock->appendNew<PatchpointValue>(m_proc, Int64, origin());
patchpoint->append(arg, ValueRep::SomeRegister);
if (isX86()) {
patchpoint->append(signBitConstant, ValueRep::SomeRegister);
patchpoint->numFPScratchRegisters = 1;
}
patchpoint->clobber(RegisterSet::macroScratchRegisters());
patchpoint->setGenerator([=] (CCallHelpers& jit, const StackmapGenerationParams& params) {
AllowMacroScratchRegisterUsage allowScratch(jit);
FPRReg scratch = InvalidFPRReg;
FPRReg constant = InvalidFPRReg;
if (isX86()) {
scratch = params.fpScratch(0);
constant = params[2].fpr();
}
jit.truncateFloatToUint64(params[1].fpr(), params[0].gpr(), scratch, constant);
});
patchpoint->effects = Effects::none();
result = patchpoint;
return { };
}
} } // namespace JSC::Wasm
#include "WasmB3IRGeneratorInlines.h"
#endif // ENABLE(WEBASSEMBLY)