darling-JavaScriptCore/assembler/LinkBuffer.cpp

527 lines
18 KiB
C++

/*
* Copyright (C) 2012-2019 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 "LinkBuffer.h"
#if ENABLE(ASSEMBLER)
#include "CodeBlock.h"
#include "Disassembler.h"
#include "JITCode.h"
#include "Options.h"
#include "WasmCompilationMode.h"
#if OS(LINUX)
#include "PerfLog.h"
#endif
namespace JSC {
bool shouldDumpDisassemblyFor(CodeBlock* codeBlock)
{
if (codeBlock && JITCode::isOptimizingJIT(codeBlock->jitType()) && Options::dumpDFGDisassembly())
return true;
return Options::dumpDisassembly();
}
bool shouldDumpDisassemblyFor(Wasm::CompilationMode mode)
{
if (Options::asyncDisassembly() || Options::dumpDisassembly() || Options::dumpWasmDisassembly())
return true;
switch (mode) {
case Wasm::CompilationMode::BBQMode:
return Options::dumpBBQDisassembly();
case Wasm::CompilationMode::OMGMode:
case Wasm::CompilationMode::OMGForOSREntryMode:
return Options::dumpOMGDisassembly();
default:
break;
}
return false;
}
LinkBuffer::CodeRef<LinkBufferPtrTag> LinkBuffer::finalizeCodeWithoutDisassemblyImpl()
{
performFinalization();
ASSERT(m_didAllocate);
if (m_executableMemory)
return CodeRef<LinkBufferPtrTag>(*m_executableMemory);
return CodeRef<LinkBufferPtrTag>::createSelfManagedCodeRef(m_code);
}
LinkBuffer::CodeRef<LinkBufferPtrTag> LinkBuffer::finalizeCodeWithDisassemblyImpl(bool dumpDisassembly, const char* format, ...)
{
CodeRef<LinkBufferPtrTag> result = finalizeCodeWithoutDisassemblyImpl();
#if OS(LINUX)
if (Options::logJITCodeForPerf()) {
StringPrintStream out;
va_list argList;
va_start(argList, format);
va_start(argList, format);
out.vprintf(format, argList);
va_end(argList);
PerfLog::log(out.toCString(), result.code().untaggedExecutableAddress<const uint8_t*>(), result.size());
}
#endif
if (!dumpDisassembly || m_alreadyDisassembled)
return result;
StringPrintStream out;
out.printf("Generated JIT code for ");
va_list argList;
va_start(argList, format);
out.vprintf(format, argList);
va_end(argList);
out.printf(":\n");
uint8_t* executableAddress = result.code().untaggedExecutableAddress<uint8_t*>();
out.printf(" Code at [%p, %p):\n", executableAddress, executableAddress + result.size());
CString header = out.toCString();
if (Options::asyncDisassembly()) {
CodeRef<DisassemblyPtrTag> codeRefForDisassembly = result.retagged<DisassemblyPtrTag>();
disassembleAsynchronously(header, WTFMove(codeRefForDisassembly), m_size, " ");
return result;
}
dataLog(header);
disassemble(result.retaggedCode<DisassemblyPtrTag>(), m_size, " ", WTF::dataFile());
return result;
}
#if ENABLE(BRANCH_COMPACTION)
class BranchCompactionLinkBuffer;
using ThreadSpecificBranchCompactionLinkBuffer = ThreadSpecific<BranchCompactionLinkBuffer, WTF::CanBeGCThread::True>;
static ThreadSpecificBranchCompactionLinkBuffer* threadSpecificBranchCompactionLinkBufferPtr;
static ThreadSpecificBranchCompactionLinkBuffer& threadSpecificBranchCompactionLinkBuffer()
{
static std::once_flag flag;
std::call_once(
flag,
[] () {
threadSpecificBranchCompactionLinkBufferPtr = new ThreadSpecificBranchCompactionLinkBuffer();
});
return *threadSpecificBranchCompactionLinkBufferPtr;
}
DECLARE_ALLOCATOR_WITH_HEAP_IDENTIFIER(BranchCompactionLinkBuffer);
class BranchCompactionLinkBuffer {
WTF_MAKE_NONCOPYABLE(BranchCompactionLinkBuffer);
public:
BranchCompactionLinkBuffer()
{
}
BranchCompactionLinkBuffer(size_t size, uint8_t* userBuffer = nullptr)
{
if (userBuffer) {
m_data = userBuffer;
m_size = size;
m_bufferProvided = true;
return;
}
auto& threadSpecific = threadSpecificBranchCompactionLinkBuffer();
if (threadSpecific->size() >= size)
takeBufferIfLarger(*threadSpecific);
else {
m_size = size;
m_data = static_cast<uint8_t*>(BranchCompactionLinkBufferMalloc::malloc(size));
}
}
~BranchCompactionLinkBuffer()
{
if (m_bufferProvided)
return;
auto& threadSpecific = threadSpecificBranchCompactionLinkBuffer();
threadSpecific->takeBufferIfLarger(*this);
if (m_data)
BranchCompactionLinkBufferMalloc::free(m_data);
}
uint8_t* data()
{
return m_data;
}
private:
void takeBufferIfLarger(BranchCompactionLinkBuffer& other)
{
if (size() >= other.size())
return;
if (m_data)
BranchCompactionLinkBufferMalloc::free(m_data);
m_data = other.m_data;
m_size = other.m_size;
other.m_data = nullptr;
other.m_size = 0;
}
size_t size()
{
return m_size;
}
uint8_t* m_data { nullptr };
size_t m_size { 0 };
bool m_bufferProvided { false };
};
static ALWAYS_INLINE void recordLinkOffsets(AssemblerData& assemblerData, int32_t regionStart, int32_t regionEnd, int32_t offset)
{
int32_t ptr = regionStart / sizeof(int32_t);
const int32_t end = regionEnd / sizeof(int32_t);
int32_t* offsets = reinterpret_cast_ptr<int32_t*>(assemblerData.buffer());
while (ptr < end)
offsets[ptr++] = offset;
}
// We use this to prevent compile errors on some platforms that are unhappy
// about the signature of the system's memcpy.
ALWAYS_INLINE void* memcpyWrapper(void* dst, const void* src, size_t bytes)
{
return memcpy(dst, src, bytes);
}
template <typename InstructionType>
void LinkBuffer::copyCompactAndLinkCode(MacroAssembler& macroAssembler, JITCompilationEffort effort)
{
allocate(macroAssembler, effort);
const size_t initialSize = macroAssembler.m_assembler.codeSize();
if (didFailToAllocate())
return;
Vector<LinkRecord, 0, UnsafeVectorOverflow>& jumpsToLink = macroAssembler.jumpsToLink();
m_assemblerStorage = macroAssembler.m_assembler.buffer().releaseAssemblerData();
uint8_t* inData = bitwise_cast<uint8_t*>(m_assemblerStorage.buffer());
#if CPU(ARM64E)
ARM64EHash verifyUncompactedHash { static_cast<uint32_t>(bitwise_cast<uint64_t>(&macroAssembler.m_assembler.buffer())) };
m_assemblerHashesStorage = macroAssembler.m_assembler.buffer().releaseAssemblerHashes();
uint32_t* inHashes = bitwise_cast<uint32_t*>(m_assemblerHashesStorage.buffer());
#endif
uint8_t* codeOutData = m_code.dataLocation<uint8_t*>();
BranchCompactionLinkBuffer outBuffer(m_size, useFastJITPermissions() ? codeOutData : 0);
uint8_t* outData = outBuffer.data();
#if CPU(ARM64)
RELEASE_ASSERT(roundUpToMultipleOf<sizeof(unsigned)>(outData) == outData);
RELEASE_ASSERT(roundUpToMultipleOf<sizeof(unsigned)>(codeOutData) == codeOutData);
#endif
int readPtr = 0;
int writePtr = 0;
unsigned jumpCount = jumpsToLink.size();
auto read = [&](const InstructionType* ptr) -> InstructionType {
InstructionType value = *ptr;
#if CPU(ARM64E)
uint32_t hash = verifyUncompactedHash.update(value);
unsigned index = (bitwise_cast<uint8_t*>(ptr) - inData) / 4;
RELEASE_ASSERT(inHashes[index] == hash);
#endif
return value;
};
if (useFastJITPermissions())
threadSelfRestrictRWXToRW();
if (m_shouldPerformBranchCompaction) {
for (unsigned i = 0; i < jumpCount; ++i) {
int offset = readPtr - writePtr;
ASSERT(!(offset & 1));
// Copy the instructions from the last jump to the current one.
size_t regionSize = jumpsToLink[i].from() - readPtr;
InstructionType* copySource = reinterpret_cast_ptr<InstructionType*>(inData + readPtr);
InstructionType* copyEnd = reinterpret_cast_ptr<InstructionType*>(inData + readPtr + regionSize);
InstructionType* copyDst = reinterpret_cast_ptr<InstructionType*>(outData + writePtr);
ASSERT(!(regionSize % 2));
ASSERT(!(readPtr % 2));
ASSERT(!(writePtr % 2));
while (copySource != copyEnd) {
InstructionType insn = read(copySource++);
*copyDst++ = insn;
}
recordLinkOffsets(m_assemblerStorage, readPtr, jumpsToLink[i].from(), offset);
readPtr += regionSize;
writePtr += regionSize;
// Calculate absolute address of the jump target, in the case of backwards
// branches we need to be precise, forward branches we are pessimistic
const uint8_t* target;
#if CPU(ARM64)
const intptr_t to = jumpsToLink[i].to(&macroAssembler.m_assembler);
#else
const intptr_t to = jumpsToLink[i].to();
#endif
if (to >= jumpsToLink[i].from())
target = codeOutData + to - offset; // Compensate for what we have collapsed so far
else
target = codeOutData + to - executableOffsetFor(to);
JumpLinkType jumpLinkType = MacroAssembler::computeJumpType(jumpsToLink[i], codeOutData + writePtr, target);
// Compact branch if we can...
if (MacroAssembler::canCompact(jumpsToLink[i].type())) {
// Step back in the write stream
int32_t delta = MacroAssembler::jumpSizeDelta(jumpsToLink[i].type(), jumpLinkType);
if (delta) {
writePtr -= delta;
recordLinkOffsets(m_assemblerStorage, jumpsToLink[i].from() - delta, readPtr, readPtr - writePtr);
}
}
#if CPU(ARM64)
jumpsToLink[i].setFrom(&macroAssembler.m_assembler, writePtr);
#else
jumpsToLink[i].setFrom(writePtr);
#endif
}
} else {
if (ASSERT_ENABLED) {
for (unsigned i = 0; i < jumpCount; ++i)
ASSERT(!MacroAssembler::canCompact(jumpsToLink[i].type()));
}
}
// Copy everything after the last jump
{
InstructionType* dst = bitwise_cast<InstructionType*>(outData + writePtr);
InstructionType* src = bitwise_cast<InstructionType*>(inData + readPtr);
size_t bytes = initialSize - readPtr;
RELEASE_ASSERT(bitwise_cast<uintptr_t>(dst) % sizeof(InstructionType) == 0);
RELEASE_ASSERT(bitwise_cast<uintptr_t>(src) % sizeof(InstructionType) == 0);
RELEASE_ASSERT(bytes % sizeof(InstructionType) == 0);
for (size_t i = 0; i < bytes; i += sizeof(InstructionType)) {
InstructionType insn = read(src++);
*dst++ = insn;
}
}
recordLinkOffsets(m_assemblerStorage, readPtr, initialSize, readPtr - writePtr);
for (unsigned i = 0; i < jumpCount; ++i) {
uint8_t* location = codeOutData + jumpsToLink[i].from();
#if CPU(ARM64)
const intptr_t to = jumpsToLink[i].to(&macroAssembler.m_assembler);
#else
const intptr_t to = jumpsToLink[i].to();
#endif
uint8_t* target = codeOutData + to - executableOffsetFor(to);
if (useFastJITPermissions())
MacroAssembler::link<memcpyWrapper>(jumpsToLink[i], outData + jumpsToLink[i].from(), location, target);
else
MacroAssembler::link<performJITMemcpy>(jumpsToLink[i], outData + jumpsToLink[i].from(), location, target);
}
size_t compactSize = writePtr + initialSize - readPtr;
if (!m_executableMemory) {
size_t nopSizeInBytes = initialSize - compactSize;
if (useFastJITPermissions())
Assembler::fillNops<memcpyWrapper>(outData + compactSize, nopSizeInBytes);
else
Assembler::fillNops<performJITMemcpy>(outData + compactSize, nopSizeInBytes);
}
if (useFastJITPermissions())
threadSelfRestrictRWXToRX();
if (m_executableMemory) {
m_size = compactSize;
m_executableMemory->shrink(m_size);
}
#if ENABLE(JIT)
if (useFastJITPermissions()) {
ASSERT(codeOutData == outData);
if (UNLIKELY(Options::dumpJITMemoryPath()))
dumpJITMemory(outData, outData, m_size);
} else {
ASSERT(codeOutData != outData);
performJITMemcpy(codeOutData, outData, m_size);
}
#else
ASSERT(codeOutData != outData);
performJITMemcpy(codeOutData, outData, m_size);
#endif
jumpsToLink.clear();
#if DUMP_LINK_STATISTICS
dumpLinkStatistics(codeOutData, initialSize, m_size);
#endif
#if DUMP_CODE
dumpCode(codeOutData, m_size);
#endif
}
#endif // ENABLE(BRANCH_COMPACTION)
void LinkBuffer::linkCode(MacroAssembler& macroAssembler, JITCompilationEffort effort)
{
// Ensure that the end of the last invalidation point does not extend beyond the end of the buffer.
macroAssembler.label();
#if !ENABLE(BRANCH_COMPACTION)
#if defined(ASSEMBLER_HAS_CONSTANT_POOL) && ASSEMBLER_HAS_CONSTANT_POOL
macroAssembler.m_assembler.buffer().flushConstantPool(false);
#endif
allocate(macroAssembler, effort);
if (!m_didAllocate)
return;
ASSERT(m_code);
AssemblerBuffer& buffer = macroAssembler.m_assembler.buffer();
void* code = m_code.dataLocation();
#if CPU(ARM64)
RELEASE_ASSERT(roundUpToMultipleOf<Assembler::instructionSize>(code) == code);
#endif
performJITMemcpy(code, buffer.data(), buffer.codeSize());
#if CPU(MIPS)
macroAssembler.m_assembler.relocateJumps(buffer.data(), code);
#endif
#elif CPU(ARM_THUMB2)
copyCompactAndLinkCode<uint16_t>(macroAssembler, effort);
#elif CPU(ARM64)
copyCompactAndLinkCode<uint32_t>(macroAssembler, effort);
#endif // !ENABLE(BRANCH_COMPACTION)
m_linkTasks = WTFMove(macroAssembler.m_linkTasks);
}
void LinkBuffer::allocate(MacroAssembler& macroAssembler, JITCompilationEffort effort)
{
size_t initialSize = macroAssembler.m_assembler.codeSize();
if (m_code) {
if (initialSize > m_size)
return;
size_t nopsToFillInBytes = m_size - initialSize;
macroAssembler.emitNops(nopsToFillInBytes);
m_didAllocate = true;
return;
}
while (initialSize % jitAllocationGranule) {
macroAssembler.breakpoint();
initialSize = macroAssembler.m_assembler.codeSize();
}
m_executableMemory = ExecutableAllocator::singleton().allocate(initialSize, effort);
if (!m_executableMemory)
return;
m_code = MacroAssemblerCodePtr<LinkBufferPtrTag>(m_executableMemory->start().retaggedPtr<LinkBufferPtrTag>());
m_size = initialSize;
m_didAllocate = true;
}
void LinkBuffer::performFinalization()
{
for (auto& task : m_linkTasks)
task->run(*this);
#ifndef NDEBUG
ASSERT(m_isJumpIsland || !isCompilationThread());
ASSERT(!m_completed);
ASSERT(isValid());
m_completed = true;
#endif
MacroAssembler::cacheFlush(code(), m_size);
}
#if DUMP_LINK_STATISTICS
void LinkBuffer::dumpLinkStatistics(void* code, size_t initializeSize, size_t finalSize)
{
static unsigned linkCount = 0;
static unsigned totalInitialSize = 0;
static unsigned totalFinalSize = 0;
linkCount++;
totalInitialSize += initialSize;
totalFinalSize += finalSize;
dataLogF("link %p: orig %u, compact %u (delta %u, %.2f%%)\n",
code, static_cast<unsigned>(initialSize), static_cast<unsigned>(finalSize),
static_cast<unsigned>(initialSize - finalSize),
100.0 * (initialSize - finalSize) / initialSize);
dataLogF("\ttotal %u: orig %u, compact %u (delta %u, %.2f%%)\n",
linkCount, totalInitialSize, totalFinalSize, totalInitialSize - totalFinalSize,
100.0 * (totalInitialSize - totalFinalSize) / totalInitialSize);
}
#endif
#if DUMP_CODE
void LinkBuffer::dumpCode(void* code, size_t size)
{
#if CPU(ARM_THUMB2)
// Dump the generated code in an asm file format that can be assembled and then disassembled
// for debugging purposes. For example, save this output as jit.s:
// gcc -arch armv7 -c jit.s
// otool -tv jit.o
static unsigned codeCount = 0;
unsigned short* tcode = static_cast<unsigned short*>(code);
size_t tsize = size / sizeof(short);
char nameBuf[128];
snprintf(nameBuf, sizeof(nameBuf), "_jsc_jit%u", codeCount++);
dataLogF("\t.syntax unified\n"
"\t.section\t__TEXT,__text,regular,pure_instructions\n"
"\t.globl\t%s\n"
"\t.align 2\n"
"\t.code 16\n"
"\t.thumb_func\t%s\n"
"# %p\n"
"%s:\n", nameBuf, nameBuf, code, nameBuf);
for (unsigned i = 0; i < tsize; i++)
dataLogF("\t.short\t0x%x\n", tcode[i]);
#endif
}
#endif
} // namespace JSC
#endif // ENABLE(ASSEMBLER)