JitCodeBuffer: Flush cache after committing code on ARM CPUs

ARM's instruction and data caches are not coherent, so we need to flush
before executing to ensure there's no stale data left over.
This commit is contained in:
Connor McLaughlin 2019-12-03 20:21:30 +10:00
parent bbe1cb5fe9
commit 98e67616ec
2 changed files with 62 additions and 33 deletions

View file

@ -7,14 +7,26 @@
#include <sys/mman.h> #include <sys/mman.h>
#endif #endif
JitCodeBuffer::JitCodeBuffer(size_t size /* = 64 * 1024 * 1024 */, size_t far_code_size /* = 0 */) static void DoCacheFlush(u8* address, u32 len)
{
#if defined(Y_PLATFORM_WINDOWS)
FlushInstructionCache(GetCurrentProcess(), address, len);
#elif defined(Y_COMPILER_GCC) || defined(Y_COMPILER_CLANG)
__clear_cache(address, address + len);
#else
#error Unknown platform.
#endif
}
JitCodeBuffer::JitCodeBuffer(u32 size /* = 64 * 1024 * 1024 */, u32 far_code_size /* = 0 */)
{ {
m_total_size = size + far_code_size; m_total_size = size + far_code_size;
#if defined(Y_PLATFORM_WINDOWS) #if defined(Y_PLATFORM_WINDOWS)
m_code_ptr = VirtualAlloc(nullptr, m_total_size, MEM_COMMIT, PAGE_EXECUTE_READWRITE); m_code_ptr = static_cast<u8*>(VirtualAlloc(nullptr, m_total_size, MEM_COMMIT, PAGE_EXECUTE_READWRITE));
#elif defined(Y_PLATFORM_LINUX) || defined(Y_PLATFORM_ANDROID) #elif defined(Y_PLATFORM_LINUX) || defined(Y_PLATFORM_ANDROID)
m_code_ptr = mmap(nullptr, m_total_size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); m_code_ptr = static_cast<u8*>(
mmap(nullptr, m_total_size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0));
#else #else
m_code_ptr = nullptr; m_code_ptr = nullptr;
#endif #endif
@ -39,28 +51,45 @@ JitCodeBuffer::~JitCodeBuffer()
#endif #endif
} }
void JitCodeBuffer::CommitCode(size_t length) void JitCodeBuffer::CommitCode(u32 length)
{ {
if (length == 0)
return;
#if defined(Y_CPU_ARM) || defined(Y_CPU_AARCH64)
// ARM instruction and data caches are not coherent, we need to flush after every block.
DoCacheFlush(m_free_code_ptr, length);
#endif
Assert(length <= (m_code_size - m_code_used)); Assert(length <= (m_code_size - m_code_used));
m_free_code_ptr = reinterpret_cast<u8*>(m_free_code_ptr) + length; m_free_code_ptr += length;
m_code_used += length; m_code_used += length;
} }
void JitCodeBuffer::CommitFarCode(size_t length) void JitCodeBuffer::CommitFarCode(u32 length)
{ {
if (length == 0)
return;
#if defined(Y_CPU_ARM) || defined(Y_CPU_AARCH64)
// ARM instruction and data caches are not coherent, we need to flush after every block.
DoCacheFlush(m_free_code_ptr, length);
#endif
Assert(length <= (m_far_code_size - m_far_code_used)); Assert(length <= (m_far_code_size - m_far_code_used));
m_free_far_code_ptr = reinterpret_cast<u8*>(m_free_far_code_ptr) + length; m_free_far_code_ptr += length;
m_far_code_used += length; m_far_code_used += length;
} }
void JitCodeBuffer::Reset() void JitCodeBuffer::Reset()
{ {
#if defined(Y_PLATFORM_WINDOWS) std::memset(m_code_ptr, 0, m_code_size);
FlushInstructionCache(GetCurrentProcess(), m_code_ptr, m_total_size); DoCacheFlush(m_code_ptr, m_code_size);
#elif defined(Y_PLATFORM_LINUX) || defined(Y_PLATFORM_ANDROID) if (m_far_code_size > 0)
// TODO {
#endif std::memset(m_far_code_ptr, 0, m_far_code_size);
DoCacheFlush(m_far_code_ptr, m_far_code_size);
}
m_free_code_ptr = m_code_ptr; m_free_code_ptr = m_code_ptr;
m_code_used = 0; m_code_used = 0;
@ -71,11 +100,11 @@ void JitCodeBuffer::Reset()
void JitCodeBuffer::Align(u32 alignment, u8 padding_value) void JitCodeBuffer::Align(u32 alignment, u8 padding_value)
{ {
DebugAssert(Common::IsPow2(alignment)); DebugAssert(Common::IsPow2(alignment));
const size_t num_padding_bytes = const u32 num_padding_bytes =
std::min(static_cast<size_t>(Common::AlignUpPow2(reinterpret_cast<uintptr_t>(m_free_code_ptr), alignment) - std::min(static_cast<u32>(Common::AlignUpPow2(reinterpret_cast<uintptr_t>(m_free_code_ptr), alignment) -
reinterpret_cast<uintptr_t>(m_free_code_ptr)), reinterpret_cast<uintptr_t>(m_free_code_ptr)),
GetFreeCodeSpace()); GetFreeCodeSpace());
std::memset(m_free_code_ptr, padding_value, num_padding_bytes); std::memset(m_free_code_ptr, padding_value, num_padding_bytes);
m_free_code_ptr = reinterpret_cast<char*>(m_free_code_ptr) + num_padding_bytes; m_free_code_ptr += num_padding_bytes;
m_code_used += num_padding_bytes; m_code_used += num_padding_bytes;
} }

View file

@ -4,34 +4,34 @@
class JitCodeBuffer class JitCodeBuffer
{ {
public: public:
JitCodeBuffer(size_t size = 64 * 1024 * 1024, size_t far_code_size = 0); JitCodeBuffer(u32 size = 64 * 1024 * 1024, u32 far_code_size = 0);
~JitCodeBuffer(); ~JitCodeBuffer();
void Reset(); void Reset();
void* GetFreeCodePointer() const { return m_free_code_ptr; } u8* GetFreeCodePointer() const { return m_free_code_ptr; }
size_t GetFreeCodeSpace() const { return (m_code_size - m_code_used); } u32 GetFreeCodeSpace() const { return static_cast<u32>(m_code_size - m_code_used); }
void CommitCode(size_t length); void CommitCode(u32 length);
void* GetFreeFarCodePointer() const { return m_free_far_code_ptr; } u8* GetFreeFarCodePointer() const { return m_free_far_code_ptr; }
size_t GetFreeFarCodeSpace() const { return (m_far_code_size - m_far_code_used); } u32 GetFreeFarCodeSpace() const { return static_cast<u32>(m_far_code_size - m_far_code_used); }
void CommitFarCode(size_t length); void CommitFarCode(u32 length);
/// Adjusts the free code pointer to the specified alignment, padding with bytes. /// Adjusts the free code pointer to the specified alignment, padding with bytes.
/// Assumes alignment is a power-of-two. /// Assumes alignment is a power-of-two.
void Align(u32 alignment, u8 padding_value); void Align(u32 alignment, u8 padding_value);
private: private:
void* m_code_ptr; u8* m_code_ptr;
void* m_free_code_ptr; u8* m_free_code_ptr;
size_t m_code_size; u32 m_code_size;
size_t m_code_used; u32 m_code_used;
void* m_far_code_ptr; u8* m_far_code_ptr;
void* m_free_far_code_ptr; u8* m_free_far_code_ptr;
size_t m_far_code_size; u32 m_far_code_size;
size_t m_far_code_used; u32 m_far_code_used;
size_t m_total_size; u32 m_total_size;
}; };