Remove platform-dependent code for unaligned loads/stores.

Snappy issues multi-byte (16/32/64-bit) loads and stores that are not
aligned, meaning the addresses are 16/32/64-bit multiples. This is
accomplished using two methods:

1) The portable method allocates a uint{16,32,64}_t on the stack, and
std::memcpy()s the bytes into/from the integer. This method relies on
well-defined behaviori (std::memcpy() works on all valid pointers,
fixed-width unsigned integer types use a pure binary representation and
therefore have no invalid values), and should compile to valid code on
all platforms.

2) The fast method reinterpret_casts the address to a pointer to a
uint{16,32,64}_t and dereferences the pointer. This is expected to
compile to one hardware instruction (mov on x86, ldr/str on arm). The
caveat is that the reinterpret_cast is undefined behavior (UB) unless the
address happened to be a valid uint{16,32,64}_t pointer. The UB shows up
as follows.
* On architectures that don't have hardware instructions for unaligned
  loads / stores, the pointer access can trigger a hardware exceptions.
  This is mitigated by #ifdef blocks that attempt to restrict the fast
  method to platforms that support it.
* On architectures that have separate instructions for aligned and
  unaligned access, the compiler may need an explicit hint to emit the
  hardware instruction for unaligned access. This is accomplished on
  Clang and GCC by wrapping the pointers into structs tagged with
  __attribute__((__packed__)).

This CL removes the fast method. Fortunately, compilers have advanced
enough that the portable method gets compiled down to the same
instructions as the fast method, without the need for the caveats
explained above. Specifically, modern Clang, GCC and MSVC optimize
std::memcpy() to a single instruction (mov / ldr / str). A test case
proving this can be seen at https://godbolt.org/z/gZg2Fk
PiperOrigin-RevId: 306342728
This commit is contained in:
Victor Costan 2020-04-14 00:19:00 +00:00
parent 27ff130ff9
commit 041c608086
1 changed files with 18 additions and 110 deletions

View File

@ -130,132 +130,40 @@ static const int64_t kint64max = std::numeric_limits<int64_t>::max();
// Potentially unaligned loads and stores.
// x86, PowerPC, and ARM64 can simply do these loads and stores native.
#if defined(__i386__) || defined(__x86_64__) || defined(__powerpc__) || \
defined(__aarch64__)
#define UNALIGNED_LOAD16(_p) (*reinterpret_cast<const uint16_t *>(_p))
#define UNALIGNED_LOAD32(_p) (*reinterpret_cast<const uint32_t *>(_p))
#define UNALIGNED_LOAD64(_p) (*reinterpret_cast<const uint64_t *>(_p))
#define UNALIGNED_STORE16(_p, _val) (*reinterpret_cast<uint16_t *>(_p) = (_val))
#define UNALIGNED_STORE32(_p, _val) (*reinterpret_cast<uint32_t *>(_p) = (_val))
#define UNALIGNED_STORE64(_p, _val) (*reinterpret_cast<uint64_t *>(_p) = (_val))
// ARMv7 and newer support native unaligned accesses, but only of 16-bit
// and 32-bit values (not 64-bit); older versions either raise a fatal signal,
// do an unaligned read and rotate the words around a bit, or do the reads very
// slowly (trip through kernel mode). There's no simple #define that says just
// “ARMv7 or higher”, so we have to filter away all ARMv5 and ARMv6
// sub-architectures.
//
// This is a mess, but there's not much we can do about it.
//
// To further complicate matters, only LDR instructions (single reads) are
// allowed to be unaligned, not LDRD (two reads) or LDM (many reads). Unless we
// explicitly tell the compiler that these accesses can be unaligned, it can and
// will combine accesses. On armcc, the way to signal this is done by accessing
// through the type (uint32_t __packed *), but GCC has no such attribute
// (it ignores __attribute__((packed)) on individual variables). However,
// we can tell it that a _struct_ is unaligned, which has the same effect,
// so we do that.
#elif defined(__arm__) && \
!defined(__ARM_ARCH_4__) && \
!defined(__ARM_ARCH_4T__) && \
!defined(__ARM_ARCH_5__) && \
!defined(__ARM_ARCH_5T__) && \
!defined(__ARM_ARCH_5TE__) && \
!defined(__ARM_ARCH_5TEJ__) && \
!defined(__ARM_ARCH_6__) && \
!defined(__ARM_ARCH_6J__) && \
!defined(__ARM_ARCH_6K__) && \
!defined(__ARM_ARCH_6Z__) && \
!defined(__ARM_ARCH_6ZK__) && \
!defined(__ARM_ARCH_6T2__)
#if __GNUC__
#define ATTRIBUTE_PACKED __attribute__((__packed__))
#else
#define ATTRIBUTE_PACKED
#endif
namespace base {
namespace internal {
struct Unaligned16Struct {
uint16_t value;
uint8_t dummy; // To make the size non-power-of-two.
} ATTRIBUTE_PACKED;
struct Unaligned32Struct {
uint32_t value;
uint8_t dummy; // To make the size non-power-of-two.
} ATTRIBUTE_PACKED;
} // namespace internal
} // namespace base
#define UNALIGNED_LOAD16(_p) \
((reinterpret_cast<const ::snappy::base::internal::Unaligned16Struct *>(_p))->value)
#define UNALIGNED_LOAD32(_p) \
((reinterpret_cast<const ::snappy::base::internal::Unaligned32Struct *>(_p))->value)
#define UNALIGNED_STORE16(_p, _val) \
((reinterpret_cast< ::snappy::base::internal::Unaligned16Struct *>(_p))->value = \
(_val))
#define UNALIGNED_STORE32(_p, _val) \
((reinterpret_cast< ::snappy::base::internal::Unaligned32Struct *>(_p))->value = \
(_val))
// TODO: NEON supports unaligned 64-bit loads and stores.
// See if that would be more efficient on platforms supporting it,
// at least for copies.
inline uint64_t UNALIGNED_LOAD64(const void *p) {
uint64_t t;
std::memcpy(&t, p, sizeof t);
return t;
}
inline void UNALIGNED_STORE64(void *p, uint64_t v) {
std::memcpy(p, &v, sizeof v);
}
#else
// These functions are provided for architectures that don't support
// unaligned loads and stores.
inline uint16_t UNALIGNED_LOAD16(const void *p) {
uint16_t t;
std::memcpy(&t, p, sizeof t);
return t;
// Compiles to a single movzx/ldrh on clang/gcc/msvc.
uint16_t v;
std::memcpy(&v, p, sizeof(v));
return v;
}
inline uint32_t UNALIGNED_LOAD32(const void *p) {
uint32_t t;
std::memcpy(&t, p, sizeof t);
return t;
// Compiles to a single mov/ldr on clang/gcc/msvc.
uint32_t v;
std::memcpy(&v, p, sizeof(v));
return v;
}
inline uint64_t UNALIGNED_LOAD64(const void *p) {
uint64_t t;
std::memcpy(&t, p, sizeof t);
return t;
// Compiles to a single mov/ldr on clang/gcc/msvc.
uint64_t v;
std::memcpy(&v, p, sizeof(v));
return v;
}
inline void UNALIGNED_STORE16(void *p, uint16_t v) {
std::memcpy(p, &v, sizeof v);
// Compiles to a single mov/strh on clang/gcc/msvc.
std::memcpy(p, &v, sizeof(v));
}
inline void UNALIGNED_STORE32(void *p, uint32_t v) {
std::memcpy(p, &v, sizeof v);
// Compiles to a single mov/str on clang/gcc/msvc.
std::memcpy(p, &v, sizeof(v));
}
inline void UNALIGNED_STORE64(void *p, uint64_t v) {
std::memcpy(p, &v, sizeof v);
// Compiles to a single mov/str on clang/gcc/msvc.
std::memcpy(p, &v, sizeof(v));
}
#endif