Export of internal Abseil changes.
-- babbb6421068af3831870fd5995444437ace6769 by Derek Mauro <dmauro@google.com>: Rollback of: Make raw_hash_set_test less flaky. Split the timing loop into chunks so that we are less suceptible to antogantistic processes. PiperOrigin-RevId: 243854490 -- a2711f17a712f6d09799bf32363d67526737b486 by CJ Johnson <johnsoncj@google.com>: Relocates IsAtLeastForwardIterator to internal/inlined_vector.h PiperOrigin-RevId: 243846090 -- 6c14cdbeb9a61022c27f8957654f930d8abf2fc1 by Matt Kulukundis <kfm@google.com>: Make raw_hash_set_test less flaky. Split the timing loop into chunks so that we are less suceptible to antogantistic processes. PiperOrigin-RevId: 243824289 -- ee6072a6b6e0ac653622524ceb09db3b9e870f96 by Samuel Benzaquen <sbenza@google.com>: Improve format parser performance. Replace the main switch with a lookup in the existing tag table. Improve the ABI of ConsumeUnboundConversion a little. PiperOrigin-RevId: 243824112 -- 24b9e6476dfa4be8d644359eab8ac6816492f187 by Abseil Team <absl-team@google.com>: Fix DR numbers: 3800 ? 3080, 3801 ? 3081. PiperOrigin-RevId: 243804213 -- 0660404074707e197684f07cc0bffe4a9c35cd2f by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 243757359 -- ba0f5bb9b8584d75c4ffc44ff3cb8c691796ffc6 by Xiaoyi Zhang <zhangxy@google.com>: Consolidate ABSL_INTERNAL_UNALIGNED_* implementation into memcpy. The compiler should be good enough to optimize these operations. See https://github.com/abseil/abseil-cpp/issues/269 for background. PiperOrigin-RevId: 243323941 -- 00853a8756548df7217513c562d604b4ee5c6ab9 by Eric Fiselier <ericwf@google.com>: Reexport memory.h from optional.h for compatibility between libc++ and libstdc++. PiperOrigin-RevId: 243313425 GitOrigin-RevId: babbb6421068af3831870fd5995444437ace6769 Change-Id: Ic53c127ad857a431ad60c98b27cc585fed50a3e3
This commit is contained in:
parent
a02f62f456
commit
d902eb869b
10 changed files with 186 additions and 328 deletions
|
@ -25,15 +25,6 @@
|
|||
// unaligned APIs
|
||||
|
||||
// Portable handling of unaligned loads, stores, and copies.
|
||||
// On some platforms, like ARM, the copy functions can be more efficient
|
||||
// then a load and a store.
|
||||
//
|
||||
// It is possible to implement all of these these using constant-length memcpy
|
||||
// calls, which is portable and will usually be inlined into simple loads and
|
||||
// stores if the architecture supports it. However, such inlining usually
|
||||
// happens in a pass that's quite late in compilation, which means the resulting
|
||||
// loads and stores cannot participate in many other optimizations, leading to
|
||||
// overall worse code.
|
||||
|
||||
// The unaligned API is C++ only. The declarations use C++ features
|
||||
// (namespaces, inline) which are absent or incompatible in C.
|
||||
|
@ -108,164 +99,8 @@ inline void UnalignedStore64(void *p, uint64_t v) {
|
|||
#define ABSL_INTERNAL_UNALIGNED_STORE64(_p, _val) \
|
||||
(absl::base_internal::UnalignedStore64(_p, _val))
|
||||
|
||||
#elif defined(UNDEFINED_BEHAVIOR_SANITIZER)
|
||||
|
||||
namespace absl {
|
||||
namespace base_internal {
|
||||
|
||||
inline uint16_t UnalignedLoad16(const void *p) {
|
||||
uint16_t t;
|
||||
memcpy(&t, p, sizeof t);
|
||||
return t;
|
||||
}
|
||||
|
||||
inline uint32_t UnalignedLoad32(const void *p) {
|
||||
uint32_t t;
|
||||
memcpy(&t, p, sizeof t);
|
||||
return t;
|
||||
}
|
||||
|
||||
inline uint64_t UnalignedLoad64(const void *p) {
|
||||
uint64_t t;
|
||||
memcpy(&t, p, sizeof t);
|
||||
return t;
|
||||
}
|
||||
|
||||
inline void UnalignedStore16(void *p, uint16_t v) { memcpy(p, &v, sizeof v); }
|
||||
|
||||
inline void UnalignedStore32(void *p, uint32_t v) { memcpy(p, &v, sizeof v); }
|
||||
|
||||
inline void UnalignedStore64(void *p, uint64_t v) { memcpy(p, &v, sizeof v); }
|
||||
|
||||
} // namespace base_internal
|
||||
} // namespace absl
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD16(_p) \
|
||||
(absl::base_internal::UnalignedLoad16(_p))
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD32(_p) \
|
||||
(absl::base_internal::UnalignedLoad32(_p))
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD64(_p) \
|
||||
(absl::base_internal::UnalignedLoad64(_p))
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE16(_p, _val) \
|
||||
(absl::base_internal::UnalignedStore16(_p, _val))
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE32(_p, _val) \
|
||||
(absl::base_internal::UnalignedStore32(_p, _val))
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE64(_p, _val) \
|
||||
(absl::base_internal::UnalignedStore64(_p, _val))
|
||||
|
||||
#elif defined(__x86_64__) || defined(_M_X64) || defined(__i386) || \
|
||||
defined(_M_IX86) || defined(__ppc__) || defined(__PPC__) || \
|
||||
defined(__ppc64__) || defined(__PPC64__)
|
||||
|
||||
// x86 and x86-64 can perform unaligned loads/stores directly;
|
||||
// modern PowerPC hardware can also do unaligned integer loads and stores;
|
||||
// but note: the FPU still sends unaligned loads and stores to a trap handler!
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD16(_p) \
|
||||
(*reinterpret_cast<const uint16_t *>(_p))
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD32(_p) \
|
||||
(*reinterpret_cast<const uint32_t *>(_p))
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD64(_p) \
|
||||
(*reinterpret_cast<const uint64_t *>(_p))
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE16(_p, _val) \
|
||||
(*reinterpret_cast<uint16_t *>(_p) = (_val))
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE32(_p, _val) \
|
||||
(*reinterpret_cast<uint32_t *>(_p) = (_val))
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE64(_p, _val) \
|
||||
(*reinterpret_cast<uint64_t *>(_p) = (_val))
|
||||
|
||||
#elif defined(__arm__) && \
|
||||
!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__)
|
||||
|
||||
|
||||
// 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. Newer gcc (>= 4.6) set an __ARM_FEATURE_ALIGNED #define,
|
||||
// so in time, maybe we can move on to that.
|
||||
//
|
||||
// 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.
|
||||
|
||||
namespace absl {
|
||||
namespace base_internal {
|
||||
|
||||
struct Unaligned16Struct {
|
||||
uint16_t value;
|
||||
uint8_t dummy; // To make the size non-power-of-two.
|
||||
} ABSL_ATTRIBUTE_PACKED;
|
||||
|
||||
struct Unaligned32Struct {
|
||||
uint32_t value;
|
||||
uint8_t dummy; // To make the size non-power-of-two.
|
||||
} ABSL_ATTRIBUTE_PACKED;
|
||||
|
||||
} // namespace base_internal
|
||||
} // namespace absl
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD16(_p) \
|
||||
((reinterpret_cast<const ::absl::base_internal::Unaligned16Struct *>(_p)) \
|
||||
->value)
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD32(_p) \
|
||||
((reinterpret_cast<const ::absl::base_internal::Unaligned32Struct *>(_p)) \
|
||||
->value)
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE16(_p, _val) \
|
||||
((reinterpret_cast< ::absl::base_internal::Unaligned16Struct *>(_p)) \
|
||||
->value = (_val))
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE32(_p, _val) \
|
||||
((reinterpret_cast< ::absl::base_internal::Unaligned32Struct *>(_p)) \
|
||||
->value = (_val))
|
||||
|
||||
namespace absl {
|
||||
namespace base_internal {
|
||||
|
||||
inline uint64_t UnalignedLoad64(const void *p) {
|
||||
uint64_t t;
|
||||
memcpy(&t, p, sizeof t);
|
||||
return t;
|
||||
}
|
||||
|
||||
inline void UnalignedStore64(void *p, uint64_t v) { memcpy(p, &v, sizeof v); }
|
||||
|
||||
} // namespace base_internal
|
||||
} // namespace absl
|
||||
|
||||
#define ABSL_INTERNAL_UNALIGNED_LOAD64(_p) \
|
||||
(absl::base_internal::UnalignedLoad64(_p))
|
||||
#define ABSL_INTERNAL_UNALIGNED_STORE64(_p, _val) \
|
||||
(absl::base_internal::UnalignedStore64(_p, _val))
|
||||
|
||||
#else
|
||||
|
||||
// ABSL_INTERNAL_NEED_ALIGNED_LOADS is defined when the underlying platform
|
||||
// doesn't support unaligned access.
|
||||
#define ABSL_INTERNAL_NEED_ALIGNED_LOADS
|
||||
|
||||
// These functions are provided for architectures that don't support
|
||||
// unaligned loads and stores.
|
||||
|
||||
namespace absl {
|
||||
namespace base_internal {
|
||||
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue