cuberite/libdeflate
1
0
Fork 0
mirror of https://github.com/cuberite/libdeflate.git synced 2025-09-10 04:41:45 -04:00
Code Issues Packages Projects Releases Wiki Activity

lib/arm/cpu_features: reorganize arm feature macros

Browse Source 
Reorganize up some confusing logic.
This commit is contained in:
Eric Biggers 2020-10-10 23:02:50 -07:00
parent 3b9775d4aa
commit 7373bdc9ff
2 changed files with 33 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
common/compiler_gcc.h
View File

@ -91,26 +91,38 @@
# define COMPILER_SUPPORTS_AVX512BW_TARGET_INTRINSICS \ # define COMPILER_SUPPORTS_AVX512BW_TARGET_INTRINSICS \
COMPILER_SUPPORTS_AVX512BW_TARGET COMPILER_SUPPORTS_AVX512BW_TARGET
# endif # endif
# elif (defined(__arm__) && defined(__ARM_FP)) || defined(__aarch64__)
/* arm: including arm_neon.h requires hardware fp support */ # elif defined(__arm__) || defined(__aarch64__)
/* /*
* Prior to gcc 6.1 (r230411 for arm, r226563 for aarch64), NEON * Determine whether NEON and crypto intrinsics are supported.
* and crypto intrinsics not available in the main target could not be
* used in 'target' attribute functions.
* *
* clang as of 5.0.1 still doesn't allow it. But, it does seem to allow * With gcc prior to 6.1, (r230411 for arm32, r226563 for arm64), neither
* the pmull intrinsics if only __ARM_NEON is enabled. * was available unless enabled in the main target.
*
* But even after that, to include <arm_neon.h> (which contains both the
* basic NEON intrinsics and the crypto intrinsics) the main target still
* needs to have:
* - gcc: hardware floating point support
* - clang: NEON support (but not necessarily crypto support)
*/ */
# define COMPILER_SUPPORTS_NEON_TARGET_INTRINSICS GCC_PREREQ(6, 1) # if (GCC_PREREQ(6, 1) && defined(__ARM_FP)) || \
# ifdef __ARM_NEON (defined(__clang__) && defined(__ARM_NEON))
# define COMPILER_SUPPORTS_PMULL_TARGET_INTRINSICS \ # define COMPILER_SUPPORTS_NEON_TARGET_INTRINSICS 1
(GCC_PREREQ(6, 1) || __has_builtin(__builtin_neon_vmull_p64)) /*
# else * The crypto intrinsics are broken on arm32 with clang, even when using
# define COMPILER_SUPPORTS_PMULL_TARGET_INTRINSICS \ * -mfpu=crypto-neon-fp-armv8, because clang's <arm_neon.h> puts them
(GCC_PREREQ(6, 1)) * behind __aarch64__. Undefine __ARM_FEATURE_CRYPTO in that case...
*/
# if defined(__clang__) && defined(__arm__)
# undef __ARM_FEATURE_CRYPTO
# elif __has_builtin(__builtin_neon_vmull_p64) || !defined(__clang__)
# define COMPILER_SUPPORTS_PMULL_TARGET_INTRINSICS 1
# endif # endif
# endif # endif
# endif /* __arm__ || __aarch64__ */
#endif /* COMPILER_SUPPORTS_TARGET_FUNCTION_ATTRIBUTE */ #endif /* COMPILER_SUPPORTS_TARGET_FUNCTION_ATTRIBUTE */
/* Newer gcc supports __BYTE_ORDER__. Older gcc doesn't. */ /* Newer gcc supports __BYTE_ORDER__. Older gcc doesn't. */

8
lib/arm/crc32_impl.h
View File

@ -35,14 +35,12 @@
* See x86/crc32_pclmul_template.h for an explanation. * See x86/crc32_pclmul_template.h for an explanation.
*/ */
#undef DISPATCH_PMULL #undef DISPATCH_PMULL
#if (defined(__ARM_FEATURE_CRYPTO) || \ #if !defined(DEFAULT_IMPL) && \
(defined(__ARM_FEATURE_CRYPTO) || \
(ARM_CPU_FEATURES_ENABLED && \ (ARM_CPU_FEATURES_ENABLED && \
COMPILER_SUPPORTS_PMULL_TARGET_INTRINSICS)) && \ COMPILER_SUPPORTS_PMULL_TARGET_INTRINSICS)) && \
/* not yet tested on big endian, probably needs changes to work there */ \ /* not yet tested on big endian, probably needs changes to work there */ \
(defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) && \ (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
/* clang as of v5.0.1 doesn't allow pmull intrinsics in 32-bit mode, even
* when compiling with -mfpu=crypto-neon-fp-armv8 */ \
!(defined(__clang__) && defined(__arm__))
# define FUNCNAME crc32_pmull # define FUNCNAME crc32_pmull
# define FUNCNAME_ALIGNED crc32_pmull_aligned # define FUNCNAME_ALIGNED crc32_pmull_aligned
# ifdef __ARM_FEATURE_CRYPTO # ifdef __ARM_FEATURE_CRYPTO