arm: add support for detecting NEON (WIP) #144
Conversation
|
These days vector seems more popular than simd. And should be generic of course. Sljit does not emit any vector instructions at the moment. |
carenas
force-pushed
the
neon
branch
3 times, most recently
from
06c92fe
to
522d11d
Compare
Requires getauxval() which is available at least in Linux/Android with recent versions of the libc and that is therefore behind a configure like macro. A similar function from FreeBSD>=12, Windows and NetBSD is used for each case. While at it, consolidate the code to use the same externally visible flag that is used in x86 for SSE2.
| #define SLJIT_DETECT_SSE2 1 | ||
| #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32) || \ | ||
| (defined SLJIT_CONFIG_ARM_V7 && SLJIT_CONFIG_ARM_V7) | ||
| /* Auto detect availability of SSE2 (using CPUID) or NEON. |
There was a problem hiding this comment.
It might be worth doing the checks here to avoid adding the "detect" support for cases that are known not to require it (ex: OpenBSD or macOS)
| return; | ||
|
|
||
| #if defined(__ARM_ARCH) && __ARM_ARCH == 8 | ||
| /* TODO: confirm if optional with armv9 */ |
There was a problem hiding this comment.
if true (because NEON is being "replaced" by SVE2, then this might also apply to the 64 bit version and therefore some of it might need to move to a "common" file
| if (simd && fpu) | ||
| strcpy(features, " (with: fpu, simd)"); | ||
| else if (fpu) | ||
| strcpy(features, " (with fpu)"); | ||
| else | ||
| strcpy(features, " (without fpu)"); |
There was a problem hiding this comment.
this might misrepreset the case where (!fpu && simd), but that is by design as that combination (while possible) is not something that is available in the market yet.
SLJIT_IS_FPU_AVAILABLE and its usecases is also missing and worth discussing, but IMHO might be able to wait until SIMD support is added to sljit proper.
Probably worth some refactoring to make it more generic and apply as well to s390x, or more extensible considering there are now multiple vector instruction sets per architecture with the most recent ones being variable width.
Is there any value on haviing implementation speciific names?, wouldn't it better to use SLJIT_HAS_SIMD?, how to differenciate if vectors would have different width?
the arm code for SIMD in PCRE2 has a lot of issues, but one of the advantages it had was that it was less verbose and offloaded some of the hand coding by using the compiler intrinsics (but should be instead using sljit intrinsics), why not add them?