diff options
| author | Nathan Moinvaziri <nathan@nathanm.com> | 2026-03-06 11:38:28 -0800 |
|---|---|---|
| committer | Hans Kristian Rosbach <hk-github@circlestorm.org> | 2026-03-08 23:37:37 +0100 |
| commit | a5b10a321d1a4011cb188e09f9d29f22fe6b5338 (patch) | |
| tree | 677824c625472cc162dad4869a969bf69906e54e | |
| parent | 30206c1cac40a8bed335405ed2e64559fb7b40bc (diff) | |
| download | Project-Tick-a5b10a321d1a4011cb188e09f9d29f22fe6b5338.tar.gz Project-Tick-a5b10a321d1a4011cb188e09f9d29f22fe6b5338.zip | |
Add NMAX_ALIGNED32 and use it in NEON adler32
Define NMAX_ALIGNED32 as NMAX rounded down to a multiple of 32 (5536)
and use it in the NEON adler32 implementation to ensure that src stays
32-byte aligned throughout the main SIMD loop. Previously, NMAX (5552)
is not a multiple of 32, so after the alignment preamble the first
iteration could process a non-32-aligned number of bytes, causing src
to lose 32-byte alignment for all subsequent iterations.
The first iteration's budget is rounded down with ALIGN_DOWN after
subtracting align_diff, ensuring k is always a multiple of 32.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
| -rw-r--r-- | adler32_p.h | 2 | ||||
| -rw-r--r-- | arch/arm/adler32_neon.c | 6 |
2 files changed, 5 insertions, 3 deletions
diff --git a/adler32_p.h b/adler32_p.h index b5d5f1615c..836029b2ab 100644 --- a/adler32_p.h +++ b/adler32_p.h @@ -11,6 +11,8 @@ #define BASE 65521U /* largest prime smaller than 65536 */ #define NMAX 5552 /* NMAX is the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1 */ +#define NMAX_ALIGNED32 (NMAX & ~31) +/* NMAX rounded down to a multiple of 32 is 5536 */ #define ADLER_DO1(sum1, sum2, buf, i) {(sum1) += buf[(i)]; (sum2) += (sum1);} #define ADLER_DO2(sum1, sum2, buf, i) {ADLER_DO1(sum1, sum2, buf, i); ADLER_DO1(sum1, sum2, buf, i+1);} diff --git a/arch/arm/adler32_neon.c b/arch/arm/adler32_neon.c index 9bbb293bd8..169a785edc 100644 --- a/arch/arm/adler32_neon.c +++ b/arch/arm/adler32_neon.c @@ -294,7 +294,7 @@ Z_FORCEINLINE static uint32_t adler32_copy_impl(uint32_t adler, uint8_t *dst, co * In the copying variant we use fallback to 4x loads and 4x stores, * as ld1x4 seems to block ILP when stores are in the mix */ size_t align_diff = MIN(ALIGN_DIFF(src, 32), len); - size_t n = NMAX; + size_t n = NMAX_ALIGNED32; if (align_diff) { adler32_copy_align(&pair[0], dst, src, align_diff, &pair[1], 31, COPY); @@ -302,7 +302,7 @@ Z_FORCEINLINE static uint32_t adler32_copy_impl(uint32_t adler, uint8_t *dst, co dst += align_diff; src += align_diff; len -= align_diff; - n -= align_diff; + n = ALIGN_DOWN(n - align_diff, 32); } while (len >= 16) { @@ -321,7 +321,7 @@ Z_FORCEINLINE static uint32_t adler32_copy_impl(uint32_t adler, uint8_t *dst, co if (COPY) dst += k; len -= k; - n = NMAX; + n = NMAX_ALIGNED32; } /* Process tail (len < 16). */ |