simdtest allocated 64 spare bytes so it could test an unaligned buffer, which
is exactly the slack that hid a 64-byte over-read in the AVX2 assembly for as
long as it existed. Add a pass that places the buffer flush against a
PROT_NONE page, so a read past the end faults in the test rather than in
somebody's transfer.
Every length from 128 to 4096, so each remainder mod 64 and both alignments
are covered, and all four implementations are checked -- the assembly was the
one at fault here, but the intrinsic paths preload too.
It fails closed. A guard this test cannot set up means it is not testing what
the caller thinks, so a failed sysconf/mmap/mprotect is a failure rather than a
pass that looks identical to a real one. And because the dispatcher falls back
on a CPU without AVX2 -- where the guard loop proves nothing about the code
under test -- it says which of the two happened rather than letting a fallback
run read as coverage.
Without the fix this segfaults; the suite's simd-checksum test reports the
non-zero exit.
Remove stores that are never read before being overwritten or going
out of scope. No behavior change except batch.c write_opt, which now
accumulates the leading-space write error into the return value
(consistent with the arg branch) instead of discarding it.
simd-checksum-x86_64.cpp, options.c, util1.c, batch.c
When built with --enable-roll-asm, get_checksum1() called the AVX2 asm
routine get_checksum1_avx2_asm() unconditionally. Unlike the intrinsic
path (get_checksum1_avx2_64), which is function-multiversioned with a
target("default") fallback and so resolves safely on any CPU, the asm
routine is a single AVX2-only symbol with no fallback. On an x86-64 host
without AVX2 (an older CPU, or a VM that does not expose AVX2) the first
block checksum executes a VEX-encoded instruction and dies with SIGILL,
which surfaces as "connection unexpectedly closed (0 bytes received so
far)" and a code-12 protocol error.
Gate the asm call on a cached __builtin_cpu_supports("avx2") check, the
same signal the intrinsic resolver uses. When AVX2 is absent we skip it
and the SSSE3/SSE2/scalar steps (safe everywhere) do the work. Apply the
same guard in the simdtest harness so it can run on non-AVX2 hosts too.
The AVX2 get_checksum1_avx2_64() read mul_one before initializing it,
which is undefined behavior. Replace the cmpeq/abs trick with
_mm256_set1_epi8(1) to match the SSSE3 and SSE2 versions.
Add a TEST_SIMD_CHECKSUM1 test mode that verifies all SIMD paths
(SSE2, SSSE3, AVX2, and the full dispatch chain) produce identical
results to the C reference, across multiple buffer sizes with both
aligned and unaligned buffers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Building with clang-16 complains with:
./simd-checksum-x86_64.cpp:204:25: warning: passing 1-byte aligned argument to
16-byte aligned parameter 1 of '_mm_store_si128' may result in an unaligned pointer
access [-Walign-mismatch]
Signed-off-by: Holger Hoffstätte <holger@applied-asynchrony.com>
- Make the SIMD ASM code off by default. Use configure --enable-simd-asm
to enable.
- Allow MD5 ASM code to be requested even when OpenSSL is handling MD4
checksums. Use configure --enable-md5-asm to enable.
* x86-64 SIMD build fixes
configure.ac was modified to detect g++ >=5 and clang++ >=7. Additionally
some script malfunctions on FreeBSD were corrected.
The get_checksum1() code has been modified to fix clang and g++ 10
compilation.
This version of the code and configure.ac has been tested on:
Ubuntu 16 - gcc 7.3.0, clang 6.0.0
Debian 10 - gcc 5.4.0, 6.4.0, 7.2.0, 8.4.0, 9.2.1, 10.0.1, clang 5.0.2,
6.0.1, 7.0.1, 8.0.0, 9.0.0, 10.0.0
ArchLinux 20200605 - gcc 10.1.0, clang 10.0.0
FreeBSD 12.1 - gcc 9.3.0, clang 8.0.1
It is unknown if it will work on gcc 5.0-5.3, but the script currently
allows it.
Additionally restructures build switches and defines from SSE2 to SIMD,
to allow potential reuse should patches become available with SIMD
instructions for other processor architectures.
(Some minor tweaks of Jorrit's patch to avoid requiring GNU make and to
avoid C++ comments in .c files.)
Andrew Tridgell