Hi,

building with the -mno-ssse3 causes build failure, e.g.:
CFLAGS="-O2 -march=bonnell -mtune=bonnell -mno-ssse3"
...
gcc  -I. -I./lib  -Ilib -I./lib -Isrc -I./src  -mavx -mpclmul   -O2
-march=bonnell -mtune=bonnell -mno-ssse3 -c -o
src/libcksum_pclmul_a-cksum_pclmul.o `test -f 'src/cksum_pclmul.c' ||
echo './'`src/cksum_pclmul.c
In file included from /usr/lib/gcc/i686-pc-linux-gnu/12/include/immintrin.h:37,
                 from /usr/lib/gcc/i686-pc-linux-gnu/12/include/x86intrin.h:32,
                 from src/cksum_pclmul.c:22:
/usr/lib/gcc/i686-pc-linux-gnu/12/include/tmmintrin.h: In function
'cksum_pclmul':
/usr/lib/gcc/i686-pc-linux-gnu/12/include/tmmintrin.h:136:1: error:
inlining failed in call to 'always_inline' '_mm_shuffle_epi8': target
specific option mismatch
  136 | _mm_shuffle_epi8 (__m128i __X, __m128i __Y)
      | ^~~~~~~~~~~~~~~~
src/cksum_pclmul.c:103:19: note: called from here
  103 |           data7 = _mm_shuffle_epi8 (data7, shuffle_constant);
      |                   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/lib/gcc/i686-pc-linux-gnu/12/include/tmmintrin.h:136:1: error:
inlining failed in call to 'always_inline' '_mm_shuffle_epi8': target
specific option mismatch
The problem is in the pclmul optimization. The configure.ac autodetect
script checks for the pclmul, but not for the ssse3 instructions which
are later utilized in the pclmul optimized code. Unfortunately, it
also seems there is no way to override the configure.ac check result
(e.g. by the ./configure --without-pclmul option). The attached patch
added the ssse3 detection and fixed the problem for me. Maybe it would
be better to split it into two checks - pclmul and ssse3, but I didn't
want to complicate it. The patch was initially created for
coreutils-9.1 where the problem initially manifested, but it also
applies to coreutils-9.3 which is also affected

thanks & regards

Jaroslav
--- ./configure.ac.old	2023-04-26 00:42:37.840034913 +0200
+++ ./configure.ac	2023-04-26 00:43:13.828730731 +0200
@@ -577,6 +577,7 @@
     {
       __m128i a, b;
       a = _mm_clmulepi64_si128 (a, b, 0x00);
+      a = _mm_shuffle_epi8 (a, b);
       return 1;
     }
   ]])

Reply via email to