wc: add AVX512 function for line counting

* configure.ac: Add detection of AVX512 intrinsics for wc.
* src/local.mk: Build AVX512 wc libraries.
* src/wc.c: Add runtime detection of AVX512 intrinsics and call
appropriate function when detected.
* src/wc.h (wc_lines_avx512): Declare function.
* tests/wc/wc-cpu.sh: Add a test that disables AVX512 intrinsics.
* src/wc_avx512.c: New file containing the wc -l implementation using
AVX512. The logic and code is reused from the AVX2 implementation with
slight adaptations. Replaced __builtin_popcount by __builtin_popcountll
and the combination of _mm256_cmpeq_epi8 and _mm256_movemask_epi8 by a
single call to _mm512_cmpeq_epi8_mask.
* NEWS: Mention the improvement.

1 files changed, 10 insertions, 2 deletions

diff --git a/tests/wc/wc-cpu.sh b/tests/wc/wc-cpu.sh
index 1118fe14e..6ad4f5f9c 100755
--- a/tests/wc/wc-cpu.sh
+++ b/tests/wc/wc-cpu.sh
@@ -19,7 +19,7 @@
 . "${srcdir=.}/tests/init.sh"; path_prepend_ ./src
 print_ver_ wc
 
-GLIBC_TUNABLES='glibc.cpu.hwcaps=-AVX2' \
+GLIBC_TUNABLES='glibc.cpu.hwcaps=-AVX2,-AVX512F' \
  wc -l --debug /dev/null 2>debug || fail=1
 grep 'using.*hardware support' debug && fail=1
 
@@ -27,8 +27,16 @@ lines=$(shuf -i 0-1000 | head -n1)  || framework_failure_
 seq 1000 | head -n "$lines" > lines || framework_failure_
 
 wc_accelerated=$(wc -l < lines) || fail=1
-wc_base=$(GLIBC_TUNABLES='glibc.cpu.hwcaps=-AVX2' wc -l < lines) || fail=1
+wc_accelerated_no_avx512=$(
+          GLIBC_TUNABLES='glibc.cpu.hwcaps=-AVX512F' \
+          wc -l < lines
+         ) || fail=1
+wc_base=$(
+          GLIBC_TUNABLES='glibc.cpu.hwcaps=-AVX2,-AVX512F' \
+          wc -l < lines
+         ) || fail=1
 
 test "$wc_accelerated" = "$wc_base" || fail=1
+test "$wc_accelerated_no_avx512" = "$wc_base" || fail=1
 
 Exit $fail