7.6.0: ugrep installation tests fail on AMD EPYC 9575F with AVX512BW

[vt-alt]

We have this test failing for i586 build on AMD build node (but it does not fail locally on Intel CPU nor does it fail for aarch64 and x86_64 builds):

+ make -j128 test
make: Entering directory '/usr/src/RPM/BUILD/ugrep-7.6.0'

*** SINGLE-THREADED TESTS ***

ugrep 7.6.0 i586-alt-linux-gnu +avx512; -P:pcre2jit; -z:zlib,bzip2,lzma,lz4,zstd,brotli,bzip3,7z,tar/pax/cpio/zip
Have libpcre2?  yes (recommended)
Have libz?       no (recommended)
Have libbz2?     no (recommended)
Have liblzma?   yes (recommended)
Have liblz4?    yes (optional)
Have libzstd?    no (optional)
Have libbrotli?  no (optional)
Have 7zip?      yes (optional)
Have libbzip3?  yes (requested)
...........................................--- -     2026-03-08 23:22:30.507010132 +0000
+++ out/Hello_Hello--max-count.out   2026-03-08 23:22:02.000000000 +0000
@@ -2,3 +2,4 @@
 Binary file Hello.class matches
-Hello.java:  { System.out.println("Hello World!");
+Hello.java:public class Hello // prints a Hello World! greeting
+Binary file Hello.pdf matches
 Hello.sh:echo "Hello World!"
Error: ugrep --sort --max-count=1 Hello Hello.bat Hello.class Hello.java Hello.pdf Hello.sh Hello.txt empty.txt emptyline.txt failed
make: *** [Makefile:1104: test] Error 1
make: Leaving directory '/usr/src/RPM/BUILD/ugrep-7.6.0'
error: Bad exit status from /usr/src/tmp/rpm-tmp.93736 (%check)

Temporary build log
https://git.altlinux.org/tasks/410525/build/100/i586/log

[vt-alt]

Another difference, on the AMD build node it says:

[00:00:11] ugrep 7.6.0 i586-alt-linux-gnu +avx512; -P:pcre2jit; -z:zlib,bzip2,lzma,lz4,zstd,brotli,bzip3,7z,tar/pax/cpio/zip
[00:00:11] Have libpcre2?  yes (recommended)
[00:00:11] Have libz?       no (recommended)
[00:00:11] Have libbz2?     no (recommended)
[00:00:11] Have liblzma?   yes (recommended)
[00:00:11] Have liblz4?    yes (optional)
[00:00:11] Have libzstd?    no (optional)
[00:00:11] Have libbrotli?  no (optional)
[00:00:11] Have 7zip?      yes (optional)
[00:00:11] Have libbzip3?  yes (requested)

And locally (on Intel) all of these are yes.

ugrep 7.6.0 i586-alt-linux-gnu +avx2; -P:pcre2jit; -z:zlib,bzip2,lzma,lz4,zstd,brotli,bzip3,7z,tar/pax/cpio/zip
Have libpcre2?  yes (recommended)
Have libz?      yes (recommended)
Have libbz2?    yes (recommended)
Have liblzma?   yes (recommended)
Have liblz4?    yes (optional)
Have libzstd?   yes (optional)
Have libbrotli? yes (optional)
Have 7zip?      yes (optional)
Have libbzip3?  yes (requested)

[genivia-inc]

Not yet sure what's causing this to fail specifically for that system config. Hope to update you on this issue as soon as I have more info to share.

[genivia-inc]

@vt-alt We have this test failing for i586 build on AMD build node...

Which Linux OS 32 bit? Clang or GCC?

[vt-alt]

We compile on Linux (docker tag alt:sisyphus) using GCC (14.3.1), CPU is AMD EPYC 9575F 64-Core Processor.

[vt-alt]

Tried on the older system with GCC 13.2.1 - the problem is not reproduced. (The host CPU is should be the same, 99% but can't ensure 100% because they are in the build cluster).

[vt-alt]

I tried GCC 13.2.1 on the new system, and it fails. So, this is perhaps not GCC version related.

[genivia-inc]

Thanks for sharing the info.

I'm trying to figure out how to best attack this issue. I'm not even sure if the problem is specifically caused by option --max-count on that specific platform, or if other options are in play. To narrow this down to the option, please change tests/verify.sh to remove the exit 1 at line 122 so all tests are performed even after an error:

function ERR() {
  echo "�[1;31mError:�[0m�[1m ugrep --sort $1 �[1;31mfailed�[0m"
  # exit 1 # temporarily removed
}

So that make test will continue after the error is reported and perhaps dump more errors. If so, that helps.

Also, could you share the config.log after building ugrep for that machine with ./build.sh?

To check if this is related to SIMD optimizations, please try this:

./build.sh --disable-avx2

Then also try:

./build.sh --disable-sse2

[vt-alt]

Temporary log with tests/verify.sh modification, cat -n config.log after build.sh https://git.altlinux.org/tasks/414827/build/100/i586/log

Temporary log as above and --disable-avx2 https://git.altlinux.org/tasks/414828/build/100/i586/log

And with --disable-sse2 https://git.altlinux.org/tasks/414830/build/100/i586/log

[genivia-inc]

Thanks for sharing the logs. From this it appears something in the pattern matcher is not working correctly when AVX512BW is enabled to build ugrep. I'm very sure the AVX2-optimized code is fine, but AVX512BW-optimized code likely has a pattern matching bug somewhere that shows up in multiple test runs with verify.sh. Perhaps the problem may depend on the compiler and CPU configuration, definitely something I need to work on.

[genivia-inc]

By the way, when ugrep is built on a machine that supports AVX2 but not AVX512BW, then the generated binary will safely run on a CPU with AVX512BW instructions supported, because the AVX512BW-optimizations are not compiled when built on the AVX2 capable machine. In addition, that binary also runs on scaled-down CPUs that don't support AVX2. It seemed a reasonable middle ground to me.

[vt-alt]

When ugrep built for distributions the build box may be significantly different from end-users boxes. So, it would be nice to have an option to build maximum range of optimizations even on the box without the required hardware capability.

[genivia-inc]

This problem is still not resolved on my end, because I cannot replicate the issue on a AWS EC2 T3 instance that supports AVX-512:

*** SINGLE-THREADED TESTS ***

ugrep 7.6.0 x86_64-pc-linux-gnu +avx512
Have libpcre2?   no (recommended)
Have libboost_regex? no (optional)
Have libz?       no (recommended)
Have libbz2?     no (recommended)
Have liblzma?    no (recommended)
Have liblz4?     no (optional)
Have libzstd?    no (optional)
Have libbrotli?  no (optional)
Have 7zip?      yes (optional)
...............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
ALL TESTS PASSED

The not installed pcre2 and other libs shown above should have no relationship to this issue, where almost all pattern matches fail in your case regardless of these libraries.

So how do we move forward on this? AWS EC2 T3 (link) uses an Intel Xeon Platinum 8000 series processor (Skylake-SP or Cascade Lake). I have no access to an AMD EPYC 9575F, where the build tests fail, to analyze and debug the problem.

Now, i586 is 32 bit so there could be a cast issue when using certain C++ types. However, I'm using stdint types when applicable. It's almost as if this is a HW failure, not a SW one, but that's silly...