trofi

trofi, 11 days ago to random

Sometimes it takes more time to refute a gcc bug than confirm it. This week I spent most time on https://gcc.gnu.org/PR114872 where sagemath SIGSEGVed on some simple inputs.

Bug updates are a bit hard to read and are missing a bit of compiler-unrelated context. I wrote something more coherent in https://trofi.github.io/posts/312-the-sagemath-saga.html

#sagemath #bug

trofi, 6 months ago to random

Today's duperemove bug is a minor accounting bug: https://github.com/markfasheh/duperemove/pull/323

$ ls -lh /nix/var/nix/db/db.sqlite
1.4G /nix/var/nix/db/db.sqlite

Before the change:

$ ./show-shared-extents /nix/var/nix/db/db.sqlite
/nix/var/nix/db/db.sqlite: 27065321263104 shared bytes

After the change:

$ ./show-shared-extents /nix/var/nix/db/db.sqlite
/nix/var/nix/db/db.sqlite: 1169276928 shared bytes

The size reduction is not as impressive as initially reported :)

#duperemove #bug

trofi, 7 months ago to random

Today's gcc bug is a gcc compilation performance problem:
https://gcc.gnu.org/PR111619

Final gcc is compiled with -O2 by default (intended) and booter gcc is compiled with -O0 (also intended).

But unoptimized gcc is very inefficient at handling huge generated files like insn-recog.cc.

#gcc #bug

trofi, 9 months ago to random

Today's gcc bug https://gcc.gnu.org/PR111051 is a case where gcc stopped recognizing some avx2 primitives on avx512 code (all on highway-1.0.6 code again):

#include <immintrin.h>

#pragma GCC target("avx512vl,avx512dq")

void bug(__m256i i) {
volatile auto v1 = _mm256_cvtepi64_pd(i);
}

and failed as error: inlining failed in call to 'always_inline' '__m256d _mm256_setzero_pd()': target specific option mismatch

#gcc #bug

trofi, 9 months ago to random

Today's gcc bug is https://gcc.gnu.org/PR111048 where gcc generated wrong code for highway-1.0.6 library in -mavx2 mode.

When handling the following loop:

u8 in_lanes[32];
for (unsigned i = 0; i < 32; i += 2) {
in_lanes[i + 0] = 0;
in_lanes[i + 1] = ((u8)0xff) >> (i & 7);
}

gcc broke i =12 iteration and instead of i[13] = 0xf; (0xff >> 4) stored something like i[13] = 0xef; there.

#gcc #bug

trofi, 11 months ago to random

Today's gcc bug is https://gcc.gnu.org/PR110271 where gcc itself crashed with ICE on pycryptodome package.

While the crash is not very interesting the change that caused the regression is very cool: https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=43a3252c42af12

gcc is now able to detect add-with-carry patterns and generate branchless code for them. On x86 these are sub -> sbb.

Bug tracker says it's 100th bug I reported to gcc!

#gcc #bug

trofi, 1 year ago to random

Another year, another ::nix-guix Gentoo overlay status update: https://trofi.github.io/posts/287-nix-and-guix-for-gentoo-in-2023.html

Tl;DR:

• the overlay installation procedure changed slightly
• the overlay still packages latest sys-apps/nix and sys-apps/guix packages ready for use
• a few workarounds are added to make nix and guix installable/usable on clang-based profiles.

trofi, 1 year ago to random

Every 6 months nixpkgs enters "Zero Hydra Failures" development phase when the main focus is to fix most known build or test failures.

It's a good opportunity to get your first contribution as fixing the package is usually straightforward: it used to work and your goal is to make it build again (or schedule it for deletion).

I usually try to fix at least one package during ZHF.

Here is an example for this cycle: https://trofi.github.io/posts/285-Zero-Hydra-Failures-towards-23.05-NixOS-release.html

#nixpkgs

trofi, 1 year ago to random

gcc-14 development has started. A few cleanup were already merged in. This week's bugs are:

• https://gcc.gnu.org/PR109643: ICE on pkg-config
• https://gcc.gnu.org/PR109652: ICE on valgrind

I'm also struggling to extract runtime failures on nlohmann_json and aws-sdk-cpp test suites that seem to be gcc-14 regressions.

#gcc #bug