Re: gcc 14 and netbsd/alpha

Hi,

On 2026/08/03 13:04, matthew green wrote:
> hi folks.
> 
> 
> in -current, alpha and m68k are the only port still using gcc 12.
> gcc 14 works fine for m68k and we will switch it sometime soon,
> but there are at least two issues in alpha that need someone to
> have a look at them.
> 
> from src/external/gpl3/gcc/README.gcc14:
> 
> [11]: alpha needs an -O0 for mopd's file.c to avoid an ICE.
> [14]: cc1 crashes by jumping into rodata segment when compiling pkgsrc/pkgtools/libnbcompat:
>        In file included from glob.c:12:
>        __glob13.c: In function 'globextend':
>        __glob13.c:800:1: internal compiler error: Illegal instruction
>          800 | globextend(path, pglob, limit)
>              | ^~~~~~~~~~
> [15]: most(?) nfs-related tests timedout. ipf tests fail with incorrect results.
>        some pic_profile variants of tests in usr.bin/c++ crash.
> 
> [11] is handled but could be looked at, but [14] and [15] are
> serious problems (without workarounds?) yet.
> 
> 
> any takers?  thanks.

I'm trying, but no luck so far :(

With recent -current and pkgsrc-2026Q2, [14] now SEGV at
gimple_simplify_CFN_BUILT_IN_NEARBYINTF32X().

Full backtrace and preprocessed source code are provided here:

https://gist.github.com/rokuyama/6c601a132120fcf3e5ef300590ea697c
https://gist.github.com/rokuyama/e91601a2a6807b0de9fb1711f67de2c5

First, I built 6 source files of GCC that appears in the
backtrace with -O0 (including main.cc!). Then, this file
could be compiled, but another ICE happened when compiling
another file.

Since there is no alpha-related changes in releases/gcc-14
branch after 14.3.0 release, I cherry-picked some random
fixes from master branch. But, it did not work so far.

I wonder what happens if in-tree gcc is upgraded to 14.4.0;
isn't this failures due to some MI bugs in 14.3.0?

Thanks,
rin

> 
> 
> .mrg.
>