Re: [PATCH v2 1/1] gdb: Enable OS generated corefiles on systems with Intel AMX support.

[Andrew Burgess <aburgess@redhat.com>]

Christina Joos <christina.joos@intel.com> writes:

> This patch addresses the issue described in PR gdb/34561.  On systems
> with Intel AMX support the xsave size is 11008.  This xsave size is not
> handled by gdb/i387-tdep.c:i387_guess_xsave_layout and is causing
> problems for corefiles generated by the linux kernel.  The problem is
> visible regardless of the AMX enablement state in the dumping process,
> so all corefiles on newer systems with AMX are broken.
>
> ~~~
> $ gdb
> GNU gdb (GDB) 19.0.50.20260814-git
> Copyright (C) 2026 Free Software Foundation, Inc.
> [...]
> (gdb) core core_main_SEGV
> [New LWP 2162880 (id 1)]
> [...]
> Core was generated by `./main'.
> Program terminated with signal SIGSEGV, Segmentation fault.
> 4         *pointer = 3;
> (gdb) p $ymm0
> $1 = void
> ~~~
>
> We should be able to print the register $ymm0 (or any other register
> belonging to a feature higher than SSE).
>
> For a live debug session we can print it:
> ~~~
> Reading symbols from main...
> (gdb) start
> Temporary breakpoint 1 at 0x1131: file main.c, line 3.
> Starting program: /tmp/main
> [Thread debugging using libthread_db enabled]
> Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
>
> Temporary breakpoint 1, main () at main.c:3
> 3         int *pointer = 0;
> (gdb) p $ymm0
> $1 = {v16_bfloat16 = {0, 2.342e-38, 0 <repeats 14 times>}, v16_half = {0, 1.5199e-05, 0 <repeats 14 times>}, v8_float = {
>     2.34180515e-38, 0, 0, 0, 0, 0, 0, 0}, v4_double = {8.256666972292243e-317, 0, 0, 0}, v32_int8 = {0, 0, -1,
>     0 <repeats 29 times>}, v16_int16 = {0, 255, 0 <repeats 14 times>}, v8_int32 = {16711680, 0, 0, 0, 0, 0, 0, 0}, v4_int64 = {
>     16711680, 0, 0, 0}, v2_int128 = {16711680, 0}}
> [...]
> ~~~
>
> Note that this is not reproducible for corefiles generated by the gcore
> command, since in such corefiles the GDB target description is
> available and gdb/i387-tdep.c:i387_fallback_xsave_layout configures
> the xsave_layout based on xcr0 derived from the target description.
>
> This patch fixes this issue by handling the new xsave size in
> i387_guess_xsave_layout.  It is necessary even though GDB does not
> support AMX yet, since we still pass the full xsave_size (11008)
> extracted from the corefile to i387_guess_xsave_layout.
> The patch does not fix i387_fallback_xsave_layout to configure the new
> xsave size.  Configuring sizeof_xsave to 2696 is fine at this point,
> since truncating to 2696 drops only the parts GDB doesn't model yet.
>
> As a quick solution for the upcoming release and CPUs supporting
> Intel AMX this patch should be sufficient.
>
> Testing:
>
> I noticed that we don't have tests for corefiles for the features AVX,
> AVX512 and PKRU in the GDB testsuite.  So this commit adds corefile tests
> for AVX, AVX512 and PKRU registers.
>
> As a result of this patch, the warning:
>
>   warning: Unexpected size of section `.reg-xstate/1373786' in core file.
>
> which was causing failure in gdb.base/coredump-filter.exp and
> gdb.arch/i386-tls-regs.exp on Intel AMX systems disappears, these tests now
> pass.
>
> Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=34561

Thanks for your work on this patch.

I added a NEWS file entry for this as it felt like something that was
NEWS worthy, especially on the gdb-18-branch.

I also changed "corefile" to "core file" in the commit message and
comments.

Then I pushed this to master and gdb-18-branch.

Thanks,
Andrew