RE: [PATCH 1/2] gdb: Enable OS generated corefiles on systems with Intel AMX support.

["Joos, Christina" <christina.joos@intel.com>]

Hi Andrew,

Thanks a lot for your quick feedback. Please find my comments on your feedback below.
For the test and the PKRU register issue, I still have some opens on which I would appreciate your feedback before I send a v2.

> -----Original Message-----
> From: Andrew Burgess <aburgess@redhat.com>
> Sent: Montag, 7. September 2026 12:06
> To: Joos, Christina <christina.joos@intel.com>; gdb-patches@sourceware.org
> Subject: Re: [PATCH 1/2] gdb: Enable OS generated corefiles on systems with
> Intel AMX support.
> 
> 
> Thanks for addressing this issue.  I have some style issues, and one deeper
> though about the test, see below.
> 
> Christina Joos <christina.joos@intel.com> writes:
> 
> > This patch addresses the issue described in
> > https://sourceware.org/bugzilla/show_bug.cgi?id=34561.
> 
> Bugs should be linked using a line at the end of the commit message (see
> gdb/MAINTAINERS) like:
> 
>   Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=34561
> 
> Given the link always gets added at the end like this, in the commit message
> body you can just write:
> 
>   This patch addresses the issue described in PR gdb/34561.  On systems
>   with Intel AMX support the .... etc ...

Ah I see, I was not aware of this guideline. Thanks!

> anyone who wants a direct link to the bug can find it in the 'Bug: '
> line at the end of the commit message.
> 
> >
> > On Systems with Intel AMX support the xsave size is 11008. This xsave
> 
> lower case 's' on 'Systems' please.

Yes, will fix.

> > size is not handled by gdb/i387-tdep.c:i387_guess_xsave_layout and is
> > causing problems for corefiles generated by the linux kernel.
> >
> > ~~~
> > $ 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
> 
> Drop the comma after 'Note'.
> 
> > 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.
> >
> > It further fixes the tests gdb.base/coredump-filter.exp and
> > gdb.arch/i386-tls-regs.exp on systems with Intel AMX support.
> > With my patch the warning
> > "warning: Unexpected size of section `.reg-xstate/1373786' in core file.^M"
> > disappears and the tests are passing again.
> 
> Can you clean up the ^M artefact please.
> 
> Also, as you're touching the commit message anyway, I think this would be
> better written as:
> 
>   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 AMX systems disappears, these tests now
>   pass.

Yes, will fix.

> 
> 
> > ---
> >  gdb/i387-tdep.c                        |  9 +++-
> >  gdb/testsuite/gdb.arch/i386-avx.c      |  6 +++
> >  gdb/testsuite/gdb.arch/i386-avx.exp    | 61 ++++++++++++++++++++++----
> >  gdb/testsuite/gdb.arch/i386-avx512.c   |  6 +++
> >  gdb/testsuite/gdb.arch/i386-avx512.exp | 56 +++++++++++++++++++++++
> >  gdb/testsuite/gdb.arch/i386-pkru.c     |  6 +++
> >  gdb/testsuite/gdb.arch/i386-pkru.exp   | 39 ++++++++++++++++
> >  7 files changed, 173 insertions(+), 10 deletions(-)
> >
> > diff --git a/gdb/i387-tdep.c b/gdb/i387-tdep.c index
> > f22a5e02bbb..19429536537 100644
> > --- a/gdb/i387-tdep.c
> > +++ b/gdb/i387-tdep.c
> > @@ -918,9 +918,14 @@ bool
> >  i387_guess_xsave_layout (uint64_t xcr0, size_t xsave_size,
> >  			 x86_xsave_layout &layout)
> >  {
> > -  if (HAS_PKRU (xcr0) && xsave_size == 2696)
> > +  if (HAS_PKRU (xcr0) && (xsave_size == 2696 || xsave_size == 11008))
> >      {
> > -      /* Intel CPUs supporting PKRU.  */
> > +      /* Intel CPUs supporting PKRU.
> > +	 Note that in this case 2 possible xsave_sizes have to be handled,
> 
> The 'Note' here is not really needed.  The comment is by definition a note.  Also,
> variables should be capitalised, or avoid using the variable name maybe, like:
>
>   In this case two possible XSAVE_SIZE values have to be handled.  As
>   GDB does not yet support Intel AMX and CPUs that support this feature
>   will have a different XSAVE_SIZE (11008).  If we don't handle this
>   size then we will not be able to handle any xsave registers from OS
>   generated core files on AMX systems.

Yes, your version sounds much better, thanks for this suggestion.

> > +	 since GDB does not support Intel AMX yet and CPUs that support
> > +	 this feature still have a different xsave size (xsave_size == 11008).
> > +	 Otherwise, we are not able to read xsave registers in OS generated
> > +	 corefiles.  */
> >        layout.avx_offset = 576;
> >        layout.k_offset = 1088;
> >        layout.zmm_h_offset = 1152;
> > diff --git a/gdb/testsuite/gdb.arch/i386-avx.c
> > b/gdb/testsuite/gdb.arch/i386-avx.c
> > index b6cd89e5bea..7edcabcec8c 100644
> > --- a/gdb/testsuite/gdb.arch/i386-avx.c
> > +++ b/gdb/testsuite/gdb.arch/i386-avx.c
> > @@ -18,11 +18,14 @@
> >     along with this program.  If not, see
> > <http://www.gnu.org/licenses/>.  */
> >
> >  #include <stdio.h>
> > +#include <stdlib.h>
> >  #include "nat/x86-cpuid.h"
> >
> >  /* Align sufficient to be able to use vmovaps.  */  #define ALIGN 32
> >
> > +volatile int should_dump_core_p = 1;
> > +
> >  typedef struct {
> >    _Alignas (ALIGN) float f[8];
> >  } v8sf_t;
> > @@ -87,6 +90,9 @@ main (int argc, char **argv)
> >
> >    asm ("nop"); /* first breakpoint here */
> >
> > +  if (should_dump_core_p) /* Crash for OS corefile. */
> > +    abort ();
> 
> We should still follow GDB/GNU style as much as possible in tests, which
> means avoiding trailing comments.  Please move the comment above the `if`
> line.  The exception in tests is when we use a comment to tag a line on which
> we wish to place a breakpoint, e.g. the 'first breakpoint here' line above.

Yes, will fix.

> I also wonder if using `abort()` here is the right approach.
> 
> You want to check the register values in the generated core file, not just C level
> variables.  Making a function call runs the risk that the callee might touch some
> or all of these registers.  I guess it's unlikely for something as simple as abort(),
> but I wonder maybe, if that's what's causing the problem your second patch is
> addressing?
> 
> Maybe it would be better to trigger the core file by reading from address zero?
> We have a testsuite proc `is_address_zero_readable` which can be used to gate
> such a test.  I think it might be worth testing this on the problematic machines
> you mention in the second patch to see if this fixes the issue.

I've tested it and it does not fix the problem, but I still think your comment makes a
lot of sense when reading registers. I'd generate the corefile as follows:

+      /* Crash for OS corefile. */
+      if (should_dump_core_p)
+       {
+         /* Generate SIGSEGV to crash.  */
+         *(volatile int *) 0;
+       }

Does that look better?

I am still trying to clarify the actual reason, but I cannot promise I will make this until Friday.
Given that this issue is unrelated to the original issue, I'd hope that this is not a blocker for the upcoming release.

At the same time, I wonder if it would be better to simply let the test fail, instead of having a second patch.
If this issue is fixed at some point, we won't have to touch the test again.
OTOH I don't like introducing new fails in the GDB testsuite.

> 
> > +
> >    asm (
> >         "vmovaps %%ymm0, 0(%0)\n\t"
> >         "vmovaps %%ymm1, 32(%0)\n\t"
> > diff --git a/gdb/testsuite/gdb.arch/i386-avx.exp
> > b/gdb/testsuite/gdb.arch/i386-avx.exp
> > index 1a786ee8a51..1740045427a 100644
> > --- a/gdb/testsuite/gdb.arch/i386-avx.exp
> > +++ b/gdb/testsuite/gdb.arch/i386-avx.exp
> > @@ -42,12 +42,26 @@ if { [gdb_compile "${srcdir}/${subdir}/${srcfile}"
> "${binfile}" executable \
> >      return
> >  }
> >
> > +# Check reading registers after the first breakpoint for ymm int8 and
> > +# float types.
> > +proc test_regs_first_bp {} {
> > +    for { set r 0 } { $r < $::nr_regs } { incr r } {
> > +	gdb_test "print \$ymm$r.v8_float" \
> > +	    ".. = \\{$r, $r.125, $r.25, $r.375, $r.5, $r.625, $r.75, $r.875\\}.*" \
> > +	    "check float contents of %ymm$r"
> > +	gdb_test "print \$ymm$r.v32_int8" \
> > +	    ".. = \\{(-?${::decimal}, ){31}-?${::decimal}\\}.*" \
> > +	    "check int8 contents of %ymm$r"
> > +    }
> > +}
> >  clean_restart ${::testfile}
> 
> Add a blank line before the clean_restart line please.

Yes, will fix.

> >
> >  if {![runto_main]} {
> >      return
> >  }
> >
> > +gdb_test_no_output "set should_dump_core_p=0"
> > +
> >  gdb_test "break [gdb_get_line_number "first breakpoint here"]" \
> >  	 "Breakpoint .* at .*i386-avx.c.*" \
> >  	 "set first breakpoint in main"
> > @@ -59,14 +73,7 @@ if {[is_amd64_regs_target]} {
> >      set nr_regs 8
> >  }
> >
> > -for { set r 0 } { $r < $nr_regs } { incr r } {
> > -    gdb_test "print \$ymm$r.v8_float" \
> > -	".. = \\{$r, $r.125, $r.25, $r.375, $r.5, $r.625, $r.75, $r.875\\}.*" \
> > -	"check float contents of %ymm$r"
> > -    gdb_test "print \$ymm$r.v32_int8" \
> > -	".. = \\{(-?${::decimal}, ){31}-?${::decimal}\\}.*" \
> > -	"check int8 contents of %ymm$r"
> > -}
> > +test_regs_first_bp
> >
> >  for { set r 0 } { $r < $nr_regs } { incr r } {
> >      gdb_test "set var \$ymm$r.v8_float\[0\] = $r + 10" "" "set %ymm$r"
> > @@ -100,9 +107,47 @@ with_test_prefix "force-disable xml descriptions" {
> >  	    return
> >  	}
> >
> > +	gdb_test_no_output "set should_dump_core_p=0"
> > +
> >  	# With qXfer:features:read disabled, we won't know anything
> >  	# about YMM registers.
> >  	gdb_test "print \$ymm0" " = void"
> >  	gdb_test "print \$xmm0" "v4_float.*"
> >      }
> >  }
> > +
> > +# Restart gdb, load the corefile generated by gcore or the OS # and
> > +check reading registers from the corefile.
> > +proc test_corefile {core_filename} {
> > +    clean_restart $::testfile
> > +
> > +    gdb_test "core $core_filename" "Core was generated by .*" \
> > +	"load corefile"
> > +
> > +    test_regs_first_bp
> > +}
> > +
> > +with_test_prefix "OS generated corefile" {
> > +    set corefile [core_find $binfile]
> > +    if { $corefile eq "" } {
> > +	unsupported "unable to generate core file"
> > +    } else {
> > +	test_corefile $corefile
> > +    }
> > +}
> > +
> > +with_test_prefix "gcore corefile" {
> > +    clean_restart ${::testfile}
> > +    if { ![runto_main] } {
> > +	return
> > +    }
> > +
> > +    set line [gdb_get_line_number "first breakpoint here"]
> > +    gdb_breakpoint $line
> > +    gdb_continue_to_breakpoint "first breakpoint here" ".*$srcfile:$line.*"
> > +
> > +    set gcorefile $binfile.gcore
> > +    if { [gdb_gcore_cmd $gcorefile "save a corefile"] } {
> > +	test_corefile $gcorefile
> > +    }
> > +}
> > diff --git a/gdb/testsuite/gdb.arch/i386-avx512.c
> > b/gdb/testsuite/gdb.arch/i386-avx512.c
> > index b1e023ce4d4..d935fab50bc 100644
> > --- a/gdb/testsuite/gdb.arch/i386-avx512.c
> > +++ b/gdb/testsuite/gdb.arch/i386-avx512.c
> > @@ -17,8 +17,11 @@
> >     You should have received a copy of the GNU General Public License
> >     along with this program.  If not, see
> > <http://www.gnu.org/licenses/>.  */
> >
> > +#include <stdlib.h>
> >  #include "x86-cpuid.h"
> >
> > +volatile int should_dump_core_p = 1;
> > +
> >  typedef struct
> >  {
> >    double f[8];
> > @@ -233,6 +236,9 @@ main (int argc, char **argv)
> >        move_zmm_data_to_reg ();
> >        asm ("nop"); /* third breakpoint here  */
> >
> > +      if (should_dump_core_p) /* Crash for OS corefile. */
> > +	abort ();
> 
> Trailing comment again.

Will fix.

> > +
> >        /* Test script incremented values,
> >  	 move back to array and check values.  */
> >        move_zmm_data_to_memory ();
> > diff --git a/gdb/testsuite/gdb.arch/i386-avx512.exp
> > b/gdb/testsuite/gdb.arch/i386-avx512.exp
> > index c79563366dc..6984c2280e9 100644
> > --- a/gdb/testsuite/gdb.arch/i386-avx512.exp
> > +++ b/gdb/testsuite/gdb.arch/i386-avx512.exp
> > @@ -35,6 +35,8 @@ if {![runto_main]} {
> >      return
> >  }
> >
> > +gdb_test_no_output "set should_dump_core_p=0"
> > +
> >  set supports_avx512 0
> >
> >  set test "probe AVX512 support"
> > @@ -181,3 +183,57 @@ gdb_test "print \$zmm0.v16_int32" "= {-1, -1, -1, -
> 1, 0 <repeats 12 times>}"
> >  if { $nr_regs >= 16 } {
> >      gdb_test "print \$zmm16.v16_int32" "= {-1 <repeats 16 times>}"
> >  }
> > +
> > +# Test reading k and zmm registers and for double and int16 types.
> > +proc test_regs_corefile {} {
> > +    for { set r 1 } { $r < 8 } { incr r } {
> > +	gdb_test "print/x \$k$r" \
> > +	    ".. = 0x[format %x $r]2[format %x $r]1" \
> > +	    "check contents of %k$r"
> > +    }
> > +
> > +    for { set r 0 } { $r < $::nr_regs } { incr r } {
> > +	gdb_test "print \$zmm$r.v8_double" \
> > +	    ".. = \\{$r, $r.125, $r.25, $r.375, $r.5, $r.625, $r.75, $r.875\\}.*" \
> > +	    "check double contents of %zmm$r"
> > +	gdb_test "print \$zmm$r.v32_int16" \
> > +	    ".. = \\{(-?${::decimal}, ){31}-?${::decimal}\\}.*" \
> > +	    "check int16 contents of %zmm$r"
> > +    }
> > +}
> > +
> > +# Restart gdb, load the corefile generated by gcore or the OS # and
> > +check reading registers from the corefile.
> > +proc test_corefile {core_filename} {
> > +    clean_restart $::testfile
> > +
> > +    gdb_test "core $core_filename" "Core was generated by .*" \
> > +	"load corefile"
> > +
> > +    test_regs_corefile
> > +}
> > +
> > +with_test_prefix "OS generated corefile" {
> > +    set corefile [core_find $binfile]
> > +    if { $corefile eq "" } {
> > +	unsupported "unable to generate core file"
> > +    } else {
> > +	test_corefile $corefile
> > +    }
> > +}
> > +
> > +with_test_prefix "gcore corefile" {
> > +    clean_restart ${::testfile}
> > +    if { ![runto_main] } {
> > +	return
> > +    }
> > +
> > +    set line [gdb_get_line_number "third breakpoint here"]
> > +    gdb_breakpoint $line
> > +    gdb_continue_to_breakpoint "third breakpoint here" ".*$srcfile:$line.*"
> > +
> > +    set gcorefile $binfile.gcore
> > +    if { [gdb_gcore_cmd $gcorefile "save a corefile"] } {
> > +	test_corefile $gcorefile
> > +    }
> > +}
> > diff --git a/gdb/testsuite/gdb.arch/i386-pkru.c
> > b/gdb/testsuite/gdb.arch/i386-pkru.c
> > index 9561d1e923a..e316ca90bba 100644
> > --- a/gdb/testsuite/gdb.arch/i386-pkru.c
> > +++ b/gdb/testsuite/gdb.arch/i386-pkru.c
> > @@ -18,12 +18,15 @@
> >     along with this program.  If not, see
> > <http://www.gnu.org/licenses/>.  */
> >
> >  #include <stddef.h>
> > +#include <stdlib.h>
> >  #include "x86-cpuid.h"
> >
> >  #ifndef NOINLINE
> >  #define NOINLINE __attribute__ ((noinline))  #endif
> >
> > +volatile int should_dump_core_p = 1;
> > +
> >  unsigned int have_pkru (void) NOINLINE;
> >
> >  static inline unsigned long
> > @@ -83,6 +86,9 @@ main (int argc, char **argv)
> >        wrpkru (wr_value);
> >        asm ("nop\n\t");	/* break here 1.  */
> >
> > +      if (should_dump_core_p) /* Crash for OS corefile. */
> > +	abort ();
> 
> Trailing comment again.

Will fix!

Kind Regards,
Christina
________________________________________
Intel Deutschland GmbH 

Registered Address: Dornacher Strasse 1, 85622 Feldkirchen, Germany 

Tel: +49 (89) 99143-0 

www.intel.de 

Managing Directors: Candice Moore, Jeffrey Schneiderman, Ramachandran Sitaraman

Chairperson of the Supervisory Board: Sonja Pierer

Registered Seat: Munich Commercial Register B: Amtsgericht Munich HRB 186928

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.