Re: aarch64 gdb failures in testsuite

Re: aarch64 gdb failures in testsuite

[Steve Ellcey]

Following up to my own mail at
ttps://sourceware.org/ml/gdb/2017-07/msg00037.html

It looks like the problem is with the GCC compiler.  If I use
GCC 4.8.4 to run the gdb.base/store.c test case it works,
if I use GCC 5.4.0 it does not.  I actually get around 50 or
more gdb failures when I use GCC 5.4.0 rather than GCC 4.8.0.
I believe that this may be due to aarch64 changing over from
reload to lra during this timeframe.

I have also found that the test case works if I remove the
register keywords from the test.  Apparently the register
keyword is ignored when optimizing but not at -O0.

Steve Ellcey
sellcey@cavium.com

Re: aarch64 gdb failures in testsuite

[Andrew Pinski]

On Tue, Jul 25, 2017 at 12:45 PM, Steve Ellcey <sellcey@cavium.com> wrote:
>
> Following up to my own mail at
> ttps://sourceware.org/ml/gdb/2017-07/msg00037.html
>
> It looks like the problem is with the GCC compiler.  If I use
> GCC 4.8.4 to run the gdb.base/store.c test case it works,
> if I use GCC 5.4.0 it does not.  I actually get around 50 or
> more gdb failures when I use GCC 5.4.0 rather than GCC 4.8.0.
> I believe that this may be due to aarch64 changing over from
> reload to lra during this timeframe.
>
> I have also found that the test case works if I remove the
> register keywords from the test.  Apparently the register
> keyword is ignored when optimizing but not at -O0.

function.c:
/* Return true if we should assign DECL a pseudo register; false if it
   should live on the local stack.  */

bool
use_register_for_decl (const_tree decl)
{

....

  if (optimize)
    return true;

  if (!DECL_REGISTER (decl))
    return false;

  /* When not optimizing, disregard register keyword for types that
     could have methods, otherwise the methods won't be callable from
     the debugger.  */
  if (RECORD_OR_UNION_TYPE_P (TREE_TYPE (decl)))
    return false;

  return true;
}

--------- CUT ------

This code has been that way since 2009 (r144939) but that function
returned true for DECL_REGISTER before that.

Thanks,
Andrew Pinski


>
> Steve Ellcey
> sellcey@cavium.com

Re: aarch64 gdb failures in testsuite

[Steve Ellcey]

On Tue, 2017-07-25 at 15:13 -0700, Andrew Pinski wrote:
> 
> function.c:
> /* Return true if we should assign DECL a pseudo register; false if it
>    should live on the local stack.  */
> 
> bool
> use_register_for_decl (const_tree decl)
> {
> 

Yes, I just tried tweaking that function to return false when not
optimizing, even if the decl has DECL_REGISTER set.  If fixed
gdb.base/store.exp but not much else.  The other things that did
not fail after the change were:

FAIL: gdb.base/step-over-syscall.exp: clone: displaced=off: single step over clone
FAIL: gdb.base/watchpoint.exp: next over global_ptr_ptr buffer set
FAIL: gdb.base/watchpoint.exp: next over global_ptr_ptr init
FAIL: gdb.base/watchpoint.exp: next over global_ptr_ptr pointer advance

And I don't actually see why those would change behaviour, I didn't see
any register variables in those tests  I could submit a patch to GCC
with this change but it seems like a rather large hammer to use on a
few fixes.  Maybe changing gdb.base/store.exp to not use register
variables would be a reasonable fix.  Originally, I thought that was a
bad idea because I thought the problem affected more tests, now I am
not so sure.

Steve Ellcey
sellcey@cavium.com

Re: aarch64 gdb failures in testsuite

[Steve Ellcey]

On Tue, 2017-07-25 at 15:26 -0700, Steve Ellcey wrote:
> 
> And I don't actually see why those would change behaviour, I didn't see
> any register variables in those tests  I could submit a patch to GCC
> with this change but it seems like a rather large hammer to use on a
> few fixes.  Maybe changing gdb.base/store.exp to not use register
> variables would be a reasonable fix.  Originally, I thought that was
> a
> bad idea because I thought the problem affected more tests, now I am
> not so sure.
> 
> Steve Ellcey
> sellcey@cavium.com

FYI: I submitted a bug against gcc that gdb.base/store.exp does not
work with GCC 5.4.0 (or later) but was working with GCC 4.8.0.

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81562

Steve Ellcey
sellcey@cavium.com

aarch64 gdb failures in testsuite

[Steve Ellcey]

I am getting some aarch64 gdb failures that do not seem to be showing up 
in other peoples test runs from the gdb-testers mailing list and I was
wondering if anyone could help me figure out why these tests are failing
for me.

As an example, here is gdb.base/store.exp.  In my test case I am
getting this failuire:

Temporary breakpoint 9, wack_doublest (u=-1, v=-2) at /home/sellcey/gdb-std/src/binutils-gdb/gdb/testsuite/gdb.base/store.c:125
125       register doublest l = u, r = v;
(gdb) next
126       l = add_doublest (l, r);
(gdb) PASS: gdb.base/store.exp: continue to wack_doublest
print l
$29 = -1
(gdb) PASS: gdb.base/store.exp: var doublest l; print old l, expecting -1
print r
$30 = <optimized out>
(gdb) FAIL: gdb.base/store.exp: var doublest l; print old r, expecting -2
set variable l = 4


But when I look at https://sourceware.org/ml/gdb-testers/2017-q3/msg01016.html
and look at the gdb.log file I see it passes there:


Temporary breakpoint 9, wack_doublest (u=-1, v=-2) at /home/qiyao/buildbot/ubuntu-aarch64-m64/build/gdb/testsuite/../../../binutils-gdb/gdb/testsuite/gdb.base/store.c:125
125       register doublest l = u, r = v;
(gdb) next
126       l = add_doublest (l, r);
(gdb) PASS: gdb.base/store.exp: continue to wack_doublest
print l
$29 = -1
(gdb) PASS: gdb.base/store.exp: var doublest l; print old l, expecting -1
print r
$30 = -2
(gdb) PASS: gdb.base/store.exp: var doublest l; print old r, expecting -2
set variable l = 4


In looking for differences in the runs I see the passing test uses GCC 4.8.4
and I am using GCC 5.4.0 (Ubuntu 16.04).  I don't think that should make
a difference but maybe it does.

I looked at the compiler options since for something to be <optimized out>
I assume we would have to be compiling with some optimization option but
both compilers are just using -g.

spawn -ignore SIGHUP gcc /home/sellcey/gdb-std/obj/binutils-gdb/gdb/testsuite/outputs/gdb.base/store/store0.o -g -lm -o /home/sellcey/gdb-std/obj/binutils-gdb/gdb/testsuite/outputs/gdb.base/store/store

I compiled the test case (gdb/testsuite/gdb.base/store.c) with -g using
the system GCC (5.4.0) and then ran the system GDB on it to see what
happened there and I get the same <optimized out> message there that I
do with the latest GDB sources.

% gcc -g store.c -o x
% gdb x

(gdb) b wack_doublest
Breakpoint 1 at 0x4008cc: file store.c, line 125.
(gdb) r
Starting program: /home/sellcey/gdb-std/x 

Breakpoint 1, wack_doublest (u=-1, v=-2) at store.c:125
125	  register doublest l = u, r = v;
(gdb) n
126	  l = add_doublest (l, r);
(gdb) print l
$1 = -1
(gdb) print r
$2 = <optimized out>


Any ideas on what is happening here?  Does anyone else get this failure?

Steve Ellcey
sellcey@cavium.com