Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[Helge Deller]

On 27.09.2016 21:51, Niko Tyni wrote:
> Anyway, I'm attaching a workaround patch that just lowers the optimization
> of op.c. I've verified that the test suite passes with this.
> We'll include this in the next perl upload unless somebody objects.

I agree that including the patch is a better approach for now.
We need a working up-to-date perl on the buildds.

Helge

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[Niko Tyni]

On Sun, Sep 25, 2016 at 03:30:45PM -0400, John David Anglin wrote:
> On 2016-09-25, at 4:09 AM, Niko Tyni wrote:
> 
> > - building with gcc-5 or at -O0 makes it go away
> > - I've bisected with gcc #pragmas that disabling optimization for just
> >  Perl_custom_op_register() in op.c makes it go away
> > - I haven't managed to find which specific optimization causes it
> 
> Attached is the diff in assembly output for Perl_custom_op_register compiled 
> with gcc-5 and gcc-6,
> respectively.  This is with 5.22.2 which I happened to have handy.
> 
> The differences are just label changes and I don't see any functional 
> difference in the assembly.

That's a bit weird. Presumably then the relevant difference in generated
code between gcc versions is somewhere else, and the optimized version
of Perl_custom_op_register() only breaks with the gcc 6 generated code.

Anyway, I'm attaching a workaround patch that just lowers the optimization
of op.c. I've verified that the test suite passes with this.

We'll include this in the next perl upload unless somebody objects.
-- 
Niko Tyni   nt...@debian.org
>From 0950188fc6a08ecb3b924fc4973c3907ecfdbdf7 Mon Sep 17 00:00:00 2001
From: Niko Tyni 
Date: Sun, 25 Sep 2016 09:59:22 +0200
Subject: [PATCH] Temporarily lower the optimization of op.c on hppa due to
 gcc-6 problems

This fixes a crash in ext/XS-APItest/t/customop.t

It's currently unclear if the problem is with the code or the
gcc-6 optimizer.

Bug-Debian: https://bugs.debian.org/838613
---
 cflags.SH | 6 ++
 1 file changed, 6 insertions(+)

diff --git a/cflags.SH b/cflags.SH
index a50044e..ba4ac89 100755
--- a/cflags.SH
+++ b/cflags.SH
@@ -495,6 +495,12 @@ for file do
 # or customize here
 
 case "$file" in
+op) : work around http://bugs.debian.org/838613
+case $archname in
+hppa-*)
+optimize="$optimize -O0";;
+esac;;
+
 regcomp) : work around http://bugs.debian.org/754054
 case $archname in
 mips-*|mipsel-*)
-- 
2.9.3

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[John David Anglin]

On 2016-09-25, at 4:09 AM, Niko Tyni wrote:

> - building with gcc-5 or at -O0 makes it go away
> - I've bisected with gcc #pragmas that disabling optimization for just
>  Perl_custom_op_register() in op.c makes it go away
> - I haven't managed to find which specific optimization causes it

Attached is the diff in assembly output for Perl_custom_op_register compiled 
with gcc-5 and gcc-6,
respectively.  This is with 5.22.2 which I happened to have handy.

The differences are just label changes and I don't see any functional 
difference in the assembly.

Dave
--
John David Anglin   dave.ang...@bell.net


--- op.s-5-op_reg   2016-09-25 15:09:24.274823706 -0400
+++ op.s-6-op_reg   2016-09-25 15:06:37.273864748 -0400
@@ -7,7 +7,7 @@
.PROC
.CALLINFO FRAME=128,CALLS,SAVE_RP,ENTRY_GR=7
.ENTRY
-.LVL1221:
+.LVL1222:
stw %r2,-20(%r30)
.loc 1 14168 0
ldi 0,%r23
@@ -21,7 +21,7 @@
copy %r24,%r6
.loc 1 14168 0
copy %r25,%r24
-.LVL1222:
+.LVL1223:
.loc 1 14162 0
stw %r4,-116(%r30)
stw %r3,-112(%r30)
@@ -32,23 +32,23 @@
.loc 1 14168 0
bl Perl_newSViv,%r2
copy %r26,%r3
-.LVL1223:
+.LVL1224:
copy %r3,%r26
bl Perl_sv_2mortal,%r2
copy %r28,%r25
-.LVL1224:
+.LVL1225:
.loc 1 14170 0
ldw 2024(%r3),%r4
-.LVL1225:
-   comib,= 0,%r4,.L2023
+.LVL1226:
+   comib,= 0,%r4,.L2026
copy %r28,%r7
-.L2021:
+.L2024:
.loc 1 14173 0
ldi 0,%r23
copy %r6,%r24
bl Perl_newSViv,%r2
copy %r3,%r26
-.LVL1226:
+.LVL1227:
stw %r0,-68(%r30)
ldi 4,%r19
stw %r28,-64(%r30)
@@ -60,44 +60,44 @@
copy %r3,%r26
bl Perl_hv_common,%r2
stw %r0,-52(%r30)
-.LVL1227:
-   comib,= 0,%r28,.L2024
+.LVL1228:
+   comib,= 0,%r28,.L2027
ldw -148(%r30),%r2
.loc 1 14175 0
ldw -124(%r30),%r6
-.LVL1228:
+.LVL1229:
ldw -120(%r30),%r5
ldw -116(%r30),%r4
ldw -112(%r30),%r3
-.LVL1229:
 .LVL1230:
+.LVL1231:
bv %r0(%r2)
ldwm -128(%r30),%r7
-.LVL1231:
-.L2023:
-.LBB2498:
+.LVL1232:
+.L2026:
+.LBB2505:
.loc 1 14171 0
ldi 12,%r25
bl Perl_newSV_type,%r2
copy %r3,%r26
-.LVL1232:
-.LBE2498:
+.LVL1233:
+.LBE2505:
stw %r28,2024(%r3)
-   b .L2021
+   b .L2024
copy %r28,%r4
-.LVL1233:
-.L2024:
-.LBB2499:
-.LBB2500:
+.LVL1234:
+.L2027:
+.LBB2506:
+.LBB2507:
.loc 1 14174 0
ldil LR'.LC61,%r25
ldw 4(%r6),%r24
ldo RR'.LC61(%r25),%r25
bl Perl_croak,%r2
copy %r3,%r26
-.LVL1234:
-.LBE2500:
-.LBE2499:
+.LVL1235:
+.LBE2507:
+.LBE2506:
nop
.EXIT
.PROCEND

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[John David Anglin]

On 2016-09-25, at 4:09 AM, Niko Tyni wrote:

> On Fri, Sep 23, 2016 at 07:44:13AM -0400, John David Anglin wrote:
>> On 2016-09-23, at 2:24 AM, Niko Tyni wrote:
>>> On Thu, Sep 22, 2016 at 08:49:42PM -0400, John David Anglin wrote:
> 
 Since 5.22.2-4, the ext/XS-APItest/t/customop test fails on hppa and perl
 fails to build from source.  For example, see:
 https://buildd.debian.org/status/fetch.php?pkg=perl=hppa=5.22.2-5=1474392510
>>> 
>>> There are earlier cases of this on experimental buildds. The oldest I
>>> see is 5.22.1-6. It looks like this regressed with gcc 6.
> 
> I got hppa shell access from Helge, some results:
> 
> - building with gcc-5 or at -O0 makes it go away
> - I've bisected with gcc #pragmas that disabling optimization for just
>  Perl_custom_op_register() in op.c makes it go away
> - I haven't managed to find which specific optimization causes it

Thanks, that's a huge step in narrowing the problem.  If the routine isn't too 
large,
maybe I can spot problem.  Of course, most optimization bugs are in the core gcc
code.

> 
> If nothing else, we can set the optimization level for op.c to -O0 in
> cflags.SH on hppa. Of course even better if somebody can find a more
> targeted deoptimization and/or the underlying reason (whether a code
> bug or a gcc optimizer one.) I don't intend to work much on that myself.
> 
> As the Perl 5.24 transition is currently ongoing in unstable, I don't
> think we want to do a perl upload for the sake of a non-release arch
> before the transition is over.  At the same time, I expect missing the
> 500+ transition binNMUs from the release team is inconvenient for the
> hppa porters.

Yes.
> 
> Maybe you could manually upload a DEB_BUILD_OPTIONS=noopt build of 5.24
> for now, unless you have a separate hppa-specific package repository
> where a sourceful fix could go?

The perl 5.24 transition is almost done except for a few packages.  I will try 
the above
this afternoon.  we don't have separate repository.

Dave

--
John David Anglin   dave.ang...@bell.net

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[Niko Tyni]

On Fri, Sep 23, 2016 at 07:44:13AM -0400, John David Anglin wrote:
> On 2016-09-23, at 2:24 AM, Niko Tyni wrote:
> > On Thu, Sep 22, 2016 at 08:49:42PM -0400, John David Anglin wrote:

> >> Since 5.22.2-4, the ext/XS-APItest/t/customop test fails on hppa and perl
> >> fails to build from source.  For example, see:
> >> https://buildd.debian.org/status/fetch.php?pkg=perl=hppa=5.22.2-5=1474392510
> > 
> > There are earlier cases of this on experimental buildds. The oldest I
> > see is 5.22.1-6. It looks like this regressed with gcc 6.

I got hppa shell access from Helge, some results:

- building with gcc-5 or at -O0 makes it go away
- I've bisected with gcc #pragmas that disabling optimization for just
  Perl_custom_op_register() in op.c makes it go away
- I haven't managed to find which specific optimization causes it

If nothing else, we can set the optimization level for op.c to -O0 in
cflags.SH on hppa. Of course even better if somebody can find a more
targeted deoptimization and/or the underlying reason (whether a code
bug or a gcc optimizer one.) I don't intend to work much on that myself.

As the Perl 5.24 transition is currently ongoing in unstable, I don't
think we want to do a perl upload for the sake of a non-release arch
before the transition is over.  At the same time, I expect missing the
500+ transition binNMUs from the release team is inconvenient for the
hppa porters.

Maybe you could manually upload a DEB_BUILD_OPTIONS=noopt build of 5.24
for now, unless you have a separate hppa-specific package repository
where a sourceful fix could go?
-- 
Niko Tyni   nt...@debian.org

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[John David Anglin]

On 2016-09-23, at 2:24 AM, Niko Tyni wrote:

> On Thu, Sep 22, 2016 at 08:49:42PM -0400, John David Anglin wrote:
>> Package: perl
>> Version: 5.22.2-3
>> Severity: normal
>> 
>> Dear Maintainer,
>> 
>> Since 5.22.2-4, the ext/XS-APItest/t/customop test fails on hppa and perl
>> fails to build from source.  For example, see:
>> https://buildd.debian.org/status/fetch.php?pkg=perl=hppa=5.22.2-5=1474392510
> 
> There are earlier cases of this on experimental buildds. The oldest I
> see is 5.22.1-6. It looks like this regressed with gcc 6.

I'm checking with Helge.  Currently, sibaris is being moved and not available.

> 
> Are there hppa porter boxes available? Can't do much without.
> You might want to test if it still builds with gcc5, and possibly
> play with different optimization levels.

I did a test build outside buildd and will try to get more info.

> -- 
> Niko Tyni   nt...@debian.org


Dave
--
John David Anglin   dave.ang...@bell.net

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[Helge Deller]

> Since 5.22.2-4, the ext/XS-APItest/t/customop test fails on hppa and perl

I do see in the kernel dmesg that perl tries to jump into nirvana (type=6 means 
ITLB fault):

[55974.502498] do_page_fault() command='perl' type=6 address=0xbe400403 in 
attributes.so[f87c3000+3000]
[55974.503017]  vm_start = 0x001d7000, vm_end = 0x0039c000

gdb output isn't helpful either:
process 8125 is executing new program: /build/perl/perl-5.22.2/perl
.././ext/XS-APItest/t/customop.t .. [Thread debugging using libthread_db 
enabled]
Using host libthread_db library "/lib/hppa-linux-gnu/libthread_db.so.1".

Thread 2.1 "perl" received signal SIGSEGV, Segmentation fault.
0xbe400400 in ?? ()
(gdb) bt
#0  0xbe400400 in ?? ()
#1  0x0004203c in Perl_rpeep (my_perl=0x1d7008, o=0x37d8d0) at op.c:14015
Backtrace stopped: previous frame identical to this frame (corrupt stack?)

Niko wrote:
> Are there hppa porter boxes available? Can't do much without.

We don't have debian porter boxes, but if you send me your public ssh key 
and preferred login name via email to "del...@gmx.de", then I can provide you 
with
access to one of the debian buildds.

Helge

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[Niko Tyni]

On Thu, Sep 22, 2016 at 08:49:42PM -0400, John David Anglin wrote:
> Package: perl
> Version: 5.22.2-3
> Severity: normal
> 
> Dear Maintainer,
> 
> Since 5.22.2-4, the ext/XS-APItest/t/customop test fails on hppa and perl
> fails to build from source.  For example, see:
> https://buildd.debian.org/status/fetch.php?pkg=perl=hppa=5.22.2-5=1474392510

There are earlier cases of this on experimental buildds. The oldest I
see is 5.22.1-6. It looks like this regressed with gcc 6.

Are there hppa porter boxes available? Can't do much without.
You might want to test if it still builds with gcc5, and possibly
play with different optimization levels.
-- 
Niko Tyni   nt...@debian.org

Bug#838613: perl: FTBFS on hppa - ext/XS-APItest/t/customop FAILED

[John David Anglin]

Package: perl
Version: 5.22.2-3
Severity: normal

Dear Maintainer,

Since 5.22.2-4, the ext/XS-APItest/t/customop test fails on hppa and perl
fails to build from source.  For example, see:
https://buildd.debian.org/status/fetch.php?pkg=perl=hppa=5.22.2-5=1474392510

Regards,
Dave Anglin

-- System Information:
Debian Release: stretch/sid
  APT prefers buildd-unstable
  APT policy: (500, 'buildd-unstable'), (500, 'unstable')
Architecture: hppa (parisc64)

Kernel: Linux 4.7.4+ (SMP w/4 CPU cores)
Locale: LANG=C, LC_CTYPE=C (charmap=UTF-8) (ignored: LC_ALL set to en_US.utf8)
Shell: /bin/sh linked to /bin/dash
Init: sysvinit (via /sbin/init)

Versions of packages perl depends on:
ii  dpkg   1.18.10
ii  libperl5.225.22.2-3
ii  perl-base  5.22.2-3
ii  perl-modules-5.22  5.22.2-5

Versions of packages perl recommends:
ii  netbase  5.3
ii  rename   0.20-4

Versions of packages perl suggests:
ii  libterm-readline-gnu-perl  1.34-1
ii  make   4.1-9
ii  perl-doc   5.22.2-5

-- no debconf information