[Bug c/46972] New: __thread storage class variable gets optimized out on ARM

[Bug c/46972] New: __thread storage class variable gets optimized out on ARM

[paulius.zaleckas at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

           Summary: __thread storage class variable gets optimized out on
                    ARM
           Product: gcc
           Version: 4.5.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: paulius.zaleckas@gmail.com


Created attachment 22775
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22775
test case source code

My host system is x86_64 and I am cross compiling this file(attached) for ARM
with following command: arm-unknown-linux-uclibcgnueabi-gcc -pipe -O1 -Wall
thread_test.c -o thread_test
Executing ./thread_test on my ARM system results in segfault. If I comment out
test_set() function or remove __thread specifier or compile it without
optimizations (-O0) test passes.
I tried this test also on my host system (Fedora 14) and it does not fail
there.
My cross compiler is vanila GCC + patches from:
http://git.buildroot.net/buildroot/tree/toolchain/gcc/4.5.1

It took me about a week to realize what is going on and make this test case out
of not working cross compiled pulseaudio-0.9.22 server :)

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[paulius.zaleckas at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #1 from Paulius Zaleckas <paulius.zaleckas at gmail dot com> 2010-12-15 21:14:16 UTC ---
Created attachment 22776
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22776
assembler

This is what I get with gcc's -S -fverbose-asm

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[mikpe at it dot uu.se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

Mikael Pettersson <mikpe at it dot uu.se> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mikpe at it dot uu.se

--- Comment #2 from Mikael Pettersson <mikpe at it dot uu.se> 2010-12-16 09:14:33 UTC ---
The test case works for me on armv5tel-linux-gnueabi with gcc 4.4, 4.5, and
4.6.
Comparing your assemby code with mine I see that your calls
__emutls_get_address while mine calls __aeabi_read_tp, and the section
allocation for .LANCHOR0 is different (your is in .bss, mine is in .section
.tdata,"awT",%progbits).

So I suspect the issue is dependent on emutls or uclibc.

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[paulius.zaleckas at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #3 from Paulius Zaleckas <paulius.zaleckas at gmail dot com> 2010-12-16 20:11:54 UTC ---
(In reply to comment #2)
> So I suspect the issue is dependent on emutls or uclibc.

Yes, you are right. I have experimented today with few toolchains and build few
variants myself and I have discovered that gcc has to be built with
--disable-tls configure option to reproduce this bug.

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[paulius.zaleckas at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #4 from Paulius Zaleckas <paulius.zaleckas at gmail dot com> 2010-12-19 16:07:15 UTC ---
Created attachment 22823
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22823
diff with and without -fsection-anchors

I have found out that it is -fsection-anchors optimization causing this bug.
Attached -S assembler output diff with and without this option.

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[paulius.zaleckas at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #5 from Paulius Zaleckas <paulius.zaleckas at gmail dot com> 2010-12-19 19:39:07 UTC ---
Did some testing through various GCC versions:

4.5.2 - buggy
4.4.5 - buggy
4.3.5 - OK, but does not have -fsection-anchors optimization for ARM

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-12-19 21:53:14 UTC ---
I think enumtls has been fixed on the trunk with respect of fsection-anchors. 
Can you try the trunk?

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[paulius.zaleckas at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #7 from Paulius Zaleckas <paulius.zaleckas at gmail dot com> 2010-12-20 11:28:14 UTC ---
(In reply to comment #6)
> I think enumtls has been fixed on the trunk with respect of fsection-anchors. 
> Can you try the trunk?

Yes, looks like this bug is fixed in trunk. Assembly output looks sane. I will
see if it works on real hardware some time later today. Do you know svn
revision it got fixed?

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> 2010-12-20 16:51:29 UTC ---
(In reply to comment #7)
> (In reply to comment #6)
> > I think enumtls has been fixed on the trunk with respect of fsection-anchors. 
> > Can you try the trunk?
> 
> Yes, looks like this bug is fixed in trunk. Assembly output looks sane. I will
> see if it works on real hardware some time later today. Do you know svn
> revision it got fixed?

No but the change was done by Iain Sandoe to fix enum TLS with LTO on darwin.

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[iains at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

Iain Sandoe <iains at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |iains at gcc dot gnu.org

--- Comment #9 from Iain Sandoe <iains at gcc dot gnu.org> 2010-12-21 13:12:53 UTC ---
(In reply to comment #8)
> (In reply to comment #7)
> > (In reply to comment #6)
> > > I think enumtls has been fixed on the trunk with respect of fsection-anchors. 
> > > Can you try the trunk?
> > 
> > Yes, looks like this bug is fixed in trunk. Assembly output looks sane. I will
> > see if it works on real hardware some time later today. Do you know svn
> > revision it got fixed?
> 
> No but the change was done by Iain Sandoe to fix enum TLS with LTO on darwin.

The fix to emuTLS on trunk is from Richard Henderson's new implementation.

http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=162549

I made some fixes to LTO-compatibility and Section anchors, but they primarily
affect ObjC* and PPC-darwin, so likely not relevant to your case.

[Bug target/46972] __thread storage class variable gets optimized out on ARM

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46972

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.6.0

--- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> 2011-12-15 01:53:39 UTC ---
Fixed in 4.6.0 from reading the comments.