[armel] Atomic intrinsics are not implemented correctly

 milestoning glib part for alpha-3 so this comes back and we remember to verify that a respin has happened after gcc is fixed.

confirming gcc issue. and tentatively putting on alpha-3 milestone. Not sure how to proceed. If its not known upstream we should probably file a bug there etc. Leaving to doko to drive this bug.

I expect to be able to post a patch soon; it may be simplest to leave this unfixed until then.

This patch should fix the issue with __sync_lock_release(): http://gcc.gnu.org/ml/gcc-patches/2009-12/msg00198.html

The __sync_synchronize() is not fixed by this— another patch should follow soon.

This patch should fix the __sync_synchronize() issue, by making sure GCC does not optimise away the call to the libgcc function which implements this:
http://gcc.gnu.org/ml/gcc-patches/2009-08/msg00600.html

These should now be fixed on the GCC trunk and await backporting to the upstream 4.4 branch.

I've referenced the relevant GCC bugzilla entry from this bug.

This bug was fixed in the package gcc-4.4 - 4.4.2-3ubuntu2

---------------
gcc-4.4 (4.4.2-3ubuntu2) lucid; urgency=low

  * Update to SVN 20091203 from the gcc-4_4-branch (r154951), Fixes:
    PR middle-end/42049, PR c++/42234, PR fortran/41278.
  * PR target/42263, fix wrong code bugs in SMP support on ARM, backport from
    the trunk. LP: #491872 .
  * Pass -mimplicit-it=thumb to as by default on ARM. LP: #488302.
 -- Matthias Klose <email address hidden> Fri, 04 Dec 2009 10:06:17 +0100

Unfortunately, the __sync_synchronize() bug appears still to be present in the updated GCC.
A call to this intrinsic still disappears in the compiler's output.

The __sync_lock_test_and_set() bug may still be present; however it is also possible that the __sync_synchronize() bug is masking the fix.

$ apt-cache policy gcc-4.4 cpp-4.4
gcc-4.4:
  Installed: 4.4.2-3ubuntu2
  Candidate: 4.4.2-3ubuntu2
  Version table:
 *** 4.4.2-3ubuntu2 0
        500 http://arm-ports-ubuntu lucid/main Packages
        100 /var/lib/dpkg/status
cpp-4.4:
  Installed: 4.4.2-3ubuntu2
  Candidate: 4.4.2-3ubuntu2
  Version table:
 *** 4.4.2-3ubuntu2 0
        500 http://arm-ports-ubuntu lucid/main Packages
        100 /var/lib/dpkg/status

$ echo 'void f(void) { __sync_synchronize(); }' | gcc -S -o- -xc -
 .syntax unified
 .arch armv7-a
 .eabi_attribute 27, 3
 .fpu vfpv3-d16
 .eabi_attribute 20, 1
 .eabi_attribute 21, 1
 .eabi_attribute 23, 3
 .eabi_attribute 24, 1
 .eabi_attribute 25, 1
 .eabi_attribute 26, 2
 .eabi_attribute 30, 6
 .eabi_attribute 18, 4
 .thumb
 .file ""
 .text
 .align 2
 .global f
 .thumb
 .thumb_func
 .type f, %function
f:
 @ args = 0, pretend = 0, frame = 0
 @ frame_needed = 1, uses_anonymous_args = 0
 @ link register save eliminated.
 push {r7}
 add r7, sp, #0
 mov sp, r7
 pop {r7}
 bx lr
 .size f, .-f
 .ident "GCC: (Ubuntu 4.4.2-3ubuntu2) 4.4.2"
 .section .note.GNU-stack,"",%progbits

__sync_lock_test_and_set_* still don't contain a barrier in the right place, though this may be due to the above bug.

00008e04 <__sync_lock_test_and_set_4>:
    8e04: e92d 41f0 stmdb sp!, {r4, r5, r6, r7, r8, lr}
    8e08: f640 76c0 movw r6, #4032 ; 0xfc0
    8e0c: 4605 mov r5, r0
    8e0e: 460f mov r7, r1
    8e10: f6cf 76ff movt r6, #65535 ; 0xffff
    8e14: 682c ldr r4, [r5, #0]
    8e16: 4639 mov r1, r7
    8e18: 462a mov r2, r5
    8e1a: 4620 mov r0, r4
    8e1c: 47b0 blx r6
    8e1e: 2800 cmp r0, #0
    8e20: d1f8 bne.n 8e14 <__sync_lock_test_and_set_4+0x10>
    8e22: 4620 mov r0, r4
    8e24: e8bd 81f0 ldmia.w sp!, {r4, r5, r6, r7, r8, pc}

-mimplicit-it=thumb does now seem to be passed correctly to the assembler, though:

$ echo 'void f(void) { asm volatile("moveq r0, #1"); }' | gcc -c -o tmp.o -xc - && objdump -d tmp.o

00000000 <f>:
...
   4: bf08 it eq
   6: 2001 moveq r0, #1

according to https://launchpad.net/ubuntu/+source/gcc-4.4/4.4.2-3ubuntu2/+build/1378632

the patch (attached here is well) is applied. I'll update it to include the testcase, which I missed in the backport.

confirmed, the testcase in the GCC testsuite still fails

fixed in 4.4.2-5ubuntu1

This bug was fixed in the package glib2.0 - 2.23.0-1ubuntu3

---------------
glib2.0 (2.23.0-1ubuntu3) lucid; urgency=low

  * Rebuild with gcc 4.4.2-5ubuntu1. LP: #491872.
 -- Matthias Klose <email address hidden> Mon, 14 Dec 2009 15:57:11 +0000

FIx confirmed in fixed in 4.4.2-5ubuntu1

Thanks!