[armel] segfaults in make check pass when built with optimization

I am out of the office until 17/04/2011.

Note: This is an automated response to your message "[Bug 758082] [NEW]
[armel] segfaults in make check pass when built with optimization" sent on
12/4/11 0:03:12.

This is the only notification you will receive while this person is away.

This bug was fixed in the package augeas - 0.8.0-0ubuntu4

---------------
augeas (0.8.0-0ubuntu4) natty; urgency=low

  * debian/rules: Work around segfault in make check by building with -O0 on armel.
    Fixes armel FTBFS (LP: #758082)
 -- Jani Monoses <email address hidden> Tue, 12 Apr 2011 00:04:20 +0300

Just poking about, it looks like just recompiling syntax.c with -O0 gets it working.
(indeed compiling it with -O1 makes it fail more tests that -O2).

It's a whole chain of passing va_args down throuhg about 5 different levels of printf type thing.

Dave

This is from a -O1 of syntax.c - note how it stamps on r2 as a temp which is the 1st variadic argument before it gets to stack the original.

        .global syntax_error <<<<< void syntax_error(struct info *info, const char *format, ...) {
        .thumb
        .thumb_func
        .type syntax_error, %function
syntax_error:
        .fnstart
.LFB100:
        .loc 1 116 0
        .cfi_startproc
        @ args = 4, pretend = 12, frame = 8
        @ frame_needed = 0, uses_anonymous_args = 1
        .loc 1 120 0
        ldr r3, [r0, #0]
        ldr r3, [r3, #0]
        subs r2, r3, #6 !!!!!
        it ne
        movne r2, #1
        cmp r3, #0
        ite eq
        moveq r3, #0
        andne r3, r2, #1
        cbnz r3, .L153
        .loc 1 116 0
        push {r1, r2, r3}
        .save {r1, r2, r3}

the effect look similar to the one affecting telepathy-glib

https://bugs.launchpad.net/ubuntu/+source/telepathy-glib/+bug/736081

Thanks to Ramana for suggesting -fno-shrink-wrap which does indeed fix it.
(L153 is an early exit).

Dave

This test case fails with -O1 but works with -fno-shrink-wrap and seems to be the same failure pattern.

lp:~rsandifo/gcc-linaro/lp-758082 fixes this but the code generated is poor:

foo:
 push {r1, r2, r3}
 ldr r3, [r0, #0]
 push {lr}
 sub sp, sp, #8
 ldr r3, [r3, #8]
 subs r2, r3, #6
 it ne
 movne r2, #1
 cmp r3, #8
 ite eq
 moveq r3, #0
 andne r3, r2, #1
 cbnz r3, .L1
 movw r2, #:lower16:stdout
 add r3, sp, #16
 movt r2, #:upper16:stdout
 ldr r1, [sp, #12]
 ldr r0, [r2, #0]
 mov r2, r3
 str r3, [sp, #4]
 bl vfprintf
.L1:
 add sp, sp, #8
 pop {lr}
 add sp, sp, #12
 bx lr

r1 doesn't need to be saved. Ignoring hazards, r0 could be used instead of r3. The epilogue is poor due to the two adds. lr could be pushed with the initial push.

Split the performance into LP: #767986