[Bug middle-end/34621] [4.3 Regression] gcc.c-torture/execute/va-arg-25.c:32: internal compiler error: in expand_call, at calls.c:2785

[ubizjak at gmail dot com]

------- Comment #9 from ubizjak at gmail dot com  2008-02-12 15:11 -------
By disabling asserts, we can compare -O2 call frame for va-arg-25.c test
vs. what would -Os call frame look like:

gcc -O2 -fomit-frame-pointer:

main:
        subl    $76, %esp
        movdqa  .LC0, %xmm0
        movl    $2, 32(%esp)
        movdqa  %xmm0, 48(%esp)
        movl    $1, (%esp)
        movdqa  .LC1, %xmm0
        movdqa  %xmm0, 16(%esp)
        call    foo

vs. gcc -Os (aka -mno-accumulate-outgoing-args) -fomit-frame-pointer:

main:
        subl    $28, %esp
        movaps  .LC0, %xmm0
        movaps  %xmm0, (%esp)
        pushl   $2
        movaps  .LC1, %xmm0
        subl    $16, %esp
        movaps  %xmm0, (%esp)
        pushl   $1
        call    foo
        ...

So, at the call, we have -O2 stack:

 +76
 +72
 +68
 +64
 +60  .LC0
 +56  .LC0
 +52  .LC0 
 +48  .LC0
 +44   pad
 +40   pad
 +36   pad
 +32  $2
 +28  .LC1
 +24  .LC1
 +20  .LC1
 +16  .LC1
 +12   pad
 +8    pad      
 +4    pad
esp   $1
--------------

For -Os stack gets misaligned after "pushl $2". It looks that gcc should
decrease stack for additional 12 bytes before $2 is pushed, so the stack
is consistent with -O2.

For i686-linux-*, we generate correct sequence for -Os:

main:
        leal    4(%esp), %ecx
        andl    $-16, %esp
        pushl   -4(%ecx)
        pushl   %ecx
        subl    $36, %esp
        movaps  .LC0, %xmm0
        movaps  %xmm0, 12(%esp)
        pushl   $2
        movaps  .LC1, %xmm0
        subl    $28, %esp
        movaps  %xmm0, 12(%esp)
        pushl   $1
        call    foo
        ...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34621