[Bug rtl-optimization/55672] [4.8 Regression] -fstack-check=generic ICEs in print_reg, at config/i386/i386.c:13868

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55672

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |rtl-optimization

--- Comment #8 from H.J. Lu <hjl.tools at gmail dot com> 2012-12-18 20:35:17 UTC ---
(In reply to comment #4)
> If stack_realign_p is true, frame_pointer_needed is also true.   So we can use
> fp to eliminate frame but i386.c::x86_can_eliminate prohibits it.  The code
> looks strange:
> 
> 
>  if (stack_realign_fp)
>     return ((from == ARG_POINTER_REGNUM
>              && to == HARD_FRAME_POINTER_REGNUM)
>             || (from == FRAME_POINTER_REGNUM
>                 && to == STACK_POINTER_REGNUM));
> 
> So we permit to change argument pointer but not frame pointer to FP which again
> is strange IMHO.   Changing the code to
> 
>  if (stack_realign_fp)
>     return ((from == ARG_POINTER_REGNUM
>              && to == HARD_FRAME_POINTER_REGNUM)
>             || (from == FRAME_POINTER_REGNUM
>                 && to == STACK_POINTER_REGNUM)
>             || (from == FRAME_POINTER_REGNUM
>                 && to == HARD_FRAME_POINTER_REGNUM));
> 
> solves the problem.

It fixes ICE, but generates questionable code:

main:
.LFB0:
    .cfi_startproc
    pushl    %ebp
    .cfi_def_cfa_offset 8
    .cfi_offset 5, -8
    movl    %esp, %ebp
    .cfi_def_cfa_register 5
    andl    $-16, %esp
    subl    $8236, %esp
    orl    $0, (%esp)
    addl    $8204, %esp
    cmpl    $4, -40(%ebp)
    je    .L2
    call    abort
.L2:
    movl    $0, %eax
    leave
    .cfi_restore 5
    .cfi_def_cfa 4, 4
    ret

Without LRA, we got

main:
.LFB0:
    .cfi_startproc
    pushl    %ebp
    .cfi_def_cfa_offset 8
    .cfi_offset 5, -8
    movl    %esp, %ebp
    .cfi_def_cfa_register 5
    andl    $-16, %esp
    subl    $8236, %esp
    orl    $0, (%esp)
    addl    $8204, %esp
    cmpl    $4, (%esp)
    je    .L2
    call    abort
.L2:
    movl    $0, %eax
    leave
    .cfi_restore 5
    .cfi_def_cfa 4, 4
    ret

The difference is

--- x.s    2012-12-18 12:24:17.072888139 -0800
+++ no-lra.s    2012-12-18 12:30:11.419157548 -0800
@@ -14,7 +14,7 @@ main:
     subl    $8236, %esp
     orl    $0, (%esp)
     addl    $8204, %esp
-    cmpl    $4, -40(%ebp)
+    cmpl    $4, (%esp)
     je    .L2
     call    abort
 .L2:

I think LRA generated code is wrong.  The reason we don't allow
converting software frame pointer to hardware frame pointer is
when stack alignment is needed, hardware frame pointer is used
to save stack pointer.  We can no longer use it for software
frame pointer.