On Sat, Apr 30, 2011 at 1:55 PM, Eric Botcazou <ebotca...@adacore.com> wrote:
> On Saturday 23 April 2011 09:35:38 Uros Bizjak wrote:
>> Index: i386.c
>> ===================================================================
>> --- i386.c (revision 172866)
>> +++ i386.c (working copy)
>> @@ -10149,7 +10149,7 @@ ix86_adjust_stack_and_probe (const HOST_
>> /* Even if the stack pointer isn't the CFA register, we need to
>> correctly describe the adjustments made to it, in particular differentiate
>> the frame-related ones from the frame-unrelated ones. */
>> - if (size > 0)
>> + if (size != 0)
>> {
>> rtx expr = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (2));
>> XVECEXP (expr, 0, 0)
>
> Probing for a negative size doesn't make much sense; we shouldn't be there if
> the size to allocate is negative. Is that really intended? The code without
> stack checking is:
>
> pushq %rbp
> .LCFI0:
> movq %rsp, %rbp
> .LCFI1:
> andq $-32, %rsp
> addq $16, %rsp
>
> which looks weird. If this is nevertheless correct, then the fix for the
> problem is attached. Tested on x86_64-suse-linux.
>
This code aligns stack to 32byte for AVX.
--
H.J.
