https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69454
--- Comment #28 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to H.J. Lu from comment #27)
> (In reply to Ilya Enkovich from comment #26)
> > (In reply to H.J. Lu from comment #25)
> > > Please add -mpreferred-stack-boundary=2 to your tests. Otherwise,
> > > you just remove a nop.
> >
> > Here is a test which crashes LRA with the path you proposed. Crash happens
> > because LRA requests a stack slot aligned by 8 and stack is aligned by 4.
> > So Jakub's patch looks the safest option for now, but probably use 64
> > requirement instead of 128.
> >
> > >cat test.i
> > long long a, b;
> > fn1() {
> > long long c = a;
> > a = b ^ a;
> > fn2();
> > a = c;
> > }
> > >gcc test.i -O2 -m32 -mpreferred-stack-boundary=2
> > test.i: In function 'fn1':
> > test.i:5:3: warning: implicit declaration of function 'fn2'
> > [-Wimplicit-function-declaration]
> > fn2();
> > ^~~
> >
> > test.i:7:1: internal compiler error: in assign_stack_local_1, at
> > function.c:409
>
> The last place to update stack alignment is ix86_update_stack_boundary.
> Try
>
> diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
> index a03a515..10a00f8 100644
> --- a/gcc/config/i386/i386.c
> +++ b/gcc/config/i386/i386.c
> @@ -3588,16 +3588,6 @@ convert_scalars_to_vector ()
> bitmap_obstack_release (NULL);
> df_process_deferred_rescans ();
>
> - /* Conversion means we may have 128bit register spills/fills
> - which require aligned stack. */
> - if (converted_insns)
> - {
> - if (crtl->stack_alignment_needed < 128)
> - crtl->stack_alignment_needed = 128;
> - if (crtl->stack_alignment_estimated < 128)
> - crtl->stack_alignment_estimated = 128;
> - }
> -
> return 0;
> }
>
> @@ -12041,6 +12031,13 @@ ix86_update_stack_boundary (void)
> if (crtl->stack_alignment_needed < 128)
> crtl->stack_alignment_needed = 128;
> }
> +
> + /* The STV pass needs 64-bit stack alignment. */
> + if (!TARGET_64BIT
> + && TARGET_STV
> + && TARGET_SSE2 && optimize > 1
> + && crtl->stack_alignment_estimated < 64)
> + crtl->stack_alignment_estimated = 64;
> }
>
> /* Handle the TARGET_GET_DRAP_RTX hook. Return NULL if no DRAP is
Which will likely penalize even code where the stv pass wouldn't do anything.
Isn't it better to just disable the stv pass if the stack isn't aligned enough
(like in the #c4 or #c5 patches)?
