On Mon, Jan 12, 2015 at 12:19 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> The 991213-3.c testcase ICEs on aarch64-linux with -mabi=ilp32
> since wide-int merge. The problem is that
> x = convert_memory_address (Pmode, x)
> is used twice on a VOIDmode CONST_INT, which is wrong.
> For non-VOIDmode rtl the second convert_memory_address
> is a NOP, but for VOIDmode the second call treats the CONST_INT
> returned by the first call as if it was again ptr_mode, rather
> than Pmode. On aarch64-linux in particular, the constant is
> zero-extended from SImode to DImode in the first call, so it
> is not valid SImode CONST_INT any longer.
>
> emit_indirect_jump always calls convert_memory_address (Pmode, ...)
> on the operand in optabs.c when handling EXPAND_ADDRESS case
> in maybe_legitimize_operand, so the first convert_memory_address
> is both unnecessary and harmful.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux
> (which do not define POINTERS_EXTEND_UNSIGNED) and tested on the
> problematic testcase with aarch64-linux cross. Can anyone with
> easy access to POINTERS_EXTEND_UNSIGNED targets (aarch64-linux ilp32,
> x86_64 -mx32, ia64-hpux) please test this?
I will test it in a next couple of days on aarch64-linux ilp32.
Thanks,
Andrew
>
> Ok for trunk if it works there?
>
> 2015-01-12 Jakub Jelinek <ja...@redhat.com>
>
> PR middle-end/63974
> * cfgexpand.c (expand_computed_goto): Don't call
> convert_memory_address here.
>
> --- gcc/cfgexpand.c.jj 2015-01-09 21:59:54.000000000 +0100
> +++ gcc/cfgexpand.c 2015-01-12 14:41:35.210705174 +0100
> @@ -3060,8 +3060,6 @@ expand_computed_goto (tree exp)
> {
> rtx x = expand_normal (exp);
>
> - x = convert_memory_address (Pmode, x);
> -
> do_pending_stack_adjust ();
> emit_indirect_jump (x);
> }
>
> Jakub
