[PATCH] combine: Another hard register problem (PR85805)

[Richard Biener]

On Thu, Jul 26, 2018 at 12:12 PM Segher Boessenkool
<segher@kernel.crashing.org> wrote:
>
> The current code in reg_nonzero_bits_for_combine allows using the
> reg_stat info when last_set_mode is a different integer mode.  This is
> completely wrong for non-pseudos.  For example, as in the PR, a value
> in a DImode hard register is set by eight writes to its constituent
> QImode parts.  The value written to the DImode is not the same as that
> written to the lowest-numbered QImode!
>
> This patch fixes it.  Committing.  Will backport later, too.

testcase?

>
> Segher
>
>
> 2018-07-26  Segher Boessenkool  <segher@kernel.crashing.org>
>
>         PR rtl-optimization/85805
>         * combine.c (reg_nonzero_bits_for_combine): Only use the last set
>         value for hard registers if that was written in the same mode.
>
> ---
>  gcc/combine.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/combine.c b/gcc/combine.c
> index 09cbad4..fe71f3a 100644
> --- a/gcc/combine.c
> +++ b/gcc/combine.c
> @@ -10186,7 +10186,8 @@ reg_nonzero_bits_for_combine (const_rtx x, scalar_int_mode xmode,
>    rsp = &reg_stat[REGNO (x)];
>    if (rsp->last_set_value != 0
>        && (rsp->last_set_mode == mode
> -         || (GET_MODE_CLASS (rsp->last_set_mode) == MODE_INT
> +         || (REGNO (x) >= FIRST_PSEUDO_REGISTER
> +             && GET_MODE_CLASS (rsp->last_set_mode) == MODE_INT
>               && GET_MODE_CLASS (mode) == MODE_INT))
>        && ((rsp->last_set_label >= label_tick_ebb_start
>            && rsp->last_set_label < label_tick)
> --
> 1.8.3.1
>