On Wed, Nov 23, 2016 at 02:26:49PM +0100, Michael Matz wrote:
> Hi,
>
> as the bug trail explains make_extraction is claiming but failing to
> handle extractions that would go outside the underlying object. So, let's
> not construct such, as the patch does.
>
> Dominik tested s390x bootstrap being recovered with this, Andreas ia64
> bootstrap, and I regstrapped this on x86-64-linux without regressions (all
> langs+ada). Okay for trunk?
Shouldn't new_rtx be set to NULL_RTX if that condition is false? Otherwise
it will be whatever make_compound_operation returned. What about the break?
Shouldn't that be done only if the condition is true too?
Anyway, I'm afraid I don't know this code enough, so deferring to Segher.
> PR bootstrap/78390
> * combine.c (make_compound_operation_int): Don't extract
> from outside underlying object.
>
> diff --git a/gcc/combine.c b/gcc/combine.c
> index 0210685..1d8bddf 100644
> --- a/gcc/combine.c
> +++ b/gcc/combine.c
> @@ -8108,9 +8108,16 @@ make_compound_operation (rtx x, enum rtx_code in_code)
> && GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (inner))
> && subreg_lowpart_p (x))
> {
> + int len = mode_width;
> new_rtx = make_compound_operation (XEXP (inner, 0), next_code);
> + /* Don't extract bits outside the underlying mode. */
> + if (CONST_INT_P (XEXP (inner, 1))
> + && (INTVAL (XEXP (inner, 1)) + len
> + > GET_MODE_PRECISION (GET_MODE (new_rtx))))
> + len = GET_MODE_PRECISION (GET_MODE (new_rtx))
> + - INTVAL (XEXP (inner, 1));
> new_rtx = make_extraction (mode, new_rtx, 0, XEXP (inner, 1),
> - mode_width, 1, 0, in_code == COMPARE);
> + len, 1, 0, in_code == COMPARE);
> break;
> }
>
Jakub
