On Tue, 4 Nov 2014, Richard Biener wrote:
>
> On the match-and-simplify branch we expose an issue in shorten_compare
> which happily transforms (double) float-var != (double) dfp-float-var
> to (float) float-var != (float) dfp-float-var which is wrong
> and causes
>
> FAIL: c-c++-common/dfp/convert-bfp-12.c -std=c++11 execution test
> FAIL: c-c++-common/dfp/convert-bfp-12.c -std=c++1y execution test
> FAIL: c-c++-common/dfp/convert-bfp-12.c -std=c++98 execution test
>
> you can expose this latent bug by making get_narrower also
> treat CONVERT_EXPRs as conversions - on trunk the DFP value
> happens to be converted to double using that.
>
> You then also run into the issue that the C frontend rejects
> the compare via its common_type implementation which complains
> about a mixed FP - DFP compare.
>
> So the following patch which I am testing right now disables
> the (premature) shorten-compare optimization on mixed
> FP - DFP operands. It also exposes the issue on trunk by
> improving get_narrower.
>
> Ok for trunk either with or without the tree.c hunk? (I'm
> going to remove that if it exposes more issues elsewhere
> revealed by testing)
I removed it - it fails very early in stage2 with a warning about an
always-true comparison. No time to sort that out at this point.
So consider this a c-common.c change only.
Thanks,
Richard.
> 2014-11-04 Richard Biener <rguent...@suse.de>
>
> * tree.c (get_narrower): Also look through CONVERT_EXPRs.
>
> c-family/
> * c-common.c (shorten_compare): Do not shorten mixed
> DFP and non-DFP compares.
>
> Index: gcc/tree.c
> ===================================================================
> --- gcc/tree.c (revision 217049)
> +++ gcc/tree.c (working copy)
> @@ -8494,7 +8494,7 @@ get_narrower (tree op, int *unsignedp_pt
> tree win = op;
> bool integral_p = INTEGRAL_TYPE_P (TREE_TYPE (op));
>
> - while (TREE_CODE (op) == NOP_EXPR)
> + while (CONVERT_EXPR_P (op))
> {
> int bitschange
> = (TYPE_PRECISION (TREE_TYPE (op))
> Index: gcc/c-family/c-common.c
> ===================================================================
> --- gcc/c-family/c-common.c (revision 217049)
> +++ gcc/c-family/c-common.c (working copy)
> @@ -4314,9 +4314,15 @@ shorten_compare (location_t loc, tree *o
> /* If either arg is decimal float and the other is float, find the
> proper common type to use for comparison. */
> else if (real1 && real2
> + && DECIMAL_FLOAT_MODE_P (TYPE_MODE (TREE_TYPE (primop0)))
> + && DECIMAL_FLOAT_MODE_P (TYPE_MODE (TREE_TYPE (primop1))))
> + type = common_type (TREE_TYPE (primop0), TREE_TYPE (primop1));
> +
> + /* If either arg is decimal float and the other is float, fail. */
> + else if (real1 && real2
> && (DECIMAL_FLOAT_MODE_P (TYPE_MODE (TREE_TYPE (primop0)))
> || DECIMAL_FLOAT_MODE_P (TYPE_MODE (TREE_TYPE (primop1)))))
> - type = common_type (TREE_TYPE (primop0), TREE_TYPE (primop1));
> + return 0;
>
> else if (real1 && real2
> && (TYPE_PRECISION (TREE_TYPE (primop0))
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Jeff Hawn, Jennifer Guild, Felix Imendoerffer, HRB 21284
(AG Nuernberg)
Maxfeldstrasse 5, 90409 Nuernberg, Germany
