This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Stop strip_float_extensions removing binary/decimal conversions
- From: "Richard Guenther" <richard dot guenther at gmail dot com>
- To: "Joseph S. Myers" <joseph at codesourcery dot com>
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Tue, 28 Oct 2008 11:23:16 +0100
- Subject: Re: Stop strip_float_extensions removing binary/decimal conversions
- References: <Pine.LNX.4.64.0810272226570.2268@digraph.polyomino.org.uk>
On Mon, Oct 27, 2008 at 11:30 PM, Joseph S. Myers
<joseph@codesourcery.com> wrote:
> In the course of looking at various optimizations relating to
> floating-point conversions I noticed that strip_float_extensions would
> remove not only genuine extensions (conversions preserving all values)
> but also conversions between binary and decimal floating-point types,
> and that in turn this could cause some optimizations using this
> function to perform bad transformations, as illustrated by the test
> case added by this patch.
>
> Even with -funsafe-math-optimizations I do not think it makes sense
> for the compiler to remove conversions between binary and decimal
> floating-point types - or to introduce such conversions, as the code
> in strip_float_extensions might have done given a decimal constant
> that can be represented in float. This patch unconditionally stops
> strip_float_extensions from doing so. Bootstrapped with no
> regressions on i686-pc-linux-gnu. OK to commit?
Ok.
Thanks,
Richard.
> 2008-10-27 Joseph Myers <joseph@codesourcery.com>
>
> * convert.c (strip_float_extensions): Do not remove or introduce
> conversions between binary and decimal floating-point types.
>
> testsuite:
> 2008-10-27 Joseph Myers <joseph@codesourcery.com>
>
> * gcc.dg/dfp/convert-bfp-12.c: New test.
>
> Index: convert.c
> ===================================================================
> --- convert.c (revision 141381)
> +++ convert.c (working copy)
> @@ -81,7 +81,7 @@
> it properly and handle it like (type)(narrowest_type)constant.
> This way we can optimize for instance a=a*2.0 where "a" is float
> but 2.0 is double constant. */
> - if (TREE_CODE (exp) == REAL_CST)
> + if (TREE_CODE (exp) == REAL_CST && !DECIMAL_FLOAT_TYPE_P (TREE_TYPE (exp)))
> {
> REAL_VALUE_TYPE orig;
> tree type = NULL;
> @@ -108,6 +108,9 @@
> if (!FLOAT_TYPE_P (subt))
> return exp;
>
> + if (DECIMAL_FLOAT_TYPE_P (expt) != DECIMAL_FLOAT_TYPE_P (subt))
> + return exp;
> +
> if (TYPE_PRECISION (subt) > TYPE_PRECISION (expt))
> return exp;
>
> Index: testsuite/gcc.dg/dfp/convert-bfp-12.c
> ===================================================================
> --- testsuite/gcc.dg/dfp/convert-bfp-12.c (revision 0)
> +++ testsuite/gcc.dg/dfp/convert-bfp-12.c (revision 0)
> @@ -0,0 +1,17 @@
> +/* Test for bug where fold wrongly removed conversions to double and
> + replaced them by conversions to float. */
> +/* { dg-options "-std=gnu99" } */
> +
> +extern void abort (void);
> +extern void exit (int);
> +
> +volatile float f = __builtin_inff ();
> +volatile _Decimal32 d32 = 1e40DF;
> +
> +int
> +main (void)
> +{
> + if ((double) f == (double) d32)
> + abort ();
> + exit (0);
> +}
>
> --
> Joseph S. Myers
> joseph@codesourcery.com
>