[PATCH] -ffinite-math-only A-A

[Christian BRUEL]

Hello,

thank you for the quick answer.

Richard Guenther wrote:
> On 8/22/07, Christian BRUEL <christian.bruel@st.com> wrote:>
> 
>>Hello,
>>
>>I believe that ieee floating point A-A can be optimized to 0 under
>>-ffinite-math-only.
>>
...
> 
> The patch is wrong:
> 

ok, let give it another try, if we agree that it's ok to relax the fold 
check on A-A -> 0


>>Index: fold-const.c
>>===================================================================
>>--- fold-const.c        (revision 127666)
>>+++ fold-const.c        (working copy)
>>@@ -10152,7 +10152,8 @@
>>       if ((! FLOAT_TYPE_P (type)
>>           || (flag_unsafe_math_optimizations
> 
> 
>   ^^^^
> this test is the restricting one
> 

yes, but it is too restrictive because it doesn't let the HONOR (thus 
the finite-math-only) checks if not in -funsafe-math-optimisations.

> 
>>               && !HONOR_NANS (TYPE_MODE (type))
>>-              && !HONOR_INFINITIES (TYPE_MODE (type))))
>>+              && !HONOR_INFINITIES (TYPE_MODE (type)))
>>+          || flag_finite_math_only)
> 
> 
> !HONOR_NANS && !HONOR_INFINITIES is exactly flag_finite_math_only.

yes

> 
> See also http://gcc.gnu.org/wiki/GeertBosch for some good overview on this
> matters.  So the correct fix would be to remove the
> flag_unsafe_math_optimizations

It all depends if we prefer to let the optimisation A-A->0 go only under 
a), b) or c) with :

a) unsafe-math-optimisations AND finite-math-only : current situation
b) unsafe-math-optimisations OR  finite-math-only : my proposed patch
c) finite-math-only only                          : attached patch

a) is too restrictive to my opinion. c) should be safe enough. I don't 
think b) is needed but I'm not sure.

> check instead.  (And check what IEEE says to -Inf - -Inf and Inf -
> Inf, just in case

-Inf - -Inf is Nan, Inf - -Inf would be Inf and Inf - Inf is NaN.
This A-A optimisation would give 0.0 so the !HONOR_INFINITY is required 
in all the cases.

> Geert is right with that only !HONOR_NANS is required).
> 
!NONOR_INFINITY was already checked, I didn't add it. So maybe it was 
missing in its wiki ? CCing him.


> Richard.
> 

Best Regards,

Christian

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: nan2.patch
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20070822/fb418e78/attachment.ksh>