This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch,fortran] Fix PR19310 unnecessary error for overflowing results

Steve Kargl wrote:
On Mon, Jun 12, 2006 at 10:41:50PM -0700, Jerry DeLisle wrote:


The attached patch expands considerably on the last one.

I think I have covered all possibilities in the correct order.

This is a request for further comment.

It occurred to me that in simplify.c there are many functions with their own error checking independent of range_check. Do we want to incorporate the -frange_check flag into these error checks?

Also I noticed that IFORT does not do all this checking by default and happily compiles a lot of stuff. We could default this the other way. Any thoughts on that? (not that intel is the only way to do this)

Regardless, the patch regtests OK, including the test case I submitted earlier.

If no further comments is this OK to commit?



I read the patch and I think it's almost ready to commit.
However, I want to study the part in arith.c where it looks
likes you disable the gradual underflow via SUBNORMAL, and
I want to check the logic the with complex types.  In a quick
read, it looks like you skip the imaginary part depending on
the checking of the real part.


If we return an ARITH_UNDERFLOW on subnormal we will get a compile time error. We want the value to continue on and not give an error.

On complex, there is no need to check the imaginary part if there is a range error on the real part. This is the original logic, I intended to just add additional cases.

Looking forward to your final review.



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]