This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Results for egcs-2.91.52 19980727 (gcc2 ss-980609 experiment
- To: Carlo Wood <carlo at runaway dot xs4all dot nl>
- Subject: Re: Results for egcs-2.91.52 19980727 (gcc2 ss-980609 experiment
- From: David Edelsohn <dje at watson dot ibm dot com>
- Date: Wed, 05 Aug 1998 23:05:59 -0400
- Cc: egcs at cygnus dot com (egcs at cygnus dot com)
>>>>> Carlo Wood writes:
Carlo> I don't think it would be a waste of time though, either it IS a bug
Carlo> that this program fails or the program should be adjusted so it
Carlo> tests the correct thing. Ruling out the PowerPC architecture seems
Carlo> unlogical to me: A user could expect (rely on ;) the accuracy one
Carlo> would expect with sizeof(float) == 4. Its a bit "fuzzy" to use
Carlo> double internally on one platform and not on others, without
Carlo> a clear way to detect this other then by knowing on what platform
Carlo> you're running.
sizeof(float) == 4 with respect to memory structures. For values
in registers, the POWER architecture provides greater precision than
required by IEEE.
If you want correct IEEE operation from IBM POWER architecture (or
common-mode within AIX), one must use the GCC -ffloat-store option to
prevent the excess precision. AIX XLC has a similar option. Maybe this
option should be used for all testcases in the execute/ieee subdirectory.
Any user who relies upon IEEE conformance without understanding
the architecture will experience problems.
Again, this is not the *PowerPC* architecture, it is the original
POWER architecture. PowerPC implements single-precision floating-point
and the test passes. When running the testsuite on AIX, one does not test
the PowerPC architecture by default.
David