execute/ieee/980619-1.c on x86

Stephen L Moshier moshier@mediaone.net
Mon Feb 7 18:47:00 GMT 2000


> This test has been XPASSing on x86 for about two months now.
> Does anyone know if the bug was definitely fixed
>
>> I guess it's the following entry:
>>   * loop.c (load_mems): Don't hoist written floating point mem
>>   if -ffloat-store.

That change was for a particular test case recently posted to the
mailing list, not for the torture test case.  A slightly different
program might still produce a compiler-generated temp variable for
comparison and thereby defeat float-store.  I believe the change is
only one of many fixes that would be needed to implement float-store
successfully.

Furthermore the loop case fails if you don't set the float-store
switch, but contrarywise it would not be hard to construct some other
test that will fail if you _do_ set the float-store switch.  i386 is
not ordinarily operated as an IEEE machine, and you ought not lose
much sleep over IEEE tests that fail on it.  If somebody remembers to
look at that subdirectory of cases once a year or so, that would be
plenty often enough.

I don't think that 980619 torture test case should even be there at
all, unless you want to establish an official policy that float-store
should actually work, or if you want to start some even more ambitious
major project effort for IEEE on i386.



More information about the Gcc mailing list