Re: Floating poing 'non-bug'? Or is this a bug?

[Jim Wilson <wilson at specifixinc dot com>]

Yuval Kfir wrote:
> Isn't this something that should be fixed in the gcc backend, then?

In theory.  But this is a very complicated problem.  It has been a known 
problem for at least 13-14 years, and it hasn't been fixed yet. 
Personally, I don't think it ever will be.

It is probably easier to fix the problem in hardware than in the 
compiler.  There is one secondary problem (the double rounding problem) 
that can't be fixed in the compiler at all.  So without a hardware fix, 
there will always be problems with x87 FP support.

Recent hardware has SSE/SSE2/etc FP instructions that do not have the 
excess precision problems that the x87 register stack has.  However, 
using the new registers requires an ABI change, which is unlikely to 
ever happen for 32-bit x86 code.

64-bit x86 code however did change the ABI to use the new FP 
instructions.  The result is that if you generate 64-bit code for an 
AMD64/EM64T processor, then you will get correct FP results.  At least 
this is true for linux.  I don't know about Windows.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com