GCC 2.95 and floating point precision

[Dara Hazeghi <sassan@cup.hp.com>]

Hello GCC users, maintainers etc.

I recently downloaded and built your gcc2.95 release (Never trusted RPMS
anyway) on an i686-pc-linux-gnu. I have a program which demands the full
80bit precision in the Pentium FPU. It's written entirely in C and built
and ran fine with gcc2.91.66 (aka egcs 1.1.2). Now however it won't run
due to a lack of floating point precision. I am almost certain of this
due to the fact that with a similar version of the program only
requiring 64bit precision, the calculations work flawleesly (albeit
slower).  This problem only happens with -O2 or -O3 optimization levels.
With -O1 or -O0 I can enable pretty much any other optimizations
supported by x86 (including -ffast-math) and the code builds and runs
fine. Is there any changes between gcc2.91.66 and gcc2.95 that restricts
this use of the Pentium FPU? The program is doing mathematical
computation, so I really need the extra 16 bits! Thank for any insight
on the topic.

On a different topic, is there any way of getting the clobber register
asm code that is so deprecated to build on gcc2.95 (maybe
--fallow-clobber-registers or something). Yes I know you're supposed to
rewrite the code, but what if you can't (I don't know asm). Thanks
again.

Dara Hazeghi