🚀 go-pugleaf

Greetings,

It appears that with 133MHz Pentium systems with pipelined burst
SRAM becoming more common, the performance of gcc on these systems
is beginning to lag considerably that of compilers that support
Pentium specific optimizations.  On a simple loop such as a
DAXPY, I am seeing about 30-33% worse performance than that
quoted for the Intel Reference Compiler and several 32bit
commercial FORTRAN compilers.  (The gcc command used is
"gcc-2.6.3 -04 -m486 -ffast-math -funroll-loops" or
"gcc-i2.6.3 -O4 -mpentium -ffast-math -funroll-loops")

I have used gcc-i2.6.3 with only modest success; I see at
most 5% speedup on some simple modules, but more often than
not things die either while compiling or executing, when
compiling at -O3 or -O4.

In order to speed up something like a DAXPY, scheduling 
optimizations that take advantage of the fxch instruction's
pairing with other floating point operations have
to be implemented.  When I look at the asm output of
gcc-i2.6.3 on simple loops, I don't see these instruction
pairings.

Is anyone already working on Pentium specific 
floating point optimizations for 2.7?

If not, who would be the interested party at FSF
to forward changes to?

Thanks,
Russell Carter