x86 g77 performance vs f2c-gcc, recent snapshot

x86 g77 performance vs f2c-gcc, recent snapshot

[N8TM]

gcc code is performing considerably better than g77 code on several of the 
higher performing Livermore Fortran Kernels (7,8,21,23).  For the most part, 
this seems to be associated with the generation of too many pointers and 
consequent spills in the g77 code.  

gcc and g77 share occasional problems with -O2 coming out much slower than 
-Os, although these have diminished in recent weeks.  That is associated with 
attempts to store floating point variables both in the floating point stack 
and in 64-bit integer registers, which also has detrimental effects on 
consistency of rounding.  If the culprit loops are isolated into a subroutine 
by themselves, they generate efficient code, not even using the entire 
floating stack, but the compiler has trouble fitting them into the stack when 
in the context of larger subroutines.  

I will be investigating whether f2c-gcc shares the failure at -ffast-math of 
g77 subroutines which incorporate both max() and min() fortran intrinsics.