Here's a revised version of the patch, incorporating comments from Richard and 
Mark.

I did some fiddling around with CSiBE to further improve the cost estimate for a 
call.  Setting the base cost to 4, rather than 3 as in the original patch, 
produced smaller code both with and without -mabicalls.  With o32 abicalls, 
which Richard indicated would be the worst case scenario for call overhead, base 
cost of 5 gave similar size results to 4 (overall a tiny decrease, but it looked 
like just as many individual tests went up as down), but code size went up for 
everything else.  So I think 4 is probably a reasonable value without having to 
conditionalize it for -mabicalls.

OK to commit this time?  I did the i686 bootstrap on this version of the patch 
as well as MIPS testing.

-Sandra