long double and i386 optimization

[Jim Wilson]

This is probably due to a long standing problem with the gcc x86 port.  The
x86 port lies about the FP hardware, which makes it difficult for gcc to emit
well optimized and correctly optimized x86 FP code.  Gcc does not know that
values in registers are always long double, so it can not take this into
account when generating or optimizing code.

This whole problem is very involved and very complicated.  Fixing it will
probably require many changes to the x86 port (to stop lying), and many
changes to the machine independent code (to better support targets where
values in registers are always in the largest supported FP format).

Jim