Floating point trouble with x86's extended precision

[Chris Lattner]

> The best pragmatic solution is probably to set the rounding control to
> 64-bits...

Actually, the easiest thing to do is to change the register allocator to
spill values in their 80-bit form.  This at least eliminates the
unpredictable extra rounding that you get because of register pressure.
We use this strategy in the LLVM x86 backend, and it seems to work pretty
well.  This obviously is not a complete solution, but it's one easy
step...

-Chris

-- 
http://llvm.cs.uiuc.edu/
http://www.nondot.org/~sabre/Projects/