new codegen

Matthew Fluet
Mon, 13 Aug 2001 15:52:26 -0700 (PDT)

> It's really too bad that reduction in code size doesn't correlate
> cleanly with runtime improvement.  Do you guys have a conjecture on
> what's causing the (modest) slowdown from MLton-stable; is it really unfold
> like Henry suggests?

The new codegen that tracks the liveness of stack slots is doing best in
all three of the benchmarks.  As to why vector-concat is faster, it's not
exclusively due to unfold, because that was incorporated into the
implementation of the basis before the 20010806 release.  Now, the speed
up may in part be due to the fact that the loop indices in the unfold are
now being carried in registers instead of being loaded and stored on each