[MLton] Shortening dependency chains in the x86 codegen

Vesa Karvonen vesa.karvonen at cs.helsinki.fi
Sun Jan 7 13:06:25 PST 2007


Quoting Florian Weimer <fw at deneb.enyo.de>:
[...]
> Ah, okay.  Have you actually measured that this is a problem?

Not in this context.  However, this is well understood issue.  See,
for example, the optimization manuals by Agner Fog

  http://www.agner.org/optimize/#manuals

or the book

  Computer Architecture - A Quantitative Approach
  Hennessy & Patterson

which discuss the related concepts (various dependencies, scheduling with
a scoreboard, etc.).

The reason I brought this up is that I believe that it could give a small
overall performance improvement for a very small effort.

> (I'm curious how you'd discover such code generation issues.)

I've spent quite a bit of time optimizing code at the assembly language
level on various processors.  So, when I look at assembly language code,
I spot issues like this pretty quickly.

-Vesa Karvonen



More information about the MLton mailing list