Re: Loop unrolling

Amelia De Vivo wrote:
>
> Could someone in this newsgroup suggest me a good reading about loop
> unrolling and related register allocation problems?
> Thank in advance
> Amelia De Vivo
> [There was an exhaustive article on software pipelining in Computing
> Surveys a year or two ago. -John]




Note that loop unrolling != software pipelining.


Software pipelining pretty much requires allocation & scheduling
during the unrolling part. You get a faster schedule, less startup
time, less code. But more compile time and more engineering
complexity (ie. bugs).


In loop unrolling, you just clone the loop body a few times (well, you
also need startup/cleanup loop handling). Then the "normal" code
schedule and register allocation takes place. Much simpler to
implement; harder to screw up (you _can_ blow out I-cache, or not
unroll enough but these are straightforward problems to avoid). Code
is a little slower; you pay pipeline penalties once per trip around
the big loop. Might be a wash on compile time, depending on how fast
you can compile big loop bodies.


Cliff


--
Cliff Click Compiler Designer and Researcher
cliffc at acm.org JavaSoft
(408) 863-3266 MS UCUP02-302
--

Return to the comp.compilers page.
Search the comp.compilers archives again.