readArray is faster than unsafeRead

[Simon Peyton-Jones]

That is indeed odd. As I recall unsafeRead doesn't do array-bound checking, and that ought to be faster, always. I don't have time to look into this myself (soon, anyway), but

a) if someone does (look at Core!) I'd be interested to know the outcome

b) unless it turns out to be a red herring, would you like to add a Trac bug report (we really ought to have a category for "performance bug") giving the smallest program you can that demonstrates the unexpected behaviour?

Simon



| -----Original Message-----
| From: glasgow-haskell-users-bounces at haskell.org [mailto:glasgow-haskell-users-bounces at haskell.org] On
| Behalf Of Scott Dillard
| Sent: 29 May 2007 06:22
| To: glasgow-haskell-users at haskell.org
| Subject: readArray is faster than unsafeRead
|
| Hello,
|
| I'm writing some matrix multiplication and inversion functions for
| small matrices (3x3 and 4x4 mostly, for 3d graphics, modeling,
| simulation, etc.)  I noticed that the matrix multiplication was a
| bottleneck so I set out to optimize and found that using unsafeRead
| instead of (!) (or readArray in stateful code) helped a lot. So then I
| went to optimize my gaussian elimination function and found just the
| opposite. unsafeRead is slower than readArray. This struck me as very
| odd considering that readArray calls unsafeRead.
|
| If there is a "good" reason why the compiler optimized readArray
| better than unsafeRead, I'd like to know what it is so that I can make
| all my array code safe as well as fast. (By "good" reason I mean
| something deterministic and repeatable, not just luck.)
|
| On the otherhand, if this is a fluke, I'm inclined to think that it's
| not the safe code which is freakishly fast, but the unsafe code which
| is needlessly slow. That is, something about my program is hindering
| optimization of the unsafe code. What is it?
|
| Attached is the profiling results and a test program with a handful of
| matrix multiplication and gaussian elimination functions to illustrate
| what I've seen. This happens both on amd64 and intel core
| architectures.
|
|
| Thanks for any insight,
| Scott