Comments on: Load-hit-stores and the __restrict keyword

By: Aliasing, the silent killer » AltDevBlogADay (Staging Site)

Aliasing, the silent killer » AltDevBlogADay (Staging Site) — Sun, 13 Feb 2011 01:18:08 +0000

[...] have been changed somewhere else. This cause a “Load hits Store” (More about it here : http://assemblyrequired.crashworks.org/2008/07/08/load-hit-stores-and-the-__restrict-keyword/). The memory containing item_index is written just before the same memory is read. The CPU just [...]

By: Elan

Elan — Sun, 11 Jul 2010 22:43:55 +0000

“x86″ is a very large family of processors with very different implementations. You can make a prediction about how an Intel Core Duo might behave, but it wouldn’t necessarily be true about an i7 or an AMD chip, because their internal pipelines are dissimilar. The upshot is you need to run your own timings on the processor you’re targeting.

One general prediction I can make is that every x86 processor I’ve timed seems to have a huge latency when shuffling data between x87 float and general-purpose registers.

By: Foo Bar

Foo Bar — Sun, 11 Jul 2010 21:59:36 +0000

Uhm, what about x86?

By: Robert

Robert — Wed, 06 Jan 2010 09:00:47 +0000

Actually, it’s assumed that any pointer can alias a variable of any type unless you compile with a strict-aliasing flag. Look up type-punning.

Here’s a good article: http://cellperformance.beyond3d.com/articles/2006/05/demystifying-the-restrict-keyword.html

By: LHS and the __restrict keyword « Un italiano nella game industry

LHS and the __restrict keyword « Un italiano nella game industry — Mon, 16 Nov 2009 17:09:34 +0000

[...] http://assemblyrequired.crashworks.org/2008/07/08/load-hit-stores-and-the-__restrict-keyword/ [...]

By: Ruskin

Ruskin — Mon, 06 Oct 2008 07:37:07 +0000

Yeah, unfortunately that's one of the many gulfs between what the C++ spec says should happen and what does happen in the compiler. The code in my other entry is what actually came out of VC++; I guess it's more paranoid and assumes the programmer might have aliased any pointer to any other unless you tell it otherwise. GCC on the PS3 does support strict aliasing rules, I'm told.

By: aruslan

aruslan — Sun, 05 Oct 2008 19:04:29 +0000

Yes, only pointers to compatible types and char* can alias.
int* can’t alias float*, and this is the reason to use unions when making dirty int-float hacks.

By: anonymouse

anonymouse — Wed, 03 Sep 2008 03:20:37 +0000

I thought that only compatable types could alias? So for example, an int array can’t alias a float array, and a CFoo* can’t alias it’s own member?

By: realtimecollisiondetection.net - the blog » Posts and links you should have read

Tue, 02 Sep 2008 15:33:38 +0000

[...] Load-hit-stores and the __restrict keyword, Elan Ruskin talks about what a gigantic performance suck load-hit-stores can be on the (pretty [...]