Re: LuaJIT with vectors

[David Given <dg@...>]

On 2010-11-19 15:35, Mike Pall wrote:
[...]
> This is often suggested, but rarely beneficial. All modern malloc
> implementations already do exactly that: they keep lists of blocks
> for common sizes and recycle them.
[...]
> Instead one should tune the malloc implementation to the usage
> patterns of the language. This is what I did with the built-in
> allocator in LuaJIT (a heavily modified dlmalloc descendant).
[...]

I'll second this: in an earlier life as an embedded systems engineer, 
when porting our product to a new platform the very first thing we would 
do would be to wrap the host malloc()/free() implementation with our own 
allocator, which grabbed big chunks from the host allocator and did our 
own allocation inside. The reason for this was that most embedded malloc 
implementations suck beyond my ability to describe in a family-friendly 
forum. We'd get substantial speed improvements *entirely* due to this 
optimisation.

So, while it's true that any decent malloc implementation should do 
this, there are an awful lot of malloc implementations that are not 
decent! (*cough*Windows*cough*) Anyone who cares about memory allocation 
speed should probably using their own memory allocator.

--
┌─── ｄｇ＠ｃｏｗｌａｒｋ．ｃｏｍ ───── http://www.cowlark.com ─────
│
│ life←{ ↑1 ⍵∨.^3 4=+/,¯1 0 1∘.⊖¯1 0 1∘.⌽⊂⍵ }
│ --- Conway's Game Of Life, in one line of APL