Re: [Tinycc-devel] vectorize the curent hash implementation

[KHMan]

On 4/23/2016 4:26 PM, Vladimir Vissoultchev wrote:
> [snip snip snip]
> (Note: SSE calc loop oversteps input string so file buffer has to
> overallocated by 8 bytes in `tcc_open_bf`)
>
> Unfortunately there was no noticable improvement, which leads me to beleave
> that hash calculation is not the main culprit for the performance of
> `next_nomacro1`

IIRC there are fast LZ compressors that are trying to use vector 
registers to speed up hashing. But those have nice and small inner 
loops.

But compilers branch all over the place, there are no nice inner 
loops to let L1 caches go hot-rodding, there are many data 
dependencies which limit a CPU's ability to schedule multiple ops. 
Branch and cache mispredictions are like train crashes these days.

Perhaps someone can test the limits on one of those academic CPU 
simulators that has an 'oracle' cache setting. How much speed can 
we further squeeze? I'm not optimistic there is a lot of 
performance left to squeeze...

> -----Original Message-----
> From: Tinycc-devel On Behalf Of Sergey Korshunoff
> Sent: Saturday, April 23, 2016 10:38 AM
> To: address@hidden
> Subject: [Tinycc-devel] vectorize the curent hash implementation
>
> Hi!
> > IMO, the way to go w/ performance improvement is to vectorize the curent
> hash implementation. Instead of computing the hash one character at a time,
> take next 4 characters, process these in parallel and compute 4 hashes
>
> How to implement this?
> There is some hash optimization for the tcc compiled by gcc.
> https://github.com/seyko2/tinycc/commit/b89f0d63af4f494a83c91fb0360d7d37c0f6
> f9a3

--
Cheers,
Kein-Hong Man (esq.)
Kuala Lumpur, Malaysia