[Mesa-dev] [PATCH 0/6] Constants through builtins evaluation rewrite

[Kenneth Graunke]

On 05/02/2012 02:11 PM, Olivier Galibert wrote:
>    Hello,
>
> Here is the second version of that patch series.  It's fundamentally
> the same thing with a number of changes:
> - respect the coding style a little better
> - index the hash table on the variable object pointer and not the name
> - handle function call in function call (tested with atan(a))
> - handle if (tested with atan(a, b))
> - handle conditional assign (tested with mix)
> - fix the access to the body when the same builtin is used multiple times
>
> What's missing is a piglit run, which I currently have technical
> problems with[1].
>
> Thanks everybody for the reviews, I hope I haven't missed anything.
>
>    OG.

These look great!  I've gone ahead and pushed them with a few tiny changes:

1. I updated a comment in patch 2, which still said the hash_table 
mapped using variable names/strings.
2. I changed the for (exec_node ...) loop in patch 4 to foreach_list, as 
it's simpler.

This /does/ cause a regression in Piglit test glsl-const-folding-01, due 
to atan(1.0) being slightly off from the value the test expects.  But I 
checked that, and it's off in the 5th digit, so it's just a precision 
issue.  I think the test is being a bit picky, so we may want to change 
it.  Or, make our built-ins more precise.  We can do that in a follow-up 
patch sometime.

Either way, it's definitely a plus that we're using the same code for 
the CPU and GPU side.  Much better to be consistent.

Thanks so much for doing this!