Re: anonymous functions optimization - wanted for 3.4.0?

[Jaroslav Hajek]

On Fri, Sep 10, 2010 at 8:56 PM, Olaf Till <address@hidden> wrote:
> On Fri, Sep 10, 2010 at 11:52:44AM +0200, Jaroslav Hajek wrote:
>> hi all,
>>
>> attached is a patch that implements an interesting (IMHO) optimization
>> of anonymous function handles. First a bit of theory:
>> A number of anonymous function handles looks like this:
>>
>> h = @(par1, par2, ...) some_func (arg1, arg2, ....)
>>
>> where each argi is either a reference to some parj, or a constant, or
>> a variable defined in the calling context. Such a function
>> is called a "binder" because it merely takes some other function or
>> handle and binds some of its parameters to certain values.
>> Callback functions are often binders.
>>
>> Examples:
>>
>> @(x) size (x, 3)
>> @(x, y) my_func (x, my_data, y)
>>
>> The attached patch optimizes calls to binders. When Octave constructs
>> an anonymous handle, it attempts to detect if it's a binder. If so, it
>> augments the handle with information how to transform the argument
>> list and the "root" function's handle (or value). It then overrides
>> the call sequence with a fast code that transforms the argument list
>> and calls the root function. This eliminates the overhead of setting
>> up an m-function, altering the call stack, and interpreting the
>> function's body, making binders only incur small additive penalty to
>> the root function's cost.
>
> If somewhere in a stack of "binders" occurs an error (reference to a
> missing argument), will the error message show in which "binder" the
> error occurs? Olaf
>

The optimized code path is only taken if all arguments are given and
thus no error can occur; otherwise, the normal interpreted code is
used.

-- 
RNDr. Jaroslav Hajek, PhD
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz