filectx: strange performance behavior

[Marc Strapetz]

On 12.05.2015 03:52, Matt Mackall wrote:
> On Fri, 2015-05-08 at 20:19 +0200, Marc Strapetz wrote:
>> As part of an extension, I'm using following code which takes several
>> seconds to execute for a large repository (e.g. setup.py of the
>> Mercurial repository):
>>
>>     filectx = repo['.'][path]
>>     filectxs = []
>>     for ctx in filectx.ancestors():
>>       filectxs.append(ctx)
>>
>>     # this loop takes a long time
>>     for ctx in filectxs:
>>       ui.write(str(ctx.rev()) + '\n')
>>
>> Interestingly, rewriting the loop to the following code gives results
>> almost instantly:
>>
>>     filectx = repo['.'][path]
>>
>>     # this loop executes almost instantly
>>     for ctx in filectx.ancestors():
>>       ui.write(str(ctx.rev()) + '\n')
>>
>> Also, when adding an additional rev(), the original loop becomes fast:
>>
>>     filectx = repo['.'][path]
>>     filectxs = []
>>     for ctx in filectx.ancestors():
>>       ctx.rev() # <-- this makes the loop fast
>>       filectxs.append(ctx)
>>
>>     # now the loop executes almost instantly
>>     for ctx in filectxs:
>>       ui.write(str(ctx.rev()) + '\n')
>>
>> Probably this is expected, but my Mercurial/Python knowledge is too
>> limited to understand that ... so any hints are appreciated :)
>
> Short answer: it's magic.
>
> Long answer: the mapping between file revisions and changesets is not
> 1:1, so determining which rev() goes with a context is not always
> trivial. One of your paths makes this work easier, because when we
> create each new filectx, we know roughly how we got there in changeset
> terms. But the other way, we don't have that information readily
> available yet because we haven't asked to compute it.

Thanks, Matt. I guess as a rule of thumb one should *create and use* a 
(file) context at the same time and use it as short/compact as possible 
to enjoy possible performance improvements?

-Marc