[fpc-devel] Return to pure functions!

J. Gareth Moreton gareth at moreton-family.com
Thu Sep 19 20:25:48 CEST 2019


Hi everyone,

The Curious Kit is back!  While I'm still waiting on my optimiser 
overhaul and the node semantic pass to be approved (or rejected), I've 
gone to taking a fresh look at pure functions, especially as I know more 
about the node pass stage (after my work on the semantic pass).

I've managed to resolve the merge conflicts and the compiler at least 
recognises the 'pure' directive and does some preliminary checks.

One thing I've fallen foul of is how to track variables in what will 
become the 'node parser' that the compiler will use to evaluate pure 
functions to calculate their return values.  Because a variable can be 
anything, I was planning on using the Variant type, mainly because it 
contains a union that has everything I need in terms of recording what 
type a variable is and how its value is stored, and I would only end up 
duplicating that. However, the bootstrapper can't find the 'variants' 
unit.  Given that no other part of the compiler uses the Variant type, 
is there a rule against using it, or a way to help Make find the 
'variants' unit?

There's still a few things I'm working out, like originally I thought 
that if a function has a 'var' parameter, then it cannot be pure, but 
this is not always the case, because if the actual parameter (what 
you're passing into it when calling said function) has a value that is 
deterministic, then you can theoretically evaluate the function at 
compile-time so long as it doesn't break any rules like accessing a 
file.  Most likely this will be an on-going project that I'll build in 
segments, and I'll make a full design specification to go with it, at 
least for my own benefit.

Submitting everything at once at the end may not be the best approach 
for this, as I or others might find future optimisations for pure 
functions later on (like ascending a pure function call outside of a 
for-loop if it doesn't depend on the loop counter, so it is only called 
once and its return value referenced as needed later on).  My thought on 
a first patch is to have the compiler at least recognise the 'pure' 
directive even if it doesn't yet do anything.  Is that permissible?

Gareth aka. Kit


---
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus



More information about the fpc-devel mailing list