Plutus Tx: avoid duplicating code for DEFAULT case bodies

[michaelpj]

Turns out we were duplicating code for a slightly subtle reason. GHC can have just a single DEFAULT case that handles a ll the non-explicitly-handled cases. But we need to make all the cases explicit, which means that we were duplicating the DEFAULT case for all the non-handled cases. And of course, the moment we see "duplicating code" it's a short hop to "exponential code", and indeed that's what I was seeing for code that has many nested default cases (which it turns out you get from using pattern synonyms, apparently).

The fix is simple: just share the code.

[zliu41]

There's a dramatic decrease in formula1 (size) but a 7.8% increase in formulaBudget. Can you look into it?

[michaelpj]

I took a look at the PIR diff. It's not at all obvious what's worse: the changes are exactly what we'd expect (there was a lot of duplicated code from this, in fact). My guess is that

• The formula PIR has a lot of default cases
• The default cases are dynamically dead, i.e. we don't actually hit them
• Since there are many default cases, the inliner doesn't inline the default body binding
• So we now pay extra for the binding and don't benefit

The inlining decisions seem correct to me, since we would pay for it with a substantial increase in program size (as we can see). So I think we basically just get the opportunity to do a cost/size tradeoff differently, and we choose to not increase size, which seems fine.

[bezirg]

Great note