Add a constant folding pass to UPLC

[ana-pantilie]

Fixes https://github.com/IntersectMBO/plutus-private/issues/2172

[github-actions]

Execution Budget Golden Diff

c8f962a (master) vs 69f54a0

output

plutus-tx-plugin/test/Budget/9.12/applicative.golden.eval

Metric	Old	New	Δ%
CPU	197_308	32_100	-83.73%
Memory	702	300	-57.26%
Flat Size	13	9	-30.77%

plutus-tx-plugin/test/Budget/9.12/monadicDo.golden.eval

Metric	Old	New	Δ%
CPU	197_308	32_100	-83.73%
Memory	702	300	-57.26%
Flat Size	13	9	-30.77%

plutus-tx-plugin/test/Budget/9.12/toFromData.golden.eval

Metric	Old	New	Δ%
CPU	3_182_422	96_100	-96.98%
Memory	13_404	700	-94.78%
Flat Size	153	15	-90.20%

plutus-tx-plugin/test/Budget/9.6/applicative.golden.eval

Metric	Old	New	Δ%
CPU	197_308	32_100	-83.73%
Memory	702	300	-57.26%
Flat Size	13	9	-30.77%

plutus-tx-plugin/test/Budget/9.6/monadicDo.golden.eval

Metric	Old	New	Δ%
CPU	197_308	32_100	-83.73%
Memory	702	300	-57.26%
Flat Size	13	9	-30.77%

plutus-tx-plugin/test/Budget/9.6/toFromData.golden.eval

Metric	Old	New	Δ%
CPU	3_182_422	96_100	-96.98%
Memory	13_404	700	-94.78%
Flat Size	153	15	-90.20%

This comment will get updated when changes are made.

[zliu41]

I don't see why everything in this module needs to be hardcoded to DefaultUni. If PIR constant folding doesn't require this, then neither should UPLC constant folding.

[ana-pantilie]

Is Plinth ever used with anything other than DefaultUni and DefaultFun? Are there any plans to add other universes of builtins? As far as I know there aren't any, and I believe abstraction just for the sake of abstraction is a code smell.

[zliu41]

If we decide that only DefaultUni is needed, the right thing to do is to open a PR and remove the uni type parameter altogether. That doesn't seem like a good idea though, since other universes are being used for testing and benchmarking.

If we keep the uni parameter, then this monomorphization just makes it inconsistent with everything else.

[ana-pantilie]

the right thing to do is to open a PR and remove the uni type parameter altogether.

I don't think removing it from Plinth means we need to remove it from all of the other packages at the same time. Although, I do need to double check if they're used in some other Plinth code as well.

That doesn't seem like a good idea though, since other universes are being used for testing and benchmarking.

I asked Claude where we use other universes, this is the response:

Custom Builtin Function Sets

ExtensionFun — plutus-core/examples/PlutusCore/Examples/Builtins.hs
Used as Either DefaultFun ExtensionFun in example programs and tests. Includes factorial, polymorphic identity functions, cost-tracking builtins, unsafe coercions, and semantic variant testing.

NopFun — cost-model/budgeting-bench/Benchmarks/Nops.hs

No-op builtins used exclusively for cost model calibration benchmarks — measure the raw overhead of calling a builtin function with varying argument types/counts.
Custom Builtin Function Sets ExtensionFun — plutus-core/examples/PlutusCore/Examples/Builtins.hs
Used as Either DefaultFun ExtensionFun in example programs and tests. Includes factorial, polymorphic identity functions, cost-tracking builtins, unsafe coercions, and semantic variant testing.

If Claude is correct, we're keeping the whole universe machinery just for an example and for cost model calibration. I think it would be worth looking into whether the cost model calibration can be represented in a different way. It doesn't make sense to keep all of the complicated universe abstraction around just for that. The current representation would be good if UPLC was used to target other platforms (not just Cardano), but that doesn't seem grounded in reality. Also, there might be performance benefits for the CEK machine. I think this would be worth exploring.

[ana-pantilie]

Btw, I'm reverting the change to Lift here since I tend to agree this would be better done separately, but I still think we should discuss ^.

[zliu41]

If Claude is correct, we're keeping the whole universe machinery just for an example and for cost model calibration.

Cost model calibration sounds quite important to me?

Btw, as a general rule, avoid posting Claude answers verbatim. They are often quite verbose and of questionable quality - inaccuracy is common. You are of course free to use Claude, but make sure to quality-check, shorten and rephrase, make it suitable for human conversation before asking others to read it.

[ana-pantilie]

Cost model calibration sounds quite important to me?

Of course it's important, I'm wondering if there's some other way to represent it.

Btw, as a general rule, avoid posting Claude answers verbatim. They are often quite verbose and of questionable quality - inaccuracy is common. You are of course free to use Claude, but make sure to quality-check, shorten and rephrase, make it suitable for human conversation before asking others to read it.

I'm 100% on the same page on this, so I'm not sure where this is coming from? If I ever post/send something for review which looks like AI slop I probably forgot to check it myself so please let me know.

[ana-pantilie]

Looks like running this pass at the UPLC level as well unlocked further optimization opportunities.

[zliu41]

If Claude is correct, we're keeping the whole universe machinery just for an example and for cost model calibration.

Cost model calibration sounds quite important to me?

Btw, as a general rule, avoid posting Claude answers verbatim. They are often quite verbose and of questionable quality - inaccuracy is common. You are of course free to use Claude, but make sure to quality-check, shorten and rephrase, make it suitable for human conversation before asking others to read it.