[ fix #2927 ] Do not eagerly evaluate expressions occurring inside `%delay` block

[AlgebraicWolf]

Description

This PR fixes #2927.

Thanks to @stefan-hoeck for his assistance in Discord. He suggested fixing the issue by using Scheme's delay and force to memoize constants, as in #2807 instead of #2817 that is currently used. However, this solution comes with a performance overhead.

To reduce the overhead, I modified CSE to track whether an expression occurs inside a %delay block. If that is the case, the expression is lifted as memoized lazy value using Scheme's delay and force. Otherwise, it becomes an eagerly evaluated top-level constant. Benchmarking the bootstrapped compiler on idris2api.ipkg showed 5% slowdown on my machine, which is an improvement compared to the 8.5% slowdown in #2807.

Should this change go in the CHANGELOG?

• [x] If this is a fix, user-facing change, a compiler change, or a new paper implementation, I have updated CHANGELOG.md (and potentially also CONTRIBUTORS.md).

[CodingCellist]

Is this related to #2791 ? : )

[AlgebraicWolf]

Is this related to #2791 ? : )

No, not really. Moreover, this PR uses proper memoization instead of relying on #2791, as it yields better performance in this specific case

I guess the only way they are related is that both the need for #2791 and discovery of #2927 happened because of an attempt to manipulate large lazily-constructed data structures

[gallais]

This makes sense to me however it would be nice to know what the cause of the slowdown is. I don't think we're using many top level lazy constants in the compiler.

[mjustus]

Seeing that the current CSE is unsound with respect to %delay, I think this should be merged unless our backend maintainers (@stefan-hoeck, in particular) have concerns.

[stefan-hoeck]

Seeing that the current CSE is unsound with respect to %delay, I think this should be merged unless our backend maintainers (@stefan-hoeck, in particular) have concerns.

To the best of my knowledge, we have (a) test(s) to verify that non-delayed top-level constants get converted to constants correctly. If these tests still pass (they do), I'm all for merging this.

[AlgebraicWolf]

I have rebased this PR, all the tests still do pass. Perhaps it's worth re-measuring compilation time as well

[mjustus]

Thanks @stefan-hoeck! I am merging this now and am happy to take the blame if something goes wrong.

[andrevidela]

Software's not about blame! Don't feel pressure to mess up something.