This is an issue I ran into working on #738 where calls to a (side-effecting) function for marking files to run or not were being removed. I think this is a minimal case that shows the issue:
let testMain = lam f. f dprint in
testMain (lam call. call 1; ())
which when running boot --debug-dead-code-elim pprints as (modulo whitespace)
let testMain = lam f. f dprint in
testMain (lam call. ())
Note that call 1; is removed. The important factor seems to be that we have a function bound via lambda parameter or pattern match, that we then call it, and that we're not in the right-hand side of a let or recursive let.
I'm not entirely sure what an appropriate fix would be. Two options/possible causes:
https://github.com/miking-lang/miking/blob/e1cba70b28112cbcac2e9b31c805f1c69374978e/src/boot/lib/deadcode.ml#L10-L36, in particular line 23, seems to assume that a variable bound in lambda or pattern match cannot contribute to a side-effect, which is true until it happens to be a function that we call, in which case it might be false. However, just saying true here means that the mere act of referencing a parameter is treated as a side-effect, which might be an issue.
I think the code is supposed to only prune top-level definitions, i.e., it should only traverse the spine of the program, which isn't what the final, default, sfold case does (it descends into anything, in this case a function-call. This could be assisted by something like #826, which would make spine-traversal easy.
Addendum: I found a smaller case, but I'm not sure it's truly illustrative. The program lam f. f 1; () is deadcode eliminated to lam f. (), but of course the two are not observably different since they're never called.
This is an issue I ran into working on #738 where calls to a (side-effecting) function for marking files to run or not were being removed. I think this is a minimal case that shows the issue:
which when running
boot --debug-dead-code-elimpprints as (modulo whitespace)Note that
call 1;is removed. The important factor seems to be that we have a function bound via lambda parameter or pattern match, that we then call it, and that we're not in the right-hand side of aletorrecursive let.I'm not entirely sure what an appropriate fix would be. Two options/possible causes:
truehere means that the mere act of referencing a parameter is treated as a side-effect, which might be an issue.sfoldcase does (it descends into anything, in this case a function-call. This could be assisted by something like #826, which would make spine-traversal easy.Addendum: I found a smaller case, but I'm not sure it's truly illustrative. The program
lam f. f 1; ()is deadcode eliminated tolam f. (), but of course the two are not observably different since they're never called.