13573 would have included one optimization which #16650, et seq., did not, namely the optimization of for-loops over integral ranges with steps other than 1 and -1 in resumable code computation expressions (like task). I figured I might as well record it here (cc @psfinaki).
It seems like it might be relatively straightforward to add this optimization now by wiring up the IntegralRange and mkOptimizedRangeLoop constructs exposed in #16650.
Supplement the existing match on IntegerForLoopExpr with another match on CompiledForEachExpr and IntegralRange:
However, the existing optimization for for-loops over int32 ranges with steps of 1 or -1 doesn't always seem to kick in — in fact, I haven't been able to get it to kick in at all, even when using the syntax for n = start to finish do … instead of for n in start..finish do …. So figuring out why that is, and what the TAST actually looks like by the time it gets to LowerStateMachines.fs, might end up being the harder problem to solve.
13573 would have included one optimization which #16650, et seq., did not, namely the optimization of
for-loops over integral ranges with steps other than1and-1in resumable code computation expressions (liketask). I figured I might as well record it here (cc @psfinaki).It seems like it might be relatively straightforward to add this optimization now by wiring up the
IntegralRangeandmkOptimizedRangeLoopconstructs exposed in #16650.Supplement the existing match on
IntegerForLoopExprwith another match onCompiledForEachExprandIntegralRange:https://github.com/dotnet/fsharp/blob/5fd6800a984c59267c406305c15e9f7efb1e8519/src/Compiler/Optimize/LowerStateMachines.fs#L485-L486
There is an existing translation using
mkIntegerForLoop; add a new one that usesmkOptimizedRangeLoop:https://github.com/dotnet/fsharp/blob/5fd6800a984c59267c406305c15e9f7efb1e8519/src/Compiler/Optimize/LowerStateMachines.fs#L692
However, the existing optimization for
for-loops overint32ranges with steps of1or-1doesn't always seem to kick in — in fact, I haven't been able to get it to kick in at all, even when using the syntaxfor n = start to finish do …instead offor n in start..finish do …. So figuring out why that is, and what the TAST actually looks like by the time it gets toLowerStateMachines.fs, might end up being the harder problem to solve.