This issue is about any expression that can yield an early return: ?, return, continue, break.

The engine has a module side_effect_utils.ml that deals with any expression with side effects: its approach is to hoist every side effect within an expression.

After applying this side effect phase, expressions are trees with let, if, match, loop nodes, and the side-effects occur only on rhs of lets.

Note early return is one of the different possible side effects. Also, one cannot rewrite early returns without considering other side effects: for instance, early returns can interact badly with mutation.

From there:

• CfIntoMonads lifts early return bottom up into various monads.
• we also want a phase that attempts to restructure control flow, to push early returns to end positions.

Road map

• resugar questions marks (Rustc rewrite ? with returns);
• add a phase that transforms early returns into monadic constructions;
  • test this monadic phase;
  • fix bugs;
• add a phase that attempts at reformulating control flow to avoid early returns (i.e. we'd better do if c {return e1} return e2 -> return (if c {e1} {e2}) than a monadic transformation) (see Add a RewriteControlFlow phase (return/continue/break) #393 and Engine: strip automatically return when in end position #254);
• wrap the monadic and reformulation phase into another phase that decides heuristically which looks best.
• support continue/return within loops: return/break/continue within loops #196

Original issue

This is related to the question mark operator, since e? actually expands to a match whose one branch is a return statement.

TODOs:

• (x) add a phase that transforms early returns into monadic constructions;
• fix the current monadic encoding Fix monadic transform for ?s #399;
• add a phase that attempts at reformulating control flow to avoid early returns (i.e. we'd better do if c {return e1} return e2 -> return (if c {e1} {e2}) than a monadic transformation) (see Add a RewriteControlFlow phase (return/continue/break) #393);
• add a phase that tries both monadic and control-flow reformulation phases and decides heuristically which seems better;
• (x) figure out how to translate code containing the ? operator nicely (some resugaring?).