transf: lower all unlabeled break statements

[zerbina]

Summary

Fix the cases where unlabeled break statements were inserted by
transf, which means that unlabeled break statements now no longer
exist after the main transf pass. This allows for:

• removing unlabeled break support from the MIR and CgNode IR
• simplifying code generation and MIR processing
• simplifying the closureiters transformation

Replacing skLabel usage in the code generators also becomes easier.

Details

Since they're directly associated with a block, labeled break
statements are easier to handle for analysis and code generation. For
this reason, transf already transforms all unlabeled break
statements into labeled ones, by:

• assigning labels to blocks that don't have a user-provided one
• wrapping while and for statements with named blocks

However, there were two occurrences where unlabeled break statements
((nkBreakStmt (nkEmpty))) were injected by transf:

• for the exit of for statements of closure iterators (liftForLoop)
• by the closureiters transformation when turning while cond loops
  into while true loops (although not strictly part of transf)

As a consequence, unlabeled break reached all the way into the code
generators, requiring dedicated support by the MIR, MIR processing, and
code generation.

While mirgen could lower unlabeled break statements, the association
of a while to its surrounding injected block no longer exists, and
mirgen is generally not well suited for non-local (i.e., not applying
to a single node) transformations.

Instead, the occurrences where unlabeled break statements are inserted
are adjusted. For the for statement, the label of the wrapper-block is
passed to liftForLoop, where the injected nkBreakStmt can then use
it. However, for the other occurrence, a more complex solution is
needed.

Since the while lowering the closureiters pass conditionally applies
(which is where the unlabeled break is injected) is the same as the
one mirgen automatically applies when generating repeat statements,
the transformation from mirgen.genWhile is turned into a PNode-based
transformation that is applied in transf.transformWhile. Some extra
adjustments are required:

• PNode doesn't have a dedicated "scope" construct like the MIR does.
  It is emulated via a nkBlockStmt
• the closure-iterator transformation doesn't support yield statements
  within if/elif conditions, so directly placing the former while
  condition into an if condition would regress on yield support. Prior
  to placing the expression into the if condition, nkStmtListExprs
  are thus first unpacked into the loop body

Only while true loops exist after the main transf pass now.

  while cond:
    body
  # is transformed into:
  block label:
    while true:
      if not cond:
        break label
      block: # a separate scope is needed for the clean-up semantics
        body

closureiters simplifications

Multiple simplifications are possible now that neither unlabeled
breaks nor while loops with non-true-literal conditions exist
at the point when the closureiters pass is run.

• remove transformBreaksAndContinuesInWhile, which handled unlabeled
  breaks and continue statements in while loops
• remove the now-obsolete statement-list lowering logic for
  nkWhileStmts (lowerStmtListExprs)
• remove the nkWhileStmt condition handling from the to-basic-block
  transformation (transformClosureIteratorBody)
• remove the nkWhileStmt handling from preprocessing (preprocess)

MIR simplifications

• remove the "unset" sentinel value support from LabelId; all breaks
  in the MIR are labeled now
• remove everything related to unlabeled MIR breaks from MIR processing
  (mirexec and cgirgen)

Code generator simplifications

Each code generator required special handling for associating unlabeled
breaks with their targeted construct (while or block), with
cgen and jsgen being affected the most. All of it is removed.

As a side effect, the output of jsgen for loops slightly improves: no
extra Label: is emitted around every while.

[zerbina]

My plan is to eventually turn the closureiters pass into a MIR-based transformation, which, besides simplifying it and fixing the places where yield is not properly supported, would also allow for removing the statement-list-expression unpacking from transf again.

For the extraneous blocks, I have a follow-up planned where mirgen will omit the mnkBlock (but not the mnkScope) for nkBlockStmt/nkBlockExpr nodes where the label doesn't have the sfUsed flag included. This will elide all mnkBlocks for blocks that are not used for control-flow.

[saem]

The increased regularity of the code is really good, it's silly looking at how much code was just devoted to carrying around that ambiguity.

[saem]

ah, due to the simpler generation the need for bookending with the endNode drops too, cool.

[saem]

/merge

[github-actions]

Merge requested by: @saem

Contents after the first section break of the PR description has been removed and preserved below:

---

To-Do

• write a proper commit message

Notes for Reviewers

• this is soft-blocked by sem: reject illegal break/continue statements #915. Without said PR, illegal break statements would now cause crashes where they previously lead to proper errors
• in addition, this is a preparation for removing/replacing skLabel usage from the code generators