Structured control flow gaps vs standard Forth

[tetsuo-cpp]

Summary

WarpForth uses structured control flow (MLIR scf dialect regions) via a recursive-descent parser, while standard Forth uses a control-flow stack that allows arbitrary interleaving of control-flow words. This means several valid standard Forth constructs cannot be expressed in WarpForth.

Root Cause

WarpForth's parser expects each control-flow construct to match a fixed pattern (IF...ELSE...THEN, BEGIN...UNTIL, BEGIN...WHILE...REPEAT, DO...LOOP). This inherits from MLIR's scf dialect, which only provides structured control flow (scf.if, scf.while, scf.for).

Standard Forth's control-flow stack is just a stack of orig (forward branch) and dest (backward branch target) entries. Words push/pop freely with no nesting requirement.

Missing Features

High impact for GPU kernels

Feature	Standard Forth	Notes
LEAVE / UNLOOP	Early exit from DO...LOOP	Very common in search/scan patterns. Workaround: restructure as BEGIN...WHILE...REPEAT with a flag
EXIT	Early return from a word	Common in utility words with multiple exit conditions. Workaround: single-exit restructuring with deeper IF/ELSE nesting
J	Outer DO...LOOP index	Important for matrix/2D operations — common GPU workloads. Workaround: manually save outer index on the data stack

Medium impact

Feature	Standard Forth	Notes
Multiple WHILEs	BEGIN c1 WHILE c2 WHILE body REPEAT THEN	Well-known idiom for multi-condition exit. Workaround: combine conditions with AND/OR
+LOOP	Arbitrary loop step	Useful for strided iteration. Workaround: BEGIN...WHILE...REPEAT with manual counter

Low/no impact for GPU kernels

Feature	Why
BEGIN...AGAIN	Infinite loops hang the GPU; trivial workaround BEGIN 0 UNTIL
?DO...LOOP	WarpForth's DO already behaves like ?DO (zero-trip safe via scf.for)
RECURSE	Recursion is dangerous on GPUs (tiny call stacks, thread divergence)
CS-PICK / CS-ROLL	Exotic metaprogramming, never used in kernel code
Cross-boundary control flow	e.g. BEGIN ... IF ... UNTIL ... THEN — essentially never written in practice

Possible Approaches

• Incremental: add the high-impact features one at a time within the structured model (e.g. J is straightforward, EXIT/LEAVE may require flag-based lowering)
• Architectural: lower to cf.br/cf.cond_br (the cf dialect) for unstructured control flow, then run a restructuring pass for GPU codegen — major change, likely not worth it given the GPU kernel use case