A recursive Forth word almost always opens with a guard that returns early,
so every leaf of the recursion costs a call whose whole body is that test.
`Call(self)` now compiles as `<guard> IF <what the guard returns> ELSE
Call(self) THEN`, which is what the callee would have done on entry anyway.
Half of fib's nodes are leaves: Fibonacci(25) 356 -> 237 us, 1.24x sf64 ->
0.83x, so all five benchmarks now beat it.
The guard runs twice along the recursive path, hence the bounds: at most six
effect-free operations, at most four call sites, never a tail call. WS-018.
Promotion was all-or-nothing per word, so one `.` or one host call put the
whole body -- hot loops included -- on the memory data stack, where a
loop-carried add costs 2.2 ns/iteration instead of 0.31. The stack simulator
now runs over each promotable stretch of a word; BEGIN/UNTIL, BEGIN/AGAIN and
BEGIN/WHILE/REPEAT join DO/LOOP as promotable when the construct is provably
stack-neutral; and the inliner no longer moves a loop-bearing callee into a
caller that can never be promoted.
Fixes a bug the BEGIN work uncovered, present since promotion was introduced
and shipped in 0.2.6: the loop fixup and the IF join copied locals one slot at
a time in index order, so a body that permutes the stack lost a value --
`: C 3 4 2 0 DO SWAP LOOP . . ;` printed `4 4` where gforth prints `4 3`.
Four of five benchmarks now beat sf64: Factorial 0.29x, Collatz 0.30x,
NestedLoops 0.27x, GCD 0.67x. Only Fibonacci is behind, at 1.24x. Also scale
GCD, Factorial and NestedLoops, which ran in 14-51 us where scatter and fixed
costs dominated -- that is what exposed GCD as a loss and pointed at BEGIN.
WS-014, WS-015, WS-016, WS-019.
Rust 1.95 promoted collapsible_match and map_unwrap_or; CI runs
-D warnings so they break the build. Collapse nested `if`s into
match guards across codegen/optimizer/export, and swap
map().unwrap_or(..) for map_or / is_ok_and.
`{: F: x F: y :}` now declares float-typed locals that live on the float
stack. `x x F* y y F* F+ FSQRT` writes real float code without manual
FSTACK juggling — previously WAFER had a 100%-compliant float wordset
but no way to name intermediate float values.
New IR ops `ForthFLocalGet(n)` / `ForthFLocalSet(n)` alongside the
existing int-local ops. Each kind has its own index namespace so mixed
declarations like `{: n F: f :}` compose cleanly. Codegen allocates f64
WASM locals after the existing f64 scratch pair; the fsp-bridge logic
mirrors the existing FDup/FSwap path.
Outer interpreter tracks a parallel `compiling_local_kinds` alongside
`compiling_locals` (keeps the 18 existing touch-points unchanged) and
extends `{:` to recognize `F:` as a per-next-name type marker. `TO` and
name resolution branch on kind to pick Int vs Float get/set ops.
Four tests: classic hypot, TO round-trip, mixed int/float args, and
uninitialized float via `|`. Inline-inhibit for the new ops added to
optimizer and is_promotable so they don't sneak into contexts that
would collide with the caller's WASM locals.
Major compliance push bringing WAFER from 3 to 10 passing Forth 2012
compliance test suites (Core, Core Extensions, Core Plus, Double,
Exception, Facility, Locals, Memory, Search Order, String).
Compiler/runtime fixes:
- DEFER: host function via pending_define, works inside colon defs
- COMPILE,: handle_pending_compile in execute_word for [...] sequences
- MARKER: full save/restore with pending_marker_restore mechanism
- IMMEDIATE: changed from XOR toggle to OR set per Forth 2012 spec
- ABORT": throw -2 via THROW, no message display when caught
- M*/: symmetric division to match WAFER's / behavior
- pending_define: single i32 flag → Vec<i32> queue for multi-action words
- Optimizer: prevent inlining words containing EXIT or ForthLocal ops
- +LOOP: corrected boundary check formula with AND step comparison
- REPEAT: accept bare BEGIN (unstructured IF...BEGIN...REPEAT)
- Auto-close unclosed IFs at ; for unstructured control flow
- _create_part_: use reserve_fn_index to preserve dictionary.latest()
Memory layout:
- Separate PICT_BUF and WORD_BUF regions to prevent PAD overlap
- Updated DEPTH hardcoded DATA_STACK_TOP in boot.fth
New word sets:
- [IF]/[ELSE]/[THEN]/[DEFINED]/[UNDEFINED]: conditional compilation
- UNESCAPE/SUBSTITUTE/REPLACES: string substitution (host functions)
- Locals {: syntax: parser, ForthLocalGet/Set IR ops, WASM local codegen
- ENVIRONMENT? support for #LOCALS (returns 16)
- N>R/NR>/SYNONYM: programming-tools extensions
- Search Order: ONLY, ALSO, PREVIOUS, DEFINITIONS, FORTH,
FORTH-WORDLIST, GET-ORDER, SET-ORDER, GET-CURRENT, SET-CURRENT,
WORDLIST, SEARCH-WORDLIST with full multi-wordlist dictionary support
via Arc<Mutex> shared state for immediate effect from compiled code
Remaining: 1 cascade error in Programming-Tools from CS-PICK/CS-ROLL
(unstructured control-flow stack manipulation, requires flat IR).
When the tail-call pass converted a Call to TailCall inside an If branch,
and the inliner subsequently inlined that word, the TailCall was not
converted back to Call in nested control-flow bodies. The TailCall codegen
emits a Return instruction, which would exit the *caller* instead of just
the inlined callee — silently corrupting the return stack.
Root cause: the inliner only converted top-level TailCalls in the body
(line-by-line iteration), missing TailCalls nested inside If/DoLoop/Begin
structures.
Fix: add detailcall() that recursively walks the entire IR tree and
converts all TailCall ops back to Call before inlining.
This unblocks defining complex Forth words (like SM/REM, FM/MOD) that
use DABS → DNEGATE → D+ chains with return-stack operations inside
conditional branches.
426 tests pass (including new regression test).
Stack-to-local promotion (Phase 1: straight-line code):
- Words with no control flow/calls use WASM locals instead of memory stack
- Stack manipulation (Swap, Rot, Nip, Tuck, Dup, Drop) emits ZERO instructions
- ~7x instruction reduction for arithmetic-heavy words like DUP *
- Pre-loads consumed items from memory, writes results back at exit
Consolidation recompiler (CONSOLIDATE word):
- Recompiles all IR-based words into single WASM module
- Direct call instructions instead of call_indirect through function table
- Cranelift can inline and optimize across word boundaries
- All control flow variants support consolidated calls
342 unit tests + 11 compliance, all passing.
Inlining: store IR bodies for all words, inline Call(id) when body <= 8 ops
and non-recursive. Convert TailCall back to Call when inlining (tail position
in callee is not tail position in caller -- found via compliance test failure
where inlined TailCall caused unreachable code after the call site).
DSP global caching: cache $dsp in WASM local 0 at function entry, use
local.get/set throughout, writeback before calls and at function exit.
Reduces global access instructions by ~30-40%.
323 unit tests + 11 compliance, all passing.
Optimizing Forth 2012 compiler targeting WebAssembly with IR-based
compilation pipeline, multi-typed stack inference, subroutine threading,
and JIT/consolidation modes. Rust kernel with ~35 primitives and Forth
standard library for core/core-ext word sets.
Oleksandr Kozachuk
ok2