Version bump plus a doc sweep: the benchmark tables in README and
docs/OPTIMIZATIONS.md were still from before the typed calling convention,
OPTIMIZATIONS listed BEGIN loop promotion as not started, and the
subroutine-threading section of docs/WAFER.md described the memory ABI as
the only one.
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.
Such a word now compiles to a fast entry (i32 x p) -> (i32 x q) carrying
its stack items as WASM values, plus the usual ( -- ) wrapper that keeps
the table slot, so EXECUTE / interpreter / host words / CATCH see the
unchanged memory ABI. Fib(25) 1035 -> 366 us, 4.3x slower than sf64 ->
1.2x; the default guards-on config 1740 -> 361 us. WS-006.
Two reporting bugs found from the browser shell.
An uncaught ABORT printed 'ABORT (throw -1)'. The standard defines ABORT
as 'empty the data stack and perform the function of QUIT', and QUIT
displays no message; gforth and SwiftForth are both silent. It now takes
the same silent path QUIT got in 0.2.5. CATCH still reports -1 and still
restores the stack depth, and ABORT" still prints its text -- different
word, different code (-2).
Compile-only constructs used in interpretation state claimed to be an
'unknown word', which is misleading for a word the system obviously
knows: ABORT", IF, THEN, LOOP, LITERAL, RECURSE and friends. They now
report 'interpreting a compile-only word: <name> (throw -14)', the
standard condition both reference engines give. The check reuses the
existing INTERPRETER_TOKENS table at the point where interpretation has
already failed, so a genuine typo still reports 'unknown word'.
Ships as v0.2.6.
The CORE word was missing. QUIT empties the return stack, enters
interpretation state, restores SOURCE-ID to the user input device and
returns to the interpreter without a message, leaving the data stack
untouched -- that last part is the whole difference to ABORT, which the
standard defines as 'empty the data stack, then QUIT'.
Implemented on the throw plumbing with the standard code -56, so nested
EVALUATE / INCLUDE frames unwind and are abandoned on the way out. Two
places treat -56 specially: CATCH lets it through (QUIT is a return to
the prompt, not an exception) and evaluate() turns it into a silent Ok
after the compile-state wipe it already performs.
Semantics checked against gforth 0.7.3 and SwiftForth sf64, which agree:
the data stack survives, nothing is printed, the rest of the input is
abandoned, and '1 2 ' QUIT CATCH .' prints nothing while leaving 1 2.
Six tests in outer.rs pin it. Deliberately NOT added to the cross-engine
corpus: what QUIT abandons is the input source, and the three engines are
fed differently there, so a comparison would measure the harness.
The gap survived because the Forth 2012 suite skips QUIT by its own
admission, and HELP's coverage lint compares dictionary against docs --
a word missing from both looks complete. docs/wafer-anki.txt had been
documenting QUIT as if it existed.
ABORT itself was already correct: executed while a definition is open it
clears both stacks and returns to interpretation state.
Ships as v0.2.5.
A host word signals failure by throwing across the JS boundary, and the
browser runtime reported that exception with its Debug form, so an
empty-stack RESIZE surfaced as
call_func(134) failed: JsValue(Error: Stack underflow ...)
with the engine's JS stack trace glued on. The thrown message IS the
Forth message, so take it verbatim: 'Stack underflow', the same text the
native CLI prints. Exceptions without a message keep the call context --
those are genuine runtime faults, not Forth throws.
CATCH is unaffected: it reads the throw code from its own channel rather
than parsing messages. Verified against a fresh VM in Node (initSync +
WaferRepl): host-word underflow, compiled-guard underflow, THROW,
unknown word and ' RESIZE CATCH . all match the native CLI.
Ships as v0.2.4.
`wasm-pack build --release` died with "can't find crate" for rustversion,
then thiserror_impl, then every other proc-macro. Cargo strips debuginfo
from release artifacts by default and on macOS that takes the metadata
proc-macro dylibs need to be loadable with it, so rustc could no longer
open them.
Debug builds are unstripped, which is why the whole test suite stayed
green while the browser REPL could not be built for production at all.
Stripping buys nothing for build scripts and proc-macros, so
[profile.release.build-override] exempts them; release binaries stay
stripped.
Also pins wafer-core to 0.2.3 in wafer-web and wafer-cli — both still
asked for 0.2.1. The caret requirement resolved, so nothing broke.
Punctuation (`,` `.` `+` `/` `:` and an embedded `-`) after the leftmost
digit now forces double-cell conversion, so `12.34`, `1,234`, `12:30:45`
and `2026-08-06` convert as doubles. Only a trailing `.` worked before,
and `1.5` was an "unknown word" error.
The punctuation is a double-cell marker, not a fractional point, so the
scale has to travel separately: DPL carries the digit count right of the
rightmost punctuation character (negative when there was none), which is
what lets `<# #>` place the point back on output. NH carries the high
cell a single-cell conversion drops, so a token that overflows a cell is
still recoverable as a double.
parse_number and parse_double_number duplicated the prefix and sign
handling and could not share a DPL counter, so they collapse into one
parse_numeric_literal that reports which kind it converted.
Verified token-for-token against sf64. One deliberate divergence: WAFER
keeps accepting a sign before a base prefix (`-$FF`), which sf64 rejects.
Oleksandr Kozachuk