rustnim — a Rust → Nim transpiler
Status
Milestone 1 is reached: all of base16ct goes through. Every one of its
source files transpiles byte-for-byte as published, alloc half included, and
its decode and encode output is byte-identical to rustc's. 33 differential
cases, 29 behavioural and 4 rejections, plus 6 unit/integration tests. All
green. Run cargo test.
Passing today: functions, impl methods, trait impls (formatting traits and
From), structs, enums (C-like and data-carrying), Option/Result with
?, closures, unsafe, slice iterators (iter/iter_mut/enumerate/zip/chunks_exact/
chunks_exact_mut/windows), borrowed slices as values and return types,
let/let mut, the full integer
and float operator set at exact widths, as casts, if/while/loop/for,
match including patterns that bind, Vec/slices/arrays, type aliases
(including generic ones), function-typed parameters (impl Fn(A) -> B),
multi-file input, #[cfg] evaluation, and println!/format! with {},
{:?}, {:x}, {:b}, positional and inline-named arguments, and
zero/space padding.
Why this exists
We tried tarekwasfy01/Code-Transpiler,
which advertises rust as a source language, on the base16ct crate. It emits
empty files and exits 0. The full investigation is in findings/
and is published at
https://rickub.com/nandi/code-transpiler-rust-frontend-findings
The decisive finding, and the reason this is a new project rather than a patch:
its Universal AST cannot represent Rust. defaultSemanticTypeContract() in
internal/backend/semantic_program.go:85 is hardcoded to
numeric: binary64, integer_width: unknown, truth: r_compatible,
ownership: unknown, index_base: 1
and semantic_document.go:1014 validates that every contract equals exactly
that, while typed_operation.go:46 rejects any value model that is not
tagged_dynamic_binary64. There is no integer width and no ownership in the
model at all. Code like base16ct's constant-time decoder —
ret += (((0x2fi16 - byte) & (byte - 0x3a)) >> 8) & (byte - 47);
— depends on exact 16-bit signed wrapping and arithmetic shift. Lowering that
into a 1-indexed dynamic float64 model produces silently wrong answers. So the
first rule of this project is the one that codebase broke:
Never approximate a semantic you cannot represent. Fail loudly instead.
src/ty.rs already does this: i128/u128 are rejected with a reason rather
than widened or truncated.
Architecture
Rust source ──syn──> syn AST ──lower──> Nim source ──nim c──> binary
The frontend is syn, deliberately. Hand-rolling a Rust grammar is how the
other project went wrong; a correct parser is not the interesting part of this
problem. The interesting part is the lowering, which is where all the work goes.
Planned modules:
| file | role | state |
|---|---|---|
src/ty.rs |
Rust type → Nim type, exact widths, explicit rejections | written |
src/lower.rs |
items, statements, expressions → Nim | written |
src/fmt.rs |
println!/format! format-string handling |
written |
src/prelude.nim |
Option/Result/panic/Display/Debug runtime |
written |
src/main.rs |
CLI: rustnim <in.rs> -o <out.nim> |
written |
tests/differential.rs |
the runner described below | written |
Enums, Option and Result
A C-like enum becomes a plain Nim enum, which compares, orders and
case-checks the way Rust's does. A data-carrying enum becomes a Nim object
variant — a discriminant enum plus one branch per variant — which is the same
shape the prelude already uses for Option and Result. Nim requires the
branches of a variant object to have distinct field names, so each payload
field is prefixed with its variant.
match takes one of two forms. Arms that neither bind nor destructure become
a Nim case, which is exhaustiveness-checked the way Rust's is. Arms that do
bind become an if/elif chain with the bindings emitted as lets, because
Nim's case cannot destructure. The chain always ends in an arm that panics:
Rust proved it unreachable, but Nim cannot see that, and leaving the chain
open would silently fall through instead.
Ok, Err and Some are emitted with their full type arguments
(rsOk[T, E](v)), because Nim cannot infer E from an Ok(v) alone. That is
why the expected type has to reach a match arm as well as a let.
? expands to statements — a temporary, a discriminant test, and an early
return — which are emitted ahead of the line being built. Rust inserts a
From::from on the error there; we accept only the case where the two error
types already agree, rather than assume a conversion is the identity. ? in a
while condition is rejected: the early return would run once before the
loop rather than on each iteration.
Trait impls
A Display impl becomes proc rsDisplay(self: T): string. Rust's Formatter
is a sink and the observable result of {} is exactly the bytes written into
it, so a write through the formatter appends to that string — a fmt body
may write repeatedly, and UpperHex writes once per byte in a loop. A body
that does anything else with the formatter — padding, precision,
debug_struct — is rejected, because those change the output and this model
does not carry them. Debug, LowerHex, UpperHex, Binary and Octal
work the same way.
Writing into a string cannot fail, so ? on a formatter write is a no-op. ?
on anything else inside a fmt body can fail, and format! panics when a
formatting impl returns an error — so that is what the error branch does, with
std's own message.
{:x} on an integer formats its two's-complement bit pattern; on any other
type it calls that type's own LowerHex impl. Those are different operations,
so a radix format on an argument of unknown type is rejected rather than
guessed.
impl From<A> for B becomes a conversion proc that .into() resolves
through. A marker trait with no items generates nothing: we do not model trait
resolution anywhere, so there is nothing for it to affect; a use that actually
needed the trait (a dyn, a bound) is rejected where it appears. Any other
trait impl is rejected.
Methods are keyed by (receiver type, name), not by name alone — two types
may define the same method, and Nim tells them apart by overload resolution on
the first parameter.
fmt::Error is not the same type as a crate's own Error. Collapsing a
qualified path to its last segment merged them, which was a real soundness
bug; core::fmt's types are now recognised by their qualified name.
Slice iterators are resolved to one index loop
Rust's slice iterators are lazy and compose. Nim's for is over one sequence,
so a chain of adaptors is resolved into a small IR and emitted as a single
index loop in which each binding is an lvalue into the original container.
That is what makes *d = v through iter_mut() write back to the caller's
slice instead of to a copy, and what lets chunks_exact(2) hand out a window
that indexes straight into the source with an offset.
Only adaptors with an exact index-loop equivalent are accepted. map,
filter and take_while are rejected rather than partially honoured:
silently dropping an adaptor would change which elements the loop visits.
zip stops at the shorter side, as Rust's does — that is a test, not an
assumption (tests/cases/023).
Borrowed slices are views, not copies
&[T] is a borrow. Nim's experimental view types model exactly that,
including returning one from a proc: writing through the returned view is
visible in the original buffer. That was probed against Nim 2.2.4 before being
relied on, because copying into a seq would print the right bytes while
silently changing aliasing.
Nim does allow a view inside an object and inside an object field — both
probed, both preserving aliasing — so Result<&[u8], E> and
HexDisplay<'a>(&'a [u8]) both work. (An earlier version of this document
claimed otherwise; that was wrong.)
Two real constraints remain. Nim will not let a let borrow out of a local,
so .unwrap()/.expect() on a Result holding a view is expanded inline and
the binding becomes an alias — a view is a reference, so there is nothing to
materialise, and the substituted expression is a plain field access that
re-evaluates nothing. And s.get(a..b) is an Option of a view whose
validity is what matters: the view and its condition travel together through
ok_or until a ? or unwrap resolves them into a bounds check plus a
binding. Keeping such an Option in a variable is rejected with a message
saying so.
A let binding a borrow keeps the view rather than copying into a seq:
let res = encode(..)? names the caller's buffer, and copying would print the
right bytes while silently breaking the aliasing.
Closures and unsafe
unsafe is a permission marker, not a semantic change: it does not alter what
the enclosed operations mean. So the block is transparent, and every operation
inside still goes through the ordinary lowering and is still rejected if it has
no faithful mapping. unsafe fn lowers like any other proc.
A closure becomes a Nim anonymous proc. Nim's closures capture by reference, as
Rust's non-move closures do; a move closure captures by value, which is a
different thing, so it is rejected rather than lowered to the same construct.
impl Fn(A) -> B is left at Nim's default calling convention, which accepts
both a plain top-level proc and a capturing closure — as Rust's impl Fn does.
.map/.and_then over an Option/Result are expanded inline with the
closure's parameter aliased to the payload, rather than handed to a generic
proc. That keeps the whole thing an expression and keeps a view a view.
&str is a borrowed view of someone else's bytes, so it maps to
openArray[char], not to an owned string. Nim accepts a string argument
for an openArray[char] parameter, so a literal still passes straight through.
from_utf8_unchecked reinterprets a byte view as a character view over the
same memory — no copy, no validation, and writes through the original are
visible, as in Rust.
Modules
Rust keeps lower::decode and mixed::decode apart by module; flattening into
one Nim module would merge them — they are different functions. So the first
input is the crate root and each later one is a module named by its file stem,
items are emitted as <module>_<name>, and a call resolves through an explicit
qualifier, then the current module, then what use brought into scope, then
the root.
Generics
Rust type parameters become Nim's. Nim instantiates a generic structurally at
the call site much as Rust does, so fn f<T>(x: T) -> T has a direct target in
proc f[T](x: T): T and no monomorphisation pass is needed.
Trait bounds and where clauses are dropped. That is sound in the
direction that matters: an operation the bound permitted either exists for the
instantiated type or is a compile error at that instantiation site. Dropping a
bound cannot make an accepted program mean something different — it only makes
rustnim accept some programs rustc would reject, which does not matter when
the input is known-good Rust. (Where it would matter is bound-directed
method selection, e.g. blanket impls choosing between candidates. We do not
model trait resolution at all, so such a program is rejected elsewhere.)
Const generic parameters have no Nim equivalent and are still rejected.
Two things need more than a rename. Nim cannot infer an object's generic
parameters from a constructor's field values, so Pair { a: 1, b: 2 } is
emitted as Pair[int32](...) using the expected type — and a generic enum's
unit variant (Holder::Empty) likewise. And a binding's annotation cannot
name a parameter Nim is still inferring, so call sites run a small unifier:
the callee's declared parameter types are matched against the actual argument
types to bind T, and the result is substituted into the return type.
Declaration order
Rust has no declaration-before-use rule and Nim does, so every proc is
forward-declared between the type definitions and the bodies. Reordering the
input instead would not handle mutual recursion.
Type propagation is load-bearing
Rust infers an unsuffixed integer literal's type from context and falls back
to i32; Nim falls back to 64-bit int. So lower.rs threads an expected
type down through every expression — into let annotations, call arguments,
match patterns, compound assignments and both operands of a binary — and
annotates every binding it emits. Without that, let x: u8 = 200; x + 100
means two different things in the two languages. With it, a width the lowering
gets wrong becomes a Nim compile error (a loud failure, reported by the
runner) rather than a wrong answer.
Mapping decisions made so far
- Integers: exact width.
i32→int32,usize→uint, etc.i128/u128
rejected. - Indexing: both 0-based. Direct.
&T→ plain value.&mut T→var Tparameter.&[T]→openArray[T]in parameter position,seq[T]when owned.
Nim::owned()performs that conversion.- Ownership/borrowck: ignored. Nim is GC'd; for safe Rust this is sound.
Option/Result→ object variants in the prelude.match→ Nimcasewhere the arms are simple,if/elifwhen arms have
guards or bindings.- Rust's expression-orientation maps well: Nim
if/caseare expressions
too, and a proc's trailing expression is its return value.
Settled empirically (Nim 2.2.4 vs rustc 1.98.1, both run)
-
Nim's
shron a signed integer is arithmetic, matching Rust.
int16(-256) shr 8=-1in Nim;(-256i16) >> 8=-1in Rust.
base16ct's decoder depends on this, so it maps directly with no helper. -
Nim's fixed-width unsigned arithmetic wraps silently, matching Rust's
wrapping_*.uint8(200) + 100=44in Nim;200u8.wrapping_add(100)
=44in Rust. Sowrapping_addon an unsigned type is just+. -
We model rustc's debug profile. Rust debug builds panic on signed
integer overflow; Nim's default build raisesOverflowDefecton it. Those
are the matching pair, so the runner invokesrustcwithout-Oandnim cwith its defaults, andtests/cases/016pins the behaviour. A Rust
panic exits 101 where a Nim Defect exits 1, so every generated module ends
with a handler that maps one to the other — otherwise the runner's
exit-status comparison would be vacuous.wrapping_*is therefore an
explicit operation on both sides: unsigned maps to the bare operator (item
2), signed is routed through the unsigned view of the same width. -
charround-trips. Rustchar→ NimRune, confirmed for ASCII and
non-ASCII scalars in both{}and{:?}, and acrossas u32
(tests/cases/014). -
Nim's integer conversion
T(x)truncates; it does not range-check.
uint8(511'u16)is255,uint8(300'i32)is44,uint8(-1'i32)is
255— the same answers ascast[uint8]. An earlier version of this
document asserted thatT(x)range-checks, and used that to justify
cast. The conclusion stands —castis the clearer spelling of
"truncate" — but the stated reason was wrong, and it had been assumed
rather than probed. Found by sabotaging the cast lowering and watching the
exhaustive proof not fail, which is what a sabotage test is for.
Still open
- Const generic parameters,
moveclosures, closure bodies with statements,
trait objects andmacro_rules!definitions are rejected with a reason. Atraitdeclaration lowers to nothing — trait resolution
is not modelled — but one giving a method a default body is rejected,
since that body is code with no impl to be emitted into.
Lifetime parameters are not a rejection: they carry no runtime meaning
and Nim is GC'd, sofn encode<'a>(..)lowers fine. saturating_*andchecked_*are implemented, detecting overflow on the
unsigned view of the same width rather than with a range check that would
itself trap.wrapping_*,overflowing_*andstrict_*are not all
covered: only add, sub and mul have the saturating and checked forms.- Float formatting matches Rust for ordinary values and for
inf/NaN, but
the exponent-form thresholds have only been checked at1e21. - Functions are scoped by module now, but types are still global: two
modules declaring the same type name would collide. Relatedly, a crate's
owntype Result<T>is told apart from the builtinResult<T, E>by
arity, which is not how Rust resolves it. - Host
#[cfg]predicates —unix,windows,target_os,target_arch,
target_family,target_pointer_width,target_endian,
target_has_atomic— are evaluated
against the machine, since the generated Nim is compiled for it. That makes
the output host-shaped: a crate branching on platform has had that branch
decided at transpile time.doc/doctest/miriare false. A custom or
build-scriptcfg(crossbeam_loom,target_has_atomic) has no value we
could know and is rejected. - Associated types (
impl Iterator { type Item = .. }) andmod
directories (specialized/mod.rs) are not implemented. String::from_utf8_uncheckedcopies, because Nim'sstringis an owned
value. Rust's consumes theVecwithout copying. Observably the same from
the caller, but it is a copy where Rust has none.
A second crate: does this generalise, or is it fitted to base16ct?
base16ct is the crate this was built toward, so passing it proves less than
it looks. adler2 2.0.1 was picked as a deliberately different shape —
a stateful struct with methods, operator-overload trait impls, a hand-unrolled
four-lane inner loop — and it now works: tests/cases/029-adler2-crate/
transpiles algo.rs byte-for-byte as published, with lib.rs's items and a
driver, and its checksums are byte-identical to rustc's across every single
byte, every length to 600 (crossing the 4-byte unrolling boundary and the
5552-chunk path), and 144 incremental-write splits.
It needed real work, which is the honest part of the answer. Ten features:
trait impls generalised beyond formatting and From (any trait's methods
become procs on the type, with the operator traits wired into +=/+
dispatch), Self, Type::method() static calls, u32::from between
primitives, tuple-destructuring let, split_at, iterators bound to
variables and .remainder(), [0; 4] as an array rather than a seq, and
the bare #[cfg] flags.
It also caught a regression I had introduced: the three-phase emission
added for forward declarations was silently dropping const items declared
inside a function body. base16ct has none, so 33 passing cases said
nothing about it.
What the other crates did
Run without fixing anything, to see where the wall is rather than to move it:
| crate | outcome |
|---|---|
adler2 2.0.1 |
works, byte-identical |
siphasher 1.0.1 |
rejected: u128 |
rustc-hash 2.1.1 |
rejected: u128 |
hex 0.4.3 |
rejected: impl Iterator needs an associated type |
crc32fast 1.5.0 |
rejected: directory modules (specialized/mod.rs), then SIMD intrinsics |
Two of the five stop at u128, which is the founding rule doing its job
rather than a gap: they are told they cannot be translated instead of being
handed a silently truncated hasher. The other two are honest missing
features — associated types, and mod directories.
How far off is a crate like libcosmic?
Measured, not guessed. Running rustnim over libcosmic's own src/:
0 of 164 files produce any translation
50,633 lines, 112 direct dependencies
with 66 generic-parameter blockers, 20 trait objects, 51 async uses, 235
where clauses, 73 associated types and 1,104 lifetime annotations. Those are
not features the crate happens to use; they are its architecture. libcosmic
is a north star, not a next step.
The blocker survey, and what it says about roadmaps
400 crates from the local registry (under 4,000 lines each), all their module
files passed together, first blocker recorded:
| blocker | start | +host-cfg |
+generics | +assoc types |
|---|---|---|---|---|
unevaluable #[cfg] |
124 | 34 | 34 | 34 |
| generic type parameter | 69 | 91 | 1 | 1 |
associated types in an impl |
54 | 63 | 87 | 2 |
| unsupported type | 20 | 29 | 60 | 71 |
| trait object | 17 | 25 | 30 | 35 |
| macro definition | 21 | 24 | 25 | 32 |
| raw pointer | 12 | 22 | 24 | 28 |
| crates fully transpiled | 2 | 2 | 2 | 3 |
This has now happened twice. Evaluating the host #[cfg] predicates cleared
90 of 124 blockers and moved the fully-working count by zero. Generics then
cleared 90 of 91 and moved it by zero again. Every crate each unblocked simply
hit its next blocker.
That is the shape of the problem: blockers are deep, not wide. A frequency
ranking of first blockers is not a roadmap — it says which feature is most
often first, not which one finishes a crate. base16ct, adler2 and
cosmic-theme's spacing model work because their whole stack was ground
through, one blocker at a time.
Does the stack have a bottom?
If the stacks are drawn from one shared finite pool of language features, then
clearing the pool clears every stack at once, and the flat counts above are
just what progress looks like before a step change. That is a real possibility
and it is testable, so it was tested. Sampling every file of the 400 crates
rather than only each crate's first blocker — 1,766 blocker observations:
| count | bounded? | |
|---|---|---|
| distinct normalised blocker kinds | 63 | yes — this is the language-feature pool |
| distinct unsupported std methods | 80 | no — this is std's surface |
| distinct unknown functions | 75 | no — these are calls into dependencies |
| distinct unsupported macros | 13 | no |
So the answer is both, split by class:
- Language features do bottom out. 63 distinct kinds, and each one cleared
is cleared for every crate forever. This part is finite and the step-change
intuition is correct for it. Associated types produced the first net gain
(2 → 3), which is what that dynamic looks like starting. - The API surface does not. 80 distinct
stdmethods appeared in 1,766
samples of small crates;stdhas thousands of items, and each needs a
verified Nim equivalent rather than a guess. That is enumerable but it has
no bottom you reach by clearing features. - Dependencies are not a pool at all. 315 of the 400 crates depend on
other crates, which have to be transpiled too, recursively, until the graph
ends atlibc, proc macros or SIMD intrinsics — which it does, and those do
not translate.
And among the 85 dependency-free crates, 1 of 84 currently transpiles in full.
So even with dependencies removed from the picture, features alone are not the
only remaining gate.
So the ranking above is not a roadmap — it says which feature is most often
first, which is not the same as which feature finishes a crate. The only
honest way to add a crate is to pick it and clear its stack, as was done
twice.
Generics and associated types are now done. By frequency the next are
unsupported types (71), trait objects (35) and macro_rules! definitions (32)
— but see above before treating that as a plan.
cosmic-theme: what was reachable
tests/cases/032-cosmic-theme-spacing/ transpiles corner.rs, spacing.rs
and layout.rs from cosmic-theme 1.0.0 — the spacing scale, corner radii
and density model a COSMIC-native UI needs to match the desktop — with output
byte-identical to rustc's, including the Density/Spacing and
Roundness/CornerRadii round trips.
Those files are the crate's own, with one mechanical change recorded here: the
use serde::{Deserialize, Serialize} line and the Serialize, Deserialize
entries in two derive lists were removed, because the oracle is plain
rustc with no dependencies available. Nothing else was touched; rustnim
ignores both anyway.
The rest of cosmic-theme — theme.rs (1,830 lines), color.rs,
cosmic_palette.rs, derivation.rs, steps.rs, composite.rs — is colour
work built on palette (40,874 lines across 122 files, plus a proc-macro
crate). mode.rs needs cosmic-config and its derive macro. Those are
dependency walls, not language gaps.
bitflags!, and why it is lowered rather than expanded
bitflags is the most-depended-on translatable crate in libcosmic's
resolved tree — 79 of its 741 crates — so it is the highest-leverage target
there. It is also 26 macro_rules! definitions across five files, which is
the one thing the lowering cannot represent.
Expanding the macro does not rescue this. RUSTC_BOOTSTRAP=1 cargo rustc -- -Zunpretty=expanded on a single-flag user produces 869 lines that still call
bitflags::{Bits, Flag, Flags, iter::Iter, iter::IterNames, parser::from_str, parser::to_writer} — items defined by those same macros. The chain does not
end in code we could lower.
What the macro means, though, is small and stable: a newtype over an integer
with named constants and set operations. So src/macros.rs parses the
invocation and the lowering emits that directly. This is a deliberate
exception to "a macro whose expansion is not known is rejected", and the
argument is that the expansion is known here — it is documented, stable, and
now pinned by a test.
tests/cases/034-bitflags.rs is that test, and it is unusual: //@ extern: bitflags makes the oracle compile against the real crate while rustnim
gets no such crate. rustnim has to reproduce bitflags' behaviour without it,
and the outputs are compared byte for byte. Two behaviours it pins that a
reimplementation would get wrong:
!xis complemented and then masked toall()—!(READ|WRITE)is
EXEC, not0xFFFFFFFC.from_bitsreturnsNonefor any bit outsideall();
from_bits_truncatemasks instead.
plus the Debug spelling, which is Perms(READ | WRITE) and Perms(0x0).
The risk this carries is version drift: a future bitflags could change what
the macro generates, and the shim would not know. The test is what would catch
it, which is why it links the real crate rather than a copy of its
documentation.
log, on the same argument
log is second by dependents in libcosmic's tree (61 of 741), and it has
the same shape as bitflags: src/macros.rs is 20 macro_rules!, and that
is what the dependents use. src/lib.rs is the facade — Level,
LevelFilter, the Log trait behind a &'static dyn, atomics and
set_logger.
So the macros are lowered directly, against behaviour pinned from the real
crate by tests/cases/035-log.rs (again //@ extern: log, so the oracle
links it and rustnim does not). What that test pins:
max_level()starts atOff.log_enabled!is false even afterset_max_levelwhen no logger is
installed, because the facade consults the logger as well as the level.Displayfor a level is upper-case (WARN),Debugis not (Warn).Level::Error as usizeis 1 throughTraceas 5;LevelFilter::Offis 0.
A record's arguments are not evaluated when the level is disabled, so the
lowering emits if rsLogEnabled(l): rsLog(l, ..) rather than computing the
message first.
The boundary is deliberate: rustnim models log's emitting side, not its
installing side. A transpiled library's info! calls work and, with no
logger, do nothing — which is exactly Rust's behaviour. Installing a logger
is an application's job and is done from Nim:
rsLogSetLogger(proc (level: RsLogLevel, target, msg: string) =
echo "[", rsDisplay(level), "] ", msg)
rsLogMaxLevel = int(rsLvlTrace)
Modelling impl Log instead would mean reproducing Record and Metadata,
which is more shim surface for something a Nim application would not write in
Rust anyway.
The facade's types are emitted as RsLogLevel and RsLogFilter, not Level
and LevelFilter. The first attempt used Rust's names and broke six existing
cases, because a crate's own Error type and an enum field named Error
cannot coexist in one Nim module.
libc: not transpiled, and should not be
libc is third by dependents in libcosmic's tree (60 of 741), but it is not
a crate to translate. Measured:
129,594 lines across 387 files
54,544 `pub const`
7,660 `pub fn`, nearly all inside `extern "C"` blocks -- declarations
1,926 `pub type`
121 actual function bodies in the entire crate
There is essentially no code in it. It is a set of declarations binding to the
platform's C library, and Nim reaches those same symbols natively — the
symbols are the same objects, not two implementations of one idea.
So what was built is the general capability instead: an extern "C" block
lowers to Nim importc declarations. Both are statements about a symbol
someone else defines, and both are bound by the C ABI, so the two declarations
describe one symbol rather than one being a translation of the other. Raw
pointers map too (*mut T to ptr T), which cleared all 38 raw-pointer
blockers in the survey.
const matters at the C level even though it does not at the Rust one
Rust emits no C prototype; Nim emits a real one. So declaring strlen as
taking *const u8 produces:
error: conflicting types for 'strlen'; have 'NU(char *)'
note: previous declaration with type 'size_t(const char *)'
which is the C compiler catching a declaration that does not match the symbol
— a loud failure, and a better outcome than Rust's silence. To make honest
declarations expressible, *const T maps to a generated const-qualified
alias:
type RsConstPtrcchar* {.importc: "const char *", nodecl.} = distinct pointer
proc strlen*(s: RsConstPtrcchar): uint {.importc: "strlen", cdecl.}
one per element type actually used. Nim's generic form (importc: "const $1*")
was tried first and does not work in 2.2.4. A *const T whose C spelling we
do not know is rejected rather than declared without the const.
tests/cases/036-extern-c.rs calls abs, labs, strlen and atoi through
this path, byte-identical to rustc.
macro_rules!: expanded, not translated
Nim has template and macro, so a shape-level correspondence with
macro_rules! exists. Across 1,833 definitions in a 400-crate sample:
| shape | share | Nim equivalent |
|---|---|---|
| single rule, no repetition | 36% | a template |
| multiple rules | 12% | a macro dispatching on shape |
$(..) repetition |
28% | varargs in a macro |
:tt token-tree munching |
22% | an interpreter; not mechanical |
So ~76% has a translatable shape. We expand instead, and the reason is the
type-directed lowering. A Nim template body is untyped: substituted first,
type-checked after. This lowering needs a type at nearly every step — to
choose div over /, to size a cast, to pick an integer literal's width.
Translating a macro body would mean lowering Rust with no type information,
which is exactly the guessing the project refuses. Expanding at the call site
yields ordinary Rust in a context where the types are known, so it lowers like
anything else. Same applicability, faithful output.
src/mrules.rs handles any number of rules, with $(..) repetition.
Measured against the 1,833 definitions in the sample:
| count | ||
|---|---|---|
| expandable | 1,369 | 74% |
nested $(..) inside $(..) |
44 | 2% |
:tt token-tree munching |
420 | 22% |
Rules are tried top to bottom and the first whose matcher matches wins, which
is Rust's rule and is also what makes a recursive macro terminate — a base
case sits above the recursive rule. Recursion falls out for free: an expansion
containing another macro call is lowered like any expression, so it expands
again, with a 64-level limit so a non-terminating macro is an error rather
than a hang.
A definition it cannot handle is recorded with its reason, so a call site
says "foo! cannot be expanded: nested $(..) repetition is not implemented
yet" rather than "unknown macro". Captured fragments are parenthesised on
substitution, so square!(2 + 3) is 25 and not 11.
Three things here are easy to get wrong, all caught by tests/cases/038 and
039 rather than by reasoning:
>closes=>and->as well as a generic argument list, so the nesting
depth used to find a fragment's end must not go negative. Without that,
$($a:expr => $b:expr),*swallows the whole invocation.- A repetition's separator may be inside the pattern rather than between
iterations —$first:expr $(, $rest:expr)*has no separator, its comma
leads each iteration. A trailing fragment therefore stops at the separator
when there is one and at whatever starts the next iteration when there is
not. - A fragment never spans a top-level
,or;. Without that, a greedy
$x:exprswallows the whole invocation anddescribe!(3, 4)matches the
one-argument rule, silently taking the wrong branch. A comma inside
brackets is within aGrouptoken and so is not at this level at all,
which is why the rule is safe.
A block expression with statements ({{ let mut m = ..; m }}, which is how
these macros are usually written) now lowers to Nim's block: expression
rather than being hoisted, so the same macro expanded at two call sites does
not collide in one scope.
Expansion moved the survey more than everything before it combined: 3 → 18
of 400 crates accepted. A macro_rules! used to be a hard stop at item
level, failing a whole crate on sight.
Repetition and multi-rule then each moved it by zero — 18 before, 18
after, 12 compilable throughout — for the reason every previous feature did:
the crates they unblocked hit their next blocker. That is now eight features
running. The capabilities are real and tested; the crate count is gated by
something else, and the remaining blockers say what: unsupported types (86)
and calls into dependencies (49), which is the unbounded tail rather than the
finite feature pool.
A stricter number
"rustnim exits 0" is not "the output is real". Of those 18, 12 produce Nim
that the Nim compiler accepts:
adler2 arrayref cfg_aliases ×3 cfg-if ×2 ctor-lite darling ×4
Compiling is still not behaving: only the cases in tests/cases/ are checked
against rustc for identical output. Three numbers, in increasing strength —
accepted 18, compiles 12, behaviourally verified only the corpus.
serde: assessed, not attempted
serde is 49 dependents in libcosmic's tree and the most generic crate
looked at here: 17,237 lines, 369 impl<, 922 where clauses, 849 uses of
'de, 324 associated types. serde_derive is another 8,975 lines of proc
macro — a program that runs at compile time, so it can be expanded (as
bitflags! was, with RUSTC_BOOTSTRAP=1) but never translated.
The derive's expansion is small and clean — thirteen lines for a two-field
struct. But it is generic over a Serializer, so unlike bitflags!
(self-contained) and log (a facade with a defined no-op default), it has no
observable behaviour at all until a format crate supplies one. A shim would
therefore have to pick a format and implement that, which is a narrower and
much larger commitment than either previous shim.
serde_json's exact output was pinned for whenever that is attempted:
declaration order, no whitespace, null for None, and a float keeps the
.0 that Display drops — {"x":-3,"ratio":2.0,"maybe":null}.
Proof of byte-identity for base16ct
PROOF.md sets out what is actually established: exhaustive
agreement over every two-byte decode input (65,536), every two-byte encode
input (65,536), every single byte through encode_str and HexDisplay, and
every length to 128 — plus a compositional argument extending those to inputs
of any length, and 20,000 pseudorandom multi-chunk cases attacking the one
step in that argument that is inspection rather than enumeration. Run it with
cargo test --test proof. It is explicit about the difference between the
exhaustive parts and the sampled ones.
Testing: differential, not golden
The bar is behavioural equivalence with rustc, not that the output looks
plausible. For each case in tests/cases/:
rustc case.rs && ./case > expected
rustnim case.rs -o case.nim && nim c -r case.nim > actual
diff expected actual
A case only counts as passing when both binaries build and produce identical
stdout and exit with the same status.
tests/differential.rs implements this, and checks each stage separately so a
failure says where it went wrong: rustnim, rustc, nim, or diff. Three
guards exist specifically because of how the other transpiler failed:
rustnimexiting 0 while writing no output file is a failure.rustnimexiting 0 while writing an empty output file is a failure.- An empty corpus is a failure, so the runner cannot pass by finding
nothing to do.
All three have been verified by deliberately breaking the transpiler and
confirming the runner goes red.
Cases carry directives in leading //@ comments:
| directive | meaning |
|---|---|
//@ reject: <substring> |
rustnim must fail, with this in its message |
//@ skip: <reason> |
not run; reported as skipped |
//@ args: <argv> |
passed to both binaries |
//@ stdin: <line> |
fed to both binaries |
//@ cfg: feature=<name> |
passed to rustnim, and to rustc as --cfg feature="<name>" |
//@ extern: <crate> |
the oracle links this crate; rustnim does not get it |
reject cases are how the "fail loudly" rule is tested rather than merely
stated: 900–904 pin the rejections of i128, an unmapped standard-library
method, a float→int cast, an unimplemented format spec, and a closure.
Run one case with RUSTNIM_CASE=005 cargo test --test differential -- --nocapture. Nim is found at .nim-toolchain/bin/nim in the repository root
or any parent, or via RUSTNIM_NIM.
Toolchain
rustc/cargo1.98.1 — system.- Nim 2.2.4 — vendored at
.nim-toolchain/(gitignored; downloaded from
nim-lang.org, not installed system-wide). Binary:.nim-toolchain/bin/nim.
Milestone 1
Transpile base16ct 1.0.0 — the crate the other transpiler failed on — and
have its decoder produce byte-identical output to the Rust original.
Reached. tests/cases/026-base16ct-crate/ transpiles every source file
of base16ct 1.0.0 — error.rs, lower.rs, upper.rs, mixed.rs and
display.rs, each byte-for-byte as published on crates.io, verified with
cmp rather than by eye — together with lib.rs's decoded_len,
encoded_len and decode_inner verbatim. The alloc half is on, via
--cfg feature=alloc. Output is byte-identical to rustc's:
lower ok abcd1234 len=4 decode: lower, upper, mixed
upper-rej err InvalidEncoding ... upper correctly rejects lowercase
oddlen err InvalidLength / invalid Base16 length <- Debug and Display
encode ok 6162636431323334 len=8 encode, both cases
encode_str ok abcd1234 len=8 closure over unsafe, borrowed &str
Ok([171, 205, 18, 52]) decode_vec \
abcd1234 encode_string > the alloc half
ABCD1234 abcd1234 HexDisplay {:X} {:x}
Everything lowers as written: dst.get_mut(..decoded_len(src)?),
src.chunks_exact(2).zip(dst.iter_mut()), *dst = byte as u8, the returned
&'a [u8] view into the caller's buffer, encode(src, dst).map(|r| unsafe { core::str::from_utf8_unchecked(r) }), and HexDisplay's UpperHex impl
writing once per byte into the formatter.
This is the crate whose six files the transpiler in findings/ emitted empty
output for, while exiting 0.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 |
|