One way down: the nightly, Bridge-IR and the rustc_ast path, and why they went

Date: September 12, 2026 Status: history. Nothing here is normative and nothing depends on it to know what a program means. The decisions that stand are ADR-001 D1, ADR-003 D1 and D2, and ADR-004 D1; the release entry is ../CHANGELOG.md. The measurements are elsewhere and are unrevised: nightly-cost.md, subprocess-cost.md, std-sysroot.md, technical_notes.md.

The specification and the ADRs are written as though none of this existed, which is what docs/README.md §2 requires of a withdrawal as against a supersession. This page is where the route went. It is here so that a person who finds bridge in a two-year-old commit message, or wonders why a language that lowers to text ever needed a nightly, gets an answer instead of an archaeology project.

---

1. What the three things were

They were one design in three parts, and they only make sense together.

Bridge-IR — crates/bridge-ir, 85 lines. A pure-data intermediate representation: BridgeModule, BridgeFunction, BridgeStruct, BridgeLetStmt, BridgeExpr with Literal, Variable and Call. No NodeId, no interned strings, spans as Range<usize> byte offsets, serde on everything. It was the narrow waist: a language frontend would target Bridge-IR and never a compiler data structure, so a rustc change reached the executor and stopped there, and a frontend could be tested by asserting on the IR it produced with no compiler in the loop. A construct with no direct Rust spelling was to be lowered in the frontend into ones that had, rather than widening the protocol.

The rustc_ast path — crates/rustc-executor. It took a BridgeModule and built a rustc_ast::Crate in one pass, so that the mapping for a construct existed in exactly one place and a syntactically invalid program was unrepresentable — the types refuse it, where a string builder emits it and fails later somewhere that does not name the cause. Text was then a print of that crate (rustc_ast_pretty::pprust) rather than a second code generator, and a rustc subprocess compiled the printed text. The exit that was kept open: hand the same crate to rustc_interface::run_compiler and skip rustc’s parser altogether.

The pinned nightly — bridge-toolchain/rust-toolchain.toml, channel nightly-2026-01-01, components rustc-dev, llvm-tools-preview, rustfmt, clippy, installed and used through scripts/bridge-toolchain.sh. It existed for one reason: rustc_ast lives behind #![feature(rustc_private)], which is nightly-only and carries no stability promise, so the crate that linked it had to be tested against exactly one snapshot and shipped as a unit with it. A second CI leg built the whole workspace on that channel, because a backend that is not built against its pin is a backend that rots.

2. What was measured, when, and by whom it was measured for

Every number below was taken for a different question and is written up in the notes named beside it. None was taken to justify the withdrawal.

what	number	where
what Bridge-IR could carry	1 of 15 corpus programs	subprocess-cost.md §5
the printed-Rust round trip	0.161 % of a full compile (2.77 M of 1 719 M Ir)	subprocess-cost.md §4
starting the rustc process	0.48 % (8.3 M Ir), two thirds of it the rustup shim	subprocess-cost.md §3–§4
the two together — what the in-memory exit would buy	0.64 %	subprocess-cost.md §4
-Zpolonius=next over the corpus	0 verdicts changed at +7.11 % of a build	nightly-cost.md §3
the pinned toolchain against a stable one	1 420 MiB / 24.7 s against 602 MiB / 12.3 s	nightly-cost.md §4
what the nightly cost a user who only runs programs	2 MiB — and a non-relocatable binary	nightly-cost.md §4
the compiler in std’s build graph	58 of 97 packages, 29 s for a hello-world	std-sysroot.md

Two of those deserve their own sentence.

One of fifteen. Bridge-IR had no control flow — no if, no while, no match, not even a binary operator. hello_world.nika reached the backend; let_assignment.nika was refused by name, because println with a variable argument had nowhere to go in BridgeExpr; and every examples/*.nika lowers to a grammar! { } macro invocation, which Bridge-IR had no node for. So the waist was narrow in the wrong dimension. A narrow waist is a discipline — this much, and constructs without a spelling get desugared before they arrive. An empty one is not a discipline; it is an absence with a rule written on it.

Not relocatable. A binary built with the backend linked librustc_driver-<per-build hash>.so by absolute RUNPATH into the build machine’s rustup home, and the hash in the soname differed per build, so a different install had a differently named file and therefore the wrong one. The same compiler built without the feature linked libc/libgcc and went wherever it was put. This is the number that did the most work, and it is not a percentage.

3. The decision, and that no further measurement was taken

The owner withdrew all three without a further measurement, knowingly. That is worth stating plainly, because this project’s habit is the opposite — a decision here usually waits for a number, and several of the records above exist only because somebody refused to decide without one.

It was the right call anyway, for a reason the numbers themselves show: there was no measurement left that could have changed the answer. Each of the four findings answers a different question, and they agree.

• If the concern was speed, the exit was worth 0.64 %, and three quarters of even that was process start rather than the text round trip.
• If the concern was reach, the path compiled one program in fifteen.
• If the concern was the borrow checker — the one language capability the nightly could have bought — the flag changed zero verdicts and cost 7 % of every build.
• If the concern was what a user installs, the pin cost two megabytes and cost them a binary that only runs beside the toolchain that built it.

A measurement is worth taking when its outcome would change what you do. Nobody could name one here. What a further round would have bought is confidence about a decision already unanimous across four independent axes, at the price of the weeks it takes to grow Bridge-IR far enough to be worth re-measuring — which is itself the work being declined.

What it costs to have decided this way, stated rather than hidden. The argument the withdrawn design made was never disproved: a narrow waist between many languages and one backend is a good shape, and if Nikaia ever hosts a second language and finds text a bad interface, this page is the record of a design that was half-built and what it would take to build again. That question is open in the ordinary sense that any unbuilt thing is open. It is not open in the records, because the records owe a reader the reason the current answer is the answer.

4. Things that were true and are now only history

Kept because each cost somebody a day to find out.

• A default nobody measures is a default nobody runs. While bridge was the default backend it panicked on every input — Ident::from_str interning a symbol with no rustc_span session installed, a scoped-tls panic before it printed a line — and this survived undetected because cargo check cannot tell a lowering that works from one that aborts on its first symbol. It was found while measuring something else.
• The AST was print-only. Every node carried NodeId::from_u32(0) and DUMMY_SP, where rustc’s parser writes DUMMY_NODE_ID and expansion asserts it before assigning a real id. So the “free to keep open” exit was not free: it needed a node-id discipline, a SourceMap in the lowering, and a second route for diagnostics, whose whole mechanism is translating rustc --error-format=json about a generated file that an in-memory compile does not write.
• The subprocess compiled at the wrong edition — no --edition was passed, so the printed crate compiled at 2015 while it had been interned and printed for 2021. Nothing Bridge-IR could express told the two apart, which is why it went unnoticed and why fixing it changed no byte of output.
• Two toolchains cannot share one target/deps. Two rustcs’ rlibs in one directory, and a test harness that reads that directory to find the crates a generated program links cannot tell a stable winnow_grammar from a nightly one. scripts/bridge-toolchain.sh gave the nightly build a target directory of its own for that reason, not for tidiness.
• The test harness held the workspace’s only -Z. rustc -Zls=root read an rlib’s dependency list, so the suite needed the nightly even where nothing it compiled did. The same question is answered without a flag by asking which candidate rlib’s filename hash occurs in the anchor’s metadata — which is what crates/nikaia/tests/common/mod.rs does now, and it is the reason a stable-only run of the corpus became possible at all.
• rustc_private internals as found on that nightly — Item not holding its own identifier, Fn holding it, P<T> being Box<T>, rustc_ast::Lit not being rustc_ast::token::Lit — are in technical_notes.md, unchanged.

5. What replaced it, in one sentence each

• The interface is Rust source text (ADR-003 D1): a frontend emits bytes, and nothing else crosses the line.
• The lowering is crates/nikaia/src/emit, one pass from the Nikaia AST to Rust (ADR-004 D1) — which is what --backend rust always was, and is now simply what the compiler does.
• The toolchain is stable, named in rust-toolchain.toml, and there is one CI leg (ADR-001 D1).
• The orchestrator stayed, because it was never Bridge-IR’s: it is crates/orchestrator now, and it owns the cache, the lockfile and the Cargo wrapping (ADR-003 D2).
• Group B.2 is answered by desugaring in the frontend onto entry-style APIs (ADR-005 D2), which is what the corpus already writes unprompted, and not by a nightly-only flag.

This site is open source. Improve this page.