A compiler is refused only when the realisation needs a token it cannot take - #671
Merged
Merged
Conversation
repositories hit the same shape within a day of 2026.9.18.1 Three regressions from 2026.9.18.1, all one shape: the engine treated a target already satisfying a [c-abi] declaration and a target being unable to realise it as the same case, so it either refused a compiler it never needed to, or refused a target it could have realised. 1. GCC on Linux was refused for a package merely DECLARING [c-abi], before realisation ran at all. openkal-musl 0.15.0's own declaration on Linux (presents = "posix", data-model = "arch-default", wchar = 32) realises to nothing -- Linux's default triple already is that -- so nothing about it needed Clang. Every GCC user of openkal-musl on Linux lost the package for a substitution nothing asked GCC to perform. Fixed by reordering: cenv::realise (a pure function of the declaration and the target, not the compiler) runs first, and the Clang-only gate fires only when the result is non-empty. Windows is unchanged: the Cygwin-flavoured substitution there is non-empty, so a non-Clang compiler is still refused, naming what it cannot do. 2. macOS does not satisfy presents = "posix" by default -- caught by the verification probe on a real build (declared __unix__ defined, measured undefined), which is exactly the class of error that probe exists to catch. Apple's clang predefines __APPLE__/__MACH__ on its default triple, never __unix__; the original design text assumed macOS already presented it the way Linux does. 3. A freestanding target had no realisation for presents = "posix" at all, refused outright, which blocked openkal-llvm-runtime on riscv64-none-elf before it reached any other target. 2 and 3 dissolve into the same one-line rule, simpler than what it replaces: realising presents = "posix" means the same observable fact everywhere -- __unix__ defined, _WIN32 not -- and targets differ only in what it costs. Linux: nothing. macOS and freestanding: one token, -D__unix__ -- neither already defines it, for the identical reason. Windows: the Cygwin-flavoured substitution. Known cost, left to the mcpp-index 30-member measurement to weigh: #ifdef __unix__ ... #elif defined(__APPLE__) now takes the Unix branch on macOS too. Tests: test_cenv.cpp pins the GCC/Linux-accepted and GCC/Windows-still- refused halves via cenv::realise directly (MacosPosixArchDefaultDefinesUnix, FreestandingPosixDefinesUnix, FreestandingWindowsIsStillRefused); test_cenv_probe.cpp adds the token cenv::realise now produces for macOS/ freestanding handed to a real compiler, proving the probe agrees with it; e2e 742 reproduces the GCC/Linux regression against a real build (fails without this fix) and pins GCC/Windows still refused. Version bump: 2026.9.18.1 shipped these regressions; downstream needs a release with the fix, so this is 2026.9.18.2.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three regressions from 2026.9.18.1, found by openkal-musl and
openkal-llvm-runtime CI within hours of that release, all one shape: the
engine treated "the target already satisfies this declaration" and "this
declaration cannot be realised" as the same case.
Separate from #670 on purpose — that PR is a release-machinery fix
(
release.yml); this is an engine fix. Mixing them would make either harderto review, and this one gates three downstream repositories, so it needs to
ship as its own release regardless of #670's timing.
The three regressions, and the one rule they dissolve into
GCC on Linux was refused for merely DECLARING
[c-abi], beforerealisation ran. openkal-musl 0.15.0's own Linux declaration
(
presents = "posix", data-model = "arch-default", wchar = 32) realisesto nothing — Linux's default triple already is that — so nothing about
it needed Clang. Every GCC user of openkal-musl on Linux lost the package
for a substitution nothing asked GCC to perform.
macOS does not satisfy
presents = "posix"by default — caught bythe verification probe on a real build: declared
__unix__defined,measured undefined. Apple's clang predefines
__APPLE__/__MACH__onits default triple, never
__unix__.A freestanding target had no realisation for
presents = "posix"atall, refused outright — blocking openkal-llvm-runtime on
riscv64-none-elfbefore it reached anything else.(1) is fixed by reordering:
cenv::realise(a pure function of thedeclaration and the target, not the compiler) runs first, and the
Clang-only gate fires only when the result is non-empty. Windows is
unchanged — its Cygwin-flavoured substitution is non-empty, so a non-Clang
compiler is still refused there.
(2) and (3) dissolve into one rule, simpler than what it replaces:
realising
presents = "posix"means the same observable fact everywhere —__unix__defined,_WIN32not — and targets differ only in what itcosts. Linux: nothing. macOS and freestanding: one token,
-D__unix__—neither already defines it, for the identical reason. Windows: the
Cygwin-flavoured substitution.
Known cost, not hidden: portable code written
#ifdef __unix__ ... #elif defined(__APPLE__)now takes the Unix branch onmacOS too. Left to the mcpp-index 30-member measurement to weigh, per the
coordinator's own framing — flip it there with data if the cost exceeds the
benefit, not here in advance.
Tests
tests/unit/test_cenv.cpp—MacosPosixArchDefaultDefinesUnix(wasMacosPosixArchDefaultIsANoOp, now asserts the-D__unix__token andexpectation pair),
FreestandingPosixDefinesUnix(new — wasFreestandingAcceptsOnlyNone, which assertedpresents = "posix"on afreestanding target REFUSED; that assertion was the bug and is gone),
FreestandingWindowsIsStillRefused(new — pins thatwindowsonfreestanding is unaffected).
tests/unit/test_cenv_probe.cpp—TheDUnixTokenCenvRealiseProducesForMacosAndFreestandingSatisfiesItsOwnExpectation:the exact
-D__unix__token + expectation paircenv::realisenowproduces, handed to a real compiler, produces no mismatch — the half
cenv::realise's own pure-function tests cannot check.tests/e2e/742_gcc_accepted_when_the_realisation_is_empty.sh(new) — legA reproduces the GCC/Linux regression against a real build with
openkal-musl's own declaration (fails without this fix, confirmed
locally before committing); leg B pins GCC (mingw) on Windows still
refused with the existing diagnostic.
Version
2026.9.18.1 shipped these regressions; bumped to 2026.9.18.2 so
downstream has a release with the fix.
check_version_pins.shpasses.Verification
Full
mcpp test: 123/123. Both e2e legs above pass. Doc lint(
check_docs_style.sh+check_docs_structure.sh) clean. docs/22 (+ zhmirror) updated with the corrected mapping table and the reasoning for both
corrections. CHANGELOG carries a
## [2026.9.18.2]section (verifiedagainst the extraction logic PR #670 hardens, so this release's notes will
not come out empty either way).