Skip to content

2026.11.22.1: the platform column refreshes only for a refused request (mcpp#996), i686-windows-msvc is a row, one name per index location, an installed toolchain's companions - #1001

Merged
Sunrisepeak merged 13 commits into
mainfrom
round/2026.11.22.1
Nov 22, 2026

Conversation

@Sunrisepeak

@Sunrisepeak Sunrisepeak commented Nov 22, 2026 •

Copy link
Copy Markdown
Member

This PR completes the round described in design 2026-11-22:

  • the platform column's refresh condition (mcpp#996);
  • i686-windows-msvc added as a row of the target table;
  • one name per index location;
  • an installed toolchain's companions, taken from the records the store follows;
  • the bounds of two RunExec cases;
  • mtest 0.11.3.

Changes

  • A miss in the platform column refreshes the index only when the request would otherwise be refused for lack of a platform (mcpp#996, SPEC-013 §14 item 6).
    • Two kinds of build need no platform: a custom triple that a [target.<triple>] section defines, and a planned row built with a named toolchain.
    • Before, each such build synced the index whenever it ran more than the debounce interval after the last sync, adding 12 to 19 s per build.
    • One predicate now decides. It reads the same section lookup and named toolchain that the refusal gates read.
    • platform = "none" in a table ends the lookup, as --platform none does.
    • e2e 1145 fails on the previous engine and passes here.
  • i686-windows-msvc is a preview row, served by the msvc provider.
    • x86_64 Windows runs its programs (WOW64), as e2e 892 and 1144 measure.
    • --target i686-windows and x86-windows-msvc now select it.
    • The target matrix, the canary matrix, the tier tables in both languages and the target-row snapshot state its cells.
  • Two index entries that name one location are refused (index-location-named-twice, SPEC-004 §9 item 12).
    • Before, a member's entry beside its workspace's entry for the same directory registered that directory twice. The install then stopped as ambiguous without naming either entry.
    • The build and mcpp index update now refuse the configuration and name both entries, the location and the remedy.
    • mcpp add reads the index files without registering them, so it is not affected.
    • Covered by unit Config/OneLocationIsOneIndex and e2e 1146. Two e2e fixtures that named one directory under two names now use one directory per name.
  • An installed toolchain's companions are the first declared version the store holds (SPEC-006 §2.9).
    • Two declarations are read in order: the descriptor the payload was installed with (<version>/.xpkg.lua), then the index's.
    • Bug in 2026.11.21.1: it read only the index's declaration. After xim's gcc recipe moved binutils from 2.42.1 to 2.47, a gcc@16.1.0 installed beside binutils 2.42.1 was left with no binutils, and its build stopped at "cannot execute 'as'" (seen in the hermetic job with a restored cache).
    • A home that links the payload from an [xlings] upstream has the upstream's descriptor beside it, but installs the companions the index names. Reading both declarations covers that case too, and the unit test covers both kinds of home.
    • The bootstrap pin stays at 2026.11.13.1. 2026.11.21.1 carries this defect, so .xlings.json moves to this release once it is published.
  • The two RunExec/Queued*OutputRenewsIdleAfterALineHandler cases now bound the property under test, not the producer's start.
    • The macOS failure on main (exit 137, timed out, 0 lines) was the self-started producer not writing its first line within 500 ms.
    • The idle bound is now 2 s, and the handler sleeps past it while lines queue.
    • A failure states when the ready line arrived.
  • mtest 0.11.3 is used for the scaffold and for this repository's tests.

Verification

  • Locally:
    • e2e index (19 cases, 1145 and 1146 among them), deps/names, deps/lock, cli/golden;
    • unit test_config, test_process_run_exec, test_toolchain_triple, test_target_set, test_toolchain_msvc, test_test_protocol, test_xlings, TargetRows/*;
    • static checks: tiers, reason tokens, matrix reasons, version pins, docs style and structure, module wiring, file lengths.
  • e2e 1145 run against the previous engine fails on all three requests: a section, a planned row with --toolchain, and a table's none.
  • Companions in a home linked from an [xlings] upstream: before, two degradations (binutils 2.42.1 and linux-headers 5.11.1 not held); now none, with binutils 2.47 and linux-headers 6.18.64.

Related: the Windows crash of the mcpp-index archive members is an lld-link defect (mcpp#1002). It is fixed in the packages, in mcpplibs/mcpp-index#873.

…uest refused without a platform (mcpp#996)

A custom triple a [target.<triple>] section defines, and a planned row built
with a named toolchain, are built without a platform; their column miss went to
the refresh policy as a refusal, so every such build more than the debounce
apart synced the index. One predicate, reading the section lookup and the named
toolchain the refusal gates read, now decides; platform = "none" in a table
ends the lookup as --platform none does (SPEC-013 §14 item 6, e2e 1145).
…by the msvc provider (mcpp#996)

The msvc probe already answered the 32-bit x86 ABI and Windows CI built and ran
it through a [target.i686-windows-msvc] section (e2e 892). The row is data: the
msvc provider serves it, and x86_64 Windows runs its programs (WOW64), which
e2e 892 and 1144 measure. The target matrix and the canary matrix declare its
cells.
…index-location-named-twice)

A member's [indices] entry beside its workspace's [workspace.indices] entry for
the same directory registered the directory twice, and each of its packages
was two candidates of one name: the install stopped at the provider as
ambiguous and named neither entry (measured on mcpp-index). The build and
mcpp index update refuse it with the location and the remedy (SPEC-004 §9,
e2e 1146).
…cer's start

The producer is the test program started again; on a loaded macOS runner it
wrote no line inside a 500 ms idle bound (exit 137, timed out, 0 lines). The
idle bound is 2 s and the handler sleeps past it while the lines queue; a
failure states when the ready line arrived and how many lines were read.
@Sunrisepeak Sunrisepeak added the ci:canaries Run the canary matrix (canaries.yml) on this pull request label Nov 22, 2026
…escriptor declares, not the index's current one

xim's gcc recipe moved binutils from 2.42.1 to 2.47 on 2026-11-21. mcpp
2026.11.21.1 read the declaration from the index, so a gcc@16.1.0 installed
beside binutils 2.42.1 was given no binutils after an index update and its
build stopped at "cannot execute 'as'" (measured on the hermetic e2e job with
a restored home). The payload's own .xpkg.lua, kept by the installer, is read
first (design 2026-11-21 §3.5: the payload's install record).
…ays until the release that fixes the companion choice
…led descriptor's and then the index's

A home that links a payload from an [xlings] upstream carries the
upstream's descriptor beside it and installs the companions the index
names: there the installed descriptor names a binutils the store does not
hold. Both records are read in order and the first the store satisfies
decides; the payload's own declaration stays the first.
…he target rows record the round's additions

names/162 and lock/203 named one directory under two index names, which
the build now refuses. The golden surface lists the reason
index-location-named-twice; the target-row snapshot records
i686-windows-msvc, which --target i686-windows and x86-windows-msvc now
select. The refusal is stated for the commands that register indices.
…lection of import slots (mcpp#1002), the fixtures, the template audit
@Sunrisepeak Sunrisepeak added the ci:full Run the full e2e set on this pull request (read when a push starts the run) label Nov 22, 2026
@Sunrisepeak Sunrisepeak changed the title 2026.11.22.1: the platform column refreshes only for a refused request (mcpp#996), i686-windows-msvc is a row, one name per index location 2026.11.22.1: the platform column refreshes only for a refused request (mcpp#996), i686-windows-msvc is a row, one name per index location, an installed toolchain's companions Nov 22, 2026
… states (WOW64); SPEC-016 §2.4 states the row

The windows unit lane's PlatformRun/TheRunnerReachesTheArtifactsArchitecture
expected a same-arch runner to refuse an i686 artifact on every system
but Linux; the execution table now states windows-x86_64 -> i686.
@Sunrisepeak
Sunrisepeak merged commit 367a593 into main Nov 22, 2026
98 checks passed
@Sunrisepeak
Sunrisepeak deleted the round/2026.11.22.1 branch November 22, 2026 12:32
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci:canaries Run the canary matrix (canaries.yml) on this pull request ci:full Run the full e2e set on this pull request (read when a push starts the run)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant