Commit acb9f52
* docs: eight reports after 2026.9.27.1, the design and the implementation plan
* fix: a host-module lib root is ordered with the units of its package (#720)
The lib root was placed at the head of the host-module compile list before
the list was sorted by imports, so a lib root that imports a sibling unit
was compiled first and failed. The lib root is now the first node of the
same sort; it is emitted first whenever it imports nothing of its package,
so existing packages keep their order. e2e 807 covers an explicit and a
conventional lib root.
* fix: an index that requires a newer mcpp is a closing tip, not an error
The read site printed the E0006 text as error: at the start of every run
that read such an index, including runs that then resolved every package
elsewhere and exited 0. The fact is now recorded without printing:
- a run that fails carries the E0006 text in the message that stops it;
- a run that refreshed an index and met a floor ends with one tip: line,
printed after the command's output (closing notices, mcpp.ui), and an
envelope reports it as the note MCPP_INDEX_REQUIRES_NEWER_MCPP;
- mcpp self doctor lists every index whose floor this mcpp does not meet;
- a tree recorded as unusable answers no later lookup either;
- the E0006 upgrade note starts on its own line.
Unit tests for the read site, the hint, the notices; e2e 812.
* docs: regenerate the design-record index
* W1 (mcpp#725): a rooted workspace reaches its own path dependency, and -p resolves the package first
A rooted workspace (root [package] + [workspace]) built as itself never set
state.wsManifest / state.runtimeWorkspaceRoot, so a member reached through the
root's own [dependencies] path entry was loaded as an ordinary path dependency:
its x.workspace = true entries were refused or silently unresolved, and it
received none of [workspace.package] / [workspace.build]. The workspace context
is now set in that branch too, before any dependency is loaded, matching the
member-switch branch and satisfying "a member is a member however it is
reached" (graph.cpp's depIsMember).
-p, --package <NAME> promised a package but both resolvers
(manifest.cpp's inline loop and project.cppm's resolve_member_dir) matched only
a member's directory basename or path. They are now one function in
project.cppm, resolving in order: a member's qualified name
(<namespace>.<name>), then its bare package name (refused, naming every match,
when shared by two or more members), then its directory path or basename (kept
as a fallback; a duplicate basename keeps today's first-match selection but now
warns naming the others). A value that is one member's package name and a
different member's directory selects the package, with a warning naming the
other member.
SPEC-004 §9 item 1 names the missing position and states that -p resolves the
package identity first; docs/07 §5.3 (en, zh) and the four `-p` help strings
state the resolution order.
Tests: six unit tests for the resolver in test_workspace_inheritance.cpp
(package/path/basename all selecting one member, same name under two
namespaces, package-outranks-directory with a warning, duplicate-basename
warning, the "not found" listing, and the unaffected no-filter cases); e2e 805
(a rooted workspace's own path dependency, with a non-latest pinned version,
an omitted package.version, a [workspace.build] flag, and a shared root
[toolchain], served from a project-local index so the build touches no
network) and e2e 806 (-p's three resolution steps end to end). Both e2e
scripts fail on the released 2026.9.27.1 binary and on 2026.9.26.1, and pass
with this fix.
Co-authored-by: speak-agent <agent@mcpp-community>
* W6 (#723): one destination, one content, one writer for a deploy target
Two or more source paths for one deploy destination no longer collide at
planning. `add_deploy` (src/build/plan.cppm) merges them into one
`DeployFile` entry instead of refusing a second source path; a `BuildPlan`
whose deploy has exactly one source still produces the byte-identical
`build.ninja` line it always has. The merged destination becomes one
`stage_file` edge with every source as an input (src/build/ninja_backend.cppm),
and `mcpp stage` places it once all sources agree byte-for-byte
(`stage_files`, src/build/stage.cppm), otherwise failing and naming every
source and the destination. `cmd_stage` (src/cli/cmd_build.cppm) accepts one
or more source positionals.
One destination, one writer: `place-dlls`'s post-link DLL placement
(src/pack/pack.cppm `place_runtime_dlls`, invoked from
src/cli/cmd_publish.cppm `cmd_place_dlls`) now receives the set of names the
merged deploy list already places beside a program, computed in
ninja_backend.cppm and passed through a per-edge `$placed` ninja variable so
the generated graph stays stable. `place_runtime_dlls` skips those names
instead of overwriting them, comparing content and reporting a difference as
a warning.
SPEC-007 R4.2 and R4.3 are amended to state the content check and the
single-writer rule; docs/04 (en, zh) is corrected to match. e2e 810 covers
the merge and the build-time refusal; e2e 811 (`# requires: windows`, not run
here) covers the single-writer rule; e2e 646's collision case is updated for
the new build-time message. Unit tests cover `stage_files`'s multi-source
behaviour and the deploy-list skip/warn logic in `place_runtime_dlls`.
Co-Authored-By: Claude <noreply@anthropic.com>
* 2026.9.27.2: on Windows an xlings invocation leaves the process environment as it found it and starts in the registry's home (#726)
On Windows, build_command_prefix prepended the registry's subos/default/bin
to the process PATH and set XLINGS_HOME process-wide, and ran xlings in
mcpp's working directory. After a build installed a payload, every action
found xim:llvm's cl/link/lib/rc shims in front of MSVC's tools, and vcpkg's
compiler detection failed; a project with a .xlings.json at its root also
received the shims of mcpp's toolchain and payloads in its own SubOS.
ScopedInvocationEnv now applies XLINGS_HOME, the scope variables and the
PATH prefix for the invocation and restores them afterwards, and the Windows
prefix starts with `cd /d "<home>" &&`, as the POSIX prefix does.
Refs #726.
* T3 (#724 W3/W4/W5): a device source is not a compile unit; a failed build
program's own diagnostic survives; emit writes no project file
W3 (src/build/plan.cppm): the compile-unit loop now skips
SourceKind::Device graph units, so build.ninja carries no dead cxx_object
edge for a rule-claimed device source, and compile_commands.json / the S1
document agree without their own filter (unit_invocations already excluded
only NASM; nothing else needed to change). The source stays in `watch` and
still reaches the package's build program through MCPP_DEVICE_SOURCES,
since both read the manifest's sources glob directly, not plan.compileUnits.
Checked every other consumer of plan.compileUnits (prepare/plan.cpp's
dependency-cache collection keys by path, not by index, so it is unaffected
beyond a smaller artifact set for a package with device sources).
W4: a package whose build program failed under `emit`'s plan_only records
MCPP_BUILD_DATABASE_PROGRAM_FAILED and applies none of its directives
(state.cppm: new PrepareState::programFailedPackages, set at both call
sites in target_side.cpp and features.cpp). The device-source orphan check
in target_side.cpp now skips such a package outright, instead of reading
every device source as unclaimed and failing the whole member. Also: notes
a phase recorded before prepare_build's own failing return are no longer
silently dropped (driver.cpp: a thread_local sink in the same per-run-sink
style as mcpp::build::refusal, exported as
mcpp::build::take_notes_on_failure); cmd_build.cppm's emit failure path
folds any such note into the one diagnostic SPEC-005 R5.2 allows a wholly-
failed member (path stays the member's mcpp.toml, exactly one entry), so
the true cause is not lost behind a downstream symptom without violating
that invariant. hasProgram's existing exists(build.mcpp) check is now
correct by construction, since a failed program's package never reaches it.
W5 (src/build/prepare/xlings.cpp): ensure_project_index_dir's two calls
under a private work_dir are collapsed into the one call the
ownerRoot==workRoot branch always made, targeting workRoot in every case.
Previously the runtime-environment half (deps/subos/workspace) went to
runtimeSelection.ownerRoot, which is always the real project root
regardless of emit's private work_dir -- so `emit build-database` wrote
<root>/.mcpp/.xlings.json into a project that declares [xlings] deps.
SPEC-005: R3.7 names device sources beside NASM units (both absent from S1
and compile_commands.json, for different reasons -- NASM is a compile unit
excluded from export, a device source is never a compile unit at all). R5.2
gains the sentence that a check whose premise is a build program's
directives does not run for a package whose program failed in this pass.
R2.1 needed no change.
Tests: e2e 808 (device source: no dead ninja edge, absent from both
databases, and the rule still compiles it and the build still runs), e2e
809 (a device source plus a build.mcpp that does not compile: PROGRAM_FAILED
with path build.mcpp, no device-source mention, package still described),
e2e 817 (688's project-tree digest repeated on a stub-xlings fixture with
[xlings] deps: byte-identical tree, no .mcpp/.xlings.json, no write-project
effect, and the private work directory does gain one naming the
dependency). Each fails against the released 2026.9.27.1 binary and passes
on the fresh build. Full regression: all 15 emit/build-database e2e
scripts, 798, and four more that exercise the compile-unit loop (asm/GAS,
NASM, object-path-collision, multi-module) all still pass; `mcpp test`
(130 unit tests) passes.
* docs: #726 and #727 join the round (W13); the split moves to the last stage
* W10 (#724): the S1 document names what a rule generates (S1 0.3.0)
A set's ide.generated lists each output of its package's source-role
actions (header or source, with the step's id, inputs, arguments and
work directory) and each generated include directory its units name,
with the path the document names and the path a mcpp build of the same
selection writes. The profile version is 0.3.0 (Sunrisepeak/
mcpp-language-server#28); compile_commands.json is unchanged. SPEC-005
R3.12, docs 50; e2e 815, e2e 688 reads the new version.
* W11: one renderer for every acquisition; plain output off a terminal
- ProgressBar prints one start line and one finish line when stdout is not
a terminal: no carriage return, no erase sequence, no repaint per frame;
an item that does not complete says so instead of reporting it done.
- The index refresh runs through xlings interface update_packages and draws
its progress and download events with the same renderer; an xlings that
emits none shows its status line as before, and its terminal text no
longer reaches the output.
- The clone of a git dependency passes --progress and draws the download
phase, read as it is redrawn (run_streaming_bounded gains an opt-in rule
that a lone carriage return ends a line); the output is kept whole for
the failure message, and a clone silent for fifteen minutes is stopped.
- The sandbox bootstrap's hand-drawn spinner is the shared bar.
e2e stubs accept the interface refresh; unit tests for the git progress
parser, both render modes and the line splitting; e2e 816; docs 09.
* W11: draw one bar per index-refresh phase, not one per event message
* T7 (#717 W8): a graph-wide dialect switch under a target condition
[target.<selector>.build] dialect_cxxflags is now accepted, parsed into a
new ConditionalConfig member kept apart from BuildInputs (the key is
graph-wide, not a per-package additive input), and merged into the same
BuildConfig::dialectCxxflags every consumer already reads: the std BMI
prebuild, the scan, every TU and the fingerprint. Only the root of the
build renders it; a dependency's own value (conditional or not) reaches
no command and is now excluded from its own fingerprint contribution,
matching the rule that a key enters a fingerprint only where it reaches
a command. The build-program directive is deliberately not added.
SPEC-004 SS3.1 and SS9 item 10 state the new rule; docs/04 and docs/zh/04
document the conditional form (since 2026.9.28.1). Unit tests cover
parsing, the emptiness gate, and root-only resolution order. e2e 813
covers reach into the std BMI/scan/TUs, a non-matching selector, the
A-B-A std BMI rebuild, and a dependency's key being fingerprint-inert;
it fails on the released 2026.9.27.1 binary and passes on the fresh one.
* W11: an index sync bar names its repository
* W6: place-dlls decides the other writer's DLLs from the directory
The one-writer rule passed the deploy list's names to place-dlls on its
command line. The plan's deploy set reads runtime search directories that
a prepare action fills, so it differs between the first and the second
plan; the command changed and every build after the first re-ran the
placement (e2e 797, found by a differential run against 2026.9.27.1).
place-dlls now treats a DLL beside the program that it did not place,
and that a runtime search directory also offers, as another writer's;
the command is the one 2026.9.27.1 wrote.
* chore: version 2026.9.28.1 in both places (mcpp.toml and MCPP_VERSION)
* T8 (#718 W9): the CRT model is a property of the MSVC ABI, not of cl.exe
Every MSVC-ABI row now receives the same CRT model: cl spells it /MT or
/MD, clang++ targeting *-windows-msvc spells it -fms-runtime-lib=static
or =dll. One helper (msvc_abi_crt_word, dialect.cppm) decides the word
for the translation units, the std/std.compat BMIs and the link command
alike, closing #649 E10 (a compile-only flag never reached the clang
driver's own link-time choice of -defaultlib:).
toolchain-coupled (the dynamic CRT, with the toolset's own
vcruntime140.dll/msvcp140.dll staged beside the artifact) is now the
default for every role on this ABI (dist::msvc_abi_default_contract,
ContractStatement::msvcAbiDefault). A toolset with no VC\Redist\MSVC
directory defaults to host-coupled silently and refuses an explicit
toolchain-coupled, naming the missing directory (prepare/plan.cpp). A
free-form CRT word in cxxflags/dialect_cxxflags is checked against the
resolved model: agreeing is a warning, contradicting is a refusal
(dialect::check_crt_word, wired in prepare/scan.cpp).
The toolset's redistributable directory is carried as its own
Toolchain field (msvcRedistDir), populated for cl from vc_redist_dir
and for the LLVM row from its sysroot's tools directory
(vc_redist_dir_for_tools_dir) rather than from linkRuntimeDirs, which
holds LLVM's own runtime directories on that row. The staging gate and
the mcpp run/test search path both read this field, gated on the MSVC
ABI rather than on which compiler is in use. mcpp pack carries the
staged DLLs by default; an explicit --mode system now resolves a
defaulted (never-declared) toolchain-coupled contract to host-coupled
instead of refusing.
e2e 703 is inverted to the new default; e2e 814 covers the LLVM row's
import table, staged DLL, clean-PATH run, self-contained round trip,
BMI switching (A, B, A) and pack modes (Windows-only, unverified here).
Unit tests cover the CRT-word derivation, the free-form-word check, the
MSVC-ABI default/redistributable resolution and a compute_flags-level
property test across both rows; the link-line half of that test is
gated on mcpp.platform.is_windows, since link_shape resolves
LinkShape::PeLld only when current_link_host() is Windows and a
Linux-built mcpp cannot reach that branch regardless of the plan's
target triple.
docs/20, docs/zh/20 and SPEC-006 record the new default and the
upgrade; docs/04 needed no change (it only points at docs/20).
131/131 unit tests pass on Linux (gcc 16.1.0 and llvm 22.1.8 rows);
docs structure/style checks pass.
* docs(specs): SPEC-004 1.9, SPEC-005 1.5, SPEC-006 0.3, SPEC-007 0.4, and the index
* docs(changelog): 2026.9.28.1
* docs: implementation readings of the round (13.5)
* docs(50): the refusal token msvc-redist-unavailable (#718)
* chore: xlings pin 2026.9.28.1 (interface protocol 1.2, the progress events)
* review: every set of a package names what its build program generates; place-dlls comment matches its command; docs/20 names the --mode system refusal
* T6 (mcpp#722, W7): split phase13_finish (plan.cpp) into sub-steps
Verbatim extraction along the sections its own banners already name:
prebuilt dependencies, link forms, make_plan, the C++ runtime checks,
graph/schedule, declared build-graph actions, assembly units, Windows
resources, the global dependency cache, mcpp.lock, runtime provider
overrides, ABI enforcement, resolution.json, and the empty-link check.
Longest resulting function: 356 lines (step13_build_graph_actions).
* T6 (mcpp#722, W7): split phase4b_graph_worklist (graph.cpp) into sub-steps
The worklist step for one item is split into identity resolution, the
already-resolved / identity-adoption handling (with its own version-merge
sub-step), acquiring a fresh dependency's source and manifest, and
finalizing it (recording the package, recursing into children); the
per-item locals that cross those boundaries move into a phase-local
struct, WorklistItemCtx, the same PrepareState pattern one level deeper.
The post-loop cycle check is its own function. Preamble closures that
captured only `state` (or nothing) become static file-scope functions,
called with an explicit PrepareState& where they used to close over it;
this also fixes the one comment that had gone factually stale (activateFeatures's
group banner said 'defined as local lambdas, not file-scope functions' --
now they are file-scope, and are still safe because a static function
carries no external linkage into the module's exported interface).
Longest resulting function: 300 lines (step4b_identity_version_merge).
* T6 (mcpp#722, W7): split phase6_features_and_host_tools (features.cpp) into sub-steps
Along the sections its own banners already name: feature activation,
device extensions and rule application, the graph's [xlings.workspace]
provisioned before build.mcpp, host-module registration, host-tool
provisioning, the dependencies' build programs, and capability binding.
aggregatedRequest (needed by two of these sections) is promoted from a
local lambda to a file-scope function of PrepareState&. Two scoping
braces that had no matching close within their own section (opened to
wrap several sections at once) are dropped as redundant once the
content is distributed across separate functions, each of which
supplies its own scope.
Longest resulting function: 425 lines (step6_provision_host_tools).
* T6 (mcpp#722, W7): split phase9_target_side (target_side.cpp) into sub-steps
Gather the graph's target-side candidates into a phase-local struct
(TargetSideGather, the PrepareState pattern one level deeper), then
resolve and realise [c-abi], broadcast the include set, check kernel-abi
interfaces and layer requirements, apply the layer-conditional config
(L1b), decide each dependency's link form (#519), define the
graph_package_entry closure, run the root build.mcpp (L3), require every
device source to reach an action, and handle re-run inputs (R1.3). Two
scoping braces that wrapped several sections at once are dropped as
redundant, as in the two previous files.
Longest resulting function: 404 lines (step9_kernel_abi_interfaces_and_requirements).
* T6 (mcpp#722, W7): split phase4a_graph_load (graph_load.cpp) into sub-steps
The closures phase4a_graph_load assigns onto state (each captures only
state) split into two groups: split/identity closures, and candidate
selection closures. LoadedDep is hoisted to file scope so it stays
visible to state.loadVersionDep, which remains where it was.
state.loadVersionDep itself (509 lines) is not split further: its six
local closures (readLuaContent, findRawInstalled,
installedLayoutMatchesIndex, revisionIsCurrent, findCompleteInstalled,
markInstalled) mutually capture nine-odd shared locals by reference:
hoisting them to free functions would mean threading all of that
through explicit parameter lists for one closure, judged higher risk
than benefit within this round; noted as a residual in the T6 report.
* review: the round's CI failures and the three review angles
CI (PR #727):
- e2e 190/191 select the program by name; bin/ also holds the staged
redistributable DLLs on an MSVC-ABI row (#718).
- A DLL found in a runtime search directory yields to a declared deploy of
the same name, and a difference is warned at planning, where a successful
build shows it (SPEC-007 R4.3, e2e 811; e2e 818 is its Linux-hosted form).
- build.mcpp that imports only a build rule compiles in the build directory
under GCC (e2e 807).
- The runtime environment half of .xlings.json goes to the runtime's owner
again; only plan_only redirects it to the planning directory (e2e 205, 817).
Review:
- Every set of a package names what its build program generates (R3.12).
- A dependency's contradicting CRT word is refused; debug CRT words are
refused; the agreeing-word message names the key of the dynamic model; the
dependency cache key carries the CRT word.
- PE deploy destinations compare without case; MSVC version directories
compare numerically; a missing deploy source is named as missing.
- A failed index refresh draws "did not complete"; an automatic refresh that
exhausts its retries warns; the floor tip names the version and the
install-aware upgrade.
- -p compares paths as paths; the help names the qualified form.
* docs: global review and the first CI run of the round (13.6)
* T6 (mcpp#722, W7): split phase1_toolchain_spec_and_axes (toolchain.cpp) into sub-steps
Split at phase1's own banner boundaries: the closures phase1 assigns onto
state, the target/--static override resolution, and the device axis plus
the L1 conditional-section merge.
phase2_define_toolchain_resolver is not split: it is a single stored
closure, state.resolve_target_toolchain, whose ~1000-line body is one
sequential toolchain-resolution flow with dozens of interdependent
locals -- the same category of residual as state.loadVersionDep in
graph_load.cpp, noted in the T6 report. step1_target_and_static_overrides
(524 lines) is a smaller residual of the same kind.
* T6 (mcpp#722, W7 follow-on): split phase0_manifest_and_workspace and phase3_xlings_before_graph
Both were found over ~400 lines by the same sweep that produced the
seven functions #722 named (the issue's own "any other function under
src/build/prepare/ over ~400 lines"), and split cleanly along their own
banners: phase0's own "Workspace handling" section becomes
step0_workspace_handling (108 lines; phase0 itself drops to 372); the
host-toolchain closures phase3 assigns onto state, plus the index-refresh
section, become step3_define_host_tc_closures_and_refresh_index (298
lines; phase3 itself drops to 235).
* T6 (mcpp#722): add the function-size gate, verified but not wired into CI yet
.github/tools/check_function_sizes.sh runs clang-tidy's
readability-function-size (LineThreshold=400) over the compile database
mcpp produces for its own LLVM build (mcpp build --toolchain
llvm@22.1.8), restricted to files under src/build/prepare/. clang-tidy
is not part of the plain xim:llvm payload; it ships in the sibling
xim:llvm-tools package at the same version, which the script locates
under the xlings store.
Measured: at 7ccbc8d (before this round's split) it reports 10
functions over 400 lines -- the seven #722 named, plus
phase0_manifest_and_workspace, phase11_scan and
phase3_xlings_before_graph, found by the same sweep. After this round's
split it reports 6: step6_provision_host_tools (421),
phase4a_graph_load (518, its loadVersionDep closure), phase11_scan (773,
untouched), step9_kernel_abi_interfaces_and_requirements (401),
step1_target_and_static_overrides (522) and
phase2_define_toolchain_resolver (1006, untouched) -- see the T6 report
for why each remains.
Not wired into CI next to check_file_lengths.sh: it does not yet pass,
so adding the workflow step now would land a gate red on day one. Wire
it once the remaining residuals are split in a follow-up; the file
gate stays the only enforced one for this round, per the design's own
fallback for a working tool over an incomplete split.
* prepare: mcpp.lock and resolution.json move to records.cpp (plan.cpp under the file-length limit after the merge; byte-identical on the seven fixtures)
* T6 (mcpp#722, W7 residual 1/6): split step6_provision_host_tools (features.cpp)
The per-tool body of the outer loop (~421 lines) becomes HostToolCtx, a
phase-local struct following the WorklistItemCtx / TargetSideGather
pattern, and six sub-steps: target resolution, the override/self-request
checks, the store-key/cache-hit resolution, the sub-build, and publish.
The `record` closure the single-function version captured per tool is a
named helper called from all four sites that used it (an override, a
cache hit, a deferred plan-only tool, and a freshly built one).
Verified: byte-identical on the seven fixtures; check_function_sizes.sh
reports no finding in this file.
* T6 (mcpp#722, W7 residual 2/6): split the loadVersionDep closure (graph_load.cpp)
phase4a_graph_load's stored closure state.loadVersionDep (~518 lines) is
split the way phase2's single stored closure is: a local struct
(LoadVersionDepCtx) holds the parameters and the locals more than one
section reads, and the body becomes three named steps (locate an
already-installed copy, fetch one if none is installed, then read its
manifest) the closure calls in sequence. The stored closure stays the
entry point with the same signature, so its own recursive call for a
preinstall hook's dependencies is unchanged.
Two of its local lambdas (readLuaContent, findRawInstalled) are read from
two of the three steps (the initial read and a later re-check, or the
initial probe and the post-install one) and become named helpers for the
same reason `record` did in the previous commit; a third (markInstalled)
is promoted for the same reason though it has only one call site, since
the closure it replaced no longer has anywhere to live once the body it
was local to is split.
Verified: byte-identical on the seven fixtures; check_function_sizes.sh
reports no finding in this file.
* T6 (mcpp#722, W7 residual 3/6): split phase11_scan (scan.cpp)
Split at its own section boundaries into eight named steps (source scan
and validation, the dependency-standard scope check, the dialect-flag
gate, the MSVC CRT-word check, the package std-module source, the Apple
SDK C++ runtime rule, the std-module availability gate, the fingerprint
computation, and the std-module prebuild). needsStdModule is the one
value several steps read; it is computed by the first step and passed as
a plain bool parameter rather than through a struct, since it is the
scan's only shared local.
Verified: byte-identical on the seven fixtures; check_function_sizes.sh
reports no finding in this file.
* T6 (mcpp#722, W7 residuals 4-5/6): split step1_target_and_static_overrides
and phase2_define_toolchain_resolver (toolchain.cpp)
step1 (~522 lines): a phase-local struct (TargetOverrideCtx) holding the
parsed triple, its vocabulary-table row and the section lookup carries
seven sub-steps through one `--target` request's resolution (resolve the
request, validate its tier, the Apple SDK check, the wasm shared-lib
check, the host_can_serve diagnosis, capturing the display name and
canonicalising, and the row-pin/capability check). The `apply_target_section`
closure is a named helper, called from the resolved-request branch and
from the host-row branch that follows it.
phase2 (~1006 lines, a single stored closure, state.resolve_target_toolchain):
a phase-local struct (ToolchainResolveCtx) holds only the parsed spec and
the Windows installed-toolset probe -- the two locals the branch-selection
half of the closure shares. Every other local (the first-run defaults, the
explicit-spec resolution's own payload/frontend locals, and so on) stays
local to the one step that declares it, as it was in the single function.
The closure becomes twelve named steps: parsing the spec, each arm of the
compiler-resolution if/else-if chain, the Windows-first-run persist, toolchain
detection, the retargetable-driver fixup, the MSVC toolset bind, the Windows
runtime identity, the MSVC-ABI-without-MSVC repair, and the musl default
linkage. The stored closure stays the entry point with the same signature,
so its own recursive call (the target pass) is unchanged.
Verified: byte-identical on the seven fixtures; check_function_sizes.sh
reports no finding in this file.
* T6 (mcpp#722, W7 residual 6/6): split step9_kernel_abi_interfaces_and_requirements
(target_side.cpp); the gate's own false failure at zero findings
step9 (~401 lines) splits into five named steps at its own section
boundaries: the kernel-abi interface enumeration, the layering and
requirement checks (over TargetSideGather's `requirements`), the
pin-and-linkage diagnostics, the same-OS check and report, and the
platform-sdk closure visibility check. No phase-local struct is needed:
each step reads and writes only `state` (plus `gather` for the one step
that needs its `requirements`).
This was the last of the six residuals check_function_sizes.sh reported at
the base of this round; with it split, the gate ran clean over the whole
of src/build/prepare/ for the first time -- and immediately failed for an
unrelated reason: `--warnings-as-errors='*'` makes clang-tidy exit non-zero
on the two pre-existing readability-function-size findings in the bundled
third-party modules/libs/src/json/json.hpp, which the script's own
`relevant` filter already excludes from the pass/fail decision but which
its exit-code fallback then reads as "a diagnostic tool problem" -- a
false failure this repository's own tree could not previously reach, since
no earlier revision had zero findings under src/build/prepare itself.
Dropping the flag leaves clang-tidy's exit code meaningful only for an
actual tool failure (a bad compile command, a crash), which a plain
warning never produces; `relevant` remains the sole pass/fail signal, as
the surrounding comments already say it is.
Verified: check_function_sizes.sh reports "ok" over the whole of
src/build/prepare/; byte-identical on the seven fixtures.
* ci: the function-size gate runs after the LLVM self-build; clang-tidy is resolved at the version that wrote the compile database
* docs: #722 completed, the function-size gate in CI (changelog, 13.6)
* review: the function-size gate is run by hand until CI builds mcpp with clang (#729); a generated deploy source is compared after the link only
- ci-linux.yml: the gate step comes out again. The LLVM toolchain job's
llvm@20.1.7 self-build has never completed (libc++ 20's std module hides
directory_iterator's comparison) and the step reads the resolution line,
not the build's exit status; clang-tidy crashed over the partial database.
Recorded as #729.
- check_function_sizes.sh: a finding is a diagnostic line ending in the
bracketed check name; a crash dump no longer reads as a finding.
- prepare/plan.cpp: a declared deploy source that an action writes is left to
the post-link comparison; at planning it may hold the previous build's bytes.
---------
Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
Co-authored-by: speak-agent <agent@mcpp-community>
Co-authored-by: Claude <noreply@anthropic.com>
1 parent b439fd9 commit acb9f52
115 files changed
Lines changed: 10490 additions & 3147 deletions
File tree
- .agents/docs
- .github
- actions
- bootstrap-mcpp
- setup-macos-llvm
- tools
- workflows
- docs
- specs
- zh
- modules
- manifest/src
- platform/src
- toolchain-model/src
- versioning/src
- src
- build
- prepare
- cli
- fetcher
- pack
- pm
- toolchain
- xlings
- tests
- e2e
- unit
Some content is hidden
Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.
Lines changed: 1417 additions & 0 deletions
Large diffs are not rendered by default.
Lines changed: 115 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
| 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 | + | |
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
18 | 18 | | |
19 | 19 | | |
20 | 20 | | |
21 | | - | |
| 21 | + | |
22 | 22 | | |
23 | 23 | | |
24 | 24 | | |
| |||
56 | 56 | | |
57 | 57 | | |
58 | 58 | | |
| 59 | + | |
59 | 60 | | |
60 | 61 | | |
61 | 62 | | |
| |||
82 | 83 | | |
83 | 84 | | |
84 | 85 | | |
| 86 | + | |
85 | 87 | | |
86 | 88 | | |
87 | 89 | | |
| |||
102 | 104 | | |
103 | 105 | | |
104 | 106 | | |
| 107 | + | |
| 108 | + | |
105 | 109 | | |
106 | 110 | | |
107 | 111 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
25 | 25 | | |
26 | 26 | | |
27 | 27 | | |
28 | | - | |
| 28 | + | |
29 | 29 | | |
30 | 30 | | |
31 | 31 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
15 | 15 | | |
16 | 16 | | |
17 | 17 | | |
18 | | - | |
| 18 | + | |
19 | 19 | | |
20 | 20 | | |
21 | 21 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
| 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 | + | |
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
17 | 17 | | |
18 | 18 | | |
19 | 19 | | |
20 | | - | |
| 20 | + | |
21 | 21 | | |
22 | 22 | | |
23 | 23 | | |
| |||
0 commit comments