Skip to content

0.21.0: a second name for a file, and the requests to end counted (openkal 0.25) - #46

Merged
Sunrisepeak merged 7 commits into
mainfrom
feat/0.25.0
Nov 9, 2026
Merged

Sunrisepeak merged 7 commits into
mainfrom
feat/0.25.0

Conversation

@Sunrisepeak

@Sunrisepeak Sunrisepeak commented Nov 9, 2026 •

Copy link
Copy Markdown
Member

Summary

Test plan

  • Old suite: CI run 42675285217 green at ab227a5 (this package's tests and the specification branch's conformance suite, as they were).

  • New suite: CI run 42679092276 green at 9f441f4 — the conformance suite through the pinned mtest engine (one case per interface, its observations within it), mcpp test 2026.11.9.1 keeping this package's runs in the mtest store, each job's report in its summary, one report over every job.

  • Tests ported to the mtest library, and the switch to openkal = "0.25.0": CI run 42695429206 green at 1ffab79 on every leg (msvc, llvm with the MSVC ABI, gcc with the GNU CRT: 13 passed; the wine cross job; the report). Every file under tests/ declares its cases as UTest objects with the assertions it had; what a test printed as not observed is an mtest skip. Copies are child objects started with child_command (null_device_at_a_position, openkal_0_23, what_a_started_program_receives, environment_read_from_two_threads). Two copies need care. The environment-race copy is entered from GetCommandLineW during the file's initialisation: mtest's state reads its variables through kal_env_var when the first case is declared, which would fill the tables the copy observes. The handle-number copy is started without MTEST_REPORT/MTEST_PROTOCOL, so the number cannot name the report file. mcpplibs.mtest 0.5.2 is a dev-dependency.

  • The gnu ABI's -fno-exceptions -fno-rtti -fno-builtin are attached to src/**. The tests import std through mtest, and mcpp refuses dialect flags that the prebuilt std was not compiled with. The objects the checks examine are compiled as before.

  • A toolchain defect on the gcc/gnu-CRT leg: GCC on PE emits the function-local static std::_Sp_make_shared_tag::_S_ti()::__tag of an imported std as a non-COMDAT definition, in every object importing std and in libstdc++.a, and every test link was refused. That leg runs mcpp test --profile gnu-odr: the dev profile (opt 0, debug) with -Wl,--allow-multiple-definition. A profile is the root's own and never reaches a consumer's link. The leg still builds and runs every test. console_input declares its four console functions and the key record beside src/win32.h, because GCC refuses <windows.h> together with the imported std ("conflicting language linkage").

…enkal 0.25)

- kal_fs_hard_link: NtSetInformationFile with the link record; KAL_FS_PROP_HARD_LINKS
  where the volume reports FILE_SUPPORTS_HARD_LINKS; a directory is
  kal_err_is_directory and STATUS_NOT_SAME_DEVICE kal_err_not_supported
- the stop word counts each console event atomically and skips zero on wrap
- openkal follows the specification's branch feat/0.25.0 until 0.25.0 is released
- the conformance suite runs through the pinned mtest engine (MTEST, set by
  .github/actions/mtest): one case per interface, its observations within it
- mcpp 2026.11.9.1: `mcpp test` keeps this package's runs in the mtest store
  (target/mtest), one case per test program
- each job's runs in its summary and as an artifact; one report over every job

The test programs stay programs of their own (protocol level 0): the mtest
library reaches mcpplibs.os, which depends on openkal and on this
implementation, so a dev-dependency on it would close a cycle through the
package under test.
Every test under tests/ declares its cases as mtest UTest objects, with
the same assertions as before; what a test reported as not observed is
an mtest skip. A test that starts a copy of itself declares what the copy
does as a child and starts it with child_command. The copy that reads the
environment first from two contexts is entered from the system's command
line before mtest reads its own variables, which would fill the tables
the copy observes; the copy that writes through a handle number is
started without the runner's report variables. mcpplibs.mtest 0.5.2 is a
dev-dependency.
-fno-exceptions and -fno-rtti applied to every unit of the build, the
tests included; the tests now import the mtest library, which imports
std, and mcpp refuses a dialect flag that the prebuilt std interface was
not compiled with. The flags are attached to src/** as openkal-linux
attaches its own; the objects the checks examine are compiled as before.
…the link accepts GCC's repeated std static

console_input declares the console functions and the key record it uses
beside src/win32.h: GCC on PE refuses <windows.h> together with the
imported std ("conflicting language linkage for imported declaration
std::__is_constant_evaluated()").

GCC on PE emits the function-local static
std::_Sp_make_shared_tag::_S_ti()::__tag of an imported std as a
definition that is not COMDAT, in every object importing std and in
libstdc++.a, so every test link was refused. The CI step of the gcc leg
writes -Wl,--allow-multiple-definition into its working copy of the
manifest for the tests and restores the manifest afterwards; the
committed manifest does not carry it.
The link option that GCC's repeated std static needs on PE is stated in a
profile, [profile.gnu-odr]: the dev profile with
-Wl,--allow-multiple-definition. A profile is the root's own and never
reaches a consumer's link. The CI step no longer edits its working copy
of the manifest. console_input's watchdog returns after ExitProcess,
which src/win32.h does not declare as not returning (msvc C4716).
@Sunrisepeak
Sunrisepeak merged commit b399b20 into main Nov 9, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant