Skip to content

Compile morph's UI layer as a static library - #1120

Merged
Yaraslaut merged 38 commits into
masterfrom
feat/compiled-ui-library
Nov 15, 2026
Merged

Yaraslaut merged 38 commits into
masterfrom
feat/compiled-ui-library

Conversation

@Yaraslaut

@Yaraslaut Yaraslaut commented Nov 14, 2026 •

Copy link
Copy Markdown
Member

morph::morph stops being header-only: it is now a STATIC library, and the UI layer's non-template code lives in src/ui/*.cpp. Every translation unit that includes a UI header used to recompile all of that code, which is what pushed single compiles past what a GitHub-hosted runner (16 GB, 4 CPUs) can hold.

Design: docs/superpowers/specs/2026-11-14-compiled-ui-library-design.md. Part of #1118. Part B, an optional shared build (MORPH_BUILD_SHARED), follows in its own PR.

What moved

  • interpreter.hpp: 11,077 → 2,840 lines.
  • document.hpp: 5,648 → 775 lines.
  • expression.hpp: 3,184 → 620 lines.
  • The service headers: service, service_client, host, workspace, shell_frame, shell_parts_model, backend, theme, explorer, navigator. Most halved or better.
  • ui/chat/service.hpp.
  • client/source.hpp no longer includes explorer.hpp.

Build plumbing

  • GenerateExportHeader provides MORPH_API, which every moved declaration carries.
  • Coverage and the sanitizers are applied to the library, and the branch-coverage gate and Codecov see src/ui.
  • The install check's consumer calls moved interpreter, document, expression and service code, so it proves the installed archive links.

Peak compile memory (measured)

TU master this branch
tests/test_ui_shell_parts.cpp, g++ -O3 (gcc-release) 5640 MB 2380 MB
tests/test_ui_interpreter_commands.cpp, g++ -O3 4683 MB 2464 MB
examples/lims/screens/tests/test_lims_ignition_screen.cpp, clang coverage 3433 MB 2176 MB
#include <morph/ui/interpreter.hpp> alone, clang coverage 1816 MB 870 MB

The lims figure misses the design's 2 GB goal. The remaining cost is in core/remote.hpp and in what interpreter.hpp still has to declare. The gcc-release goal (3.5 GB per UI test TU) is met.

Interpreter benchmark (clang -O3, [.benchmark][interpreter]), within noise:

  • Mount of a 200-row table: 1.43 → 1.42 ms.
  • Click and settle: 5.0 → 5.1 µs.

Verified locally, on the rebased head

  • clang linux-everything: every target builds, including Qt, Qt Quick, TUI and the lims rungs. ctest passes 5712/5712, excluding lims|Stress|load|soak|fuzz; the lims suites pass 622/622 separately.

  • gcc-debug: full build with no warnings, and ctest passes 4170/4170 with the same exclusion.

  • *_verify_interface_header_sets targets build, and scripts/check_install_export.sh passes.

  • llvm-cov shows the moved files counted:

    File Line coverage
    document.cpp 89%
    expression.cpp 94%
    theme.cpp 88%
    service.cpp 70%

    This is from a 4-file subset of the tests, not the whole suite.

  • Mutation controls: one moved function each in the interpreter, document, service and chat code was broken on purpose, and a test failed every time.

Not verified locally

  • Windows/MSVC and WASM.
  • clang-tidy 22, which CI runs. Local clang-tidy 23 still flags readability-trailing-comma in four moved files, and flags the same in untouched files.
  • The full coverage gate.
  • gcc-release ctest.

🤖 Generated with Claude Code

https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj

Yaraslaut and others added 26 commits November 15, 2026 10:05
…urce

morph::morph is now add_library(morph STATIC) instead of an INTERFACE
target. Include dirs, link libraries and compile definitions are PUBLIC,
both header file sets are PUBLIC and still verified, and the install rule
gains an ARCHIVE destination. GenerateExportHeader writes morph/export.hpp
(MORPH_API) into the binary dir, shipped through its own FILE_SET so build
and install both find it; MORPH_STATIC_DEFINE is PUBLIC. The package
version file is no longer ARCH_INDEPENDENT, since the package now holds an
archive.

src/ui/library.cpp holds fragmentOf and routeOfFragment, moved out of
route_history.hpp (declarations stay, marked MORPH_API).

apply_warnings, apply_sanitizers and apply_coverage are applied to morph.
Coverage machinery now sees src/ui: coverage.sh SOURCES, the Codecov
framework component, and check_branch_coverage.py, which scores src/ui
under the include/morph/ui floor (self-tests extended; the new case fails
when the prefix handling is removed).

check_install_export.sh's consumer now calls a function from the compiled
archive. With library.cpp reduced to an empty namespace the consumer fails
to link (undefined reference to morph::ui::fragmentOf), so the check
proves compiled code is linked.

Sanitizer consumer audit (clang, AF_SANITIZER=asan, linux-everything
options): 54 executables link libmorph.a. Six had no -fsanitize at link:
the four examples/vetted_hmac executables and the two
lims_cross_module_effects probes. They linked only because no archive
member of libmorph.a was pulled in; the first moved symbol they use would
fail on __asan_init. They now call apply_sanitizers; the re-audit shows
54 of 54 instrumented. Examples, benches, fuzz and soak targets already
carried an AF_SANITIZER block. morph_guard_probe_deps and the try_compile
probes do not link morph::morph (they link its dependencies or only take
its include directories), so they are unaffected.

Docs and comments that called morph header-only now describe the compiled
static library.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Every out-of-class `inline` member definition, every in-class member body of
two or more lines, and the `detail` free functions move to the compiled
library. Classes with a moved member, and the structs among them, carry
MORPH_API. A free function that something outside the interpreter calls
(tests, host.hpp, the inspector, the workspace) keeps its documented
declaration in the header with MORPH_API. One nothing else calls moves into
an unnamed namespace in the .cpp, together with the helper types and the
named constants only moved bodies read.

One-line accessors stay in the header so a hot reader keeps inlining.
Templates and constexpr stay; kMaxInlineFileBytes is unchanged.

tests/bench_interpreter.cpp (hidden, tag [.benchmark][interpreter]) mounts a
ui/1 document with a 200-row table and times a state change; bench_mount
never reaches the interpreter.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Includes. The compiled bodies carry the includes only they need: the
bridge, the forms engine, the wire envelope, the message and locale tables,
the chart scale and the backend's mount layer. The header forward-declares
what its declarations take by reference or pointer (bridge::Bridge,
bridge::BindSharing, the wire bodies, the forms sessions, ChartRange,
Mounted).

ElementState's constructor and destructor move out of line, because it owns
forms sessions, and its copy and move are deleted. BridgeDispatcher::Channel
is defined in the .cpp. Screen::_mounted becomes a unique_ptr, so the header
needs Mounted only declared. kMaxInlineFileBytes is spelled as its value, and
the .cpp asserts it equals the wire envelope formula it came from.

host.hpp, preference_sync.hpp, ui_interpreter_support.hpp, test_push_cache.cpp
and test_ui_interpreter_commands.cpp relied on interpreter.hpp for an include
and now name it. preference_sync.hpp only failed the standalone
interface-header-set check, which is not part of the default build.

check_install_export.sh's consumer also calls an interpreter function
(detail::base64Size), so an archive without interpreter.cpp fails to link.

clang-tidy. Moving a body makes every line of it lintable, so the .cpp gets
the const-correctness, trailing-comma, redundant lambda parameter list and
string-view-conversion fixes, NOLINT reasons where a body implements an
interface (the InterpreterHost defaults, the recursive menu walks, the two
long element builders), and ScreenCore::chartSeries becomes static.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The parser, the validator, the kind and key tables, and the loader's entry
points (`loadDocument`, `loadDocumentText`, `nodeKinds`, `documentLookup`) move
to the compiled library. The header keeps the document model, the key and kind
vocabulary types the interpreter reads (`KeySpec`, `KindSpec`, their flags),
`NameCategory`, `listed`, and documented MORPH_API declarations of `findKind`,
`pageGeometryOf` and `regionGeometryOf`, which the interpreter calls.

The parser, the validator and every table they read are in an unnamed namespace
in the .cpp; nothing else uses them. Includes the header no longer needs are
dropped: a TU that includes only document.hpp peaks at 446 MB of front-end
memory, from 728 MB.

The page `margins` reading is split out of `pageGeometryOf`, and the table
regions carry a NOLINT for readability-trailing-comma beside the existing one
for designated initializers, so clang-tidy is clean on the moved code. The
branch-coverage allowlist entries for document.hpp now name src/ui/document.cpp.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The parser, the checker, the evaluator, the function table and the unit and
date helpers move to the compiled library. `parseExpression`,
`checkExpression`, `evaluate`, `functionNames`, `walk`, `readDateValue`,
`runRecord`, `splitPath`, `typeAlong`, `numeric`, `convertUnit` and
`unitDisplay` keep documented MORPH_API declarations: the document loader and
the interpreter call them. `StaticEnvironment` carries MORPH_API, with `control`
and `report` out of line.

The evaluator's body lives in an unnamed namespace as `EvaluatorCore`. The
interpreter drives an evaluator directly to group and order table rows, so
`detail::Evaluator` is a handle over it with `eval`, `sortOrder`, `sortRows`
and `sameRun`; `evaluate` uses the core without the handle's allocation.

Includes the header no longer needs are dropped: a TU that includes only
expression.hpp peaks at 417 MB of front-end memory, from 544 MB. The
branch-coverage allowlist entries for expression.hpp now name
src/ui/expression.cpp, and the install check's consumer loads a document and
evaluates an expression so it links the moved code.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
UiRegistry's and UiService's member bodies of two or more lines move to the
compiled library, and both classes carry MORPH_API. The `detail` helpers
(checkServable, checkEntry, DocumentActions, permitsAll) have no user outside
the moved bodies, so they live in the .cpp's unnamed namespace. UiService's
screenAdmits is declared in the header and defined in the .cpp, because it
calls one of them.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…ient.cpp

The member bodies of two or more lines of MemoryBundleStore, VerifiedCache,
UiClient and FileTransfer, and the free functions exchangeOver,
evaluationValue, evaluatorOver and watcherOver, move to the compiled library.
Their declarations keep their Doxygen blocks in the header and carry
MORPH_API.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
DocumentScreenHost's and the nested screen classes' member bodies of two or
more lines move to the compiled library, along with identityKey, sameIdentity
and titleText, which keep documented declarations in the header because
workspace_model.hpp, the tests and the header's own lambdas call them.
ParamsEnv has no user outside the moved bodies and lives in the .cpp's unnamed
namespace.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Workspace's member bodies of two or more lines move to the compiled library,
and the class carries MORPH_API. One-line accessors stay in the header.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…s_model.cpp

ShellPartsModel's member bodies of two or more lines move to the compiled
library, along with shellText and shellLabels, which shell_parts_binding.hpp
calls and so keep documented declarations. builtinShellText has no other user
and lives in the .cpp's unnamed namespace.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
ShellFrame's member bodies of two or more lines move to the compiled library,
and the class carries MORPH_API.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
TableWidget::setRealizedWindow, IViewBackend::createCustom and widthClassOf
move to the compiled library; the widgets' one-line accessors stay in the
header. TableWidget and IViewBackend carry MORPH_API.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Every `inline` function of the token and theme readers moves to the compiled
library. The ones that tests, the Qt Quick and TUI renderers, the service or
the interpreter call (colorOf, parseTokens, resolveTokens, parseThemeIndex,
resolveTheme, checkTheme, themeValues, themeProblems, themeContrastProblems,
unreadThemeTokens, iconGlyph, isIconName, isThemeIndex, and the detail
numberOf, hexColor, channelOf, aliasOf, pixelsOf, shadowOf, familiesOf and the
colour-maths helpers) keep documented declarations with MORPH_API. The
readers, the checks and AliasResolver have no other user and live in the .cpp's
unnamed namespace.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Explorer's member bodies of two or more lines and the free functions
actionUses, actionFields, tryItScreenId, tryItDocument, schemasOf and
tryItCatalog move to the compiled library. The scope walk and the text
helpers have no other user and live in the .cpp's unnamed namespace.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Navigator's member bodies of two or more lines move to the compiled library,
and the class carries MORPH_API. The one member template stays in the header.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…c's header-only claim

theme.hpp no longer includes bundle.hpp and control_shape.hpp (884 MB alone,
front end), and service.hpp drops the logger, wire, session, catalogue and
shell-parts includes its remaining declarations do not use. The reload note in
docs/spec/ui/service.md says the functions live in the morph library.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Const-correctness, redundant lambda parameter lists and trailing commas are
applied, with the designated initialisers the trailing-comma fix-it mangled
restored by hand; the check's empty-brace misfires, the menu walks' recursion
and the interface-bound signatures carry NOLINT with their reason.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
A stray semicolon after the closing namespace brace in three .cpp files, and a
second declaration of detail::weightOf in theme.hpp.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Placement of the suppressions that the move left behind (a theme NOLINTBEGIN
without its END in either file), plus the findings the new translation units
raise on their own: short names, nested conditionals, unchecked optionals,
range loops and emplace.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The consumer already links a call into the interpreter, the document loader and
the expression language; a registry that refuses a file of an unknown kind and
builds a bundle proves the service's compiled half is exported as well.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The turn state, the shared state a turn's callbacks keep alive, and every
service entry point move to the compiled library; the header keeps the class,
its options and forward declarations of the private nested types. The header
no longer includes the document, the bundle, the protocol, the screen tools or
the logger, so a translation unit that names ChatService stops paying for them.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Source only needed the alias for a model's schemas document, which now lives
beside the other bundle types; the explorer pulls in the forms engine, which a
translation unit that only names a source has no use for.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
@Yaraslaut
Yaraslaut force-pushed the feat/compiled-ui-library branch from e04db35 to fd83fb7 Compare November 15, 2026 09:49
@Yaraslaut Yaraslaut changed the title Compile morph's UI layer as a library, optionally shared Compile morph's UI layer as a static library Nov 15, 2026
@Yaraslaut
Yaraslaut marked this pull request as ready for review November 15, 2026 10:04
…d code

clang-format-22 flagged a doubled blank line in bundle.hpp and an
over-long comment line in chat_service.cpp. clang-tidy-diff flagged three
changed declarations: Handle's two bases both carry default bodies, so
the multiple-inheritance check is suppressed with that reason; confirm's
`no` parameter is too short and becomes `noLabel`; channelOf's
declaration drops a top-level const.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Yaraslaut and others added 11 commits November 15, 2026 13:39
GCC 15 at -O3 reports -Walloc-zero inside libstdc++'s regex executor, and
gcc-release treats warnings as errors. A small hand-written parser reads
the same `role: "#digits"` lines.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The table names every module's actions, so its header pulled in all five
authorization headers and through them every model: about 7.7 GB of GCC
memory at -O0 -g in each screen test that includes lims_authorizer.hpp.
The header now declares the lookups; the table and its includes live in
app/src/declared_requirements.cpp.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The model headers declare each model and action; the registrars compile
once, in the model's source file, instead of in every translation unit
that includes the header. The org authorization header no longer includes
the org model.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
The model headers declare each model and action; the registrars compile
once, in the model's source file, instead of in every translation unit
that includes the header. The lab authorization header no longer includes
the module's models, so a model header and its authorization header form
no include cycle.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…aders

The model headers declare each model and action; the registrars compile
once, in the model's source file, instead of in every translation unit
that includes the header. The billing authorization header no longer includes
the module's models, so a model header and its authorization header form
no include cycle.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…ders

The model headers declare each model and action; the registrars compile
once, in the model's source file, instead of in every translation unit
that includes the header. The ledger authorization header no longer includes
the module's models, so a model header and its authorization header form
no include cycle.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…headers

The model headers declare each model and action; the registrars compile
once, in the model's source file, instead of in every translation unit
that includes the header. The inventory authorization header no longer includes
the module's models, so a model header and its authorization header form
no include cycle.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…ation unit includes fewer models

The shell test's probe bundles, its seeding of people and the approval
inbox's lab and bundle move into translation units of their own, so the
tests that drive the screen harness no longer include every module's
models. The shell and screens tests are each split in two. The declared
requirements table includes the models whose action ids it names, now
that an authorization header no longer brings them in.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…world's helpers static

The shell test's JSON scan added an int to a size_t; GCC's sign-conversion
error stops the build. Two of the approval-inbox world's helpers use no
member state.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
Trailing commas, designated initializers, a longer identifier than `id`,
const fixtures, no unused using-declarations, and the shell test's German
menu lookup in a function of its own to keep signedIn under the
cognitive-complexity threshold.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
…parses the page query

The collator's fold was one expression of thirty nested replace calls over
lower(). SQLite before 3.46 parses with a fixed 100-entry stack, and that
nesting overflows it ("parser stack overflow"): on the Ubuntu 24.04
runners (SQLite 3.45.1) every lims sample and ledger page with a text
filter or text sort failed. Each folded expression is now computed as a
column of a three-stage WITH chain, ten replace calls per stage, and the
filter and ordering name that column.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No1Nzo5wbYPDDynGBphgqj
@Yaraslaut
Yaraslaut merged commit 5b6a431 into master Nov 15, 2026
17 of 19 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