Convert non-Windows apphost to C

[Copilot]

Convert non-Windows apphost to C. On windows, we still have apphost.windows.cpp and it remains C++. dotnet and singlefilehost also remain C++.

• Separate entry points for apphost and dotnet
  • apphost entry point is apphost.c, shared by the standalone apphost and singlefilehost.
  • dotnet entry point is dotnet.cpp
  • Remove corehost.cpp
• hostfxr resolvers:
  • apphost resolver is now C
  • singlefilehost keeps its C++ resolver for statically-linked hostfxr
  • dotnet resolver is separate and under the dotnet folder

cc @dotnet/appmodel @AaronRobinsonMSFT

linux x64 Release

Binary	main	PR	Δ bytes	Δ%
apphost	41,896	32,240	−9,656	−23.0%
dotnet	36,184	34,280	−1,904	−5.3%
singlefilehost	16,801,512	16,784,704	−16,808	−0.10%

macOS arm64 Release

Binary	main	PR	Δ bytes	Δ%
apphost	96,176	91,232	−4,944	−5.1%
dotnet	95,872	95,152	−720	−0.8%
singlefilehost	10,299,280	10,281,104	−18,176	−0.2%

[dotnet-policy-service]

Tagging subscribers to this area: @agocke, @jeffschwMSFT, @elinor-fung
See info in area-owners.md if you want to be subscribed.

[Copilot]

@copilot rename the hostmisc C files prefixed with apphost_ to drop the apphost_ prefix. They should be .c versions living next to their corresponding .cpp files. For the header files, they can ha...

Done in 7b1a73c. Renamed all hostmisc C files to drop the apphost_ prefix (e.g., apphost_trace.c → trace.c) and headers to use _c suffix (e.g., apphost_trace.h → trace_c.h). Updated all #include directives, CMakeLists.txt references, and header guards.

[Copilot]

Copilot's findings

• Files reviewed: 22/22 changed files
• Comments generated: 2

[Copilot]

Copilot's findings

• Files reviewed: 22/22 changed files
• Comments generated: 1

[Copilot]

Copilot's findings

• Files reviewed: 22/22 changed files
• Comments generated: 0 new

[elinor-fung]

@copilot what is the size impact of this change on non-Windows apphosts?

[Copilot]

@copilot what is the size impact of this change on non-Windows apphosts?

On a local linux-x64 Release build, comparing the current branch to the pre-PR baseline (ad0692a2a5, before the apphost conversion commits), the non-Windows apphost is smaller now:

• apphost: 39,352 -> 31,944 bytes (-7,408, -18.8%)
• .text in apphost: 18,330 -> 14,493 (-3,837)
• singlefilehost: 10,809,016 -> 10,808,312 bytes (-704, effectively flat)

So the earlier size regression was transient while dn-containers was still in the picture. With the latest branch state, I’m not seeing a net size increase for the non-Windows apphost on Linux.

[am11]

On windows, we still have apphost.windows.cpp and it remains C++.

Missing the 'reason' why we can't avoid this divergence?

[elinor-fung]

On windows, we still have apphost.windows.cpp and it remains C++.

Missing the 'reason' why we can't avoid this divergence?

Honestly, no real reason - apphost for non-Windows is the scope Jeremy started with and I didn't change it. But since I started with the scope of nethost for all platforms, most of the shared helpers and such should be converted now - doing the rest of apphost for Windows shouldn't be much more.

[MichalPetryka]

Can we ask copilot for a rebase and remeasure with #129762? The previous values seem to be without that.

[am11]

most of the shared helpers and such should be converted now - doing the rest of apphost for Windows shouldn't be much more.

I've the same feeling, it'll be easier to reason about when they are on the same plan.

[Copilot]

Copilot's findings

• Files reviewed: 21/21 changed files
• Comments generated: 3

[Copilot]

Copilot's findings

• Files reviewed: 26/26 changed files
• Comments generated: 0 new

[Copilot]

Copilot's findings

• Files reviewed: 26/26 changed files
• Comments generated: 2

[am11]

LGTM!

[github-actions]

Workflow state for the Holistic Review Orchestrator.

{
  "version": 5,
  "last_dispatched_commit": "a2d778f3b75100e2ce3062f55b28eb5c6279acc5",
  "last_dispatched_base_ref": "main",
  "last_dispatched_base_sha": "e220bf0be5ec6d43445c6af075a8c7729ba01f7d",
  "last_reviewed_commit": "a2d778f3b75100e2ce3062f55b28eb5c6279acc5",
  "last_reviewed_base_ref": "main",
  "last_reviewed_base_sha": "e220bf0be5ec6d43445c6af075a8c7729ba01f7d",
  "last_recorded_worker_run_id": "29689447368",
  "review_attempt_commit": "",
  "review_attempt_base_ref": "",
  "review_attempt_count": 0,
  "max_review_attempts": 5,
  "review_history_format": "holistic-review-disclosure-v1",
  "review_history": [
    {
      "commit": "a2d778f3b75100e2ce3062f55b28eb5c6279acc5",
      "review_id": 4730875069
    }
  ]
}

[github-actions]

Holistic Review

Motivation: Clear and well-justified. The apphost/corehost was a mix of C++ that pulls in the STL and C++ runtime; rewriting the non-Windows (and now Windows) apphost as pure C shrinks the statically-linked host and removes C++ runtime dependencies. This is a legitimate, host-team-driven cleanup, and the maintainers who own this area (elinor-fung, jkoritzinsky, am11) drove the direction through many review rounds.

Approach: Sound. The PR consolidates duplicated print/string/library-loading logic into C pal_* helpers and makes the C++ pal:: wrappers delegate to them, avoiding two parallel implementations. Ownership contracts for the new heap-allocated buffers (host_path, app_path, app_root, resolver dotnet_root/fxr_path) are documented and the free() paths in exe_start and hostfxr_resolver_cleanup look correct and leak-free on the paths I traced. The static/standalone/dotnet resolver split cleanly separates the singlefilehost, dynamic-apphost, and muxer cases.

Summary: ⚠️ Needs Human Review. I found no concrete correctness bugs in the changed code, and the refactor is faithful to the original behavior. However, this is a large (26 files, ~1800/-1300) change to the critical, hard-to-unit-test apphost/host startup path across all platforms, and the ADO runtime CI leg currently reports failures on the head commit. A host-area maintainer should confirm those CI failures are unrelated/infra (Build Analysis + a runtime leg reporting build errors) and that single-file, GUI-apphost error dialog, and Windows event-log paths were exercised, since those are exactly the paths with no automated coverage. Minor observations below are non-blocking.

---

Detailed Findings

✅ PAL deduplication and ownership

The move of file_vprintf/err_print_line/out_vprint_line, load_library/get_symbol/unload_library, and is_path_fully_qualified into C pal_* functions with the C++ pal:: layer delegating is a clean, correct-by-construction dedup. pal_is_path_fully_qualified correctly folds in the old Windows is_path_not_fully_qualified logic (UNC, DOS-device \\?\, and drive-absolute) and adds the missing NULL guard used by callers.

✅ Buffer and encoding handling

pal_utf8_to_palstr is length-checked on both platforms (required > out_len on Unix; MultiByteToWideChar size probe on Windows), and the dynamic buffer growth in pal_load_library's trace path and format_alloc/buffering_trace_writer in apphost.windows.c handle realloc failure without leaking. report_outdated_framework_error and the bundle/startupinfo dispatch preserve the original error-writer propagation semantics.

💡 Minor: dead length check in apphost.c

In is_exe_enabled_for_execution, the early guard binding_len == 0 || binding_len >= app_dll_len (with app_dll_len == EMBED_MAX) already guarantees binding_len < EMBED_MAX, so the subsequent if (binding_len > EMBED_MAX - 1) check is unreachable. Harmless, but could be dropped for clarity. Non-blocking.

💡 Follow-up: test coverage for host startup

The host startup paths reworked here (single-file bundle detection, GUI apphost Task Dialog / applaunch URL, Windows Event Log reporting, v1 fallback) are not covered by unit tests and rely on the existing host functional tests plus manual validation. Given the blast radius, a maintainer should confirm the relevant host functional tests pass on all target platforms before merge.

Note

This review was generated by this repository's Holistic Review agentic workflow to complement the built-in Copilot review.

Generated by Holistic Review · 196.9 AIC · ⌖ 10.8 AIC · ⊞ 10K