[Debian/Build] Build ml-api on Ubuntu 26.04 (gcc fallback, C++17)

[myungjoo]

Why

ppa:nnstreamer/ppa now publishes nnstreamer, nnstreamer-edge, ssat, tensorflow2-lite and edgetpu for Ubuntu 26.04 (resolute), but ml-api cannot follow: the Launchpad recipe nnstreamer-ml-api-daily stops at the build-dependency stage on 24.10 and later because every compiler alternative in debian/control (gcc-9 | gcc-8 | ... | gcc-5) has been dropped from the archive. Once past that, all four gtest binaries fail to compile on 26.04 because GoogleTest 1.17 refuses anything older than C++17.

What

• debian/control: append the unversioned gcc as the last compiler alternative, the same fallback nnstreamer (#4895) and nnstreamer-edge ([Build] Compile and package on Ubuntu 26.04 (gcc 15) nnstreamer-edge#256) already carry. 22.04 and 24.04 still resolve gcc-9 first, so nothing changes there.
• meson.build: cpp_std=c++14 -> c++17. The Android build already passes -std=c++17 and nnstreamer has been on C++17 for years.

How it was verified

Full packaging build in an ubuntu:26.04 container with ppa:nnstreamer/ppa enabled, mk-build-deps from the modified control, then dpkg-buildpackage -us -uc -b (gcc 15.2.0, meson 1.10, openjdk 11 from the archive). Without the second commit the build stops with

/usr/include/gtest/internal/gtest-port.h:273:2: error: #error C++ versions less than C++17 are not supported.

in unittest_capi_inference, unittest_capi_inference_single, unittest_capi_inference_latency and unittest_capi_datatype_consistency. With both commits the build completes, override_dh_auto_test (packaging/run_unittests.sh) runs, and all eight .debs are produced.

No functional change, so no new test case; the existing unit tests are what the C++17 change makes buildable again on 26.04.

Part of the 26.04 PPA work tracked in nnstreamer/nnstreamer#4902. After this merges, the Launchpad recipe needs Resolute added and a build requested (I can do that).

🤖 Generated with Claude Code

[myungjoo]

[Automated review by a separate Claude agent, relayed to this PR by the author's account]

Reviewed both commits (7ffa297, 492e55e) against a fresh checkout of the branch.

Verdict: approve. No blocking findings; one minor commit-message nit.

What was checked

• C++ surface affected by the standard bump. The only C++ sources in the repo are the nine gtest binaries under tests/capi/*.cc; every shipped library (c/src, java/android/.../jni) is C. So cpp_std=c++17 changes nothing about installed artifacts or ABI. A grep for constructs removed in C++17 (std::auto_ptr, random_shuffle, dynamic exception specifications, register, unary_function/bind1st/ptr_fun) finds nothing in the test sources.
• Other build files do not disagree. java/android/nnstreamer/src/main/jni/Application.mk already sets APP_CPPFLAGS := -std=c++17; there is no CMake, and no cpp_args/override_options anywhere in the meson tree that would pin a different standard.
• Tizen GBS. packaging/machine-learning-api.spec strips -std=gnu++11 from CXXFLAGS explicitly so that meson decides the standard, so the bump does propagate to Tizen builds. That is the same toolchain nnstreamer has been building with cpp_std=c++17 for years, so no new risk there.
• 22.04 / 24.04. Both series still carry gcc-9 in universe, and the new | gcc is appended last, so apt resolution on jammy/noble is unchanged. gcc-9 with -std=c++17 -Werror is exercised today by nnstreamer's own jammy pdebuild, and this PR's jammy pdebuild job will confirm it for these test files.
• meson floor. meson_version: '>=0.50.0' accepts c++17 as a cpp_std value, and meson has had the c++17 -> c++1z fallback for older compilers since before 0.50.
• Bisectability. Commit 1 alone only lets 26.04 get further (to the gtest error); it cannot regress any series that built before. Commit 2 alone is harmless on every series. Either order is safe.
• Test policy. Build-script-only change; the PR body states why no new test case is added, as CLAUDE.md requires.
• Commit hygiene. Subjects follow [Component] Summary ([Debian], [Build]); each commit has exactly one Signed-off-by for the author plus a Co-Authored-By trailer; no duplicate sign-off.

Findings (ranked)

1. Minor / commit message accuracy (492e55e). The body says "gcc 7 and later, the oldest compiler the packaging accepts, support it", but debian/control still lists gcc-6 | gcc-5 (>=5.4) as alternatives. On those compilers meson would fall back to -std=c++1z, which gcc 5/6 only partially implement. In practice this is moot (no supported series ships gcc-5/6, and the gtest that needs C++17 is not on any series that has them), but the sentence is wrong as written. Suggest rewording to something like "every compiler a current series can resolve (gcc-9 and later) supports it", or, separately, pruning the gcc-5/6 alternatives if you want the claim to hold literally. Not worth blocking on.

2. Informational, not a defect. No CI job builds on 24.10+/26.04, so the 26.04 verification is manual-only and will stay that way until the Launchpad recipe gets Resolute added (which the PR body already plans). If you want it covered in-repo later, the pdebuild matrix in .github/workflows/pdebuild.yml would need a second distroname entry; out of scope for this PR.

Nothing else. The change is the minimum needed and mirrors what nnstreamer (#4895) and nnstreamer-edge (#256) already merged.

[myungjoo]

[Relayed on behalf of the author by a Claude agent]

Thanks. Finding 1 is fixed: the second commit's message now says "every compiler a current series can resolve from the packaging, gcc-9 and later, supports it" instead of the wrong gcc-7 claim (commit message only, no content change; force-pushed to the same branch). The gcc-5/6 alternatives are left as they are in this PR since no series that still has them is affected.

Finding 2 is understood: 26.04 coverage comes from the Launchpad recipe once Resolute is added there, which is the next step after this merges.

[myungjoo]

[Automated re-review by a separate Claude agent, relayed to this PR by the author's account]

Re-reviewed after the force-push that rewords commit 792c886's message (7ffa297 unchanged). Verified directly against origin/fix/ubuntu-26.04-gcc rebased on upstream/main, not just the GitHub diff view.

Verdict: approve. The prior minor finding is fixed, and no new issues were found on independent re-verification.

Prior finding re-checked

• Finding 1 (commit message accuracy) — confirmed fixed. Commit 792c886 now reads "Every compiler a current series can resolve from the packaging, gcc-9 and later, supports it," replacing the earlier "gcc 7 and later" claim. This is accurate: apt resolves gcc-9 first on 22.04/24.04, and falls through to the newly-added unversioned gcc (→ gcc-13/14/15) only on 24.10+/26.04 — every path lands on gcc-9 or newer. The stale gcc-6 | gcc-5 (>=5.4) alternatives are left untouched, as the author noted, since no currently supported series resolves them; this doesn't affect the truth of the reworded sentence. Message-only change, oid list confirms no content diff: debian/control/meson.build are byte-identical to the first review.
• Finding 2 (26.04 not covered by in-repo CI) — unchanged, non-blocking, acknowledged. Still true; out of scope per the author's plan to add Resolute to the Launchpad recipe after merge.

Independent verification performed this pass

• Diff is exactly the same 2-line change as before: debian/control build-dep alternative, meson.build cpp_std=c++14 → c++17. Nothing else touched — proportionate to the stated goal.
• Grepped all 9 tests/capi/*.cc files (the only C++ sources in the repo besides the Android JNI glue) for constructs removed/restricted in C++17 (auto_ptr, random_shuffle, unary_function/binary_function, bind1st/bind2nd/ptr_fun, dynamic exception specs, register as a storage-class keyword). No real hits — only benign uses of the English word "register" in comments.
• Repo-wide grep for cpp_std, c++14, c++11, gnu++ found only the two expected sites: meson.build (now c++17) and packaging/machine-learning-api.spec (strips -std=gnu++11 from CXXFLAGS so meson's cpp_std wins on Tizen GBS too). No other subproject or build file pins a conflicting standard.
• java/android/nnstreamer/src/main/jni/Application.mk already hardcodes APP_CPPFLAGS := -std=c++17, so the NDK build is unaffected and was already ahead of this change.
• meson_version: '>=0.50.0' in meson.build is compatible with cpp_std=c++17.
• .github/workflows/pdebuild.yml only builds ubuntu-22.04 (jammy) with gcc-9-resolvable toolchain; it runs on every PR and will build the 9 gtest binaries under the new cpp_std=c++17, so any real C++17 incompatibility introduced in this or a future PR would fail that CI job before merge. There is no 24.04 or 26.04 job in-repo (matches the PR's own statement and its follow-up plan), so 26.04 verification stays manual until the Launchpad recipe change lands.
• No architecture or public API change; no documentation update needed.
• Build-script-only change with no functional/behavioral change to installed libraries — the PR body's stated rationale for adding no new test case holds up under inspection (existing gtest suite is what re-validates the C++17 bump; the fallback compiler line has no test surface of its own beyond "the package still builds").

Note

The PR is still marked DRAFT with a DO NOT MERGE label — presumably intentional pending the Launchpad recipe follow-up mentioned in the PR body, not a code issue.

Nothing else to add; this remains the minimal change needed and continues to mirror the fallback pattern already merged in nnstreamer and nnstreamer-edge.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR appends the unversioned gcc as the last Build-Depends alternative in debian/control and raises meson's cpp_std from c++14 to c++17 so the gtest binaries compile against GoogleTest 1.17 on Ubuntu 26.04. Verified: Debian alternative syntax and archive facts via the Launchpad API (jammy/noble still publish gcc-9 in universe; oracular, plucky, questing and resolute do not; resolute's default gcc is 4:15.2.0 and libgtest-dev is 1.17.0); no other file pins a C++ standard (Android Application.mk already uses -std=c++17, the Tizen spec strips -std=gnu++11 so meson's value wins, no cpp_args / override_options elsewhere); the only C++ sources are the nine tests/capi/*.cc files and none use C++17-removed constructs; the Tizen GBS x86_64 lane with unit_test 1 built the PR head 792c886 with GCC 14.2.0, compiled all C++ objects, linked all nine gtest binaries and ran them; merge-tree against main is clean; both commits carry Signed-off-by with accurate messages. The sibling changes nnstreamer/nnstreamer#4895 and nnstreamer/nnstreamer-edge#256 apply the byte-identical | gcc suffix, and nnstreamer's meson.build is already at cpp_std=c++17. Approving; the items below are non-blocking.

1. [Low] The Ubuntu pdebuild path is not CI-verified for this PR because the workflow is disabled — .github/workflows/pdebuild.yml:3-5 declares pull_request: branches: [main], but the Actions API reports the workflow as disabled_inactivity (last run: scheduled, 2026-08-09 on e7fedae), and no pdebuild run exists for 792c886. The earlier automated comments on this PR saying "this PR's jammy pdebuild job will confirm it" are therefore not accurate. For a 2-line change this is acceptable given the GBS lane compiled every gtest binary under cpp_std=c++17 and the author reports a full dpkg-buildpackage in an ubuntu:26.04 container, but please re-enable the workflow (gh workflow enable pdebuild.yml --repo nnstreamer/api) and ideally re-trigger it on this PR so the jammy path (gcc-11 + gtest 1.11 + -Werror) is covered too. daily-build-gbs.yml and nns_api_testhub.yml are also disabled_inactivity.
2. [Low] The compiler alternative list is vestigial; no inconsistency, but the commit text slightly overstates its effect — debian/rules never sets CC / CXX and there is no g++ alternative, so meson picks cc / c++ from build-essential. On jammy/noble apt installs gcc-9 first as the commit says, but the actual compiler is gcc-11 / gcc-13, and gcc-15 on resolute. The fallback is safe and mirrors nnstreamer / nnstreamer-edge exactly, so no change requested. Optional follow-up across all three repos: replace the gcc-9 | ... | gcc-5 (>= 5.4) | gcc group with plain gcc, g++ (or rely on build-essential).
3. [Low] Docs / changelog: nothing to update. No README or Documentation file mentions the C++ standard or compiler version, and debian/changelog is bumped only at release commits.

No back-door or suspicious behavior found: the diff is exactly two lines in two build files, with no scripts, URLs, dependencies, or workflow changes.

[myungjoo]

Libraries we use are requiring C++17 these days :(