Use two-pass IPO for OpenMP offload as well

[sbryngelson]

The NVHPC two-pass IPO (-Mextract/-Minline) was applied to OpenACC and skipped wholesale for OpenMP offload, so that lane gave up cross-file inlining entirely. The reason turns out to be a single compiler bug rather than anything structural:

nvfortran-Fatal-.../tools/fort2 TERMINATED by signal 11
gmake[3]: *** [.../m_data_output.fpp.f90.o] Error 2

fort2 segfaults inlining into m_data_output under OpenMP offload. OpenACC compiles the same file cleanly.

This is the same failure class as the existing m_start_up/m_cbc exclusions — OpenMP offload just reaches it in one more place — so it gets the same treatment, scoped with MFC_OpenMP so the OpenACC build keeps full inlining there and loses nothing.

Verification

On A100 / NVHPC 25.11:

• builds clean, all four targets, zero fort2 crashes
• 1D sweep: 175 passed, 0 tolerance mismatches
• the six cases corrupted by the OpenACC IPO bug (Stop NVHPC IPO silently miscompiling the six-equation relaxation kernel #1878) all pass here

(The 10 execution failed entries in my local sweep are a pre-existing Chemistry environment issue on this box — identical UUIDs in every configuration, including unmodified builds.)

Only one extra exclusion was needed, not a cascade.

Worth weighing before merging

OpenMP offload has been the one NVHPC GPU path not subject to this pass, and that independence is exactly what identified the OpenACC miscompile in #1878 — omp passing 713/713 where acc failed 23 is what ruled out the physics and pointed at IPO.

Enabling IPO here buys inlining on that lane at the cost of that cross-check, on a pass with a track record of three separate defects, one of them silent. Reasonable either way, but it should be a deliberate call rather than a side effect of the fix — which is why it's split out.

If you'd rather keep the independent lane, closing this is a perfectly good outcome; the fort2 segfault documented above is worth keeping on record regardless, since it's the actual reason the exclusion existed.

Merge note

Touches the same block as #1878's m_pressure_relaxation exclusion. Whichever lands second needs a one-line rebase.

https://claude.ai/code/session_017zrZooJPhZtZYgg9fJiYhg

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Pull request overview

Enables NVHPC two-pass IPO (-Mextract/-Minline) for OpenMP offload builds while adding a targeted per-source exclusion to avoid a known fort2 crash during inlining.

Changes:

• Apply NVHPC two-pass IPO to OpenMP offload targets (previously skipped).
• Add conditional no-inline exclusion for m_data_output only under the OpenMP offload configuration.
• Refactor the no-inline list to allow conditional extensions cleanly.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.