[workflows] Run extended ROCgdb tests in parallel on PRs

[lumachad]

Summary

Add five extended test configurations that run in parallel with the
standard tests on every pull request. All jobs reuse the single build
from therock-build-linux (no extra build cost) and fan out together
once the build completes.

The extended configurations are:

Job	Options	Architectures
Test GPU (-O3)	--optimization=-O3	all (expands with ci:run-all-archs)
Test GPU (-O3 -flto)	--optimization='-O3 -flto'	all (expands with ci:run-all-archs)
Test check-read1	--check-type check-read1	gfx94X only
Test check-readmore	--check-type check-readmore	gfx94X only
Test GPU (hip board)	--runtestflags=--target_board=hip --tests gdb.base/...	gfx94X only (GPU only)

GPU runner budget: without ci:run-all-archs, all five extended jobs
run on gfx94X only (5 extra GPU jobs). With ci:run-all-archs, the two
optimization jobs each fan out to all six architecture families (12 extra
GPU jobs from optimization alone, plus 3 for check-type/hip-board). The
multi-arch coverage for optimization jobs is intentional: the goal is to
catch compiler-specific failures across all supported architectures.

To support per-config scoping, therock-test-packages.yml gains one
new input and updates another:

• test_options — appended to each component's test_script (passed
  via a TEST_OPTIONS env var in the jq patch step to avoid injection).
  When test_options contains --tests, any pre-existing --tests in
  the upstream test_script is stripped first to make the override
  explicit rather than relying on argparse last-wins.
• projects_to_test — restricts the component matrix (already existed;
  gains a description). The two optimization jobs use
  rocgdb-gpu,rocgdb-corefile since the options are compiler-specific;
  the hip board job uses rocgdb-gpu only.

Commits

1. [workflows] Drop now-redundant test_rocgdb.py path rewrite — TheRock
   now emits the installed script path directly; the gsub rewrite was
   already a no-op at the pinned commit. Removes OUTPUT_ARTIFACTS_DIR.
2. [workflows] Run extended ROCgdb tests in parallel on PRs — the
   extended jobs and updated therock-test-packages.yml inputs.

[lumachad]

Blocked on failures fixed by #247.

[lumachad]

Blocked on #261.

[lumachad]

I think this is a good first cut. Once we merge #296, we should be able to run this one clean.

I'm still unsure if we should always run these or if we should put the extra tests behind a selectable label. Given we want to run these optimization tests, I think we should do it for every ROCgdb PR.

[lumachad]

@spatrang Could you please eyeball this one?

[lumachad]

@spatrang Could you please eyeball this one?

@spatrang ping?

[spatrang]

Reviewed at head a06e86a. Good direction and the mechanism is sound, but there is one blocker plus some redundancy against options TheRock already sets. Details inline.

Highlights:

• Blocker: the hip board job runs 1 of its 10 tests, and that one reports unsupported, so it passes while testing nothing.
• CI is red: check-read1 and check-readmore both fail on gdb.rocm/addr-bp-gpu-no-deb-info.exp.
• base_test_options and base_gpu_test_options duplicate --parallel -f 0.25 --toolchain llvm, which TheRock already puts in every rocgdb test_script.

Not line-anchorable: the PR description is stale. It lists six configurations (there are five, since --parallel became the shared default) and says the hip board job runs only gdb.base/break.exp when the code lists ten.

[lumachad]

Updated with the fixes suggested. I'll let CI run again.

[spatrang]

Second round, reviewed at head 94425ff. Thanks for the thorough turnaround — the blocker is genuinely fixed, and dropping base_test_options / base_gpu_test_options outright was better than what I asked for, since it also reverts the unintended change to the three pre-existing test jobs.

I verified the new patch-scripts step against the real component JSON and the hip board case now produces a single --tests with the full list. I also confirmed that removing OUTPUT_ARTIFACTS_DIR is safe: therock-test-component.yml defines its own copy (line 42) and passes it to the test step (line 71), and workflow-level env: never propagated into the called workflow anyway.

Two remaining items inline, both of which I think should block merge.

[spatrang]

LGTM — approving at 1282d323. Both blocking items are resolved, and I verified them rather than relying on the replies.

Verified the patching logic. I replayed the new patch-scripts step against all three real rocgdb components across five test_options cases (empty, --check-type, --optimization=-O3, --optimization='-O3 -flto', and the hip board case), then fed every resulting command line through a mirror of test_rocgdb.py's parser. All parse cleanly:

• Empty test_options is a proper no-op, so the three standard jobs are untouched.
• rocgdb-corefile keeps its full five-file --tests list when the caller does not pass --tests.
• The hip board case strips the upstream --tests completely, including the multi-value corefile form that broke the previous regex, and lands a single --tests with all ten files.

Verified on hardware. Run 33069089574 is green across all 26 jobs including TheRock CI Summary. The hip board job now reports:

Number of tests to run: 10
# of expected passes		38
# of unsupported tests		3
[✓] OVERALL STATUS: PASS

Up from the single skipped test that started this review. check-read1 and check-readmore also passed across all three components in this run, which supports the flaky-runner explanation over a systematic failure.

Thanks for the fast and thorough turnaround, particularly for dropping base_test_options / base_gpu_test_options outright instead of just documenting them — that removed the unintended behaviour change to the three pre-existing jobs and made the whole diff smaller.

One non-blocking question inline about the break.exp swap, plus two cosmetic notes below. None of these need to hold the merge.

Cosmetic, for a follow-up whenever convenient:

• Test GPU (check-read1) and Test GPU (check-readmore) do not set projects_to_test, so they also run rocgdb-cpu and rocgdb-corefile — visible in the run as Test GPU (check-read1) / Test rocgdb-cpu. The names claim GPU. Either scope them to rocgdb-gpu or rename to Test check-read1. This one is partly my fault for suggesting the Test GPU (...) pattern.
• The description says the hip board job uses rocgdb-gpu,rocgdb-corefile (the code uses rocgdb-gpu only) and that the workflow "gains two new inputs", though projects_to_test already existed and only gained a description.

[lumachad]

LGTM — approving at 1282d323. Both blocking items are resolved, and I verified them rather than relying on the replies.

Verified the patching logic. I replayed the new patch-scripts step against all three real rocgdb components across five test_options cases (empty, --check-type, --optimization=-O3, --optimization='-O3 -flto', and the hip board case), then fed every resulting command line through a mirror of test_rocgdb.py's parser. All parse cleanly:

• Empty test_options is a proper no-op, so the three standard jobs are untouched.
• rocgdb-corefile keeps its full five-file --tests list when the caller does not pass --tests.
• The hip board case strips the upstream --tests completely, including the multi-value corefile form that broke the previous regex, and lands a single --tests with all ten files.

Verified on hardware. Run 33069089574 is green across all 26 jobs including TheRock CI Summary. The hip board job now reports:

Number of tests to run: 10
# of expected passes		38
# of unsupported tests		3
[✓] OVERALL STATUS: PASS

Up from the single skipped test that started this review. check-read1 and check-readmore also passed across all three components in this run, which supports the flaky-runner explanation over a systematic failure.

Thanks for the fast and thorough turnaround, particularly for dropping base_test_options / base_gpu_test_options outright instead of just documenting them — that removed the unintended behaviour change to the three pre-existing jobs and made the whole diff smaller.

One non-blocking question inline about the break.exp swap, plus two cosmetic notes below. None of these need to hold the merge.

Cosmetic, for a follow-up whenever convenient:

• Test GPU (check-read1) and Test GPU (check-readmore) do not set projects_to_test, so they also run rocgdb-cpu and rocgdb-corefile — visible in the run as Test GPU (check-read1) / Test rocgdb-cpu. The names claim GPU. Either scope them to rocgdb-gpu or rename to Test check-read1. This one is partly my fault for suggesting the Test GPU (...) pattern.
• The description says the hip board job uses rocgdb-gpu,rocgdb-corefile (the code uses rocgdb-gpu only) and that the workflow "gains two new inputs", though projects_to_test already existed and only gained a description.

Addressed these cosmetic issues as well in a new push.