gpu: the PC-sampling label set

[dpsoft]

Task 9 of the GPU PC sampling plan, minus gpu_serialized (deferred — see below).

ProjectExecutionsWith(snap, ProjectionConfig) beside the unchanged ProjectExecutions, same split shape as CountingSink.Snapshot/SnapshotWith. Per PC-derived sample: gpu_pc_attrib and gpu_src_status unconditionally, gpu_src_file/_line/_func only under resolved. gpu_pc and gpu_stall unchanged in form. Frames gain nothing — a negative test asserts no frame name contains gpu:pc, gpu:src or any stall-reason string.

The anti-forgery rule needed strengthening, and the reason is new

Overwriting reserved names was sufficient while every label was unconditional. It is not once they are conditional: a producer tag named gpu_src_file survives exactly where the projection derives nothing — a forged source location sitting beside gpu_src_status="no-module".

projectionLabels now clears all seven reserved per-sample names right after the Tags copy. Reserved names win by absence as well as by value. That also closes the same pre-existing hole for gpu_stall when the producer named no reason, and for all of them on an execution with no PC samples.

Three decisions

A nil is answered by an empty store, not a branch. Task 4's "the store is the single place gpu_src_status is decided" stays structurally true, and "no store configured" yields the same honest no-module as "the cubin never arrived". Both cmd/ drivers are still nil today, so their profiles read no-module everywhere — which points at the missing wiring instead of looking like a pre-Phase-6 profile.

gpu_src_file is the basename, via path.Base — not filepath.Base, because the separator belongs to the build host, not the agent. Full paths vary per build, inflate the string table for no reader benefit, and leak build-environment layout into a profile that may be shared. The fixtures resolve to /tmp/perf-agent-cubin-fixtures/single.cu, so the test fails the moment a directory survives.

gpu_pc_attrib renders "unset-pc-attrib" for a PC-bearing view the join left undecided, mirroring SrcStatus's "unset-src-status". Omitting it would hide a join bug behind the one reading (exact) that claims vendor truth.

The cap drops the numerous label, not the useful one

ProjectionConfig.MaxDistinctPCLabels, default 20,000 — reasoned, not measured: the top of the range the plan's own size estimate calls tolerable. Past it only gpu_pc gives way; the sample keeps its weight and every other label, because gpu_stall and gpu_src_* are coarser and more actionable. An offset already emitted is always readmitted, since the cap bounds the pprof string table and a repeat costs it nothing.

Suppressions land in ProjectionStats.PCLabelsSuppressed, are raised as an anomaly by the new JoinHealthWith, and both drivers were switched over so the counter is actually printed — a profile that silently lost its PC labels looks identical to one that never had them.

is deferred, and absent

Not emitted, not stubbed, no placeholder. Its values come from Task 10's sampling windows, and the only default available is "false" — which the design names as the one answer that must never be reachable by accident. Report §7 states the follow-up: it rides on every execution, so it belongs in projectionLabels' shared map, set after the Tags copy, with the anti-forgery test extended.

Verification

All four gpu_src_status values, all four gpu_pc_attrib values, kernel-ambiguous never coinciding with gpu_ambiguous="true", a Tags entry losing to the derived value, the cap suppressing gpu_pc and only gpu_pc with an exact count, and the negative frame test.

Seven mutations applied one at a time, each caught — including the cap dropping gpu_stall, the status made conditional, the directory kept, and the reserved-name clearing removed.

Cannot verify

The real distinct-value counts per label, which is what would turn the 20,000 ceiling into a measured number — DistinctPCLabels beside PCLabelCap is the instrument.

Inherited and load-bearing: that functionIndex is the .symtab index, and that pcOffset is function-relative in the sense the line table is — the one failure in this set that is not self-announcing, because a wrong resolved line looks exactly like a right one.