[Yaraslaut]

The next batch from the compile-cache audit filed as #1688–#1708. Each change closes a way the cache served a wrong or incomplete result, or one way its reports hid that. Stacked on #1687 (based on its branch); it retargets to master once #1687 merges.

Changes

• -march=native is keyed by the CPU it resolves to (closes fastcache-cc: -march=native is keyed as literal text, so one CPU's object is served to another #1688).
  • The key held the literal argument, so two machines shared an object built for one of them; an AVX-512 object on a CPU without AVX-512 dies with SIGILL.
  • The key now folds what the driver's -### listing of the unit's own line expands native to: gcc's -march=<cpu>, feature switches and --param cache sizes; clang's -target-cpu/-target-feature. None of the listing's paths go in.
  • A line naming native is never dispatched, because the worker would resolve it to its own CPU.
• A depfile asked for without -MF is restored by a hit (closes fastcache-cc: -MD/-MMD without -MF -- a cache hit writes no depfile #1689).
  • -MD/-MMD alone, and Kbuild's -Wp,-MMD,<file>, make the driver write a depfile the launcher did not know about. A hit restored the object and no depfile, so a Makefile build lost every header dependency.
  • The parser now derives the depfile the driver writes: the -o path with its last extension replaced by .d, as measured on gcc and clang.
• A compile that writes a file beside its object is not cached (closes fastcache-cc: compiles that write a second artefact are cached locally (-gsplit-dwarf, --coverage, -ftime-trace) #1691): -gsplit-dwarf, --coverage/-ftest-coverage, -ftime-trace, -fstack-usage, -save-temps, -fdump-*, -fsave-optimization-record, -fcallgraph-info, -aux-info.
• A compile that reads a file the key cannot see is not cached (closes fastcache-cc: files named by codegen flags are not keyed (-fprofile-use, plugins, -specs, -B) #1692): -fprofile-use and the other profile-use flags, plugins, -specs=, -B.
• The reports name what a plain MISS or HIT hid (refs Cache reports hide correctness problems: failed STORE as MISS, stale hit as MISS, failed verify as HIT #1707):
  • UNSTORED: the store did not land.
  • STALE: a hit refused as stale and recompiled.
  • VERIFY-FAILED: a sampled hit whose fresh compile failed. That case was reported "inconclusive" and tallied as a HIT.
• The remaining open findings are linked from the rulebooks' ## Open work. fastcached disk tier: flush_all is undone by a restart #1690 (flush_all undone by a restart) needs a disk-format decision and is not in this PR.

What is missing (why this is a draft)

The full build and all three unit suites pass locally (FastCacheTest, fastcache-cc-tests, fastcache-compile-node-tests). Two of the 258 hygiene checks fail and are not fixed yet:

• test-name-hygiene: the new case "-march=native is keyed by the CPU the driver resolves it to, and by nothing else" in src/apps/fastcache-cc/CacheKey_test.cpp starts with -, which Catch2 reads as an option. It needs renaming, e.g. "Naming -march=native keys the CPU the driver resolves it to".
• table-totals: a table in .agent/rules/distributed-compilation.md (line ~1104) has no <!-- table-total: ... --> marker. This check fails only on this branch, so the cause sits in this branch's change to that file; it is not yet narrowed down.

Also not done:

• No end-to-end Docker or cross-machine re-measurement for this batch. Each fix was checked with a targeted launcher reproduction instead.
• Cache reports hide correctness problems: failed STORE as MISS, stale hit as MISS, failed verify as HIT #1707 is only partly fixed. The verifier still compares only the object, not the depfile or the replayed streams, and the servers still count hits on values the launcher rejects.
• fastcached disk tier: flush_all is undone by a restart #1690 (flush_all undone by a restart) is open and needs a decision. Persisting the flush generation changes the disk format, so existing stores would need conversion; resetting the store file on flush is the alternative.
• The remaining audit findings (fastcache-cc: environment variables that add compiler arguments are not keyed (CL, _CL_, CCC_OVERRIDE_OPTIONS, COMPILER_PATH) #1693–compile node: a fetch's upstream write-back can overwrite a newer local store #1706, Lower-severity cache robustness findings (atomic temp names, unchecked object, flush delay, decompression) #1708) are filed and linked from the rulebooks, not fixed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CnCkwgtUsaSM5HcgB3APNU