Silence the Windows HIP build's 83,362 warnings, and keep the Linux ones

[bernardladenthin]

Measured in run 38172107721, and the comparison is the whole argument:

job	log lines	warnings
Windows ROCm/HIP	714,345	83,362
Windows SYCL	11,981	724
Windows CPU (clang)	2,627	407
Linux CUDA	3,266	9
Linux ROCm/HIP	2,455	9
C++ Tests Windows	1,655	66
Windows arm64	1,499	24
Windows CUDA	1,068	0

The Windows ROCm job is 60× larger than the next biggest and carries 98.5 % of every warning
its Windows jobs emit. The Linux ROCm job compiles the same upstream sources with the same TheRock
clang for the same 23 GPU targets — and produces nine warnings.

The cause is the language, not the arch count

Upstream compiles the HIP sources as plain C++ on Windows (CXX_IS_HIPCC on WIN32) and as the
HIP language on Linux. Clang's C++ frontend reports what the HIP one does not:

class	count
-Wsign-compare	29,673
-Wnested-anon-types	17,964
-Wmissing-noreturn	16,378
-Wunused-variable	6,404
-Wmissing-braces	5,988
-Wunused-parameter	3,386
-Wcast-qual / -Wunreachable-code-break / 2 more	1,529

Every one is in ggml/src/ggml-cuda/*.cu — upstream's own code. Not one line of ours, and nothing
we would ever act on.

-w on ggml-hip, Windows only

Suppressing a whole warning set is normally the wrong instrument. What makes it right here is the
Linux job: it compiles the same sources and keeps every warning, in a log short enough to read. The
information is preserved where it is legible and dropped where it is multiplied by 23.

It also ends the alternative, which was chasing one more class per toolchain bump — eight today. And
it subsumes the -Wno-ignored-attributes added for the dllimport flood, so that becomes one flag
instead of a growing list.

Errors are untouched: a real compile failure still fails the job.

Whether other jobs need the same: measured, not assumed

No. Outside the ROCm job every log is between 1,068 and 11,981 lines, which is readable. SYCL's
724 warnings come from icx on upstream code and may carry information worth keeping; nothing there
is multiplied into six figures.

Upstream

Worth an issue with that table — their own windows-rocm job pays the identical cost, and the fix is
theirs: either the warnings or the C++-on-Windows choice. Recorded in TODO.md next to the other
three measured upstream findings.

check-natives.py 0 disagreements. The effect is only observable in a ROCm job: this is one compile
flag on one target, and the next run's line count is the measurement.

🤖 Generated with Claude Code

https://claude.ai/code/session_01E2h8gXyyE5UeimkL9vQv9G

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud