Add lizard cyclomatic complexity check to pre-commit

[Hardcode84]

Gate new code on CCN <= 50 and length <= 300. Existing violations are grandfathered via whitelizard.txt. Excludes vendored LLVM headers, generated code, tests, and examples.

[tgymnich]

CCN is a bad metric for code quality. What problem does this solve? Do we have serious intentions to fix all the exceptions?

[Hardcode84]

CCN is a bad metric for code quality. What problem does this solve? Do we have serious intentions to fix all the exceptions?

CCN 50 is extremely generous and all the whitelisted functions are legit candidates for refactoring. Even if we don't plan to fix them immediately, this will prevent us from creating the new ones.

[tgymnich]

That's 50 2 unvetted python dependencies running arbitrary code on my machine just to prevent nested if statements.
IMO the cost use benefit is not there for me. This can also be done using copilot reviews, which may or may not provide more value.

[Hardcode84]

Lizard Complexity Report

Generated: 2026-04-02
Thresholds: CCN > 50 or length > 300

Whitelisted Violations (sorted by CCN)

CCN	NLOC	Length	Tokens	File	Function
96	577	658	5002	waveasm/lib/Transforms/AssemblyEmitter.cpp:342	KernelGenerator::generateOp
91	294	457	2538	waveasm/lib/Transforms/Liveness.cpp:265	computeLiveness
88	148	158	1573	waveasm/lib/Transforms/handlers/AMDGPUHandlers.cpp:86	handleAMDGPUMfma
82	377	459	2849	waveasm/lib/Transforms/handlers/AMDGPUHandlers.cpp:686	handleGatherToLds
75	561	666	3419	wave_lang/kernel/wave/perf/bench_mxfp4_comparison.py:1798	main
71	362	469	3086	waveasm/lib/Transforms/handlers/AffineHandlers.cpp:398	handleAffineApply
69	338	498	1860	wave_lang/kernel/wave/mlir_converter/water_emitter.py:592	_emit_ops_from_graph
66	215	348	1849	waveasm/lib/Transforms/LinearScanPass.cpp:144	LinearScanPass::processProgram
59	257	338	1883	waveasm/lib/Transforms/TranslateFromMLIR.cpp:1373	handleVectorStore
55	169	253	1360	waveasm/lib/Transforms/handlers/AMDGPUHandlers.cpp:430	handleFatRawBufferCast
51	222	272	1445	wave_lang/kernel/compiler/host_codegen.py:127	isolated_test_call
32	213	523	1716	wave_lang/kernel/compiler/wave_codegen/emitter.py:641	gen_sympy_index
26	200	306	1835	waveasm/lib/Transforms/TranslateFromMLIR.cpp:154	emitSRDPrologue
25	260	306	1227	wave_lang/kernel/wave/asm/scripts/compare_backends.py:1642	main
10	137	433	933	wave_lang/kernel/wave/templates/extend_attention.py:28	get_extend_attention_kernel
7	158	346	1204	wave_lang/kernel/wave/templates/extend_attention_rpe.py:28	get_extend_attention_rpe_kernel
6	139	447	833	wave_lang/kernel/wave/templates/paged_decode_attention.py:56	get_paged_decode_attention_kernels
4	179	420	1002	wave_lang/kernel/wave/templates/gemm.py:873	get_hybrid_streamk_gemm_kernel
3	169	407	929	wave_lang/kernel/wave/templates/gemm.py:464	get_streamk_gemm_kernel
3	318	401	1519	wave_lang/kernel/wave/schedules/gemm_mxfp4_double_buffer.py:1606	get_mxfp4_asymmetric_schedule.mxfp4_dbuf_schedule
2	270	344	1473	wave_lang/kernel/wave/schedules/gemm_mxfp4_double_buffer.py:917	get_mxfp4_dbuf_mixed_pingpong_schedule.mxfp4_dbuf_schedule
1	141	469	781	wave_lang/kernel/wave/templates/speculative_decoding.py:14	get_speculative_sampling_kernel
1	321	367	2521	waveasm/lib/Target/AMDGCN/InstructionInfo.cpp:19	InstructionRegistry::InstructionRegistry
1	7	434	33	wave_lang/kernel/wave/schedules/gemm_mxfp4_double_buffer.py:1575	get_mxfp4_asymmetric_schedule
1	7	367	41	wave_lang/kernel/wave/schedules/gemm_mxfp4_double_buffer.py:896	get_mxfp4_dbuf_mixed_pingpong_schedule
1	7	308	41	wave_lang/kernel/wave/schedules/gemm_mxfp4_double_buffer.py:1265	get_mxfp4_dbuf_mixed_pingpong_shuffle_schedule

Summary

Metric	Count
Total whitelisted functions	26
CCN > 80 (critical)	4
CCN > 50 (high)	11
Length-only violations (CCN <= 50)	15
Worst offender CCN	96 (KernelGenerator::generateOp)
Worst offender length	666 (bench_mxfp4_comparison.py:main)

Breakdown by Component

Component	Violations
waveasm C++	12
wave_lang Python	14
water C++	0

Notes

• Templates and schedules trip the length threshold but have very low CCN (1-10). These are long declarative kernel definitions, not complex logic.
• The waveasm handlers (handleAMDGPUMfma, handleGatherToLds, handleAffineApply) are type-dispatch switches -- inherently branchy but typical for code generators.
• KernelGenerator::generateOp (CCN 96, 577 NLOC) is the single worst function and a strong refactoring candidate.

[ftynse]

Breakdown by component shows that the more robust code review policy that I enacted in water works. Very basically, if people have to actually review your code, they will complain about cognitive complexity much sooner than tools.

I agree with Tim on the intent: if there is no intent to rearchitect this code beyond asking an LLM to refactor it which will just try to game an artificial metric, let's not bother. If somebody wants to run the tool, maybe we can have it as an optional CI?

[harsh-nod]

If this is a pre-commit hook, then we can catch this even before submitting the PR? Doesnt that help reduce the cognitive load on the reviewer?