fix(cpu-ep): gate x86_64-only prefill fan-out symbols for non-x86 targets

[justinchuby]

Fixes #1415 — the aarch64 clippy gate is red on main.

WIDE_PREFILL_MACS, prefill_fan_out and prefill_column_grain are only reachable from the x86_64 prefill route, so on other architectures they are genuinely dead and -D warnings fails the build. This adds #[cfg(target_arch = "x86_64")] to the three symbols and to the six tests that exercise them, matching the treatment their neighbours already have.

Verified both sides, because gating dead code is only half the job — the risk is silently removing something x86 needs:

• aarch64: reproduced the three never used diagnostics on unmodified main first, then confirmed cargo clippy -p onnx-runtime-ep-cpu --target aarch64-pc-windows-msvc --all-targets -- -D warnings exits 0 with this change.
  Caveat, stated honestly: aarch64 clippy gate is red on main: three prefill fan-out symbols are x86-only but not cfg-gated #1415 cites aarch64-unknown-linux-gnu, whose std is not installed on this box (can't find crate for core). I used aarch64-pc-windows-msvc instead. The symbols are dead on any non-x86_64 target, and that target reproduces the identical three diagnostics and clears them, so it is a valid stand-in — but the exact triple in the issue was not the one I ran.
• x86_64: cargo clippy -p onnx-runtime-ep-cpu --all-targets -- -D warnings exits 0, and the six gated tests still run: 26 passed / 0 failed / 2 ignored.

No logic changes — attributes only.

[justinchuby]

Heads-up from validation (Pris, Tester): this fix removes the three symbols rather than silencing the lint, and that breaks a configuration the aarch64 cross gate cannot see.

prefill_fan_out / WIDE_PREFILL_MACS have two callers gated on different things:

call site	fn	gate
matmul_nbits.rs:2457	run_mlas_shards	#[cfg(feature = "mlas")]
matmul_nbits.rs:6914	borrowed_affine_int4_matmul_prefill	#[cfg(target_arch = "x86_64")]

Off x86 the second disappears but the first does not, so on aarch64 + mlas the caller is compiled and the callee is not:

error[E0425]: cannot find function `prefill_fan_out` in this scope
  --> crates/onnx-runtime-ep-cpu/src/kernels/matmul_nbits.rs:2457:27

Reproduced by mirroring the cfg resolution on the host: retarget the three item gates off x86_64 on main, leave callers alone, cargo check -p onnx-runtime-ep-cpu --features mlas --lib → exit 101. check_cross_compile.sh's aarch64 pass builds default features and mlas is off by default, which is why it stayed green. The rust-coverage macOS-arm64 leg does build it (-D warnings, --features mlas, ci.yml L434/439/501-503) — that is the shipped-wheel path.

Second effect: gating the items forced #[cfg(target_arch = "x86_64")] onto six unit tests, all of which are pure functions of literal arguments. The #1363 fan-out policy now has no aarch64 coverage.

#1382 replaces the gates with per-symbol cfg_attr(.., allow(dead_code)) (the idiom prefill_tile_grain already uses) and ungates the tests. Predicates differ per symbol — prefill_column_grain's only caller is the x86 one, so it gets not(target_arch = "x86_64") while the other two get the union; using the union for all three just relocates the failure to never used on aarch64+mlas. Verified across all four (mlas) x (arch) configs.

No action needed from you — the repair is in #1382, waiting on required CI.

[github-actions]

🔴 Benchmark Regression Detected

Comparison of criterion micro-benchmarks: PR head vs merge-base, measured on the same runner in the same job (base first → PR second).

ℹ️ Absolute times are informational only — they vary with runner load. The % change column is the reliable signal because both sides ran under identical conditions.

Status	Scenario	Base	PR	Change
🔴	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	126.39 µs	223.74 µs	+77.0%
🔴	tokenization/encode_tokens_per_second	376.52 µs	611.63 µs	+62.4%
🔴	sampling_latency/top_k_per_token	52.31 µs	76.79 µs	+46.8%
🔴	sampling_latency/top_p_per_token	387.92 µs	559.88 µs	+44.3%
🔴	matmul/large_generic_f32_threads=8/32x1024x1024	4.42 ms	6.31 ms	+42.8%
🔴	matmul/small_generic_f32_threads=8/1x256x256	47.38 µs	64.08 µs	+35.2%
🔴	matmul/large_generic_bf16_threads=8/32x1024x1024	1.36 ms	1.82 ms	+33.8%
🔴	matmul/large_generic_bf16_threads=1/32x1024x1024	2.23 ms	2.97 ms	+32.8%
🔴	add/medium_f32_threads=1-internal/262144	25.45 µs	33.58 µs	+32.0%
🔴	tokenization/decode_tokens_per_second	6.29 ms	8.22 ms	+30.6%
⚠️	matmul/medium_generic_bf16_threads=8/32x512x512	578.17 µs	727.07 µs	+25.8%
⚠️	gather/large_bf16_threads=1-internal/131072	16.02 µs	19.47 µs	+21.5%
⚠️	sampling_latency/greedy_per_token	3.21 µs	3.89 µs	+21.1%
⚠️	matmul/large_generic_f32_threads=1/32x1024x1024	10.03 ms	12.01 ms	+19.7%
⚠️	add/small_bf16_threads=1-internal/1024	498.8 ns	584.9 ns	+17.3%
⚠️	matmul/medium_generic_f16_threads=8/32x512x512	39.42 µs	45.88 µs	+16.4%
⚠️	add/small_f16_threads=1-internal/1024	510.1 ns	592.8 ns	+16.2%
✅	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	667.13 µs	766.69 µs	+14.9%
✅	gather/small_f16_threads=1-internal/4096	515.8 ns	583.4 ns	+13.1%
✅	matmul/small_generic_f16_threads=1/1x256x256	39.85 µs	44.42 µs	+11.5%
✅	reduce_mean/small_f32_threads=1-internal/4096	16.10 µs	17.87 µs	+11.0%
✅	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	4.27 ms	4.72 ms	+10.5%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	87.73 µs	96.18 µs	+9.6%
✅	qwen3_sampling_processors/top_k_partial_selection	180.20 µs	194.36 µs	+7.9%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	638.00 µs	687.31 µs	+7.7%
✅	reduce_mean/medium_f32_threads=1-internal/65536	252.09 µs	268.20 µs	+6.4%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.68 ms	2.79 ms	+4.2%
✅	matmul/small_generic_bf16_threads=8/1x256x256	39.24 µs	40.86 µs	+4.1%
✅	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	584.33 µs	601.72 µs	+3.0%
✅	gather/small_f32_threads=1-internal/4096	742.2 ns	760.1 ns	+2.4%
✅	add/small_f32_threads=1-internal/1024	223.1 ns	228.4 ns	+2.4%
✅	gather/small_bf16_threads=1-internal/4096	537.4 ns	545.0 ns	+1.4%
✅	logit_processing/seven_processor_chain_per_step	355.82 µs	357.29 µs	+0.4%
✅	matmul/small_generic_f32_threads=1/1x256x256	51.71 µs	51.27 µs	-0.8%
✅	matmul/small_generic_bf16_threads=1/1x256x256	36.82 µs	36.41 µs	-1.1%
✅	qwen3_sampling_processors/top_p_fast_after_top_k	605.53 µs	596.42 µs	-1.5%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.08 ms	1.06 ms	-1.9%
✅	matmul/small_generic_f16_threads=8/1x256x256	40.67 µs	39.67 µs	-2.5%
✅	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	6.67 ms	6.49 ms	-2.8%
✅	qwen3_sampling_processors/top_k_top_p_fast	758.57 µs	734.53 µs	-3.2%
✅	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	140.23 µs	134.64 µs	-4.0%
✅	matmul/large_generic_f16_threads=8/32x1024x1024	106.95 µs	102.58 µs	-4.1%
✅	gather/medium_f16_threads=1-internal/32768	2.98 µs	2.85 µs	-4.1%
✅	qwen3_sampling_processors/top_k_full_sort_baseline	2.62 ms	2.47 ms	-5.6%
✅	add/large_bf16_threads=1-internal/4194304	1.92 ms	1.77 ms	-7.6%
✅	gather/medium_f32_threads=1-internal/32768	4.50 µs	4.12 µs	-8.4%
✅	gather/medium_bf16_threads=1-internal/32768	2.86 µs	2.55 µs	-11.1%
✅	gather/large_f16_threads=1-internal/131072	15.36 µs	13.48 µs	-12.2%
✅	add/large_f16_threads=1-internal/4194304	2.07 ms	1.80 ms	-12.8%
✅	add/medium_f16_threads=1-internal/262144	126.38 µs	109.28 µs	-13.5%
✅	sampling_latency/min_p_per_token	291.19 µs	251.58 µs	-13.6%
✅	grammar_masking/llguidance_compute_mask/32	97.82 µs	83.95 µs	-14.2%
🟢	kv_cache/alloc_dealloc_pages	48.64 µs	41.14 µs	-15.4%
🟢	add/medium_bf16_threads=1-internal/262144	126.62 µs	106.11 µs	-16.2%
🟢	matmul/medium_generic_f16_threads=1/32x512x512	43.39 µs	36.01 µs	-17.0%
🟢	gather/large_f32_threads=1-internal/131072	33.12 µs	27.11 µs	-18.1%
🟢	add/large_f32_threads=1-internal/4194304	907.47 µs	736.71 µs	-18.8%
🟢	matmul/medium_generic_f32_threads=8/32x512x512	1.62 ms	1.25 ms	-22.8%
🟢	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	114.05 µs	68.00 µs	-40.4%

Visual flags: ⚠️ ≥ 15% slower, 🔴 ≥ 30% slower — calibrated against measured runner noise (~27% worst-case on multi-threaded matmul)

Host info

CPU: Apple M1 (Virtual)
Cores: 3
OS: Darwin 25.5.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 4.45 4.47 5.57 }

What this cannot catch

• Regressions in code paths not covered by these benchmarks (e.g., end-to-end decode with a real model)
• Sub-threshold regressions that compound over multiple PRs
• Performance changes that only manifest under GPU execution
• Latency changes in the ORT integration path (these benchmarks exercise the native Rust kernels)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.10%. Comparing base (e19697f) to head (e211f03).
⚠️ Report is 104 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1443       +/-   ##
===========================================
+ Coverage   80.14%   82.10%    +1.96%     
===========================================
  Files         364       12      -352     
  Lines      160978     5471   -155507     
  Branches   160978     5471   -155507     
===========================================
- Hits       129012     4492   -124520     
+ Misses      27310      780    -26530     
+ Partials     4656      199     -4457     

Flag	Coverage Δ
cli-ort-windows	82.10% <ø> (?)
mlas	?
offline	?

Flags with carried forward coverage won't be shown. Click here to find out more.
see 376 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.