[tap]

What this changes

This is the roadmap's A8 (row P10; Q6 decided a hard gate, two-sided at ±3 %). Each engine gets host workloads (bench/icount/host_main.cpp, binaries tap_sr_<engine>_hostcount_*):

• The float converter streams 128-frame blocks at 2 and 12 channels.
• scripts/icount.py --target host counts them under valgrind's callgrind and gates them against a new host key in each bench/baselines.json, with the same mechanics as the QEMU legs.
• A new CI job, Instruction-count ratchet (host), runs the gate.

Why

The channel-parallel host path sat outside every ratchet. TAP_DSP_CHANNEL_PARALLEL is off under Helium, the M33's DSP extension and Hexagon, so no QEMU leg compiles it, and a host regression there could land unseen. async's 12-channel rows run that path.

Verification

What pins the count to the code, the compiler and the C library:

• One function counted. Only tap_sr_host_counted(), the streaming loop, is counted (--toggle-collect). Construction, the fixture's sin() calls, process start-up and the environment stay outside.
• Eager binding. LD_BIND_NOW=1, so no PLT resolution lands inside the counted window.
• glibc pinned. GLIBC_TUNABLES holds glibc's CPU-dispatched routines to their x86-64 baseline variants and pins the copy thresholds glibc derives from cache sizes. Before the pin, this machine's memcpy resolved to the AVX variant, 1 % of the 12-channel async count.
• No -march. The compiled code is the x86-64 baseline whatever the runner.

Recorded with g++ 13.3.0 (13.3.0-6ubuntu2~24.04.1), glibc 2.39 (2.39-0ubuntu8.9) and valgrind 3.22; the job prints the runner's versions. All of these count identically:

• two runs of one binary;
• a second build with another generator;
• a replay of the CI job's own build and gate steps.

On CI: the runner reproduced all ten counts exactly, +0 against every baseline with the same checksums (job 115919836408, run 38583667902), and all 32 checks are green.

Engine	Workload	2 ch	12 ch
async	pipeline_float	109,526,808	362,078,536
bridge	up_float	79,352,295	470,621,528
bridge	down_float	99,664,796	593,181,679
rational	up2_float	57,296,876	288,223,796
rational	down2_float	29,983,981	161,054,642

async at 12 channels runs about 30 M instructions per channel over the stream, against 55 M at 2: that is what the channel-parallel path buys. bridge has no frame-major path and costs the same per channel at both counts.

Also checked:

• clang 18 -Werror build, clang-tidy and clang-format are clean on the three new files.
• check_retired_words.py passes.
• The M33 cross configure defines no host workloads, and the README tables are regenerated.

Notes for the reviewer

• The QEMU tables in the READMEs are byte-identical. The host figures are a second table from the same generator, between the same markers.
• The book's icount chapter gains "The host, under callgrind"; the root CLAUDE.md and the roadmap (A8 landed, P10 green) describe it.

🤖 Generated with Claude Code

https://claude.ai/code/session_016GKzq1VSdcMYCRMhddiMaf