perf: in-place select — elide keep-val2 move (#209, VCR-SEL-002)

[avrabe]

What

The stack selector's Select lowering emitted three instructions into a fresh dst:

cmp cond, #0
selectmove dst, val1, NE     ; cond != 0 -> dst = val1
selectmove dst, val2, EQ     ; cond == 0 -> dst = val2

Since select consumes all three operands, val2's register is dead after the op. We reuse it as dst, so the EQ "keep val2" branch becomes a no-op move on itself and is elided — leaving exactly one conditional move (the NE override), which is what native emits for a clamp (x>k)?k:x. Each removed SelectMove is one fewer IT;MOV: −2 bytes and −1 cycle (the M4 executes a predicated-false MOV).

Soundness

In-place only when val2 is safely overwritable, else fall back to the original two-move form:

• not a live param register — the arm regalloc: select_with_stack allocates param registers (r0-r3) for temps/results before live param reads — i64_lowering fuzz class #193 param-clobber class
• distinct from cond — already consumed by the CMP
• distinct from val1 — degenerate; fresh dst keeps it simple

No instruction sits between the CMP and the conditional move in the in-place path, so the flags survive (the same property the two-move form relied on to avoid a flag-clobbering MOVS).

Oracle — RESULT-identical on every frozen differential

This is an optimization: bytes change, results do not.

fixture	result	status
control_step	13/13 → 0x00210A55	PASS
flight_seam inlined	0x07FDF307	MATCH
flight_seam flat	0x07FDF307	MATCH
div_const	338/338	PASS

Measured .text reduction

fixture	before	after	Δ
control_step	378 B	354 B	−24 (−6.3%)
flight_seam inlined	1058 B	1020 B	−38 (−3.6%)
flight_seam flat	1294 B	1244 B	−50 (−3.9%)

Tests

• Updated test_control_flow_select / test_select_stack_mode_with_constants / test_select_with_global_values to assert the NE override (emitted in both forms; the EQ move is what in-place elides).
• New test_select_in_place_elides_keep_move_209 pins one-NE-zero-EQ for a reusable val2.
• 323/323 synth-synthesis lib tests pass, clippy clean.

Falsification

This change is wrong if any module produces a different result than wasmtime where val2's register was actually still needed after the select — i.e. the live-param / aliasing guards missed a case. Watched by the four differentials above + gale on-target against the 5 frozen baselines.

Traceability: VCR-SEL-002 (filed in #282). Part of the #209 codegen-quality program.

🤖 Generated with Claude Code

[avrabe]

On-target results for the in-place select (#283) — this is the biggest lever so far, and it's a clean non-regressing win (eliding moves only):

bench	before (v0.11.34)	after #283	Δ	instrs	seam
flat_flight	255	241	−14 cyc	170→157	0x07FDF307 ✅
controller_step	162	150	−12 cyc	110→99	0x05e33e81 ✅
control_step	158	151	−7 cyc	121→113	2165333 ✅
filter_axis	37	37	0 (no select)	—	—

The clamps were the dominant cost (your "18 IT-blocks vs native 6" from the gap decomposition), and eliding the keep-val2 move per Select cuts directly into it — −33 cyc across the three clamp-heavy benches, vastly more than const-CSE (−1) or the allocator's expected modest gains. This confirms your read that instruction selection (clamp/IT-block lowering) is the larger slice of the 255→103 gap, not the allocator.

flat_flight now 241 / 103 = 2.34× (was 2.54× a few days ago, 3.18× at v0.11.18). All seams bit-identical. If this is the new default (no flag), I'd update the acceptance gate to flat_flight 241 / controller 150 / control_step 151 / filter 37 once it merges — and the const-CSE/allocator deltas then stack on top of this lower baseline. Nice — the select lowering was hiding a lot. 🎯

[codecov]

Codecov Report

❌ Patch coverage is 75.60976% with 10 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-synthesis/src/instruction_selector.rs	75.60%	10 Missing ⚠️

📢 Thoughts on this report? Let us know!