perf(codegen): MSB/LSB → clz/ctz peephole family + if [Br N] [] → br_if

[ggreif]

Summary

Extends the existing LSB-and-1 → ctz peephole family (PRs #3873 / #6103) with the symmetric MSB-extraction → clz direction, the LSB-via-shl (N-1) route, and a new if [Br N] [] → br_if (N-1) rule that finally lets moc emit BrIf at all. Also unsilences the FileCheck gate (commit 5265cd4) so the new // CHECK directives are actually validated end-to-end — that fix is also shipping standalone as #6116.

MSB-extraction → Clz family (cf729f046)

• i32.(shr_u|shr_s) 31; if t e → i32.clz; if e t (or-pattern on the shift opcode)
• i32.and 0x80000000; if t e → i32.clz; if e t
• i32.(shr_u|shr_s) 31; eqz; br_if l → i32.clz; br_if l
• i32.and 0x80000000; eqz; br_if l → i32.clz; br_if l
• i64 mirrors with wrap_i64 (EOP path) and the i64.eqz-route variants
• Drops the dead i64.and 1; if rule (fix(codegen): see through i32.wrap_i64 in LSB-and-1 → ctz peephole #6103's commit message claimed removal; diff didn't)

LSB-via-shl (N-1) family (cf729f046)

• i32.shl 31; if t e → i32.ctz; if e t — fires from compile_classical.ml's pow body
• i32.shl 31; eqz; br_if l → i32.ctz; br_if l
• i64.shl 63; i64.eqz; if t e → i64.ctz; wrap; if t e — fires from compile_enhanced.ml's pow body (Word64 + TaggedSmallWord variants)
• i64.shl 63; i64.eqz; br_if l → i64.ctz; wrap; br_if l

Originally landed with failwith probes; the probes fired immediately on aardvark.mo (i64) and assign-ops.mo (i32), pinning the source sites to each backend's pow square-and-multiply LSB-test. Comments now name those exact emit sites.

if [Br N] [] → br_if (N-1) (09bd1b652)

moc never emitted BrIf directly, so the pre-existing LSB eqz; br_if rules + the new MSB br_if rules were correct-by-symmetry but had no input source. This adds the single rule that materialises BrIf from moc's preferred if (br N) end shape. DeBruijn: removing the if's implicit label decrements every reachable br target by 1; Br 0 (targets the if's own end) is a no-op fall-through and collapses to Drop.

Fires today on internal codegen with no-result blocks — every trans_state{N} state-equality check goes from if-br-end to a single br_if. Does not yet fire on user-level break/continue/let-else because those push a unit value before the br (if [Const 0; Br N] []); rewriting that needs a local, out of scope here.

i64.load*; i32.wrap_i64 → i32.load* (c60e99159)

Wasm is little-endian, so i64.load addr reads bytes [addr, addr+8) with the low 32 bits at [addr, addr+4) — exactly what i32.load addr would load. Three cases:

• i64.load (sz=None) → i32.load, align clamped to ≤ 2.
• i64.load32_{u,s} → i32.load (Pack32 + i32 is degenerate).
• i64.load{8,16}_{u,s} → i32.load{8,16}_{u,s} (sz preserved).

Packed variants are fully spec-preserving (same byte count read either way; wrap just discards the i64 extension). The full i64.load → i32.load narrows the read from 8 to 4 bytes — same OOM caveat as the existing Load; Drop → drop rule.

Impact — EOP corpus scan over test/run/_out/ (268 wasm files):

Stat	Before	After
Full i64.load; i32.wrap_i64 instances	201 (in 12 files)	0
Top offenders	candid.wasm 53, idl-opt-tests.wasm 31, idl-decoding-bug.wasm 23, idl-extra-variant-bug.wasm 22, idl.wasm / idl-ops.wasm / issue-5606.wasm 16 each, …	—

The ~3 packed-load+wrap per file that remain are RTS-side (Rust toolchain output, linked verbatim) and outside the OCaml peephole's reach. CI's auto-bench captured a small cycle win on typtbl.drun-run.ok (−2 cycles uniformly across 5 measurements) attributable to this rule.

i64.and (-1<<n); i64.eqz → i64.shrU n; i64.eqz (30426ab0a)

When the constant is a "high mask" (-1 << n, n ∈ [1, 63]), eqz tests "are bits [n, 63] all zero?". i64.shrU n brings those same bits down to positions [0, 63−n] and eqz asks the identical question. Same predicate, fewer bytes — sleb128 of small n is 1 byte; sleb128 of -1 << n is up to 9.

The detector c ≠ 0 ∧ c ≠ -1 ∧ ctz(c) + clz(~c) = 64 (via Wasm.I64.{ctz, clz}) is the compact characterisation of a contiguous high mask. The c ∈ {0, -1} edge cases both pass the bare formula (Wasm spec: ctz(0) = clz(0) = 64) but yield meaningless n; they're ruled out explicitly.

Placement matters: must come after the existing MSB and (1<<63); eqz; (if|br_if) → clz; wrap; (if|br_if) rules. For n = 63 in a branch context the MSB rule is strictly better (produces clz directly, no shift). After all the MSB rules fail to match, this rule catches the remaining high-mask-and-eqz shapes.

Impact (EOP build, selected files):

File	hits
candid.wasm	18
words-enhanced.wasm	7
issue-5606.wasm	5
assign-ops.wasm	4

Each hit saves ~7 bytes of code and one mask materialisation. Fires on every enforce_unsigned_bits env n in compile_enhanced.ml's small-word arithmetic kernels (Nat8/Int8/Nat16/Int16/Nat32/Int32 wrapping arithmetic + overflow check).

Tests (cf729f046, e8d8b80e2, 5265cd41e)

• test/run/top-bit.mo — classical MSB i32 (Nat32/Int32) → clz; if
• test/run/top-bit-i64.mo — EOP MSB i64 (Nat64/Int64) → clz; wrap; if
• test/run/shl-lsb.mo — classical LSB-via-shl i32 in wpow_nat<Nat32> → ctz; if
• test/run/shl-lsb-i64.mo — EOP LSB-via-shl i64 in wpow_nat<Nat64> → ctz; wrap; if

The FileCheck-gate fix in 5265cd41e (= #6116) is what makes those // CHECK directives actually enforce.

Test plan

• make -C test/run -j8 quick green.
• EXTRA_MOC_ARGS="--enhanced-orthogonal-persistence" make -C test/run top-bit-i64.only shl-lsb-i64.only green.
• Manually verified each ~~> rewrite in disassembled wat (wasm2wat) before adding it.
• All four new tests' // CHECK lines pass FileCheck after the gate fix lands.
• Run-eop / drun / mo-idl on CI.

[github-actions]

Comparing from 5d3c2f3 to 8ad7469:
In terms of gas, 5 tests improved and the mean change is -0.0%.
In terms of size, 5 tests improved and the mean change is -0.0%.