optimizer: branchless abs via bit manipulation strength reduction

[yevbar]

Summary

Adds a strength reduction optimization that converts branching abs(x) patterns into branchless bit manipulation.

Before (branching)

if(x >= 0, x, 0 - x)    ;; generates WASM if/else block

After (branchless)

mask = x >> 63           ;; arithmetic shift: 0 for positive, -1 for negative
(x ^ mask) - mask        ;; 4 ALU instructions, no branches

How it works

The classic branchless abs trick for signed 64-bit integers:

• x >> 63 produces a sign mask: 0 for non-negative, -1 (all ones) for negative
• x ^ mask flips all bits when negative (one's complement), no-op when positive
• - mask adds 1 when negative (completing two's complement negation), no-op when positive

WASM impact

Before (branching abs):

local.get $p0
i64.const 0
i64.ge_s
if (result i64)
  local.get $p0
else
  i64.const 0
  local.get $p0
  i64.sub
end

After (branchless abs):

local.get $p0
i64.const 63
i64.shr_s          ;; mask = x >> 63
local.tee $mask
local.get $p0
local.get $mask
i64.xor            ;; x ^ mask
local.get $mask
i64.sub            ;; - mask

Eliminates branch misprediction and reduces code size.

Patterns matched

1. Trivial operands (var/literal): if(x >= 0, x, 0 - x) → direct branchless form
2. Block-wrapped (non-trivial operands like a+b): detects let-binding + if pattern from IRGen's abs lowering

Tests

10 new tests covering:

• Optimizer pattern matching (trivial and block-wrapped forms)
• Negative cases (non-abs patterns preserved unchanged)
• End-to-end WAT generation (verifies branchless output, no if/else)
• Correctness via Wasmex (positive, negative, zero, manhattan distance)