flowchart TD
    H["HLSL: while (true) {<br/>if (key == WaveReadLaneFirst(key)) {<br/>sum = WaveActiveSum(v); break; } }"] --> D["DXC: the break becomes<br/>if (dx.break.cond[0] == 0) break<br/>dx.break.cond = constant zero"]
    D -->|before| P["dxil-spirv: Private array with initializer<br/>SPIRV-Cross: promoted to a constant array"]
    D -->|now| V["dxil-spirv patch 0006: decorated Volatile<br/>SPIRV-Cross patch 0043: read through<br/>*(volatile thread uint*)&amp;_14[0u]<br/>(_14 is the array's generated name)"]
    P --> F["Metal compiler folds the test:<br/>the break is always taken,<br/>simd_sum moves after the loop"]
    F --> X["every lane gets the whole wave's sum<br/>25 25 25 25 ..."]
    V --> K["Metal compiler can't fold the load:<br/>simd_sum stays in the loop"]
    K --> OK["each iteration sums its own lanes<br/>12 9 4 9 12 0 0 4 ... (as on D3D)"]