D106 - Auto fidelity reads the shader and warns

Status: decided Date: 2026-08-20

The lane model refuses to write the execution mask. Fidelity::Auto inspects a shader's operands, chooses the wavefront model when the shader touches the mask, and reports the choice as a warning carrying the subgroup width that would serve it faster.

Why: a model with no inactive lanes runs every lane regardless, which is plausible and wrong. Refusing whenever a shader masks would refuse nearly every real shader. The fallback costs a factor of the wavefront width, so it is a warning a caller must decide to ignore, not a field a caller must know to read.

Rejected:

• Ignoring mask writes in the lane model: silently wrong.
• Choosing by opcode: forces every 64-bit move onto the slow model.
• Choosing the subgroup level automatically: the translator never sees the device.