11 Jan
2024
11 Jan
'24
10:07 p.m.
On Thu Jan 11 22:07:15 2024 +0000, Henri Verbeet wrote:
That probably makes sense. I imagine it may be helpful to have that clarification somewhere in the code as well though, for future reference.
I hope the new proposal is better.
Let me point out that I'm not completely convinced myself of what is the right choice here. I can see good reasons for mandating the TEMPs are always vec4. So for the moment I'm sticking with the more permissive choice and we'll see in the future. The option of introducing a pass that converts all TEMPs to vec4 remains on the plate, also (though it is already validated that scalar registers always have the obvious write mask and swizzle, so the backend can probably just ignore the dimension if so they wish).