tgsi to glsl conversion miscompile some write-masks

We currently mis-compile instructions with less than four components in the write-mask, that doesn't simply use the first n components. For example:

MIN OUT[1].z, TEMP[0].xyxw, CONST[0].xyyw

becomes:

gl_FragDepth = float((min( temp0[0].xyxw , uintBitsToFloat(fsconst0[0].xyyw))));

instead of something like:

gl_FragDepth = (min( temp0[0].xyxw , uintBitsToFloat(fsconst0[0].xyyw))).z;

or even:

gl_FragDepth = min( temp0[0].x , uintBitsToFloat(fsconst0[0].y))));

The problem stems from the combination of two things:

1. We always narrow the result by calling the constructor of the type, who will simply discard the unused components from the end (see get_destination_info() where we compute dstconv).
2. We don't cull unused components from the source-swizzle before the operation (see get_source_info(), ).

The problem could be fixed by changing either of these two. However:

• Fixing 1) is a massive undertaking; this would require changing the format-string of every opcode. That smells super bug-prone to me.
• Fixing 2) seems fragile; every format-string would have to be vetted if it's OK to change their inputs. For instance TGSI_OPCODE_LIT depends on the input-order, and would have to become much, much more complex than it is now.

So yeah, it seems a bit like we've done an early design-mistake here, and that has propagated into a lot of code. And it's pretty hard to fix.

I'd love to hear alternative ideas on how to fix this.