> (1) discard (fast but not a great user experience)
What are we discarding here, and why is it fast but not a great user experience?
> (2) primitive synthesis
I assume this is retriangulating clipped triangles that are now no longer triangles?
> reverse transformed clipping rectangle
Which spaces does this reverse transformation map from and to? I assume the clipping rectangle here is the triangle's AABB in raster space (or as you say, barycentric space).
> integer fixed point, but that is historical — modern fixed point can be handled with careful control of the fp unit in the mantissa
So we are no longer doing 16.16 fixed point, but tweaking the FP representation itself?
> The major issue is regenerating the Z and the 1/Z values for the new vertices
Why is this a major issue?
> deferred attribute synthesis rasterizer
I assume this means attributes are perspective-correct interpolated in raster space.
On the flip side, if a triangle is 'really big' we need a guard band to either reject or subdivide triangles. The first one is fairly cheap -- we're throwing away the triangle! -- the second one is a better user experience, but requires synthesizing primitives. (The worst case is that a single triangle becomes 5 triangles, I think.) Each of those triangles needs its Z and 1/Z calculated in fixed precision. The precision of that fixed precision (though) can be clamped to the local tile; so, even though the global precision might need to be 25.25 (or whatever), the tile-local precision is only 4.9 (or whatever), with an intermediate 24.24 that can be handled with a float-float patch-up. The computation should all occur in the triangle's barycentric space: that means you need the inverted barycentric mapping to invert the guard band into the triangle's barycentric space. You do that because it lets you control the fixed point calculations better. (You can leave off the inverted determinant multiplication until the last moment.)
When I say "deferred attribute synthesis" I mean that we don't calculate attributes in the vertex shader. Instead, we calculate the barycentric, Z, and 1/Z values and pass those along. When we fire up the tile walker for the triangle (in general we only need 1-3 tiles), we calculate the attributes "on the fly, as they're used" and then let the compiler do CSE to fold down the replicated constructions.