Reading past the end of a vertex buffer, the Direct3D way

What changed

Two gaps in plain drawing closed this session.

First, the four "adjacency" topologies (primitive types that send each line or triangle together with its neighbours' vertices, for geometry shaders) now draw when there is no geometry shader. Direct3D 12 then ignores the neighbours. mtl12 picks out the real vertices in the same GPU pre-pass that already rewrites index buffers for strip cuts and triangle fans.

Second, vertex fetch bounds. A vertex buffer view in Direct3D 12 is an address, a stride and a size. When a draw reads an element that ends past the size, the spec says the shader gets 0 in the format's components and the default (0, 0, 0, 1) in the rest (D3D11.3 functional spec 8.19.2). Metal 4 binds a vertex buffer as an address and a stride only, with no length. A read past the end returns whatever memory follows.

mtl12 now rewrites the vertex function that SPIRV-Cross generates. The inputs Metal fetched are copied first. Then each one is checked: its index × stride + offset + format size must fit in the view, or it is replaced by the defaults. The index is the vertex id, or for per-instance data the start instance plus the instances since it divided by the step rate. The views travel to the shader as a small table, bound at a buffer slot that would otherwise belong to input slot 19. Layouts that use slot 19 are left unchecked. Indirect draws bind the table from their own records.

The interesting part

I break each check on purpose once and watch the test fail. Six such breaks failed the new test straight away. A seventh did not: after a checked pipeline has put its table where slot 19's buffer goes, the next pipeline must rebind slot 19, and skipping that rebind passed every check. Nothing in the test read slot 19. A second pipeline that does fails 40 checks with the break. A cache that forgets to invalidate is a missing guard too.

The cost was the other worry. Earlier measurements put a per-draw CPU price on vertex functions that read the instance builtins. A stash-and-rebuild A/B of bgfx's drawstress example showed none: 8.74 ms of layer CPU per frame without the checks and 8.62 ms with them.

Numbers

• Full run: 858 pass, 27 fail, 73 skip of 958, up from 856. The only per-test changes are the two new tests passing.
• api/vertex_bounds: 1793 checks pass. Without the checks, 632 of its first 1281 failed. Nine deliberate breaks each fail it.
• api/adjacency: 1629 checks pass, and 812 of its first 1309 failed before the fix.