long-vector-in-cbuffer-or-signature

Status: shipped (Phase 3) — see CHANGELOG.

(via ADR 0010)

What it detects

A vector<T, N> with N >= 5 declared as a member of a cbuffer / ConstantBuffer<T>, or as the type of a vertex / pixel / hull / domain / geometry IO signature element. The SM 6.9 long-vector specification (DXIL vectors, proposal 0030) restricts long vectors to in-shader compute use; cbuffer layout and stage-IO signatures are explicitly out of scope. Slang reflection identifies the binding kind and signature stage; the rule fires on the first long-vector member encountered at one of those boundaries.

Why it matters on a GPU

The cbuffer packing rules and the inter-stage IO packing rules predate the long-vector feature. Cbuffer layout is fixed by HLSL packing (16-byte slots, float4-shaped vector members, scalar tail packing), and the runtime maps cbuffer fetches onto the IHV-specific scalar / constant-data path. Inter-stage IO is packed into per-vertex slots (NVIDIA Ada: 32 four-component slots; AMD RDNA 2/3: parameter-cache entries; Intel Xe-HPG: URB-style slots); the slot allocation is fixed at PSO compile time and assumes 1/2/3/4-wide vector types only.

Long vectors break both surfaces: a vector<float, 8> cbuffer member has no defined packoffset because the member crosses two 16-byte slots in a way the packing rules don't enumerate, and a vector<float, 8> IO signature element doesn't fit any IHV's per-vertex slot. The DXC validator catches both forms and emits a precise error; the lint replaces the validator round trip with a source-located diagnostic and points the author at the right shape.

The fix is to split the long vector into chunks that respect the packing rules: an array of float4 for cbuffer storage, multiple TEXCOORD slots for IO signature. The rule emits a candidate rewrite as a comment so the author can review.

Examples

Bad

hlsl

// Long vector in cbuffer — packing undefined.
cbuffer Frame : register(b0) {
    vector<float, 8> Weights;   // ERROR: long vector at cbuffer boundary
};

// Long vector as vertex output — IO signature does not pack.
struct VSOut {
    float4           pos     : SV_Position;
    vector<float, 8> attribs : TEXCOORD0;  // ERROR: long vector in IO sig
};

Good

hlsl

// Cbuffer: split into float4 array.
cbuffer Frame : register(b0) {
    float4 Weights[2];
};

// IO signature: split into multiple float4 slots.
struct VSOut {
    float4 pos      : SV_Position;
    float4 attrib0  : TEXCOORD0;
    float4 attrib1  : TEXCOORD1;
};

Options

none

Fix availability

suggestion — Splitting the long vector requires updating every reader to use the new layout, which the rule cannot do safely without a full project view. The diagnostic emits the candidate split as a comment.

long-vector-in-cbuffer-or-signature ​

What it detects ​

Why it matters on a GPU ​

Examples ​

Bad ​

Good ​

Options ​

Fix availability ​

See also ​