forcecase-missing-on-ps-switch

Status: shipped (Phase 4) — see CHANGELOG.

(via ADR 0011)

What it detects

A switch statement inside a pixel-shader entry point where at least one case body contains a derivative-bearing operation — Sample, SampleBias, SampleGrad, ddx, ddy, ddx_fine, ddy_fine, or any function call that transitively invokes one — and the switch lacks the [forcecase] attribute. The rule fires only on PS entry points (the hazard is specific to the quad / derivative model); compute and other stages are not flagged.

Why it matters on a GPU

The HLSL switch statement has multiple valid lowerings: a true jump-table (one indirect branch, one taken target per lane), a chained if/else ladder (linear scan over case labels), or — in some compiler versions — a branch-free predicate fan that evaluates every case body and selects the right result. The [forcecase] attribute pins the compiler to the jump-table form. Without [forcecase], the compiler is free to pick the chained-if form, and that's where the pixel-shader hazard appears: chained ifs mean each case introduces a new branch, and per-pixel divergence on case selection retires quad lanes one-arm-at-a-time. Derivatives sampled inside the case body then see undefined neighbour values from retired lanes — exactly the same failure mode as a per-lane if containing a Sample (see derivative-in-divergent-cf and quadany-quadall-opportunity).

[forcecase] keeps the cases as a single jump-table dispatch, so the quad either takes one case (when uniform within the quad) or branches once with all four lanes still active. On AMD RDNA 2/3 the jump-table form preserves quad coherence at the SX (sequencer) level; NVIDIA Ada keeps the warp's quad participation intact through the indirect-branch instruction; Intel Xe-HPG's compiler honours [forcecase] similarly. The visible artefact when the hint is missing is mip-aliasing or seams at the boundary between case arms — a function-of-material-id switch that samples per-material textures is the canonical bad case.

The fix is one token: [forcecase] switch (matId) { ... }. The cost is essentially zero on modern compilers (the jump-table form is rarely worse than the chained-if form even when both would be correct) and the safety win is significant: derivative correctness becomes a property of the source code rather than a property of which compiler version was used. The rule fires on the missing-attribute case rather than dictating that every PS switch use the attribute; some switch statements have no derivative-bearing case bodies and the attribute is unnecessary there.

Examples

Bad

hlsl

Texture2D<float4> g_AlbedoA : register(t0);
Texture2D<float4> g_AlbedoB : register(t1);
SamplerState g_Samp : register(s0);

float4 ps_per_material(float2 uv : TEXCOORD0, uint matId : MAT_ID) : SV_Target {
    // No [forcecase] — compiler may emit chained-if; per-lane case selection
    // retires quad lanes individually and the Sample inside each arm sees
    // undefined derivative inputs from retired lanes.
    switch (matId) {
        case 0: return g_AlbedoA.Sample(g_Samp, uv);
        case 1: return g_AlbedoB.Sample(g_Samp, uv);
        default: return float4(0, 0, 0, 1);
    }
}

Good

hlsl

Texture2D<float4> g_AlbedoA : register(t0);
Texture2D<float4> g_AlbedoB : register(t1);
SamplerState g_Samp : register(s0);

float4 ps_per_material_safe(float2 uv : TEXCOORD0, uint matId : MAT_ID) : SV_Target {
    // [forcecase] pins the lowering to a jump-table dispatch; quad lanes
    // remain active through the indirect branch and derivatives stay valid.
    [forcecase]
    switch (matId) {
        case 0: return g_AlbedoA.Sample(g_Samp, uv);
        case 1: return g_AlbedoB.Sample(g_Samp, uv);
        default: return float4(0, 0, 0, 1);
    }
}

Options

none

Fix availability

suggestion — Adding [forcecase] is a one-token change but the developer may have a reason for the chained-if lowering (e.g., compiler-specific code-size considerations on a hot small-case switch). The diagnostic identifies the switch and the derivative-bearing case body so the author can apply the attribute deliberately.

forcecase-missing-on-ps-switch ​

What it detects ​

Why it matters on a GPU ​

Examples ​

Bad ​

Good ​

Options ​

Fix availability ​

See also ​