TS8038

When it fires

An emulated double (f64, a vec64) on an entry's IO boundary — a @location parameter, a @location field of an IO struct, or an entry's return. A double is a pair of f32 words after lowering, and a varying interpolates each word on its own, which is not the interpolation of the double they encode; a vec64 attribute would need two slots. Its own code rather than the UNSUPPORTED bucket because the remedy is specific and an author can look it up: narrow with f32(x), or read the double in the stage that needs it, since a uniform or a storage binding carries an f64 and every stage can see one. There is deliberately no author-facing way to split a double into its two f32 words — they are the emulation's business (§39).

Example

"use typeshade"

@fragment
export function main(@location(0) uv: vec2<f64>): vec4 {
  return vec4(f32(uv.x), f32(uv.y), 0., 1.)
}

Compiled at build time against the pinned compiler, this program gets:

How to fix it

The same program, corrected. It compiles with no diagnostic.

"use typeshade"

@fragment
export function main(@location(0) uv: vec2): vec4 {
  return vec4(uv.x, uv.y, 0., 1.)
}

The rules this enforces

Rule 2.3 Test

A value sometimes has to cross between a representation the compiler chose and one the target has: an f64 and the two f32 halves that carry it across an interface of f32.

Rule 12.6 Test

A requirement the front end can check must be checked at the front end, in the author’s words, and not left to Tint or a driver.

See also

SD0044

f64 in an interpolated @location IO field

TS8099 UNSUPPORTED

UNSUPPORTED (TS8099) is the one deliberate exception to "sequential": it is the catch-all for a diagnostic whose site does not yet deserve its own code, so it stays parked past the sequential range instead of at its head.

• Emulated double
• f64
• @location
• f32
• Error codes

Source

Where the compiler raises this code at commit 26de7be8, one line per file:

• src/compiler/ts/lower/function.ts:2501