Validation error "DXIL intrinsic overload must be valid"

[SirTimothy]

Hey, I encountered a problem while compiling some shaders through dxcompile (normally we're using d3dcompile), with a cryptic validation error message.

Managed to narrow it down, and can reproduce with the attached shader: test.txt

Gives for output:

dxc -E ps_main -T ps_6_0 test.hlsl

error: validation errors
at 0x280ab609bc0 inside block #0 of function ps_main DXIL intrinsic overload must be valid

Validation failed.

My guess is that it can't decide whether to implicitly convert the ?: result from float3 down to float, or bump the 1.0 second dot argument up to float3. I think some improved error message for this would be useful (didn't get any line number or anything) :)

[tristanlabelle]

Thanks for reporting! It looks like the ternary operator with two literal float values resolves to the double type, which there is no overload of dot for. The poor error message would be because the semantic analysis thinks there is a dot overload for double but the codegen disagrees. A workaround is to suffix the two values with f.

[llvm-beanz]

Godbolt link: https://godbolt.org/z/e54rbcoPn

Clang does not exhibit this bug, but it is still present in DXC.