[SPIR-V] Casting 16bit integral data types

[ChristianReinbold]

Description

DXC generates invalid spir-v code when casting 16bit data types.

Steps to Reproduce

https://godbolt.org/z/5vKcqxT76

Compile the following shader with options -T lib_6_4 -spirv -fspv-target-env=vulkan1.2 -enable-16bit-types -HV 2021

[[vk::binding(0,4)]] RWStructuredBuffer<uint16_t> a;

[shader("raygeneration")]
void main() {
    int16_t v = -1;
    uint16_t vu = uint16_t(v);
    a[0] = vu;
}

Actual Behavior Shader compilation fails with

fatal error: generated SPIR-V is invalid: The high-order bits of a literal number in instruction 15 must be 0 for a floating-point type, or 0 for an integer type with Signedness of 0, or sign extended when Signedness is 1 %ushort_4294967295 = OpConstant %ushort 4294967295

Note that both

void main() {
    uint16_t vu = uint16_t(-1);
    a[0] = vu;
}

and

void main() {
    uint16_t vu = uint16_t(int16_t(-1));
    a[0] = vu;
}

results in compiling to the expected result of storing an ushort constant of 65535 (2^16-1) to a[0], as would happen with a regular underflow. Also note that the generated constant in spir-v resembles 2^32 - 1. It seems like the underflow is incorrectly handled with 32bit precision instead of 16bit precision.

Environment

• DXC version: libdxcompiler.so: 1.8(dev;1-df588beb)
• Host Operating System: Compiler Explorer

[cassiebeckley]

I believe this is actually an issue with how SPIRV-Tools folds the OpBitcast instruction for unsigned 16-bit integers. I've sent in a PR to fix it upstream.