Incorrect SPIR-V generated when firstbithigh is used on a uint64_t

[manon-traverse]

generated SPIR-V is invalid: GLSL.std.450 FindUMsb: expected all operands to have the same bit width as Result Type
  %137 = OpExtInst %uint %1 FindUMsb %136

note: please file a bug report on https://github.com/Microsoft/DirectXShaderCompiler/issues with source code if possible

Calling this function causes the error when targeting SPIR-V. It works fine when targeting DXIL.

uint clz64(uint64_t v) { return 63 - firstbithigh(v); }

It does work with the following workaround:

uint clz64(uint64_t v) {
    uint2 workaround = uint2(v, v >> 32);

    uint high = firstbithigh(workaround.y);
    if (high != -1) {
        return 31 - high;
    } else {
        return 63 - firstbithigh(workaround.x);
    }
}

[godlikepanos]

FindUMsb and a few similar GLSL derived instructions do not support 64bit (see https://registry.khronos.org/SPIR-V/specs/unified1/GLSL.std.450.html). glslang has (or had) similar issues. I think Khronos should extend the instructions to support 64bit and in the meantime DXC should apply your workaround.

[sudonatalie]

DXC now emits an appropriate error message when firstbithigh is used on a uint64_t. Eventually we should support this, but it's not currently a priority to implement so the workaround in HLSL should be used until then.

int firstbithigh64(uint64_t v) {
    int high = firstbithigh(uint(v >> 32));
    if (high != -1) {
      return high + 32;
    }

    return firstbithigh(uint(v));
}

[LukasBanana]

The fix b7a50ba introduced a new issue at astContext.getTypeSize(argType) == 64 when used on int2 types. This reproduces the issue (see godbolt.org):

// -T vs_6_0 -E VSMain -spirv
int2 Foo;
void VSMain() {
    firstbithigh(Foo);
}

[sudonatalie]

Thanks for catching that @LukasBanana! Submitting a fix now.

[damyanp]

@s-perron - comments here seem to suggest this is already fixed?

[sudonatalie]

Just to clarify, the original issue in the bug is not yet fixed. I added some error messaging to make the failure more explicit, and in the process was accidentally overzealous and errored on some valid cases, so I fixed that, but the SPIR-V backend still doesn't support firstbithigh on 64 bit ints.

[s-perron]

We will not be implementing the workaround in the compiler. We will reconsider in the clang implementation.

#6864 is open to make sure the error messages are reasonable.