SPIR-V compiler regression in 410.73 involving NaNs

I have a SPIR-V shader which stores the bit pattern 0xffffffff in the four components of a float4 vector. Later, this vector is bitcast to a vector of 4 uints. Somewhere along the line, the conversion fails.

If I use the old SPIR-V compiler via export __GL_NextGenCompiler=0 things work as expected.

I suspect the new compiler mis-handles NaN values somewhere.

I can provide a sample SPIR-V shader if it’s helpful.