October 12, 2021 1:59 AM, "Henri Verbeet" hverbeet@gmail.com wrote:
Is there any particular reason to avoid handling VKD3DSPR_IMMCONST shifts? I suppose immediate shifts are unlikely to be larger than 31, but it should also be easy to verify that.
It turns out fxc will happily compile a shift > 31 and store the constants unmasked. The simplest solution is emit the OpBitwiseAnd for immediate shifts too, which drivers should easily be able to optimise out. I'll send a new patch for this.