IssueTracker

When narrowing a vector after a swizzle (e.g. during the casting of Int4 to Byte4), we throw away the upper part of the source vector and thus the swizzling of those components doesn't matter. LLVM can express and optimize for these "don't care" components by using 'undef' shuffle mask values.

This probably has no effect on x86, where since SSSE3 a single pshufb instruction can permute the bytes arbitrarily. On ARM it might have a significant effect: https://www.cnx-software.com/2017/08/07/how-arm-nerfed-neon-permute-instructions-in-armv8/