MSVC generates crap code for SSE which causes it to produce erroneous results

[cmuratori]

Although in theory the 128-bit routines in Blandwidth are written to ensure they can use 2 read and 2 write ports on every cycle where available, and CLANG produces ASM that appears to work properly, MSVC produces really bizarre ASM that does register shuffles in the middle of the inner loop. Although I have not investigated this matter yet, it would appear that the 128-bit op bandwidth reported by the CLANG build is probably the accurate (higher) one, whereas 128-bit op bandwidth reported by MSVC is probably wrong.

- Casey