Inline assembly (asm!) blocks are fed directly into LLVM and have horrible error checking, tread lightly and verify the output.
LLVM can end up generating dodgy versions of certain instructions on certain versions, it may for example turn "lgdt" into the utterly useless 16-bit version (which truncates the upper 8 bits of the address). Specify operand sizes.
An "intel" option is supported, but it seems to be somewhat dodgy (for example, it seemed to d