Conversation moved from CL 109716.

Consider this code from the runtime:

// Information from the compiler about the layout of stack frames.
type bitvector struct {
	n        int32 // # of bits
	bytedata *uint8
}

// ptrbit returns the i'th bit in bv.
func (bv *bitvector) ptrbit(i uintptr) uint8 {
	b := *(addb(bv.bytedata, i/8))
	return (b >> (i % 8)) & 1
}

ptrbit should generate a BTL instruction on amd64, but instead it generates RSH and AND. (Related work: CL 105037.) IIRC, we intentionally don't generate BTL instructions for constant-offset bit tests for bytes, since the equivalent TEST instruction is shorter or faster.

cc @rasky