Right. The idea being that the bottom two bits of words are tag bits. tadd dictactes that they have to be 0 for integers which means pointers are non-zero in their tags. Let's say the pointer tag is 1. That means to load a word from an object will have to compensate, eg.

   car: ld r2, [r3-1]
        ret
        nop

If it happens that r3 _wasn't_ an object pointer, that load above would not be word aligned and thus trap. In other words, the hardware does the tag checking for you. (As an aside, the 486 introduced the option for making unaligned access trap, but AFAIK, no one has ever used it because you'd have to fix all the alignment assumptions in library code).

But unaligned load is very important to a lot of code like LZ compression.

Thanks for the explanation.

Loads/stores not "naturally" aligned: 4-byte boundaries for 4-byte data, 8-byte boundaries for 8-byte data.

In the late 90s I would routinely get bus errors while running Netscape on SPARC Solaris machines, presumably due to corner-case unaligned memory accesses. x86 processors perform unaligned loads and stores, but at a slight speed penalty, and with the loss of atomicity.