In x86 terminology/documentation, a "word" is 16 bits because x86 evolved out of 16-bit 8086. Changing the meaning of the term as extensions were added would have just been confusing, because Intel still had to document 16-bit mode and everything, and instruction mnemonics like cwd (sign-extend word to dword) bake the terminology into the ISA.
- x86 word = 2 bytes
- x86 dword = 4 bytes (double word)
- x86 qword = 8 bytes (quad word)
- x86 double-quad or xmmword = 16 bytes, e.g.
movdqa xmm0, [rdi].
Also in the cqo mnemonic, oct-word. (Sign-extend RAX into RDX:RAX, e.g. before idiv)
And then we have fun instruction like punpcklqdq: shuffle together two qwords into a dqword, or pclmulqdq for carry-less multiplication of qwords, producing a dq full result. But beyond that, SIMD mnemonics tend to be AVX vextracti128 or AVX512 (with optional per-element masking) vextractf64x4 to extract the high 256 bits of a ZMM register.
Not to mention stuff like "tbyte" = 10 byte x87 extended-precision float; x86 is weird and not everything is a power of 2. Also 48-bit seg:off 16:32 far pointers in Protected mode. (Basically never used, just the 32-bit offset part.)
Most other 64-bit ISAs evolved out of 32-bit ISAs (AArch64, MIPS64, PowerPC64, etc.), or were 64-bit from the start (Alpha), so "word" means 32 bits in that context.
- 32-bit word = 4 bytes
- dword = 8 bytes (double word), e.g. MIPS
daddu is 64-bit integer add
- qword = 16 bytes (quad word), if supported at all.
"Machine word" and putting labels on architectures.
The whole concept of "machine word" doesn't really apply to x86, with its machine-code format being a byte stream, and equal support for multiple operand-sizes, and unaligned loads/stores that mostly don't care about naturally aligned stuff, only cache line boundaries for normal cacheable memory.
Even "word oriented" RISCs can have a different natural size for registers and cache accesses than their instruction width, or what their documentation uses as a "word".
The whole concept of "word size" is over-rated in general, not just on x86. Even 64-bit RISC ISAs can load/store aligned 32-bit or 64-bit memory with equal efficiency, so pick whichever is most useful for what you're doing. Don't base your choice on figuring out which one is the machine's "word size", unless there's only one maximally efficient size (e.g. 32-bit on some 32-bit RISCs), then you can usefully call that the word size.
A "word" doesn't mean 64 bits on any 64-bit machine I've heard of. Even DEC Alpha AXP, which was designed from the ground up to be aggressively 64-bit, uses 32-bit instruction words. IIRC, the manual calls a word 32-bit bits.
Being able to load 64-bits into an integer register with a single instruction does not make that the "word size". Bitness and word size don't have hard specific technical meanings; most CPUs have multiple different sizes internally. (e.g. 64 byte buses between L2 and L1d cache on Intel since Haswell, along with 32-byte SIMD load/store.)
So it's basically up to the CPU vendor's documentation authors to choose what "word" (and thus dword / qword) mean for their ISA.
Fun fact: SPARC64 talks about "short word" (32 bits) vs. "long word" (64 bits), rather than word / double-word. I don't know if just "word" without any qualifier has any meaning in 64-bit SPARC documentation.
