On Sunday, 30 March 2025 04:11:14 Eastern Daylight Time Bo Persson via Std-
Proposals wrote:
> The standard fails to say what should be faster for a "fast" type. If a
> certain type uses more data cache, but less i-cache (because of a
> compact instruction encoding), does that make it "fast"?

That wouldn't be the case for x86-64 because encoding the 64-bit instructions
requires the REX.W bit to be set, so a REX prefix is required. 64-bit
instructions are longer than their 32-bit equivalent.

Notably, 16-bit instructions are also longer. Moreover, the Intel Optimization
Manual does warn that the 16-bit Length Changing Prefix does affect decoding
time, as the front-end can only decode one such per 16 bytes, and has a one-
cycle stall if the LCP is used with a 16-bit immediate (if you ever
disassemble code using uint16_t, you may note compilers often emit 8- or 32-
bit immediates).

> And "obviously" a 64-bit type must be the best on a 64-bit CPU. That's
> in the name. :-)

Yeah, this must be it.

-- 
Thiago Macieira - thiago (AT) macieira.info - thiago (AT) kde.org
  Principal Engineer - Intel DCAI Platform & System Engineering