[PATCH 1/2] x86: optimize XCHG to MOV for same-register forms

[Jan Beulich]

On 07.07.2026 19:18, Christian Ludloff wrote:
>> Assembler optimization is intended primarily for hand-written code.
> 
> Uhm, hand-written assembler code is the prime example
> for *not* wanting specified instruction FOO suddenly turn
> into emitted instruction BAR.

Then you simply don't pass -O..., and you'll get exactly what you've
written. Well, almost - already prior to the introduction of -O we did a
very small set (one or two) of optimizations. Iirc even nowadays we
continue to do these irrespective of whether -O is passed.

> If I code XCHG, then I mean to get XCHG. Had I wanted
> MOV, I'd have coded MOV.

XCHG may be a less likely example, but as I said elsewhere - I've seen all
sorts of code (in particular in BIOSes) that is inefficient. The most
prominent perhaps being LEA where MOV can be used.

> What is the expected worthwhile benefit of the proposed
> transformations in practice?

Less execution overhead (uniformly for all optimizations), due to varying
aspects (code size, decode, execution resources needed). Imo long term all
inefficiencies in encodings may (and perhaps should) be subject to
replacement. If this isn't wanted, stringent criteria for optimization to
actually carry out would need writing down.

Jan