Re: [PATCH] x86/uaccess: fix code generation in put_user()

[Linus Torvalds]

On Fri, Oct 23, 2020 at 2:00 PM <hpa@xxxxxxxxx> wrote:
>
> There is no same reason to mess around with hacks when we are talking about dx:ax, though.

Sure there is.

"A" doesn't actually mean %edx:%eax like you seem to think.

It actually means %eax OR %edx, and then if given a 64-bit value, it
will use the combination (with %edx being the high bits).

So using "A" unconditionally doesn't work - it gives random behavior
for 32-bit (or smaller) types.

Or you'd have to cast the value to always be 64-bit, and have the
extra code generation.

IOW, an unconditional "A" is wrong.

And the alternative is to just duplicate things, and go back to the
explicit size testing, but honestly, I really think that's much worse
than relying on a documented feature of "register asm()" that gcc
_documents_ is for this kind of inline asm use.

So the "don't do pointless conditional duplication" is certainly a
very sane reason for the code.

            Linus