[PATCH v3] powerpc/64: Fix memcmp reading past the end of src/dest

[Michael Ellerman]

Segher Boessenkool <segher at kernel.crashing.org> writes:
> On Mon, Mar 25, 2019 at 11:33:56PM +1100, Michael Ellerman wrote:
>> Segher Boessenkool <segher at kernel.crashing.org> writes:
>> > On Fri, Mar 22, 2019 at 11:37:24PM +1100, Michael Ellerman wrote:
>> >> +	clrldi	r6,r4,(64-12)	// r6 = r4 & 0xfff
>> >
>> > You can just write
>> >   rlwinm r6,r4,0,0x0fff
>> 
>> > if that is clearer?  Or do you still want a comment with that :-)
>> 
>> I don't think it's clearer doing a rotate of zero bits :)
>> 
>> And yeah I'd probably still leave the comment, so I'm inclined to stick
>> with the clrldi?
>
> I always have to think what the clrldi etc. do exactly, while with rlwinm
> it is obvious.  But yeah this may be different for other people who are
> used to different idiom.

Interesting, I'm the opposite. You know ppc assembler better than me so
I guess I just need to spend more time on it and embrace the zen of the
rotate instructions.

>> Would be nice if the assembler could support:
>> 
>> 	andi	r6, r4, 0x0fff
>> 
>> And turn it into the rlwinm, or rldicl :)
>
> The extended mnemonics are *simple*, *one-to-one* mappings.

It would still be simple and 1:1, but would only be valid for certain
constants :)

> Having "andi. 6,4,0x0f0f" a valid insn, but an extended mnemonic "andi 6,4,0x0f0f"
> that is not (and the other way around for say 0xff0000ff) would violate that.

I agree that's a bit of a foot gun.

I'm not sure it's vastly more hostile though than `andi. 6,4,0xffff`
being valid but `andi. 6,4,0x1ffff` being not valid.

The assembler could print a nice error saying you need to use a
contiguous mask. And I mean how often do you andi. with a mask that
isn't contiguous?

> You could do some assembler macro, that can also expand to multiple insns
> where that is useful.  Also one for loading constants, etc.  The downside
> to that is you often do care how many insns are generated.
>
> Instead you could do a macro for only those cases that can be done with *one*
> insn.  But that then is pretty restricted in use, and people have to learn
> what values are valid.
>
> I don't see a perfect solution.

Yeah, I guess a new `andi` instruction is the only real answer :)

cheers