Re: [PATCH 3/4] x86/insn: Extract more information about instructions

[Sasha Levin <sasha.levin@oracle.com>]

On 04/15/2014 11:26 PM, H. Peter Anvin wrote:
> On 04/15/2014 08:10 AM, Sasha Levin wrote:
>>
>> Mnemonics don't have 1:1 relationship with opcodes. So, for example,
>> if kmemcheck needs to check (and it does) whether a given instruction
>> is an "ADD", it would need to compare it to 9 different opcodes.
>>
> 
> Excuse me, but on what planet does, for example, it makes sense if a
> particular instruction is a "MOV", for example?  The trend in x86
> opcodes have varied over the years and at some points it seems to have
> been trendy to have very general mnemonics (consider MOV CR, MOV DR) and
> at some points quite the opposite (hence MOVD, MOVQ, MOVDQA, MOVDQU,
> MOVAPS, MOVUPS, MOVAPD, MOVUPD, VMOVxxx).
> 
> So it is not at all clear that this makes any kind of sense whatsoever,
> and is more likely just going to be abused.

Looking at kmemcheck, and "AND" vs "MOV" for example, we need to know if a
given instruction is AND because AND may operate on only part of the memory
it's accessing to. So some accesses to what kmemcheck sees as "uninitialized
memory" are actually valid ones because we don't touch the uninitialized
part.

So for kmemcheck, AND and MOV (for example) are different because ANDing
a value and MOVing a value mean different things wrt to uninitialized memory.

Yes, if kmemcheck for some reason needs to figure out if an instruction
is a MOV variant we'll need to list quite a few mnemonics, but that list
will be much shorter and more readable than a corresponding list of opcodes.


Thanks,
Sasha