Re: 0001-Part-1.-Add-generic-part-for-Intel-CET-enabling

[Jeff Law <law@redhat.com>]

On 09/13/2017 11:07 AM, Tsimbalist, Igor V wrote:
>> -----Original Message-----
>> From: Tsimbalist, Igor V
>> Sent: Tuesday, September 12, 2017 5:59 PM
>> To: 'Jeff Law' <law@redhat.com>; 'gcc-patches@gcc.gnu.org' <gcc-
>> patches@gcc.gnu.org>
>> Cc: Tsimbalist, Igor V <igor.v.tsimbalist@intel.com>
>> Subject: RE: 0001-Part-1.-Add-generic-part-for-Intel-CET-enabling
>>
>>
>>> -----Original Message-----
>>> From: Jeff Law [mailto:law@redhat.com]
>>> Sent: Friday, August 25, 2017 10:50 PM
>>> To: Tsimbalist, Igor V <igor.v.tsimbalist@intel.com>; 'gcc-
>>> patches@gcc.gnu.org' <gcc-patches@gcc.gnu.org>
>>> Subject: Re: 0001-Part-1.-Add-generic-part-for-Intel-CET-enabling
>>>
>>> On 08/01/2017 02:56 AM, Tsimbalist, Igor V wrote:
>>>> Part#1. Add generic part for Intel CET enabling.
>>>>
>>
>>> Q. Do we need to do anything with ICF (identical code folding) and CFE?
>>> Given two functions which have the same implementation in gimple,
>>> except that one has a notrack indirect call and the other has a
>>> tracked indirect call, what is proper behavior?  I think we'd keep
>>> them separate which implies we need to make sure the notrack attribute
>>> is part of the ICF hashing implementation.  It'd probably even be
>>> worth building a test for this :-)
>> Are you talking about a case when such two functions are inlined? Or there is
>> a possibility to merge function bodies if they are identical?
>>
>> I agree with you that the functions should be kept separate. I haven't looked
>> into such optimization in gcc so I need to learn it.
> I thought over this case and my conclusion is that nothing has to be done regarding ICF.
> 
> First of all let's sync on a case we are talking about. A code template could look like
> 
> fn1 definition
> {
>   <some code>
> }
> 
> fn2 definition with notrack attribute
> {
>   <the same code as in fn1>
> }
> 
> func definition
> {
>   <calling fn1>...<calling fn2>
> }
> 
> Is it the case you are talking about? Let's consider different scenarios:
> 
> 1) calls to fn1 and fn2 are direct calls. In that case 'notrack' has no effect on direct calls as they are
> assumed to be save (it applies to indirect calls only). ICF can be done here;
> 2) one of calls is an indirect call or both calls are indirect calls. If compiler can prove what exact functions
> are called then indirect call(s) can be replaced by direct call(s) and that gives us the case 1);
> 3) if compiler cannot prove what function is called it will keep the indirect call and so there is nothing
> to do for ICF here. 
No, not the case I'm worried about.  Instead

fn1()
{
  indirect call where the signature is marked with notrack
}

fn2()
{
  indirect call where the signature is not marked with notrack
}


fn1 and fn2 would be subject to ICF which I think is wrong.

Essentially we're carrying semantic information in attributes that are
part of the type of the function pointer.  I think we need to include
those attributes when we hash and compare two objects for equality
within ICF.

Jeff