GCC bug ? Re: [PATCH v2 10/10] powerpc/32s: Implement Kernel Userspace Access Protection

[Christophe Leroy]

Le 22/01/2020 à 14:36, Segher Boessenkool a écrit :
> On Wed, Jan 22, 2020 at 07:52:02AM +0100, Christophe Leroy wrote:
>> Le 21/01/2020 à 20:55, Segher Boessenkool a écrit :
>>> On Tue, Jan 21, 2020 at 05:22:32PM +0000, Christophe Leroy wrote:
>>>> g1() should return 3, not 5.
>>>
>>> What makes you say that?
>>
>> What makes me say that is that NULL is obviously a constant pointer and
>> I think we are all expecting gcc to see it as a constant during kernel
>> build, ie at -O2
> 
> But apparently at the point where the builtin was checked it did not
> yet know it is passed a null pointer.
> 
> Please make a self-contained test case if we need further investigation?

The test in my original mail is self-contained:


#define NULL (void*)0

static inline int f1(void *to)
{
     if (__builtin_constant_p(to) && to == NULL)
         return 3;
     return 5;
}

int g1(void)
{
     return f1(NULL);
}


Build the above with -O2 then objdump:

00000000 <g1>:
    0:    38 60 00 05     li      r3,5
    4:    4e 80 00 20     blr

It returns 5 so that shows __builtin_constant_p(to) was evaluated as false.


> 
>>> "A return of 0 does not indicate that the
>>>   value is _not_ a constant, but merely that GCC cannot prove it is a
>>>   constant with the specified value of the '-O' option."
>>>
>>> (And the rules it uses for this are *not* the same as C "constant
>>> expressions" or C "integer constant expression" or C "arithmetic
>>> constant expression" or anything like that -- which should be already
>>> obvious from that it changes with different -Ox).
>>>
>>> You can use builtin_constant_p to have the compiler do something better
>>> if the compiler feels like it, but not anything more.  Often people
>>> want stronger guarantees, but when they see how much less often it then
>>> returns "true", they do not want that either.
> 
>> If GCC doesn't see NULL as a constant, then the above doesn't work as
>> expected.
> 
> That's not the question.  Of course GCC sees it as a null pointer
> constant, because it is one.  But this builtin does its work very
> early, during preprocessing already.  Its concept of "constant" is
> very different.
> 
> Does it work if you write just "0" instead of "NULL", btw?  "0" is
> also a null pointer constant eventually (here, that is).

No it doesn't.

It works if you change the 'void *to' to 'unsigned long to'

> 
> The question is why (and if, it still needs verification after all)
> builtin_constant_p didn't return true.

I sent a patch to overcome the problem. See 
https://patchwork.ozlabs.org/patch/1227249/

Christophe