Null pointers with a non-0 representation

I feel a bit silly asking this, but here goes.

The C spec does not require the null pointer to be implemented as an integer with the 0 bit-pattern. The implementation may arbitrarily choose any other (integer, but not necessarily) representation, as long as it’s distinct from any legal pointer. The only requirement is that casting an integer 0 to a pointer type results in a null pointer.

LLVM IR appears to have similar the casting behavior. That is, if I try to create a pointer with integer value 0 (e.g. by constructing an inttoptr constant expression with a i32 0 argument), it gets immediately folded into a null pointer constant. Furthermore, a PtrToInt from a null pointer results in 0 (which is even “stricter” than the C spec.)

So, that raises a few question:

a) Are those the desired semantics? I guess if clang generates this kind of IntToPtr instructions (as opposed to directly resolving the cast to a null pointer) this is unavoidable. But otherwise, why does LLVM treat null pointers and pointers with the integer value 0 as equivalent?
It’s completely natural for most platforms, but isn’t necessarily the right thing to do.

b) Assuming this is really desirable, should the fact that casting an integer 0 to a pointer type results in a null pointer be documented in the langref (as part of the inttoptr/ptrtoint documentation)? Or is it too “low-level”?

c) Let’s say I want to create a pointer with the numeric value 0, distinct from the null pointer, because 0 really does represent a valid pointer on my platform. Should this possible at the IR level? If it should be, then how?

Anyone has any inputs, except laughing me out of the (virtual) room? :slight_smile:

Thanks,

Michael

1 Like

I ran into the same problem in OpenCL with Local/Private address spaces and I believe it was discussed here in the past. So I would suggest looking through the archives as I believe some of your questions are answered there.

This post might be a good starting point:

http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/044101.html

Also the SPIR provisional spec as of last November deals with this issue if I remember correctly, so maybe you can discuss with Boaz Ouriel if he is still at Intel Israel about how they went about handling this problem, or look at the SPIR related posts from last year.

Hope this helps,

Micah

Hi Micah,

Thanks a lot for the reference.

Unfortunately, it looks like the discussion there is dealing with a somewhat different issue – modeling OpenCL address spaces. It’s true that the null pointer issue may arise in a similar context (e.g. null pointers in different address spaces may have different internal representations), but it’s not entirely related. Or have I missed something in the thread?

Thanks,

Michael

The reason why I suggested looking there is because the same questions came up as address 0 in some of the OpenCL address spaces are legal addresses. So in order to solve this problem, they would have to find answers to your questions, mainly your ‘c)’ question.

I took a look at LangRef and it does look like there is a ‘null’ representation of the null pointer constant, which wasn’t part of LLVM back when I worked on this problem. So maybe you are just running into an issue where the code hasn’t been updated yet because it works on everyones platform but yours.

Micah

+llvm-dev

Yes. The question then becomes, does LLVM handle the case where the cast of the value 0 to a non-void* pointer?

Think of this case:

int ptr = (int)0;

based on a strict reading of the spec, ptr itself technically is not the null pointer constant. If ptr points to an object in the local address space in OpenCL, or any address space where 0 is a valid address and memory exists, does LLVM handle this correctly?

I don’t believe that it does, but I can’t say for certain.

Micah

+llvm-dev
Yes. The question then becomes, does LLVM handle the case where the cast of the value 0 to a non-void* pointer?

Think of this case:
int *ptr = (int*)0;

based on a strict reading of the spec, ptr itself technically is not the null pointer constant.

Not true, the literal zero *is* a null pointer constant, which is then converted to a null pointer.

Ahh yeah duh, your right. I misread that.

That doesn’t solve this problem however, which is a different example of what I was trying to show but failed.

kernel void test(local int* a) {

local int*ptr = NULL;

if (a != ptr) *a = 0;

}

from host code:

test(0);

Micah