member function void foo() __restrict__; incorrectly applies restrict to type, not pointer

[silvasean]

Bugzilla Link	17747
Version	trunk
OS	Linux
CC	@majnemer,@DougGregor,@imallett,@rnk

Extended Description

struct Foo {
  void member() __restrict__ {
    Foo *__restrict__ This = this;
  }
};

Clang gives the diagnostic:

test.cpp:3:23: error: cannot initialize a variable of type 'Foo *restrict' with
      an rvalue of type 'restrict Foo *'
    Foo *__restrict__ This = this;
                      ^      ~~~~
1 error generated.

Clang seems to think that this has type restrict Foo * (restrict on Foo, rather than the pointer)

The GCC extension says:

http://gcc.gnu.org/onlinedocs/gcc/Restricted-Pointers.html#Restricted-Pointers

"Within the body of T::fn, this has the effective definition T *restrict const this."

[zygoloid]

*** Bug llvm/llvm-bugzilla-archive#17793 has been marked as a duplicate of this bug. ***

[rnk]

I ran into this again while working on debug info, and wrote up a long bug report to discover that it was a duplicate.

See also #11039 .

[imallett]

I also ran into this bug, and also wrote up a long bug report to discover that it was a duplicate. So, I'll try to add a bit more information. This bug is almost 7 years old now!

So first, this still occurs in Clang trunk 12.0.0 (but way earlier also; I checked as far back as 3.0.0). I can't give a cleaner example than the OP, but I will comment that Clang's error message is, specifically:

<source>:3:23: error: cannot initialize a variable of type 'Foo *__restrict' with an rvalue of type '__restrict Foo *'
    Foo *__restrict__ This = this;
                      ^      ~~~~

It is also worth noting the GCC, MSVC, ICC, and possibly others all do not suffer from this problem.

As has been mentioned, the bug is that Clang appears to think that the type of this is Foo __restrict__* (that is, a pointer to a restrict-qualified Foo object, not a restrict-qualified pointer to a Foo object). This, of course, makes no sense. Restrict-qualification is a pointer thing. Indeed, in other contexts, even writing such a type is sensibly impossible. E.g. the following line will correctly fail:

Foo __restrict* foo = nullptr; //fails (correctly)

Note: while __restrict is not a C++ keyword, when applied to C++ member functions, it is widely implemented as the this pointer becoming restrict-qualified. That seems to be what Clang attempted here, but got the type incorrect.

Note: a C-style cast works around this problem, but obviously this is not a solution.

[zygoloid]

mentioned in issue llvm/llvm-bugzilla-archive#17793

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Sean Silva (silvasean)

| | | | --- | --- | | Bugzilla Link | [17747](https://llvm.org/bz17747) | | Version | trunk | | OS | Linux | | CC | @majnemer,@DougGregor,@imallett,@rnk |

Extended Description

struct Foo {
void member() restrict {
Foo *restrict This = this;
}
};

Clang gives the diagnostic:

test.cpp:3:23: error: cannot initialize a variable of type 'Foo *restrict' with
an rvalue of type 'restrict Foo *'
Foo *restrict This = this;
^ ~~~~
1 error generated.

Clang seems to think that this has type restrict Foo * (restrict on Foo, rather than the pointer)

The GCC extension says:

<http://gcc.gnu.org/onlinedocs/gcc/Restricted-Pointers.html#Restricted-Pointers>

"Within the body of T::fn, this has the effective definition T *restrict const this."