clang does not implement __assume__ attribute, fails libstdc++ build

[trofi]

gcc implements __assume__ attribute as: https://gcc.gnu.org/onlinedocs/gcc/Statement-Attributes.html, quoting the example:

int
foo (int x, int y)
{
  __attribute__((assume(x == 42)));
  __attribute__((assume(++y == 43)));
  return x + y;
}

clang-16 fails to build it as:

$ clang -c a.c
a.c:4:18: error: 'assume' attribute cannot be applied to a statement
  __attribute__((assume(x == 42)));
                 ^                ~
a.c:5:18: error: 'assume' attribute cannot be applied to a statement
  __attribute__((assume(++y == 43)));
                 ^                  ~
2 errors generated.

Recently gcc started using it in libstdc++ https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=e39b3e02c27bd7 as

    void
    _M_assume_normalized() const
    {
      unsigned int __ofst = _M_offset;
      __attribute__ ((__assume__ (__ofst < unsigned(_S_word_bit))));
    }

and clang++ now fails to parse <vector> from libstdc++.

Is there a change clang would support this style of __assume__ attribute or at least skipped it in stead of producing build failure?

[trofi]

Meanwhile filed https://gcc.gnu.org/PR112467 on gcc side to consider using this style of the attribute only for gcc.

[pinskia]

Note GCC's assume attribute is the same as C++23 assume standard attribute.
That is: [[assume(x == 0)]]; and __attribute__((assume(x == 0) )); are exactly the same.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Sergei Trofimovich (trofi)

`gcc` implements `__assume__` attribute as: https://gcc.gnu.org/onlinedocs/gcc/Statement-Attributes.html, quoting the example:

int
foo (int x, int y)
{
  __attribute__((assume(x == 42)));
  __attribute__((assume(++y == 43)));
  return x + y;
}

clang-16 fails to build it as:

$ clang -c a.c
a.c:4:18: error: 'assume' attribute cannot be applied to a statement
  __attribute__((assume(x == 42)));
                 ^                ~
a.c:5:18: error: 'assume' attribute cannot be applied to a statement
  __attribute__((assume(++y == 43)));
                 ^                  ~
2 errors generated.

Recently gcc started using it in libstdc++ https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=e39b3e02c27bd7 as

    void
    _M_assume_normalized() const
    {
      unsigned int __ofst = _M_offset;
      __attribute__ ((__assume__ (__ofst < unsigned(_S_word_bit))));
    }

and clang++ now fails to parse <vector> from libstdc++.

Is there a change clang would support this style of __assume__ attribute or at least skipped it in stead of producing build failure?

[shafik]

CC @erichkeane @AaronBallman

[erichkeane]

@AaronBallman is particularly opinionated about attribute assume so he should set our policy here. If it is ok with him, I'd be ok reviewing a statement assume attribute.

[AaronBallman]

@AaronBallman is particularly opinionated about attribute assume so he should set our policy here. If it is ok with him, I'd be ok reviewing a statement assume attribute.

We had a Discourse thread on this, and my thoughts at the time were: https://discourse.llvm.org/t/llvm-assume-blocks-optimization/71609/23 -- my thinking really hasn't changed all that much since then. I think we need LLVM changes before we can effectively implement the attribute and so we should report 0 for __has_cpp_attribute(assume) until we're ready to support the attribute.