[C++20][Modules] Incorrectly distincted static local vars

[Lancern]

Given the following set of translation units:

export module a;

export int __attribute__((always_inline)) foo() {
  static int x;
  return ++x;
}

import a;

int test1() { return foo(); }

import a;

int test2() { return foo(); }

#include <iostream>

int test1();
int test2();
int main() {
  std::cout << test1();
  std::cout << test2();
}

The program outputs 11 instead of the expected result 12. Function test1 and function test2 access distinct static int x objects which are mistakenly inlined into their bodies.

Up to clang 17.0.1, this behavior can happen even if foo is not marked with always_inline (with optimizations enabled). Seems like this problem is partially fixed by PR #71031 , but the PR misses always_inline functions.

[llvmbot]

@llvm/issue-subscribers-clang-modules

Author: Sirui Mu (Lancern)

Given the following set of translation units:

export module a;

export int __attribute__((always_inline)) foo() {
  static int x;
  return ++x;
}

import a;

int test1() { return foo(); }

import a;

int test2() { return foo(); }

#include <iostream>

int test1();
int test2();
int main() {
  std::cout << test1();
  std::cout << test2();
}

The program outputs 11 instead of the expected result 12. Function test1 and function test2 access distinct static int x objects which are mistakenly inlined into their bodies.

Up to clang 17.0.1, this behavior can happen even if foo is not marked with always_inline (with optimizations enabled). Seems like this problem is partially fixed by PR #71031 , but the PR misses always_inline functions.

[ChuanqiXu9]

Although we may fix the issue by treating the imported static locals as available externally, it looks more consistency to not import non-inline function bodies at all no matter if it is always_inline or not. Since we already decided to not import the non-inline function bodies: https://isocpp.org/files/papers/P3092R0.html

[ChuanqiXu9]

The previous solution gets some complaints so we probably need some workaround here.