Clang omits variable with "used" attribute

[pogo59]

GCC specifies the "used" attribute as:

This attribute, attached to a variable with static storage, means that the variable must be emitted even if it appears that the variable is not referenced.

Inside a constant "if" condition, clang will fail to emit a "used" variable in the not-taken branch. This differs from GCC behavior.

void foo() {
  if (sizeof(double) == sizeof(long double)) {
    __attribute__((used)) static int my_item(__LINE__);
  } else {
    __attribute__((used)) static int my_item(__LINE__);
  }
}

Clang allocates only one variable, GCC allocates both. Given it's a GCC-defined attribute, we should behave the way GCC does. This appears to be a Clang codegen bug, as -emit-llvm -disable-llvm-passes shows only one item.

(Why would I need this? It's part of the instrumentation for my Rotten Green Tests project. My static variables are actually allocated into a custom section, but that's not needed to trigger the bug.)

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Paul T Robinson (pogo59)

[GCC](https://gcc.gnu.org/onlinedocs/gcc/Common-Variable-Attributes.html) specifies the "used" attribute as: > This attribute, attached to a variable with static storage, means that the variable must be emitted even if it appears that the variable is not referenced.

Inside a constant "if" condition, clang will fail to emit a "used" variable in the not-taken branch. This differs from GCC behavior.

void foo() {
  if (sizeof(double) == sizeof(long double)) {
    __attribute__((used)) static int my_item(__LINE__);
  } else {
    __attribute__((used)) static int my_item(__LINE__);
  }
}

Clang allocates only one variable, GCC allocates both. Given it's a GCC-defined attribute, we should behave the way GCC does. This appears to be a Clang codegen bug, as -emit-llvm -disable-llvm-passes shows only one item.

(Why would I need this? It's part of the instrumentation for my Rotten Green Tests project. My static variables are actually allocated into a custom section, but that's not needed to trigger the bug.)

[efriedma-quic]

See also #82994 (comment) . (Not the same issue, but it's in the same part of the implementation.)