[Bug gold/17704] New: icf doesn't respect section alignment

[rafael.espindola at gmail dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=17704

            Bug ID: 17704
           Summary: icf doesn't respect section alignment
           Product: binutils
           Version: unspecified
            Status: NEW
          Severity: normal
          Priority: P2
         Component: gold
          Assignee: ccoutant at google dot com
          Reporter: rafael.espindola at gmail dot com
                CC: ian at airs dot com

Created attachment 8008
  --> https://sourceware.org/bugzilla/attachment.cgi?id=8008&action=edit
testcase

With the attached testcase, running

gcc -c t1.s
ld -o foo t1.o
./foo
echo $?
ld -o foo t1.o --icf=safe
./foo
echo $?

Will print

0
1

The problem is that icf is merging foo and bar but not enforcing the stricter
alignment requirements of bar.

While this reduced testcase is assembly, the original was c++.

The check of the least significant bit of a function pointer comes from how
member functions are implemented:

* For non virtual functions, a plain pointer is passed.
* For virtual functions it is 1 + the vtable offset

To ensure that this works, the compiler has to align member functions, but has
no such requirement for regular functions. For example, given

struct S {
  void f() { }
};
inline void g() {}
void f(S *x) {
  g();
  x->f();
}

gcc will align f but not g.

It is possible for a member function to codegen to the same bits as a
non-member function, which is how this bug was found.

-- 
You are receiving this mail because:
You are on the CC list for the bug.