RFC: -Wstrict-aliasing extension

[Jim Wilson <wilson at specifixinc dot com>]

A customer moving from gcc-2.9x to gcc-3.x ran into problems with
-fstrict-aliasing, and was disappointed that -Wstrict-aliasing did not
help.

The problem is that -Wstrict-aliasing does not warn about ambiguous
cases.  If you have a structure that contains a char in addition to
other types, then it is valid to cast a pointer to the structure type
and dereference the char, but not valid to dereference one of the other
types.  Since a cast to this structure type might be valid, the current
-Wstrict-aliasing code gives no warning at all.  However, this makes it
impossible for users to find some of their invalid code.

I wrote a patch that warns for the ambiguous cases, adding "may" to the
message, to indicate that the code may or may not be OK.  This allowed
the customer to find the problems in their code.

Consider this testcase:
struct foo {
  char c;
  char d;
  short s;
  int i;
} bar;

int
sub1 (long long int foobar)
{
  struct foo *tmp = (struct foo *) &foobar;
  return tmp->i;
}

short
sub2 (long long int foobar)
{
  struct foo *tmp = (struct foo *) &foobar;
  return tmp->c;
}

The function sub2 is OK, but the function sub1 is not, and gets
miscompiled at -O2 on x86-linux.  -Wstrict-aliasing gives no warning for
either function.

With the following patch, I get[wilson@leaf gcc]$ ./xgcc -B./ -O2
-Wstrict-aliasing -S tmp.c
tmp.c: In function `sub1':
tmp.c:11: warning: dereferencing type-punned pointer may break
strict-aliasing rules
tmp.c: In function `sub2':
tmp.c:18: warning: dereferencing type-punned pointer may break
strict-aliasing rules

The warning for sub2 is a false positive, but the warning for sub1 is
accurate, and allows the user to find the bug in their code.

I am not sure if others think this addition is a good idea, which is why
I am asking for comments.  If people don't want this in
-Wstrict-aliasing, then perhaps another option can be added for it,
-Wmaybe-strict-aliasing perhaps?  Either way, this needs a doc change
which I haven't written yet.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com

Attachment: patch.warn.alias
Description: Text document