[Bug middle-end/95903] gcc 10: wrong code with -fwrapv

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95903

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 48788
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48788&action=edit
gcc11-pr95903.patch

Untested fix.  I don't really care what clang generates or what you find
expected on ilp32 targets, what matters is if what GCC generates is fast, and
unless somebody finds a wrong-code issue with that, it doesn't need to be
changed.
For the lp64 case, I can imagine a testcase, though it will be pretty rare in
practice and we need to slow down -fwrapv generated code because of that (isn't
e.g. Linux kernel built with -fwrapv?), just don't want to add it into the
testsuite as that would require allocating at least 4GB of memory.

On the other side, on:
extern char arr[];

char
baz (int off)
{
  off += 1;
  return arr[off];
}

char
qux (int off)
{
  return arr[off + 1];
}
we do generate the addition and then sign extension, and we should not, because
we would get better code otherwise.  arr[INT_MIN] is in this case certainly
invalid.  But that is about different code.