[Bug tree-optimization/48702] [4.6/4.7 Regression] optimization regression with gcc-4.6 on x86_64-unknown-linux-gnu

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48702

--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-21 11:05:09 UTC ---
Before the patch:

        movq    %rsp, %rdx
        leaq    -16(%rsp), %rcx
        movl    $0, %eax
.L5:
        addl    %eax, %eax
        addl    12(%rdx), %eax
        subq    $4, %rdx
        cmpq    %rcx, %rdx
        jne     .L5

after the patch:

        movl    $0, %edx
        movl    $0, %eax
.L5:
        addl    %eax, %eax
        addl    12(%rsp,%rdx), %eax
        subq    $4, %rdx
        cmpq    $-16, %rdx
        jne     .L5

not 100% clear which to prefer.

Alternatively alias-analysis can be dumbed down, never exploiting knowledge
of a constant offset added to a pointer before dereferencing it
(thus, ignoring the constant offset operand in MEM_REFs and TARGET_MEM_REFs).
FYI, the disambiguation that triggers is

  /* If only one reference is based on a variable, they cannot alias if
     the pointer access is beyond the extent of the variable access.
     (the pointer base cannot validly point to an offset less than zero
     of the variable).
     They also cannot alias if the pointer may not point to the decl.  */
  if ((TREE_CODE (base1) != TARGET_MEM_REF
       || (!TMR_INDEX (base1) && !TMR_INDEX2 (base1)))
      && !ranges_overlap_p (MAX (0, offset1p), -1, offset2p, max_size2))
    return false;

thus, for "the pointer base cannod validly point to an offset less than
zero" say that this only holds for ptr + constant offset, not ptr in
isolation.