[Bug tree-optimization/65206] Vectorized version of loop is removed.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65206

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
We apply versioning for aliasing but compute it as always aliasing in some way,
thus the runtime check gets immediately folded and thus the vectorized loop
removed:

t.c:7:3: note: create runtime check for data references a1[i_16] and *_34
t.c:7:3: note: created 1 versioning for alias checks.
t.c:7:3: note: loop versioned for vectorization because of possible aliasing
...

but I see the alias runtime check nowhere.

The DRs are

(gdb) p debug_data_reference (dr_a.dr)
#(Data Ref: 
#  bb: 4 
#  stmt: _5 = a1[i_16];
#  ref: a1[i_16];
#  base_object: a1;
#  Access function 0: {0, +, 1}_1
#)
$17 = void
(gdb) p debug_data_reference (dr_b.dr)
#(Data Ref: 
#  bb: 4 
#  stmt: MASK_STORE (_34, 0B, _ifc__32, _9);
#  ref: *_34;
#  base_object: MEM[(float *)&a1];
#  Access function 0: {0B, +, 4}_1
#)

so maybe the code doing masked loads/stores updates the DRs in a way that
will later confuse runtime alias checking.  Or for some reason it doesn't
update it enough to make data-dependence analysis handle it.

Clearly this is a must-dependence (but with known distance), so sth
that data dependence analysis should handle and sth that the runtime
alias checking isn't handling.