[Bug other/96281] New: TBAA does not work as expected for a simple test case

[Bug other/96281] New: TBAA does not work as expected for a simple test case

[felix.yang at huawei dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96281

            Bug ID: 96281
           Summary: TBAA does not work as expected for a simple test case
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Keywords: alias
          Severity: normal
          Priority: P3
         Component: other
          Assignee: unassigned at gcc dot gnu.org
          Reporter: felix.yang at huawei dot com
  Target Milestone: ---
            Target: aarch64

Test case: foo.c

typedef struct state_t {
    int threadid;
} state_t;

int history_h[8][12][64];

void history_good (state_t *s) {
    int i, j;

    if (s->threadid >= 0 && s->threadid < 8) {
        for (i = 0; i < 12; i++) {
            for (j = 0; j < 64; j++) {
                history_h[s->threadid][i][j] = (history_h[s->threadid][i][j] +
1) >> 1;
            }
        }
    }
}

$ gcc -S -O2 -ftree-loop-vectorize -funroll-loops foo.c -fopt-info
foo.c:14:13: optimized: loop unrolled 6 times

When the input parameter s is specified to be unaliased with __restrict__ type
qualifier like:
void history_good (state_t * __restrict__ s)

The inner loop could be auto-vectorized:
$ gcc -S -O2 -ftree-loop-vectorize -funroll-loops foo.c -fopt-info
foo.c:14:13: optimized: loop vectorized using 16 byte vectors
foo.c:8:6: optimized: loop with 15 iterations completely unrolled (header
execution count 16535624)

Looks like TBAA is not working here for this case.  Then I noticed the
following logic in tree-ssa-alias.c:

1939   /* When we are trying to disambiguate an access with a pointer
dereference
1940      as base versus one with a decl as base we can use both the size
1941      of the decl and its dynamic type for extra disambiguation.
1942      ???  We do not know anything about the dynamic type of the decl
1943      other than that its alias-set contains base2_alias_set as a subset
1944      which does not help us here.  */
1945   /* As we know nothing useful about the dynamic type of the decl just
1946      use the usual conflict check rather than a subset test.
1947      ???  We could introduce -fvery-strict-aliasing when the language
1948      does not allow decls to have a dynamic type that differs from their
1949      static type.  Then we can check
1950      !alias_set_subset_of (base1_alias_set, base2_alias_set) instead.  */
1951   if (base1_alias_set != base2_alias_set
1952       && !alias_sets_conflict_p (base1_alias_set, base2_alias_set))
1953     return false;

This was introduced by: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=42834 
>From the comments, this depends on the language of the source code. 
So at least for C & C++, could we check !alias_set_subset_of (base1_alias_set,
base2_alias_set) instead here?
Any other languages supported by GCC that makes a difference?

[Bug middle-end/96281] TBAA does not work as expected for a simple test case

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96281

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|other                       |middle-end
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |missed-optimization
                 CC|                            |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-07-22

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Memory reference 1: s_13(D)->threadid
Memory reference 2: history_h[_3][i_24][j_25]
Querying dependency of refs 1 and 2: dependent.
Querying RAW dependencies of ref 1 in loop 2: dependent
Querying RAW dependencies of ref 1 in loop 1: dependent

so in the inner loop we see

  <bb 9> [local count: 1040670577]:

  <bb 3> [local count: 1057206201]:
  # j_25 = PHI <0(5), j_16(9)>
  _3 = s_13(D)->threadid;
  _4 = history_h[_3][i_24][j_25];
  _5 = _4 + 1;
  _6 = _5 >> 1;
  history_h[_3][i_24][j_25] = _6;
  j_16 = j_25 + 1;
  if (j_16 != 64)
    goto <bb 9>; [98.44%]
  else
    goto <bb 4>; [1.56%]

and while on the entry edge we can CSE s_13(D)->threadid (later PRE does this)
we cannot prove that history_h[_3][i_24][j_25] = _6 does not alters its value.

This looks like a missed optimization as the very same reference is
read before the write.  We might also be able to use path-based
disambiguation here.  That said, 's_13(D)->threadid' should have the
base_alias_set of struct state but then the comment suggests we
know nothing about the dynamic type.

Needs more thoughts and analysis.

[Bug tree-optimization/96281] TBAA does not work as expected for a simple test case

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96281

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|middle-end                  |tree-optimization
           Severity|normal                      |enhancement