[Bug tree-optimization/84512] [8 Regression] Missed optimization: should be precalculated in compile-time

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84512

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
            Version|tree-ssa                    |8.0
           Keywords|                            |missed-optimization
   Last reconfirmed|                            |2018-02-27
             Blocks|                            |53947
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1
            Summary|Missed optimization: should |[8 Regression] Missed
                   |be precalculated in         |optimization: should be
                   |compile-time                |precalculated in
                   |                            |compile-time
   Target Milestone|---                         |8.0

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  This is another case where we vectorize one loop but not the other
and DOM doesn't handle removing vector loads against scalar stores.  Later
store-merging aggregates the stores but nothig performs CSE after it.

The vectorizer decides that vectorizing the reduction is profitable while
vectorizing the init is not:

t.c:4:3: note: Cost model analysis:
  Vector inside of loop cost: 68
  Vector prologue cost: 8
  Vector epilogue cost: 128
  Scalar iteration cost: 16
  Scalar outside cost: 0
  Vector outside cost: 136
  prologue iterations: 0
  epilogue iterations: 2
t.c:4:3: note: cost model: the vector iteration cost = 68 divided by the scalar
iteration cost = 16 is greater or equal to the vectorization factor = 4.
t.c:4:3: note: not vectorized: vectorization not profitable.

With -fno-vect-cost-model we vectorize both loops and optimize the function
like clang does.

The issue with the cost model here is that for the scalar iteration cost
we end up using builtin_vectorization_cost () while for the vector cost
we use add_stmt_cost.  Only the latter makes a difference between the
different kind of operations.

I have a patch.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations