Handle SLP of call pattern statements
commitd3f09bd8151666f5a180fc38337ea835e109caaf
authorrsandifo <rsandifo@138bc75d-0d04-0410-961f-82ee72b054a4>
Fri, 3 Aug 2018 12:56:55 +0000 (3 12:56 +0000)
committerrsandifo <rsandifo@138bc75d-0d04-0410-961f-82ee72b054a4>
Fri, 3 Aug 2018 12:56:55 +0000 (3 12:56 +0000)
treec5cea056f707c9b77bae206e6dad99911f65d947
parenta4771bd58f53e475d2e22e4fafe52df232cccb92
Handle SLP of call pattern statements

We couldn't vectorise:

  for (int j = 0; j < n; ++j)
    {
      for (int i = 0; i < 16; ++i)
a[i] = (b[i] + c[i]) >> 1;
      a += step;
      b += step;
      c += step;
    }

at -O3 because cunrolli unrolled the inner loop and SLP couldn't handle
AVG_FLOOR patterns (see also PR86504).  The problem was some overly
strict checking of pattern statements compared to normal statements
in vect_get_and_check_slp_defs:

          switch (gimple_code (def_stmt))
            {
            case GIMPLE_PHI:
            case GIMPLE_ASSIGN:
      break;

    default:
      if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
 "unsupported defining stmt:\n");
      return -1;
            }

The easy fix would have been to add GIMPLE_CALL to the switch,
but I don't think the switch is doing anything useful.  We only create
pattern statements that the rest of the vectoriser can handle, and the
other checks in this function and elsewhere check whether SLP is possible.

I'm also not sure why:

          if (!first && !oprnd_info->first_pattern
      /* Allow different pattern state for the defs of the
 first stmt in reduction chains.  */
      && (oprnd_info->first_dt != vect_reduction_def

is necessary.  All that should matter is that the statements in the
node are "similar enough".  It turned out to be quite hard to find a
convincing example that used a mixture of pattern and non-pattern
statements, so bb-slp-pow-1.c is the best I could come up with.
But it does show that the combination of "xi * xi" statements and
"pow (xj, 2) -> xj * xj" patterns are handled correctly.

The patch therefore just removes the whole if block.

The loop also needed commutative swapping to be extended to at least
AVG_FLOOR.

This gives +3.9% on 525.x264_r at -O3.

2018-08-03  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* internal-fn.h (first_commutative_argument): Declare.
* internal-fn.c (first_commutative_argument): New function.
* tree-vect-slp.c (vect_get_and_check_slp_defs): Remove extra
restrictions for pattern statements.  Use first_commutative_argument
to look for commutative operands in calls to internal functions.

gcc/testsuite/
* gcc.dg/vect/bb-slp-over-widen-1.c: Expect AVG_FLOOR to be used
on vect_avg_qi targets.
* gcc.dg/vect/bb-slp-over-widen-2.c: Likewise.
* gcc.dg/vect/bb-slp-pow-1.c: New test.
* gcc.dg/vect/vect-avg-15.c: Likewise.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@263290 138bc75d-0d04-0410-961f-82ee72b054a4
gcc/ChangeLog
gcc/internal-fn.c
gcc/internal-fn.h
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.dg/vect/bb-slp-over-widen-1.c
gcc/testsuite/gcc.dg/vect/bb-slp-over-widen-2.c
gcc/testsuite/gcc.dg/vect/bb-slp-pow-1.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/vect/vect-avg-15.c [new file with mode: 0644]
gcc/tree-vect-slp.c