[Bug tree-optimization/107717] New: [13 Regression] ICEs expanding permutes after g:dc95e1e9702f2f6367bbc108c8d01169be1b66d2

["tnfchris at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107717

            Bug ID: 107717
           Summary: [13 Regression] ICEs expanding permutes after
                    g:dc95e1e9702f2f6367bbc108c8d01169be1b66d2
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tnfchris at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64*

After

commit dc95e1e9702f2f6367bbc108c8d01169be1b66d2 (origin/trunk, origin/master,
origin/HEAD)
Author: Hongyu Wang <hongyu.wang@intel.com>
Date:   Mon Jan 17 13:01:51 2022 +0800

    Optimize VEC_PERM_EXPR with same permutation index and operation

    The sequence
         c1 = VEC_PERM_EXPR (a, a, mask)
         c2 = VEC_PERM_EXPR (b, b, mask)
         c3 = c1 op c2
    can be optimized to
         c = a op b
         c3 = VEC_PERM_EXPR (c, c, mask)
    for all integer vector operation, and float operation with
    full permutation.

    gcc/ChangeLog:

            PR target/98167
            * match.pd: New perm + vector op patterns for int and fp vector.

    gcc/testsuite/ChangeLog:

            PR target/98167
            * gcc.target/i386/pr98167.c: New test.

We see various ICEs, an example is

void foo(int n, char *restrict out, char *restrict in) {
  for (int i=n; i-->0; ) {
    out[i] += in[i];
  }
}

compiled with

aarch64-none-linux-gnu -O3 -march=armv8-a+sve2

The problem is that the match.pd pattern as written causes the permute to
switch from a single register permute to a two register one.

The reason is that when the folded result is expanded in SSA form

vec_perm (op @0 @1) (op @0 @1)

the result of applying op twice results in two distinct SSA names. This fails
because expand_vec_perm_const now tries to use a two operand expansion because
there's no easy way to tell that these two operands are the same.

If it happens early enough we can CSE the operands, but when this happens after
vec_lower it generated something the target does not support.

I tried getting expand_vec_perm_const to recognize that they are the same, but
that's quite hard.

It's best to prevent the generation of the two SSA names to begin with, or add
an additional rule for match.pd that's able to CSE this.

I'm making this issue because I don't know which approach upstream would like
so it's easier to ask first.