tree-optimization/96022 - fix ICE with vectorized shift

This fixes lane extraction for internal def vectorized shifts
with an effective scalar shift operand by always using lane zero
of the first vector stmt.

It also fixes a SLP build issue noticed on the testcase where
we end up building unary vector ops with the only operand built
form scalars which isn't profitable by itself.  The exception
is for stores.

2020-07-02  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/96022
	* tree-vect-stmts.c (vectorizable_shift): Only use the
	first vector stmt when extracting the scalar shift amount.
	* tree-vect-slp.c (vect_build_slp_tree_2): Also build unary
	nodes with all-scalar children from scalars but not stores.
	(vect_analyze_slp_instance): Mark the node not failed.

	* g++.dg/vect/pr96022.cc: New testcase.

3 files changed, 22 insertions, 7 deletions

diff --git a/gcc/testsuite/g++.dg/vect/pr96022.cc b/gcc/testsuite/g++.dg/vect/pr96022.cc
new file mode 100644
index 0000000..ca6b276
--- /dev/null
+++ b/gcc/testsuite/g++.dg/vect/pr96022.cc
@@ -0,0 +1,12 @@
+// { dg-do compile }
+// { dg-additional-options "-O3" }
+
+extern int arr_6[];
+extern char arr_7[] __attribute__((aligned));
+void test(short a, bool, int p8) {
+  for (bool b = 0; b < (bool)p8; b = 1)
+    for (short c = 0; c < 5; c++) {
+      arr_6[c] = (long)2 << a - 30574;
+      arr_7[c] = 0;
+    }
+}
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index 532809d..af123b5 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -1530,11 +1530,11 @@ fail:
 
   vect_free_oprnd_info (oprnds_info);
 
-  /* If we have all children of a non-unary child built up from
-     uniform scalars then just throw that away, causing it built up
-     from scalars.  */
-  if (nops > 1
-      && is_a <bb_vec_info> (vinfo)
+  /* If we have all children of a child built up from uniform scalars
+     then just throw that away, causing it built up from scalars.
+     The exception is the SLP node for the vector store.  */
+  if (is_a <bb_vec_info> (vinfo)
+      && !STMT_VINFO_GROUPED_ACCESS (stmt_info)
       /* ???  Rejecting patterns this way doesn't work.  We'd have to
 	 do extra work to cancel the pattern so the uses see the
 	 scalar version.  */
@@ -2230,6 +2230,7 @@ vect_analyze_slp_instance (vec_info *vinfo,
 	      return false;
 	    }
 	  /* Fatal mismatch.  */
+	  matches[0] = true;
 	  matches[group_size / const_max_nunits * const_max_nunits] = false;
 	  vect_free_slp_tree (node, false);
 	}
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index 9b7b04c..d68547e 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -5403,7 +5403,7 @@ vectorizable_shift (vec_info *vinfo,
 	      if (!op1_vectype)
 		op1_vectype = get_vectype_for_scalar_type (vinfo,
 							   TREE_TYPE (op1),
-							   slp_node);
+							   slp_op1);
 
               /* Unlike the other binary operators, shifts/rotates have
                  the rhs being int, instead of the same type as the lhs,
@@ -5575,11 +5575,11 @@ vectorizable_shift (vec_info *vinfo,
   /* Arguments are ready.  Create the new vector stmt.  */
   FOR_EACH_VEC_ELT (vec_oprnds0, i, vop0)
     {
-      vop1 = vec_oprnds1[i];
       /* For internal defs where we need to use a scalar shift arg
 	 extract the first lane.  */
       if (scalar_shift_arg && dt[1] == vect_internal_def)
 	{
+	  vop1 = vec_oprnds1[0];
 	  new_temp = make_ssa_name (TREE_TYPE (TREE_TYPE (vop1)));
 	  gassign *new_stmt
 	    = gimple_build_assign (new_temp,
@@ -5590,6 +5590,8 @@ vectorizable_shift (vec_info *vinfo,
 	  vect_finish_stmt_generation (vinfo, stmt_info, new_stmt, gsi);
 	  vop1 = new_temp;
 	}
+      else
+	vop1 = vec_oprnds1[i];
       gassign *new_stmt = gimple_build_assign (vec_dest, code, vop0, vop1);
       new_temp = make_ssa_name (vec_dest, new_stmt);
       gimple_assign_set_lhs (new_stmt, new_temp);